WP-AI-0002: repository scaffold, licence governance, PMC S3 ingest, JATS parser, segmentation, dedup, splits, Claude rewrite engine (mock/live/batch), validation, HTR M1/M2/M3 models, trainer, evaluation, Gate A report, literature
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
102 changed files +9,962 −0
added
.gitignore
+44 −0
@@ -0,0 +1,44 @@ | ||
| 1 | +# Data: raw licensed papers, parsed text, pairs with API outputs — never committed by default | |
| 2 | +data/raw/* | |
| 3 | +data/parsed/* | |
| 4 | +data/segments/* | |
| 5 | +data/pairs/* | |
| 6 | +data/splits/* | |
| 7 | +!data/*/.gitkeep | |
| 8 | +data/manifests/*.jsonl | |
| 9 | +data/manifests/*.parquet | |
| 10 | +!data/manifests/.gitkeep | |
| 11 | +# Artifacts: checkpoints are large; metrics/figures/tables/reports ARE committed (small, reproducible) | |
| 12 | +artifacts/checkpoints/* | |
| 13 | +!artifacts/checkpoints/.gitkeep | |
| 14 | +artifacts/runs/ | |
| 15 | +# Secrets | |
| 16 | +.env | |
| 17 | +*.key | |
| 18 | +# Python | |
| 19 | +__pycache__/ | |
| 20 | +*.pyc | |
| 21 | +.pytest_cache/ | |
| 22 | +.venv/ | |
| 23 | +*.egg-info/ | |
| 24 | +build/ | |
| 25 | +dist/ | |
| 26 | +# LaTeX build artifacts (paper/main.pdf IS committed) | |
| 27 | +paper/*.aux | |
| 28 | +paper/*.log | |
| 29 | +paper/*.out | |
| 30 | +paper/*.toc | |
| 31 | +paper/*.lof | |
| 32 | +paper/*.lot | |
| 33 | +paper/*.fls | |
| 34 | +paper/*.fdb_latexmk | |
| 35 | +paper/*.bbl | |
| 36 | +paper/*.blg | |
| 37 | +paper/*.bcf | |
| 38 | +paper/*.run.xml | |
| 39 | +paper/*.synctex.gz | |
| 40 | +# Logs / OS | |
| 41 | +logs/ | |
| 42 | +.DS_Store | |
| 43 | +# demo | |
| 44 | +demo/node_modules/ | |
added
CLAUDE.md
+36 −0
@@ -0,0 +1,36 @@ | ||
| 1 | +# CLAUDE.md — WP-AI-0002: Human Text Reconstruction Transformer (HTR) | |
| 2 | + | |
| 3 | +**Can a Transformer reconstruct characteristics of pre-2018 human-authored scientific prose from LLM-rewritten | |
| 4 | +passages?** Research home: UQO Research (`Recherche-UQO/ia/wp-ai-0002`). Language: English for code, docs, | |
| 5 | +paper. Identifier: **AI-0002** (series "AI", after AI-0001). Status: see `docs/STATUS.md`. | |
| 6 | + | |
| 7 | +The author's full mandate (mission, principles, RQ1–RQ6, phases 0–8, gates A–F, definition of done) is | |
| 8 | +reproduced verbatim in `docs/MANDATE.md` and binds every session. Author additions (2026-11-09): the corpus must | |
| 9 | +be **large** (target the "Medium" stage, ≥ 10 000 documents, ≥ 200 000 eligible passages), must include | |
| 10 | +**economics and finance** articles and **English reference books published before 2018**, and the released | |
| 11 | +model must be **genuinely usable afterwards** (model release is a first-class deliverable, not a by-product). | |
| 12 | +PyTorch may be installed on other MacLustr nodes if training needs them. | |
| 13 | + | |
| 14 | +Operational rules that follow from the mandate: | |
| 15 | + | |
| 16 | +1. **Never fabricate** data, metrics, sources, licences or checkpoints. Every number in `docs/`, `artifacts/` | |
| 17 | + and `paper/` traces to a run id in `artifacts/runs/<run>/manifest.json` (hashes, seeds, versions, costs). | |
| 18 | +2. **Rights first.** Open access ≠ open licence. `htr.licenses.policy` is the single source of truth; a document | |
| 19 | + is trained on only if its canonical licence code is in the configured training set, and released only if in | |
| 20 | + the release set (`CC0`, `PUBLIC-DOMAIN`, `CC-BY` by default). Dates are verified from the document or an | |
| 21 | + official metadata record, never inferred from upload timestamps. | |
| 22 | +3. **Honest claims.** Outputs are synthetic reconstructions. No claim of "restoring human authorship" or of AI | |
| 23 | + detection. Separate verbatim reconstruction, stylistic distribution matching and factual fidelity. Test | |
| 24 | + gains against identity (B0) and generic-paraphrase (B1) baselines; report null/negative findings. | |
| 25 | +4. **No leakage.** Splits are by document, near-duplicate cluster and author group (`htr.data.splits`); held-out | |
| 26 | + discipline, period and rewriting prompt; originals of held-out items are never used at inference. | |
| 27 | +5. **Spend is opt-in.** The Anthropic adapter runs in `mock`/`dry-run` by default; real calls need | |
| 28 | + `--live` and a budget ceiling; the pilot ceiling is in `configs/rewriting/*.yaml` and `docs/COMPUTE_BUDGET.md`. | |
| 29 | + Ask before exceeding it, before mass acquisition beyond the planned stages, before any public upload. | |
| 30 | +6. **Gates in order** (A repository + governance + 100-paper smoke → B validated pairs → C splits/baselines → | |
| 31 | + D M1/M2/M3 pilot → E scale → F release + paper). Each gate writes `docs/gates/GATE_<X>.md` with GO/NO-GO. | |
| 32 | +7. **Remote** `origin = gitsrv:srv/git/wpai0002_uqo.git` (spbgit on M1M32). Commit after each deliverable. | |
| 33 | + Raw data, pairs and checkpoints are never committed; `paper/main.pdf` is. | |
| 34 | + | |
| 35 | +Quick start: `.venv/bin/python -m pytest tests -q -m "not network"` · `scripts/vertical_slice.sh` · | |
| 36 | +`.venv/bin/python scripts/discover_pmc.py --max-docs 100` · `docs/REPRODUCIBILITY.md`. | |
added
artifacts/checkpoints/.gitkeep
+0 −0
added
artifacts/figures/.gitkeep
+0 −0
added
artifacts/metrics/.gitkeep
+0 −0
added
artifacts/reports/.gitkeep
+0 −0
added
artifacts/tables/.gitkeep
+0 −0
added
configs/rewriting/pilot.json
+14 −0
@@ -0,0 +1,14 @@ | ||
| 1 | +{ | |
| 2 | + "name": "pilot-v1", | |
| 3 | + "provider": "anthropic", | |
| 4 | + "model": "claude-opus-5-5", | |
| 5 | + "heldout_models": ["claude-sonnet-5-5", "claude-haiku-5-5"], | |
| 6 | + "variants": ["A", "B", "C", "D", "E"], | |
| 7 | + "samples_per_variant": {"A": 1, "B": 1, "C": 1, "D": 1, "E": 2}, | |
| 8 | + "max_tokens": 1500, | |
| 9 | + "effort": "low", | |
| 10 | + "mode": "batch", | |
| 11 | + "concurrency": 4, | |
| 12 | + "budget_usd_ceiling": 30.0, | |
| 13 | + "notes": "Pilot (Gate B): ~1000 passages x 6 generations. Live spend requires --live. Budget is a hard stop." | |
| 14 | +} | |
added
configs/rewriting/pricing.json
+13 −0
@@ -0,0 +1,13 @@ | ||
| 1 | +{ | |
| 2 | + "_comment": "USD per 1M tokens, Anthropic first-party API, retrieved 2026-11-09 from the claude-api reference (cached 2026-10-06). Batch API = 50% of these. Re-verify before every paid run; the engine refuses to run live if a model is missing here.", | |
| 3 | + "retrieved_at": "2026-11-09", | |
| 4 | + "source": "Anthropic pricing via claude-api skill reference (platform.claude.com/docs pricing page)", | |
| 5 | + "models": { | |
| 6 | + "claude-opus-5-5": {"input": 4.00, "output": 20.00, "cache_read": 0.20}, | |
| 7 | + "claude-sonnet-5-5": {"input": 2.00, "output": 10.00, "cache_read": 0.20}, | |
| 8 | + "claude-haiku-5-5": {"input": 0.10, "output": 0.50, "cache_read": 0.01}, | |
| 9 | + "claude-opus-5": {"input": 5.00, "output": 25.00, "cache_read": 0.50}, | |
| 10 | + "claude-sonnet-5": {"input": 2.00, "output": 10.00, "cache_read": 0.20} | |
| 11 | + }, | |
| 12 | + "batch_discount": 0.5 | |
| 13 | +} | |
added
data/manifests/.gitkeep
+0 −0
added
data/manifests/discover_books_bookshelf_seed1.json
+74 −0
@@ -0,0 +1,74 @@ | ||
| 1 | +{ | |
| 2 | + "created_at": "2026-11-09T17:22:26Z", | |
| 3 | + "git": { | |
| 4 | + "commit": "HEAD", | |
| 5 | + "branch": "HEAD", | |
| 6 | + "dirty": true | |
| 7 | + }, | |
| 8 | + "libraries": { | |
| 9 | + "python": "3.14.4", | |
| 10 | + "torch": "2.12.0", | |
| 11 | + "numpy": "2.5.4", | |
| 12 | + "anthropic": "2.3.0", | |
| 13 | + "transformers": "5.8.1", | |
| 14 | + "sentencepiece": "0.2.2", | |
| 15 | + "lxml": "6.1.3", | |
| 16 | + "pyarrow": "24.0.0", | |
| 17 | + "sacrebleu": "2.6.0", | |
| 18 | + "datasketch": "2.0.0", | |
| 19 | + "rapidfuzz": "3.14.6" | |
| 20 | + }, | |
| 21 | + "hardware": { | |
| 22 | + "platform": "macOS-27.0-arm64-arm-64bit-Mach-O", | |
| 23 | + "machine": "arm64", | |
| 24 | + "cpu_count": 18, | |
| 25 | + "chip": "Apple M5 Max", | |
| 26 | + "ram_gb": 48.0, | |
| 27 | + "torch_mps": true, | |
| 28 | + "torch_cuda": false | |
| 29 | + }, | |
| 30 | + "script": "discover_books", | |
| 31 | + "args": { | |
| 32 | + "source": "bookshelf", | |
| 33 | + "max_docs": 3, | |
| 34 | + "max_attempts": 12, | |
| 35 | + "seed": 1, | |
| 36 | + "out": "data/raw/books", | |
| 37 | + "codes": "CC0,PUBLIC-DOMAIN,CC-BY,CC-BY-SA,CC-BY-NC,CC-BY-NC-SA", | |
| 38 | + "locc": "Q,R,T,H,G,B,L,Z", | |
| 39 | + "max_mb": 25.0, | |
| 40 | + "pool_only": false, | |
| 41 | + "report": null | |
| 42 | + }, | |
| 43 | + "stats": { | |
| 44 | + "list_rows": 9733, | |
| 45 | + "pre2018_rows": 3941, | |
| 46 | + "attempted": 4, | |
| 47 | + "too_large": 0, | |
| 48 | + "books_kept": 3, | |
| 49 | + "chapters_written": 17, | |
| 50 | + "license_excluded": 0, | |
| 51 | + "license_unknown": 1, | |
| 52 | + "not_pre2018_xml": 0, | |
| 53 | + "no_body_chapters": 0, | |
| 54 | + "errors": 0, | |
| 55 | + "by_license": { | |
| 56 | + "PUBLIC-DOMAIN": 2, | |
| 57 | + "UNKNOWN": 1, | |
| 58 | + "CC-BY": 1 | |
| 59 | + }, | |
| 60 | + "by_publisher": { | |
| 61 | + "Agency for Healthcare Research and Quality (US)": 2, | |
| 62 | + "Springer": 1 | |
| 63 | + }, | |
| 64 | + "by_era": { | |
| 65 | + "2010-2017": 3 | |
| 66 | + }, | |
| 67 | + "ftp": { | |
| 68 | + "requests": 8, | |
| 69 | + "retries": 0, | |
| 70 | + "bytes": 8529230 | |
| 71 | + }, | |
| 72 | + "elapsed_s": 14.5 | |
| 73 | + } | |
| 74 | +} | |
added
data/manifests/discover_books_gutenberg_pool_seed1.json
+154 −0
@@ -0,0 +1,154 @@ | ||
| 1 | +{ | |
| 2 | + "created_at": "2026-11-09T17:21:53Z", | |
| 3 | + "git": { | |
| 4 | + "commit": "HEAD", | |
| 5 | + "branch": "HEAD", | |
| 6 | + "dirty": true | |
| 7 | + }, | |
| 8 | + "libraries": { | |
| 9 | + "python": "3.14.4", | |
| 10 | + "torch": "2.12.0", | |
| 11 | + "numpy": "2.5.4", | |
| 12 | + "anthropic": "2.3.0", | |
| 13 | + "transformers": "5.8.1", | |
| 14 | + "sentencepiece": "0.2.2", | |
| 15 | + "lxml": "6.1.3", | |
| 16 | + "pyarrow": "24.0.0", | |
| 17 | + "sacrebleu": "2.6.0", | |
| 18 | + "datasketch": "2.0.0", | |
| 19 | + "rapidfuzz": "3.14.6" | |
| 20 | + }, | |
| 21 | + "hardware": { | |
| 22 | + "platform": "macOS-27.0-arm64-arm-64bit-Mach-O", | |
| 23 | + "machine": "arm64", | |
| 24 | + "cpu_count": 18, | |
| 25 | + "chip": "Apple M5 Max", | |
| 26 | + "ram_gb": 48.0, | |
| 27 | + "torch_mps": true, | |
| 28 | + "torch_cuda": false | |
| 29 | + }, | |
| 30 | + "script": "discover_books", | |
| 31 | + "args": { | |
| 32 | + "source": "gutenberg", | |
| 33 | + "max_docs": 12, | |
| 34 | + "max_attempts": null, | |
| 35 | + "seed": 1, | |
| 36 | + "out": "data/raw/books", | |
| 37 | + "codes": "CC0,PUBLIC-DOMAIN,CC-BY,CC-BY-SA,CC-BY-NC,CC-BY-NC-SA", | |
| 38 | + "locc": "Q,R,T,H,G,B,L,Z", | |
| 39 | + "max_mb": 25.0, | |
| 40 | + "pool_only": true, | |
| 41 | + "report": null | |
| 42 | + }, | |
| 43 | + "stats": { | |
| 44 | + "catalog": { | |
| 45 | + "url": "https://www.gutenberg.org/cache/epub/feeds/pg_catalog.csv", | |
| 46 | + "path": "data/raw/books/_catalog/pg_catalog.csv", | |
| 47 | + "cached": true, | |
| 48 | + "sha256": "8c06ec1c46751dc184b640b8a8ed5842c873ae3b621d575f8e76837f12f33eb6", | |
| 49 | + "retrieved_or_cached_at": "2026-11-09T17:21:52Z" | |
| 50 | + }, | |
| 51 | + "catalog_rows": 79769, | |
| 52 | + "text_en": 61977, | |
| 53 | + "eligible_pool": 7799, | |
| 54 | + "pool_by_letter": { | |
| 55 | + "Q": 2186, | |
| 56 | + "H": 1653, | |
| 57 | + "T": 1051, | |
| 58 | + "B": 1016, | |
| 59 | + "R": 620, | |
| 60 | + "Z": 490, | |
| 61 | + "G": 481, | |
| 62 | + "L": 302 | |
| 63 | + }, | |
| 64 | + "pool_by_class": { | |
| 65 | + "QH": 543, | |
| 66 | + "QL": 491, | |
| 67 | + "Z": 490, | |
| 68 | + "BF": 372, | |
| 69 | + "HV": 356, | |
| 70 | + "B": 344, | |
| 71 | + "HQ": 341, | |
| 72 | + "G": 256, | |
| 73 | + "BJ": 234, | |
| 74 | + "Q": 202, | |
| 75 | + "HD": 190, | |
| 76 | + "RC": 190, | |
| 77 | + "QC": 182, | |
| 78 | + "QK": 180, | |
| 79 | + "T": 167, | |
| 80 | + "TT": 164, | |
| 81 | + "QE": 161, | |
| 82 | + "HE": 156, | |
| 83 | + "QB": 139, | |
| 84 | + "LB": 132, | |
| 85 | + "TS": 130, | |
| 86 | + "QA": 120, | |
| 87 | + "RA": 112, | |
| 88 | + "HX": 99, | |
| 89 | + "GN": 99, | |
| 90 | + "TL": 96, | |
| 91 | + "TK": 89, | |
| 92 | + "HF": 87, | |
| 93 | + "RM": 84, | |
| 94 | + "HB": 81, | |
| 95 | + "GT": 81, | |
| 96 | + "TP": 80, | |
| 97 | + "QD": 78, | |
| 98 | + "R": 77, | |
| 99 | + "HG": 76, | |
| 100 | + "TJ": 63, | |
| 101 | + "HN": 62, | |
| 102 | + "QP": 61, | |
| 103 | + "HT": 55, | |
| 104 | + "LC": 55, | |
| 105 | + "TA": 50, | |
| 106 | + "LA": 48, | |
| 107 | + "HM": 47, | |
| 108 | + "HC": 43, | |
| 109 | + "TN": 42, | |
| 110 | + "TH": 38, | |
| 111 | + "TR": 37, | |
| 112 | + "RD": 34, | |
| 113 | + "LF": 33, | |
| 114 | + "BD": 33, | |
| 115 | + "RG": 32, | |
| 116 | + "TC": 32, | |
| 117 | + "TF": 27, | |
| 118 | + "HS": 26, | |
| 119 | + "TD": 25, | |
| 120 | + "BC": 24, | |
| 121 | + "GB": 19, | |
| 122 | + "LD": 16, | |
| 123 | + "RJ": 16, | |
| 124 | + "H": 16, | |
| 125 | + "RS": 16, | |
| 126 | + "QR": 15, | |
| 127 | + "HJ": 14, | |
| 128 | + "QM": 14, | |
| 129 | + "RT": 11, | |
| 130 | + "RE": 11, | |
| 131 | + "GA": 10, | |
| 132 | + "BH": 9, | |
| 133 | + "RZ": 8, | |
| 134 | + "RK": 8, | |
| 135 | + "RX": 7, | |
| 136 | + "RB": 7, | |
| 137 | + "LT": 7, | |
| 138 | + "GC": 6, | |
| 139 | + "TE": 6, | |
| 140 | + "LH": 6, | |
| 141 | + "GF": 5, | |
| 142 | + "GR": 5, | |
| 143 | + "TG": 5, | |
| 144 | + "HA": 4, | |
| 145 | + "L": 4, | |
| 146 | + "RL": 3, | |
| 147 | + "RF": 3, | |
| 148 | + "RV": 1, | |
| 149 | + "LE": 1 | |
| 150 | + }, | |
| 151 | + "pool_with_any_author_life_dates": 5929, | |
| 152 | + "elapsed_s": 0.3 | |
| 153 | + } | |
| 154 | +} | |
added
data/manifests/discover_books_gutenberg_seed1.json
+187 −0
@@ -0,0 +1,187 @@ | ||
| 1 | +{ | |
| 2 | + "created_at": "2026-11-09T17:20:58Z", | |
| 3 | + "git": { | |
| 4 | + "commit": "HEAD", | |
| 5 | + "branch": "HEAD", | |
| 6 | + "dirty": true | |
| 7 | + }, | |
| 8 | + "libraries": { | |
| 9 | + "python": "3.14.4", | |
| 10 | + "torch": "2.12.0", | |
| 11 | + "numpy": "2.5.4", | |
| 12 | + "anthropic": "2.3.0", | |
| 13 | + "transformers": "5.8.1", | |
| 14 | + "sentencepiece": "0.2.2", | |
| 15 | + "lxml": "6.1.3", | |
| 16 | + "pyarrow": "24.0.0", | |
| 17 | + "sacrebleu": "2.6.0", | |
| 18 | + "datasketch": "2.0.0", | |
| 19 | + "rapidfuzz": "3.14.6" | |
| 20 | + }, | |
| 21 | + "hardware": { | |
| 22 | + "platform": "macOS-27.0-arm64-arm-64bit-Mach-O", | |
| 23 | + "machine": "arm64", | |
| 24 | + "cpu_count": 18, | |
| 25 | + "chip": "Apple M5 Max", | |
| 26 | + "ram_gb": 48.0, | |
| 27 | + "torch_mps": true, | |
| 28 | + "torch_cuda": false | |
| 29 | + }, | |
| 30 | + "script": "discover_books", | |
| 31 | + "args": { | |
| 32 | + "source": "gutenberg", | |
| 33 | + "max_docs": 12, | |
| 34 | + "max_attempts": null, | |
| 35 | + "seed": 1, | |
| 36 | + "out": "data/raw/books", | |
| 37 | + "codes": "CC0,PUBLIC-DOMAIN,CC-BY,CC-BY-SA,CC-BY-NC,CC-BY-NC-SA", | |
| 38 | + "locc": "Q,R,T,H,G,B,L,Z", | |
| 39 | + "max_mb": 25.0, | |
| 40 | + "pool_only": false, | |
| 41 | + "report": null | |
| 42 | + }, | |
| 43 | + "stats": { | |
| 44 | + "catalog_rows": 79769, | |
| 45 | + "candidates": 7799, | |
| 46 | + "attempted": 12, | |
| 47 | + "downloaded": 12, | |
| 48 | + "already_present": 0, | |
| 49 | + "not_found": 0, | |
| 50 | + "too_large": 0, | |
| 51 | + "copyrighted_header": 0, | |
| 52 | + "no_pd_header": 0, | |
| 53 | + "no_date": 1, | |
| 54 | + "date_from_frontmatter": 10, | |
| 55 | + "date_from_death_year": 1, | |
| 56 | + "errors": 0, | |
| 57 | + "by_locc": { | |
| 58 | + "LH": 1, | |
| 59 | + "TS": 1, | |
| 60 | + "GT": 1, | |
| 61 | + "RC": 1, | |
| 62 | + "Z": 1, | |
| 63 | + "BJ": 1, | |
| 64 | + "QH": 1, | |
| 65 | + "HB": 1, | |
| 66 | + "LC": 1, | |
| 67 | + "TJ": 1, | |
| 68 | + "G": 1, | |
| 69 | + "RT": 1 | |
| 70 | + }, | |
| 71 | + "by_era": { | |
| 72 | + "unknown": 1, | |
| 73 | + "pre1900": 5, | |
| 74 | + "1900-1949": 5, | |
| 75 | + "1950-1979": 1 | |
| 76 | + }, | |
| 77 | + "by_date_rule": { | |
| 78 | + "none": 1, | |
| 79 | + "frontmatter:imprint": 5, | |
| 80 | + "frontmatter:copyright": 3, | |
| 81 | + "frontmatter:published": 1, | |
| 82 | + "frontmatter:roman": 1, | |
| 83 | + "death_year": 1 | |
| 84 | + }, | |
| 85 | + "pool_by_class": { | |
| 86 | + "QH": 543, | |
| 87 | + "QL": 491, | |
| 88 | + "Z": 490, | |
| 89 | + "BF": 372, | |
| 90 | + "HV": 356, | |
| 91 | + "B": 344, | |
| 92 | + "HQ": 341, | |
| 93 | + "G": 256, | |
| 94 | + "BJ": 234, | |
| 95 | + "Q": 202, | |
| 96 | + "HD": 190, | |
| 97 | + "RC": 190, | |
| 98 | + "QC": 182, | |
| 99 | + "QK": 180, | |
| 100 | + "T": 167, | |
| 101 | + "TT": 164, | |
| 102 | + "QE": 161, | |
| 103 | + "HE": 156, | |
| 104 | + "QB": 139, | |
| 105 | + "LB": 132, | |
| 106 | + "TS": 130, | |
| 107 | + "QA": 120, | |
| 108 | + "RA": 112, | |
| 109 | + "HX": 99, | |
| 110 | + "GN": 99, | |
| 111 | + "TL": 96, | |
| 112 | + "TK": 89, | |
| 113 | + "HF": 87, | |
| 114 | + "RM": 84, | |
| 115 | + "HB": 81, | |
| 116 | + "GT": 81, | |
| 117 | + "TP": 80, | |
| 118 | + "QD": 78, | |
| 119 | + "R": 77, | |
| 120 | + "HG": 76, | |
| 121 | + "TJ": 63, | |
| 122 | + "HN": 62, | |
| 123 | + "QP": 61, | |
| 124 | + "HT": 55, | |
| 125 | + "LC": 55, | |
| 126 | + "TA": 50, | |
| 127 | + "LA": 48, | |
| 128 | + "HM": 47, | |
| 129 | + "HC": 43, | |
| 130 | + "TN": 42, | |
| 131 | + "TH": 38, | |
| 132 | + "TR": 37, | |
| 133 | + "RD": 34, | |
| 134 | + "LF": 33, | |
| 135 | + "BD": 33, | |
| 136 | + "RG": 32, | |
| 137 | + "TC": 32, | |
| 138 | + "TF": 27, | |
| 139 | + "HS": 26, | |
| 140 | + "TD": 25, | |
| 141 | + "BC": 24, | |
| 142 | + "GB": 19, | |
| 143 | + "LD": 16, | |
| 144 | + "RJ": 16, | |
| 145 | + "H": 16, | |
| 146 | + "RS": 16, | |
| 147 | + "QR": 15, | |
| 148 | + "HJ": 14, | |
| 149 | + "QM": 14, | |
| 150 | + "RT": 11, | |
| 151 | + "RE": 11, | |
| 152 | + "GA": 10, | |
| 153 | + "BH": 9, | |
| 154 | + "RZ": 8, | |
| 155 | + "RK": 8, | |
| 156 | + "RX": 7, | |
| 157 | + "RB": 7, | |
| 158 | + "LT": 7, | |
| 159 | + "GC": 6, | |
| 160 | + "TE": 6, | |
| 161 | + "LH": 6, | |
| 162 | + "GF": 5, | |
| 163 | + "GR": 5, | |
| 164 | + "TG": 5, | |
| 165 | + "HA": 4, | |
| 166 | + "L": 4, | |
| 167 | + "RL": 3, | |
| 168 | + "RF": 3, | |
| 169 | + "RV": 1, | |
| 170 | + "LE": 1 | |
| 171 | + }, | |
| 172 | + "catalog": { | |
| 173 | + "url": "https://www.gutenberg.org/cache/epub/feeds/pg_catalog.csv", | |
| 174 | + "path": "data/raw/books/_catalog/pg_catalog.csv", | |
| 175 | + "cached": false, | |
| 176 | + "last_modified": "Sun, 08 Nov 2026 22:43:51 GMT", | |
| 177 | + "sha256": "8c06ec1c46751dc184b640b8a8ed5842c873ae3b621d575f8e76837f12f33eb6", | |
| 178 | + "retrieved_or_cached_at": "2026-11-09T17:20:17Z" | |
| 179 | + }, | |
| 180 | + "http": { | |
| 181 | + "requests": 13, | |
| 182 | + "retries": 0, | |
| 183 | + "bytes": 25370564 | |
| 184 | + }, | |
| 185 | + "elapsed_s": 43.3 | |
| 186 | + } | |
| 187 | +} | |
added
data/manifests/discover_pmc_seed1.json
+65 −0
@@ -0,0 +1,65 @@ | ||
| 1 | +{ | |
| 2 | + "created_at": "2026-11-09T17:02:18Z", | |
| 3 | + "git": { | |
| 4 | + "commit": null, | |
| 5 | + "branch": null, | |
| 6 | + "dirty": false | |
| 7 | + }, | |
| 8 | + "libraries": { | |
| 9 | + "python": "3.14.4", | |
| 10 | + "torch": "2.12.0", | |
| 11 | + "numpy": "2.5.4", | |
| 12 | + "anthropic": "2.3.0", | |
| 13 | + "transformers": "5.8.1", | |
| 14 | + "sentencepiece": "0.2.2", | |
| 15 | + "lxml": "6.1.3", | |
| 16 | + "pyarrow": "24.0.0", | |
| 17 | + "sacrebleu": "2.6.0", | |
| 18 | + "datasketch": "2.0.0", | |
| 19 | + "rapidfuzz": "3.14.6" | |
| 20 | + }, | |
| 21 | + "hardware": { | |
| 22 | + "platform": "macOS-27.0-arm64-arm-64bit-Mach-O", | |
| 23 | + "machine": "arm64", | |
| 24 | + "cpu_count": 18, | |
| 25 | + "chip": "Apple M5 Max", | |
| 26 | + "ram_gb": 48.0, | |
| 27 | + "torch_mps": true, | |
| 28 | + "torch_cuda": false | |
| 29 | + }, | |
| 30 | + "script": "discover_pmc", | |
| 31 | + "args": { | |
| 32 | + "candidates": 500, | |
| 33 | + "max_docs": 100, | |
| 34 | + "seed": 1, | |
| 35 | + "out": "data/raw/pmc_smoke", | |
| 36 | + "codes": "CC0,PUBLIC-DOMAIN,CC-BY,CC-BY-SA,CC-BY-NC,CC-BY-NC-SA", | |
| 37 | + "report": null | |
| 38 | + }, | |
| 39 | + "stats": { | |
| 40 | + "candidates": 347, | |
| 41 | + "missing": 200, | |
| 42 | + "not_open_access": 30, | |
| 43 | + "retracted": 0, | |
| 44 | + "license_excluded": 16, | |
| 45 | + "manuscript": 1, | |
| 46 | + "fetched": 100, | |
| 47 | + "md5_fail": 0, | |
| 48 | + "errors": 0, | |
| 49 | + "by_license": { | |
| 50 | + "CC-BY-NC-ND": 4, | |
| 51 | + "CC-BY": 75, | |
| 52 | + "CC-BY-NC-SA": 11, | |
| 53 | + "UNKNOWN": 12, | |
| 54 | + "TDM": 31, | |
| 55 | + "CC0": 1, | |
| 56 | + "CC-BY-NC": 13 | |
| 57 | + }, | |
| 58 | + "http": { | |
| 59 | + "requests": 447, | |
| 60 | + "retries": 0, | |
| 61 | + "bytes": 8326427 | |
| 62 | + }, | |
| 63 | + "elapsed_s": 39.4 | |
| 64 | + } | |
| 65 | +} | |
added
data/manifests/discover_pmc_seed2.json
+66 −0
@@ -0,0 +1,66 @@ | ||
| 1 | +{ | |
| 2 | + "created_at": "2026-11-09T17:19:27Z", | |
| 3 | + "git": { | |
| 4 | + "commit": "HEAD", | |
| 5 | + "branch": "HEAD", | |
| 6 | + "dirty": true | |
| 7 | + }, | |
| 8 | + "libraries": { | |
| 9 | + "python": "3.14.4", | |
| 10 | + "torch": "2.12.0", | |
| 11 | + "numpy": "2.5.4", | |
| 12 | + "anthropic": "2.3.0", | |
| 13 | + "transformers": "5.8.1", | |
| 14 | + "sentencepiece": "0.2.2", | |
| 15 | + "lxml": "6.1.3", | |
| 16 | + "pyarrow": "24.0.0", | |
| 17 | + "sacrebleu": "2.6.0", | |
| 18 | + "datasketch": "2.0.0", | |
| 19 | + "rapidfuzz": "3.14.6" | |
| 20 | + }, | |
| 21 | + "hardware": { | |
| 22 | + "platform": "macOS-27.0-arm64-arm-64bit-Mach-O", | |
| 23 | + "machine": "arm64", | |
| 24 | + "cpu_count": 18, | |
| 25 | + "chip": "Apple M5 Max", | |
| 26 | + "ram_gb": 48.0, | |
| 27 | + "torch_mps": true, | |
| 28 | + "torch_cuda": false | |
| 29 | + }, | |
| 30 | + "script": "discover_pmc", | |
| 31 | + "args": { | |
| 32 | + "candidates": 6000, | |
| 33 | + "max_docs": 1200, | |
| 34 | + "seed": 2, | |
| 35 | + "out": "data/raw/pmc", | |
| 36 | + "codes": "CC0,PUBLIC-DOMAIN,CC-BY,CC-BY-SA,CC-BY-NC,CC-BY-NC-SA", | |
| 37 | + "report": null | |
| 38 | + }, | |
| 39 | + "stats": { | |
| 40 | + "candidates": 4187, | |
| 41 | + "missing": 2367, | |
| 42 | + "not_open_access": 395, | |
| 43 | + "retracted": 1, | |
| 44 | + "license_excluded": 210, | |
| 45 | + "manuscript": 14, | |
| 46 | + "fetched": 1200, | |
| 47 | + "md5_fail": 0, | |
| 48 | + "errors": 0, | |
| 49 | + "by_license": { | |
| 50 | + "CC0": 26, | |
| 51 | + "CC-BY": 842, | |
| 52 | + "CC-BY-NC": 210, | |
| 53 | + "TDM": 405, | |
| 54 | + "CC-BY-NC-SA": 126, | |
| 55 | + "UNKNOWN": 137, | |
| 56 | + "CC-BY-NC-ND": 73, | |
| 57 | + "CC-BY-ND": 1 | |
| 58 | + }, | |
| 59 | + "http": { | |
| 60 | + "requests": 5387, | |
| 61 | + "retries": 0, | |
| 62 | + "bytes": 104199229 | |
| 63 | + }, | |
| 64 | + "elapsed_s": 391.1 | |
| 65 | + } | |
| 66 | +} | |
added
data/pairs/.gitkeep
+0 −0
added
data/parsed/.gitkeep
+0 −0
added
data/raw/.gitkeep
+0 −0
added
data/segments/.gitkeep
+0 −0
added
data/splits/.gitkeep
+0 −0
added
docs/MANDATE.md
+195 −0
@@ -0,0 +1,195 @@ | ||
| 1 | +# Author mandate — WP-AI-0002 (verbatim, supplied 2026-11-09) | |
| 2 | + | |
| 3 | +> Reproduced verbatim from the author's CLAUDE.md brief. Author additions during the first session (2026-11-09): | |
| 4 | +> "mets aussi accent sur le fait que le modèle sera utilisé par la suite" (the released model must be genuinely | |
| 5 | +> usable), "mets aussi article en économie et finance" (include economics and finance articles), "ça doit être | |
| 6 | +> quand même large" (the corpus must be large), "tu peux installer torch ailleurs si nécessaire" (PyTorch may be | |
| 7 | +> installed on other cluster nodes), "ça peut aussi être des livres de référence en anglais pré-2018" (English | |
| 8 | +> reference books published before 2018 are in scope). | |
| 9 | + | |
| 10 | +CLAUDE.md — WP-AI-0002: Human Text Reconstruction Transformer (HTR) | |
| 11 | + | |
| 12 | +Mission | |
| 13 | + | |
| 14 | +Build a rigorous, reproducible research program and releasable model family investigating whether a Transformer can reconstruct characteristics of pre-2018 human-authored scientific prose from LLM-rewritten passages. Work under the UQO Research project, in a dedicated wp-ai-0002/ directory. Never overwrite other research projects. Deliver BOTH a peer-review-ready paper and a usable, documented model release. | |
| 15 | + | |
| 16 | +Scientific caution: Exact recovery of an original passage from a paraphrase is generally non-identifiable. Do not claim restoration of actual human authorship or reliable AI detection. Separate verbatim reconstruction, stylistic distribution matching, and factual fidelity. Test whether the apparent gains exceed simple copying and generic paraphrasing. | |
| 17 | + | |
| 18 | +Non-negotiable principles | |
| 19 | + | |
| 20 | +1. Research integrity over impressive-looking results. Never fabricate data, metrics, experiments, sources, licenses, or model checkpoints. | |
| 21 | +2. Open access is not synonymous with open licensing. Ingest, transform, train on, and redistribute only where applicable licenses, provider terms, and institutional policies permit. Track rights at the document and derivative level. | |
| 22 | +3. Preserve equations, citations, numerical values, units, negation, uncertainty, and domain terminology; identify and report failures. | |
| 23 | +4. Avoid leakage across training, validation, and test: split at least by article, and add author-group, publication-family, and near-duplicate controls. | |
| 24 | +5. Store all runs, configurations, provenance, random seeds, data hashes, environment versions, and checkpoint metadata. | |
| 25 | +6. Make all expensive API calls and large training runs opt-in, with dry runs, quotas, budgets, retries, and resumability. | |
| 26 | +7. Use local resources first where appropriate, but do not assume available compute or hardware. Detect macOS/MLX/MPS and CUDA and benchmark before scaling. | |
| 27 | +8. Respect robots.txt, rate limits, and repository bulk-access policies. Prefer official bulk datasets and APIs. | |
| 28 | +9. Do not claim that generated text is human-authored; outputs are synthetic reconstructions. | |
| 29 | +10. A paper and an independently runnable model are mandatory final artifacts. | |
| 30 | + | |
| 31 | +Research questions | |
| 32 | + | |
| 33 | +• RQ1: What syntactic, lexical, rhetorical, and discourse features change systematically when an LLM rewrites historical scientific prose? | |
| 34 | +• RQ2: Which pre-rewrite features can be recovered from rewritten text beyond an identity baseline? | |
| 35 | +• RQ3: How does recovery vary across disciplines, publication periods, writing styles, and paragraph/section types? | |
| 36 | +• RQ4: Does a model trained on Claude rewrites generalize to rewrites from held-out prompts and other LLM families? | |
| 37 | +• RQ5: How does performance change with data scale, model capacity, and style-conditioning? | |
| 38 | +• RQ6: What factual distortions, memorization risks, and irreversible information losses remain? | |
| 39 | + | |
| 40 | +Repository structure | |
| 41 | + | |
| 42 | +wp-ai-0002/ | |
| 43 | + CLAUDE.md | |
| 44 | + README.md | |
| 45 | + pyproject.toml | |
| 46 | + configs/{data,rewriting,training,evaluation,release}/ | |
| 47 | + src/htr/{ingest,licenses,parse,segment,dedup,rewrite,validate,data, | |
| 48 | + models,training,inference,evaluation,release,utils}/ | |
| 49 | + scripts/ | |
| 50 | + tests/{unit,integration,regression}/ | |
| 51 | + data/{manifests,raw,parsed,segments,pairs,splits}/ | |
| 52 | + artifacts/{checkpoints,metrics,figures,tables,reports}/ | |
| 53 | + paper/{main.tex,references.bib,figures,tables,supplement.tex} | |
| 54 | + model_cards/ | |
| 55 | + demo/ | |
| 56 | + docs/{architecture,data_governance,experiments,reproducibility}/ | |
| 57 | + | |
| 58 | +Do not commit raw licensed papers, credentials, paid API outputs, or large checkpoints by default. Use gitignore and configurable storage paths. | |
| 59 | + | |
| 60 | +Phase 0 — Audit and feasibility | |
| 61 | + | |
| 62 | +Inspect existing UQO Research structure and available hardware, Python environments, storage, credentials (never print secrets), scholarly datasets, and project conventions. Create a dependency/compute/cost feasibility report. Identify any preexisting project ID collisions before proceeding. Write an executable roadmap with decision gates and acceptance criteria. | |
| 63 | + | |
| 64 | +Phase 1 — Historical corpus | |
| 65 | + | |
| 66 | +Target scientific papers published before 2018-01-01; publication date must be verified, not inferred from an upload or revision timestamp. Prioritize clearly reusable sources, e.g., appropriately licensed arXiv material and PMC Open Access subsets; verify each item and version individually. Record DOI, title, authors, venue, publication year, subject area, source URL, license URI, license evidence, acquisition time, document hash, and permitted operations. | |
| 67 | + | |
| 68 | +Implement metadata-first discovery; license filtering; official API/bulk acquisition; PDF/XML/LaTeX parsing; section and paragraph extraction; language identification; citation/math/table preservation; reference-list exclusion; deduplication; and quality scoring. Preserve document, section, paragraph, and sentence hierarchy. Keep raw and normalized text with offsets and provenance. | |
| 69 | + | |
| 70 | +Target stages, not promised counts: | |
| 71 | + | |
| 72 | +• Pilot: ~1,000 papers / ~20,000 eligible passages. | |
| 73 | +• Medium: ~10,000 papers / ~200,000 passages. | |
| 74 | +• Large: ~100,000 papers / ~2,000,000 passages. | |
| 75 | +• Stretch: ~500,000 papers / ~10,000,000 passages. | |
| 76 | +Actual eligible counts depend on rights, parsing quality, costs, and budget. | |
| 77 | + | |
| 78 | +Build a sampling plan stratified by discipline, year, section role, length, language, and venue. Avoid domination by one archive. Create document/author/near-duplicate-safe splits before generating rewrites. Hold out entire domains and, if possible, venue/author groups. | |
| 79 | + | |
| 80 | +Phase 2 — Claude rewriting engine | |
| 81 | + | |
| 82 | +Build a configurable, versioned provider adapter for Anthropic Claude. Pin and log the actual model identifier, prompt version, sampling parameters, request ID, token usage, cost, and timestamps. Do not assume any model identifier or price remains current: retrieve current provider documentation during implementation. | |
| 83 | + | |
| 84 | +Generate controlled rewrite variants: | |
| 85 | +A. Neutral academic paraphrase. | |
| 86 | +B. More polished and standardized academic prose. | |
| 87 | +C. Concise rewrite. | |
| 88 | +D. Structure-preserving rewrite. | |
| 89 | +E. Diverse sampling with repeated independent generations. | |
| 90 | + | |
| 91 | +Explicitly instruct preservation of all scientific propositions, numbers, units, named entities, equations, references, uncertainty, and negation. Retain both original and rewrite and never overwrite either. Record failures and abstentions. Support concurrency limits, exponential backoff, checkpointing, deduped requests, caching, cost ceilings, and pause/resume. No paid batch beyond the approved pilot budget without explicit authorization. | |
| 92 | + | |
| 93 | +Phase 3 — Pair quality and integrity | |
| 94 | + | |
| 95 | +Implement automated checks: language consistency, semantic similarity, bidirectional entailment where feasible, numeric and unit invariance, named entity checks, equation/citation consistency, negation and hedging checks, length ratios, and parsing integrity. Use manual expert review on stratified samples. Flag, quarantine, and report invalid pairs rather than silently dropping them. Report acceptance rates and selection bias. Version the data schema (JSONL/Parquet) with immutable IDs and source traceability. | |
| 96 | + | |
| 97 | +Phase 4 — Architectures | |
| 98 | + | |
| 99 | +Implement strong baselines before novel architectures: | |
| 100 | + | |
| 101 | +• B0: Identity/copy input. | |
| 102 | +• B1: Generic paraphrase or minimally trained rewrite baseline. | |
| 103 | +• B2: Standard pretrained encoder-decoder fine-tuning (e.g. T5-family where rights and resources permit). | |
| 104 | +• M1: From-scratch encoder-decoder Transformer at pilot scale. | |
| 105 | +• M2: Style-conditioned encoder-decoder with learned discipline, era, and section-role embeddings. | |
| 106 | +• M3: Dual-latent Human Text Reconstruction Transformer (main research hypothesis). | |
| 107 | + | |
| 108 | +M3 design: | |
| 109 | + | |
| 110 | +• Semantic encoder E_c(x) extracts content representations from LLM-rewritten input. | |
| 111 | +• Style branch E_s(x, metadata) predicts a distribution over historical writing styles using training-only style supervision; at inference never feed the unknown original target. | |
| 112 | +• Cross-attentive Transformer decoder D(E_c, E_s, metadata) produces reconstructed historical-style prose. | |
| 113 | +• Consider gradient reversal, orthogonality regularization, or contrastive objectives to discourage style-content entanglement; verify with ablations rather than assume disentanglement. | |
| 114 | +• Include optional pointer/copy mechanisms and constrained preservation of citations, equations, and numerical spans. | |
| 115 | +• For section-level work, implement hierarchical chunking, context windows, overlap management, and consistency checks across paragraphs. | |
| 116 | + | |
| 117 | +Candidate training objective: | |
| 118 | +L = λ_ce L_token + λ_sem L_semantic + λ_style L_style + λ_con L_contrastive + λ_fact L_factual. | |
| 119 | +Define each term mathematically and operationally, document differentiability, optimize weights only on validation, and compare with token-loss-only training. No use of held-out originals in inference. | |
| 120 | + | |
| 121 | +Phase 5 — Model family and training | |
| 122 | + | |
| 123 | +Target sizes are approximate design points, not promised parameter counts: | |
| 124 | + | |
| 125 | +• HTR-Tiny: ~30M | |
| 126 | +• HTR-Small: ~125M | |
| 127 | +• HTR-Base: ~350M (primary release target) | |
| 128 | +• HTR-Large: ~1B (conditional on measured feasibility) | |
| 129 | + | |
| 130 | +Use PyTorch as the portable reference; support MPS and CUDA, and investigate MLX only where justified by performance and compatibility. Implement mixed precision when safe, gradient accumulation, clipping, efficient batching, activation checkpointing, distributed training only when available, early stopping, LR scheduling, and resumable checkpoints. Include data-loader tests, one-batch overfit test, smoke training, profiling, memory estimation, and loss/gradient diagnostics. Train baselines under comparable compute/data budgets; publish FLOPs, accelerator-hours, energy estimates if measurable, and API expenditure. | |
| 131 | + | |
| 132 | +Phase 6 — Evaluation | |
| 133 | + | |
| 134 | +Report both instance-level and corpus-level outcomes: | |
| 135 | + | |
| 136 | +• Exact-match, edit distance, chrF, ROUGE/BLEU (with clear limitations). | |
| 137 | +• Semantic preservation: embedding similarity, entailment, expert judgments. | |
| 138 | +• Factual integrity: numerical/citation/equation consistency, contradiction rates. | |
| 139 | +• Style fidelity: lexical diversity, sentence-length distributions, syntactic complexity, discourse markers, calibrated historical-style classifiers, and distributional distances. | |
| 140 | +• Authorship/style attribution only as exploratory analysis; never equate classifier scores with proof of origin. | |
| 141 | +• Memorization: training-set nearest-neighbor overlap, canary checks where lawful, contamination audits. | |
| 142 | +• Human blind evaluations on fidelity, fluency, authenticity of historical style, and factual safety. | |
| 143 | + | |
| 144 | +Core tests: document-safe held-out evaluation; author-safe splits; held-out discipline and time; held-out rewriting prompt; cross-provider generalization (only with legally/contractually permitted outputs); model scaling; rewrite strength; input length; error propagation; confidence calibration; ablations of all style and loss components; multiple random seeds and confidence intervals. Report null and negative findings. | |
| 145 | + | |
| 146 | +Explicitly test non-identifiability: generate multiple valid rewrites from distinct originals and measure ambiguity. Do not frame lower lexical overlap as proof of a failure to recover style. | |
| 147 | + | |
| 148 | +Phase 7 — Model and benchmark release | |
| 149 | + | |
| 150 | +Prepare a real usable model, not only a paper figure: | |
| 151 | + | |
| 152 | +• SafeTensors weights, tokenizer, configs, model architecture, inference and batch scripts. | |
| 153 | +• Hugging Face-compatible repository structure and model cards for each checkpoint. | |
| 154 | +• Versioned API/CLI, installation instructions, examples, deterministic inference options. | |
| 155 | +• Reproducible train/eval scripts and manifests. | |
| 156 | +• Separate benchmark documentation, data cards, and licensing matrix. | |
| 157 | +• Lightweight web demo showing input, reconstruction, differences, and preservation warnings. | |
| 158 | +• Explicit limitations, intended research uses, provenance, ethical risks, and license restrictions. | |
| 159 | +• Security/privacy/copyright review before publishing weights or any source/derivative text. | |
| 160 | +• If full training data cannot be shared, release only permitted subsets and metadata/scripts that comply with source and provider conditions. | |
| 161 | +• Do not upload models, datasets, or spend cloud money without explicit user approval. | |
| 162 | + | |
| 163 | +Proposed naming: HTR-Tiny, HTR-Small, HTR-Base, HTR-Large, HTR-Benchmark. Model names and hosting namespace require final confirmation. | |
| 164 | + | |
| 165 | +Phase 8 — Scientific paper | |
| 166 | + | |
| 167 | +Write a complete English LaTeX paper, working title: | |
| 168 | +"Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing." | |
| 169 | + | |
| 170 | +Required sections: Abstract; Introduction; Related Work (historical style, paraphrase inversion, style transfer, authorship, synthetic text effects); Research Questions and Hypotheses; Data and Rights; Rewrite Protocol; Models and Objectives; Experimental Design; Results; Ablations; Robustness; Human Evaluation; Threats to Validity; Ethics and Licensing; Limitations; Conclusion; Reproducibility Checklist. Add architecture diagrams, dataset flowchart, sample audit, scaling plots, error taxonomy, quantitative tables, confidence intervals, compute/cost disclosures, and appendices with prompts/hyperparameters. Cite verified primary sources; never fabricate references or results. Assess ACL/EMNLP/TACL suitability only after genuine experiments. | |
| 171 | + | |
| 172 | +Implementation milestones / acceptance gates | |
| 173 | + | |
| 174 | +Gate A: repository + license governance + 100-paper ingestion smoke test + parsed quality audit. | |
| 175 | +Gate B: 500–1,000 validated Claude pairs with documented spend and human review. | |
| 176 | +Gate C: leakage-safe splits, reproducible baselines, initial benchmark and statistical plan. | |
| 177 | +Gate D: working M1/M2/M3 training, one-batch overfit, reproducible pilot results. | |
| 178 | +Gate E: larger training only if pilots justify cost and quality; publish all negative findings. | |
| 179 | +Gate F: independent inference from a clean environment, full model card, audited release candidate, and compiled paper. | |
| 180 | + | |
| 181 | +Each gate must output a brief GO/NO-GO report, measured metrics, issues, and next action. Never mark a gate complete without evidence. | |
| 182 | + | |
| 183 | +First execution instructions for Claude Code | |
| 184 | + | |
| 185 | +1. Inspect the existing repository and identify the correct UQO Research location. | |
| 186 | +2. Create wp-ai-0002/ without modifying unrelated projects. | |
| 187 | +3. Produce docs/architecture/RESEARCH_PLAN.md, docs/data_governance/LICENSE_MATRIX.md, docs/experiments/EXPERIMENT_MATRIX.md, docs/reproducibility/COMPUTE_BUDGET.md, and an implementation checklist. | |
| 188 | +4. Implement the minimum vertical slice: one licensed document → parsed paragraphs → versioned Claude rewrite (dry-run/mock first) → validated pair → train/val/test split → baseline inference → metrics. | |
| 189 | +5. Run unit/integration tests and report actual outputs, failures, and costs. | |
| 190 | +6. Request approval before any material API spend, mass acquisition, public upload, or long training job. | |
| 191 | +7. Continue iteratively toward trained checkpoints and a scientifically defensible paper. | |
| 192 | + | |
| 193 | +Definition of done | |
| 194 | + | |
| 195 | +The project is complete only when (i) data rights and provenance are auditable, (ii) results are reproducible and beat appropriate baselines on pre-registered primary metrics or report convincingly why not, (iii) an independently loadable trained model and complete model card exist, (iv) the benchmark is legally distributable or an explicitly documented compliant substitute exists, and (v) the English LaTeX manuscript compiles with genuine figures, tables, citations, and honest limitations. | |
added
docs/RESEARCH_GAP.md
+124 −0
@@ -0,0 +1,124 @@ | ||
| 1 | +# WP-AI-0002 — Research gap and positioning | |
| 2 | + | |
| 3 | +*Human Text Reconstruction: Learning the Inverse of LLM-Induced Transformations in Scientific Writing* | |
| 4 | + | |
| 5 | +Literature snapshot: 2026-11-09. All bibkeys refer to `paper/references.bib`; each entry was checked against an arXiv API or OpenAlex record, logged in `docs/literature/verified_sources.jsonl` (85 entries). Claims below about a paper's findings come from the abstract returned by the API at verification time. Claims about full-text details are marked *to check*. | |
| 6 | + | |
| 7 | +--- | |
| 8 | + | |
| 9 | +## 1. How LLM rewriting changes scientific prose | |
| 10 | + | |
| 11 | +**Prevalence is large and rising, and it is measured through style rather than content.** Population-level estimators show steady growth in LLM-modified text in arXiv, bioRxiv and Nature-portfolio papers, with the fastest growth in computer science [liang2025quantifying]. The same framework applied to peer reviews finds a similar effect [liang2024monitoring]. Excess-vocabulary analysis of more than 15 million PubMed abstracts gives a lower bound of 13.5% of 2024 abstracts processed with LLMs, reaching 40% in some subcorpora [kobak2025delving]. Word-frequency models of about one million arXiv abstracts put the share of "LLM-style" computer-science abstracts at about 35%, using GPT-3.5's response to "revise the following sentences" as the baseline [geng2024chatgpt]. Keyword counts give a conservative estimate of at least 60,000 LLM-assisted papers in 2023 [gray2024chatgpt]. | |
| 12 | + | |
| 13 | +**The signature is lexical, grammatical and rhetorical.** | |
| 14 | +- **Lexical.** Specific "focal" words (e.g. *delve*, *intricate*, *underscore*) are over-represented. Their source is unresolved, and RLHF is a plausible contributor [juzek2024delve]. Authors appear to co-evolve with detectors: *delve* fell after it was publicised, while *significant* kept rising [geng2025humanllm]. | |
| 15 | +- **Grammatical and rhetorical.** Biber-feature analysis shows systematic human–LLM differences. These differences are larger for instruction-tuned than for base models [reinhart2025llms]. Comparable contrasts appear in news text [munozortiz2023contrasting]. | |
| 16 | +- **Long-run arXiv shifts.** Across 823,798 arXiv abstracts, lexical complexity rises and syntactic complexity falls [bao2025examining]. | |
| 17 | +- **Recent ML venues (2020–2025).** Formulaic phrases rose by up to 10× and hedging declined by 23% [jang2026scizoom]. | |
| 18 | + | |
| 19 | +**Homogenisation is plausible but contested.** Co-writing with an instruction-tuned model reduces lexical and content diversity across authors [padmakumar2023writing]. In contrast, a 44-year analysis of CHI papers finds that prose did *not* homogenise: LLMs accelerated a word-level shift that had already started [kosch2026chi]. These results matter for us because they imply an era confound. Pre-2018 originals differ from post-2022 human text for reasons unrelated to LLMs, such as long-run readability decline [plavensigray2017readability], rising informality [hyland2016informal] and changing stance marking [hyland2016stance]. That confound is the reason WP-AI-0002 fixes originals to pre-2018 rights-verified text and generates the LLM side under control. Human reviewers and detectors also struggle to tell ChatGPT-generated scientific abstracts from real ones [gao2023comparing]. | |
| 20 | + | |
| 21 | +**Gap (i).** Almost all of this evidence is *observational and corpus-level*: word frequencies, aggregate shares and feature means. No study we found measures, for the *same* passage, which features an LLM rewrite changes, by how much, and conditionally on discipline, section and rewrite instruction. RQ1 needs exactly that. The paired, controlled design (one original, five rewrite variants) closes this gap. | |
| 22 | + | |
| 23 | +## 2. Detection, paraphrase attacks and inversion attempts | |
| 24 | + | |
| 25 | +**Detectors are brittle under rewriting.** Zero-shot and trained detectors are well studied: [mitchell2023detectgpt], [bao2023fastdetectgpt], [hans2024binoculars], with watermarking in [kirchenbauer2023watermark]. Their limits are also well documented: | |
| 26 | +- **Paraphrasing defeats detection.** DIPPER cut DetectGPT's detection accuracy from 70.3% to 4.6% at 1% FPR [krishna2023paraphrasing]. Recursive paraphrasing evades watermarking, neural and retrieval detectors [sadasivan2023reliably]. | |
| 27 | +- **Detectors are biased against non-native writers** [liang2023detectors]. | |
| 28 | +- **Robustness collapses across generators and attacks** on shared benchmarks [dugan2024raid]. "Humanising" attacks are evaluated systematically in [zheng2025thbench]. | |
| 29 | + | |
| 30 | +**Polished and revised text is the hard case.** AI-*polished* human text — the regime closest to our rewrite variants — is frequently flagged as AI-generated even when edits are minimal. Detectors also cannot separate degrees of AI involvement [saha2025polished]. Machine-*revised* text (rewriting, expansion, polishing) defeats standard detectors. Style-preference alignment (ImBD) improves AUC by 5–19 points [chen2024imitate]. Expert-edited machine outputs evade detection, while LLM-edited texts are rarely judged human [artemova2024beemo]. | |
| 31 | + | |
| 32 | +**Inversion, the direct antecedent.** Rivera Soto, Chen and Andrews introduce **paraphrase inversion**: given a paraphrase, recover an approximation of the original. They treat it as translation from paraphrase back to source, trained on synthetic (original, paraphrase) pairs [riverasoto2024paraphrase]. They report three findings: | |
| 33 | +1. GPT-4 and Llama-3 paraphrase with learnable biases. | |
| 34 | +2. Inversion models generalise to paraphrasers unseen during training. | |
| 35 | +3. Combining inversion with detectors gives an average +22% AUROC across seven detectors and three domains. | |
| 36 | + | |
| 37 | +A 2026 follow-up trains a style encoder to *reconstruct human-authored text from its machine paraphrase* while a semantic encoder is frozen. The frozen encoder pushes style into a separate channel. The learned style representations transfer to authorship verification [riverasoto2026unsupervised]. This is architecturally very close to our dual-latent content/style model. Both papers, however, treat inversion as a *means to detection* and report AUROC. They do not evaluate the reconstruction itself: | |
| 38 | +- no factual-integrity audit (numbers, citations, equations, negation, hedging); | |
| 39 | +- no stylometric distribution matching; | |
| 40 | +- no memorisation controls; | |
| 41 | +- no discipline or era stratification *(to check in the full text)*. | |
| 42 | + | |
| 43 | +**Related inversions.** Inversion of non-textual channels is now routine. Text can be reconstructed from embeddings [morris2023embeddings] and prompts from next-token distributions [morris2023lminversion]. Both show that "lossy" representations keep more recoverable information than expected, which supports the RQ2 prior that some pre-rewrite signal survives. | |
| 44 | + | |
| 45 | +**Adjacent style tasks.** Authorship obfuscation and style rewriting are the forward direction of our problem [fisher2024styleremix]. They rely on style representations that are separated from topic [riverasoto2021luar], [wegmann2022same], and on classical stylometry [burrows2002delta], [stamatatos2008survey], [koppel2008computational], [juola2008authorship]. Style-transfer research supplies: | |
| 46 | +- the paraphrase-based formulation we invert [krishna2020reformulating]; | |
| 47 | +- disentangled and few-shot restyling [shen2017style], [fu2017style], [john2018disentangled], [riley2020textsettr]; | |
| 48 | +- low-resource *authorship* style transfer, the closest analogue to "restore this author's voice" [patel2022lowresource]; | |
| 49 | +- evaluation practice for style, content and fluency [mir2019evaluating], [jin2020survey]; | |
| 50 | +- formality transfer data [rao2018gyafc]. | |
| 51 | + | |
| 52 | +**Gap (ii).** Inversion exists as a detection auxiliary, but no published work, as far as we found: | |
| 53 | +- (a) evaluates **reconstruction fidelity as the target**, beyond identity and copy baselines; | |
| 54 | +- (b) uses **rights-verified, pre-LLM scientific originals** with controlled rewrite variants; | |
| 55 | +- (c) audits **factual drift** introduced or repaired by inversion; | |
| 56 | +- (d) tests **generalisation across rewrite prompts and LLM families** as an explicit research question (RQ4) rather than a robustness footnote. | |
| 57 | + | |
| 58 | +## 3. Non-identifiability: what cannot be recovered | |
| 59 | + | |
| 60 | +Inversion of a many-to-one map is ill-posed. Four strands of theory bound what any model can recover: | |
| 61 | +- **Unsupervised disentanglement is impossible without inductive biases on both model and data** [locatello2018challenging]. Nonlinear ICA is non-identifiable in general [hyvarinen1999nonlinear]. | |
| 62 | +- **Identifiability returns with paired views.** If content is invariant across augmentations, the content block is identifiable up to an invertible map [vonkugelgen2021selfsupervised]. Contrastive objectives invert the generating process under similar assumptions [zimmermann2021contrastive]. Our (original, rewrite₁…₅) tuples are such paired views. In this setting *content* is the invariant part and *style* is what the LLM changes. The theory therefore suggests that content is identifiable, while the original style is identifiable only as far as the rewrite leaves traces of it. | |
| 63 | +- **Adversarial removal of an attribute is often incomplete.** Gradient reversal reaches chance-level adversary accuracy while a post-hoc probe still recovers the attribute [elazar2018adversarial]. This warns against reading "disentangled" off the adversary loss alone in our gradient-reversal / domain-separation objectives [ganin2014unsupervised], [bousmalis2016domain], [hu2017controlled]. | |
| 64 | +- **Recursive training on synthetic text erodes the tails of the distribution** [shumailov2024collapse], [alemohammad2023mad], [dohmatob2024tails]. If a rewrite destroys tail stylistic variation, no inverse can restore it beyond the prior. A reconstruction model will then regress toward a "mean human style". Our stylometric distribution-matching metrics, rather than point metrics, are designed to detect this. | |
| 65 | + | |
| 66 | +**Gap (iii).** Prior inversion work does not separate *recoverable* from *non-identifiable* components, and it does not report where reconstruction amounts to hallucinating a plausible human style. WP-AI-0002 makes this an explicit evaluation axis through: | |
| 67 | +- the identity baseline; | |
| 68 | +- the multiple-rewrite consistency test; | |
| 69 | +- distributional (not only per-sample) stylometry; | |
| 70 | +- memorisation probes, because a pretrained decoder could "reconstruct" by recalling a memorised original [carlini2020extracting], [carlini2022quantifying], [lee2021dedup], [magar2022contamination], [sainz2023contamination]. | |
| 71 | + | |
| 72 | +## 4. What WP-AI-0002 adds | |
| 73 | + | |
| 74 | +1. **Inverse-mapping benchmark with rights-verified originals.** Pre-2018 human prose is drawn from open-licence scientific sources, with licence provenance logged per passage. Our corpus differs from general LLM corpora on two counts: licensing audits [longpre2023provenance] and documentation of large text corpora [gao2020pile], [soldaini2024dolma]. Scientific full-text corpora provide the parsing precedent [lo2019s2orc], [saier2023unarxive]. Five controlled Claude rewrite variants per passage support paired, per-passage measurement of LLM-induced change (RQ1, RQ3). | |
| 75 | +2. **Dual-latent reconstruction model.** An encoder–decoder [vaswani2017attention], [raffel2019t5], [lewis2019bart] with copy and preservation mechanisms [see2017pointer], [gu2016copying]. Content and style latents are separated by gradient reversal, orthogonality [bousmalis2016domain] and contrastive [oord2018cpc], [chen2020simclr] objectives. Ablations go against from-scratch and style-conditioned baselines (RQ2, RQ5). Implementation choices follow [kudo2018sentencepiece], [xiong2020layernorm], [su2021roformer]. | |
| 76 | +3. **Factual-integrity evaluation.** We evaluate numbers, citations, equations, negation and hedging alongside lexical metrics ([popovic2015chrf], [papineni2002bleu], [lin2004rouge]) and semantic metrics ([zhang2019bertscore], [reimers2019sentencebert]). This answers RQ6 on factual distortions, a dimension none of the inversion papers report. Hedging is singled out because LLM-era prose shows reduced hedging [jang2026scizoom] and because stance marking is itself era-dependent [hyland2016stance]. | |
| 77 | +4. **Explicit non-identifiability and memorisation controls** (Section 3), and **cross-prompt / cross-family generalisation** as a primary RQ (RQ4). The generalisation result reported by [riverasoto2024paraphrase] is the hypothesis to replicate on scientific text. | |
| 78 | + | |
| 79 | +## 5. Closest prior work | |
| 80 | + | |
| 81 | +| bibkey | What they did | What they reported | Difference from WP-AI-0002 | | |
| 82 | +|---|---|---|---| | |
| 83 | +| riverasoto2024paraphrase | Paraphrase inversion as paraphrase→original translation on synthetic pairs | LLM paraphrase biases are learnable; generalises to unseen paraphrasers; +22% AUROC for detectors | Goal is detection, not fidelity; not scientific-specific; no factual / stylometric / memorisation audit | | |
| 84 | +| riverasoto2026unsupervised | Style encoder trained to reconstruct human text from machine paraphrase with frozen semantic encoder | Strong few/zero-shot AI-text detection; transfers to authorship verification | Near-identical content/style split, but reconstruction is a pretext task; no rights-verified scientific corpus or fidelity evaluation | | |
| 85 | +| krishna2020reformulating | Style transfer as paraphrase-then-restyle (STRAP) | Strong unsupervised style transfer via inverse paraphrasing per style | Target style is a corpus style, not the specific original passage | | |
| 86 | +| krishna2023paraphrasing | DIPPER paraphraser vs. detectors; retrieval defence | DetectGPT 70.3%→4.6% at 1% FPR | Forward attack; no inversion | | |
| 87 | +| saha2025polished | APT-Eval: 14.7k AI-polished human texts at graded involvement | Detectors over-flag minimal polishing and miss degrees | Measures detectability, not recoverability | | |
| 88 | +| chen2024imitate | ImBD: style-preference-aligned detection of machine-revised text | +13% AUC (open models), +5/+19% (GPT-3.5/4o) | Detects revision; does not undo it | | |
| 89 | +| artemova2024beemo | Expert-edited vs LLM-edited machine outputs benchmark | Expert editing evades detectors | Edit direction is human-over-machine; no inversion | | |
| 90 | +| patel2022lowresource | Low-resource authorship style transfer | Non-famous authors can be partially imitated | Imitates an author's style, not a specific original text | | |
| 91 | +| morris2023embeddings | Iterative inversion of text from embeddings | Embeddings reveal near-exact text | Inverts a representation, not an LLM rewrite | | |
| 92 | +| geng2024chatgpt / kobak2025delving | Corpus-level excess-word estimation of LLM-style in abstracts | ~35% LLM-style CS abstracts; ≥13.5% PubMed 2024 | Observational aggregate; no per-passage paired measurement or inversion | | |
| 93 | +| vonkugelgen2021selfsupervised | Identifiability of content vs style from paired views | Content block identifiable up to invertible map | Theory our design instantiates; no text inversion | | |
| 94 | + | |
| 95 | +**Assessment.** Rivera Soto et al. (2024, 2026) already show that LLM paraphrases can be partly inverted and that a reconstruction objective with a frozen semantic channel separates style. WP-AI-0002 must cite both prominently. Its novelty claim has to rest on four points: | |
| 96 | +- reconstruction fidelity as the object of study; | |
| 97 | +- rights-verified pre-2018 *scientific* originals with controlled variants; | |
| 98 | +- factual-integrity and non-identifiability evaluation; | |
| 99 | +- the discipline, era and section analysis. | |
| 100 | + | |
| 101 | +It should not claim to be first to invert LLM paraphrases. | |
| 102 | + | |
| 103 | +## 6. Unverified or rejected candidates (not in `references.bib`) | |
| 104 | + | |
| 105 | +- **Koehn (2004), "Statistical significance tests for machine translation evaluation"** (EMNLP; bootstrap resampling). Neither the OpenAlex nor the Crossref search returned the record. Verify via ACL Anthology W04-3250 before citing. | |
| 106 | +- **"Lu et al. (2019), 'Scientific writing is becoming harder to read'".** No matching record found. This probably conflates Plavén-Sigray et al. (2017), which *is* verified as `plavensigray2017readability`. | |
| 107 | +- **Hyland & Jiang, DOI 10.1016/j.esp.2017.06.001.** The DOI resolves to Larsson (2017), a different paper. It was replaced by the verified `hyland2016informal` (10.1016/j.esp.2016.09.001). | |
| 108 | +- **arXiv 2105.07263.** It resolved to Khan et al., "A Deep Metric Learning Approach to Account Linking", not LUAR. LUAR was verified instead via its EMNLP 2021 DOI (`riverasoto2021luar`). | |
| 109 | +- **GPTZero-type evaluations.** No peer-reviewed primary record was searched. GPTZero is covered indirectly in [krishna2023paraphrasing] and [chen2024imitate]. | |
| 110 | +- **Non-bibliographic resources (no API record possible), to cite as URLs/footnotes after manual check:** | |
| 111 | + - PMC Open Access Subset documentation; | |
| 112 | + - arXiv licence documentation; | |
| 113 | + - Creative Commons legal code (CC BY 4.0, CC0); | |
| 114 | + - Claude model documentation. | |
| 115 | +- **Verified but left out for scope** (re-add if needed; records available from the verification run): | |
| 116 | + - text-generation and dataset papers: Bender et al. 2021; Herbold et al. 2023 (Sci. Rep.); Ghostbuster (2305.15047); M4 (2305.14902); GLTR (1906.04043); Wu et al. survey (2310.14724); | |
| 117 | + - paraphrase and style resources: ParaNMT (1711.05732); SCPN (1804.06059); INLP (2004.07667); Subramanian et al. (1811.00552); Prabhumoye et al. (1804.09000); | |
| 118 | + - theory and evaluation: Khemakhem et al. iVAE (1907.04809); Mosteller & Wallace 1963; sacreBLEU (1804.08771); | |
| 119 | + - homogenisation, collapse and attacks: Guo et al. linguistic-diversity collapse (2311.09807); Agarwal et al. homogenisation (2409.11360); JAMDEC (2402.08761); Lu et al. evasion (2305.10847); Kirchenbauer et al. watermark reliability (2306.04634); PADBen (2511.00416); Zhou et al. humanising attack (2404.01907); | |
| 120 | + - recent prevalence work: Gray 2025 prevalence (2512.01560); Lee 2026 Korean excess vocabulary (2609.07447). | |
| 121 | + | |
| 122 | +**Metadata caveats.** | |
| 123 | +- arXiv-only entries carry the arXiv year and no venue. Several are published (e.g. T5 in JMLR, BART at ACL 2020, LUAR at EMNLP), so upgrade venues after a Crossref check before submission. | |
| 124 | +- OpenAlex years for Stamatatos (2008) and Juola (2008) are online-first or volume years. They differ from the commonly cited 2009 and 2006. | |
added
docs/gates/GATE_A.md
+50 −0
@@ -0,0 +1,50 @@ | ||
| 1 | +# Gate A — Repository, licence governance, 100-paper ingestion smoke test, parsed quality audit | |
| 2 | + | |
| 3 | +**Date:** 2026-11-09 · **Decision: GO** (with the caveats listed below) · Evidence: `data/manifests/discover_pmc_seed1.json`, | |
| 4 | +`data/parsed/smoke/build_summary.json`, `data/splits/smoke/splits_summary.json`, `data/pairs/mock_smoke/validated/validation_summary.json`, | |
| 5 | +unit tests (`.venv/bin/python -m pytest tests/unit -q`). | |
| 6 | + | |
| 7 | +## Measured | |
| 8 | + | |
| 9 | +| Item | Value | | |
| 10 | +|---|---| | |
| 11 | +| PMCID candidates sampled (stratified over PMC10000–PMC5760000) | 347 | | |
| 12 | +| Metadata objects found (404 = id unused/removed) | 147 | | |
| 13 | +| Licence distribution of found objects | {'CC-BY-NC-ND': 4, 'CC-BY': 75, 'CC-BY-NC-SA': 11, 'UNKNOWN': 12, 'TDM': 31, 'CC0': 1, 'CC-BY-NC': 13} | | |
| 14 | +| Excluded by licence policy (ND / TDM / unknown) | 16 | | |
| 15 | +| Documents fetched (JATS XML, MD5 verified against provider digest) | 100 in 39.4 s | | |
| 16 | +| Parsed without error | 100 / 100 | | |
| 17 | +| Publication date verified from JATS `pub-date` and < 2018-01-01 | 100 / 100 (0 undated, 0 post-2017) | | |
| 18 | +| Licences of kept documents | {'CC-BY': 75, 'CC-BY-NC-SA': 11, 'CC-BY-NC': 13, 'CC0': 1} | | |
| 19 | +| Discipline (coarse, from journal/subject hints) | {'biomedicine': 75, 'life_sciences': 9, 'social_sciences': 7, 'computer_science': 4, 'chemistry': 4, 'physics': 1} | | |
| 20 | +| Paragraphs extracted / eligible passages (60–400 words, ≥ 2 sentences, English ≥ 0.9, quality ≥ 0.6) | 2964 / 2013 | | |
| 21 | +| Eligible passages by section role | {'abstract': 102, 'background': 74, 'results': 475, 'discussion': 389, 'conclusion': 36, 'methods': 367, 'other': 473, 'data': 50, 'theory': 21, 'related_work': 4, 'limitations': 7, 'introduction': 15} | | |
| 22 | +| Near-duplicate clusters (MinHash, Jaccard ≥ 0.8) | exact 0, near 0 | | |
| 23 | +| Leakage-safe split of the 100 docs (seed 20261109, author-group safe, physics + pre-1990 held out) | {'heldout_domain': 1, 'heldout_period': 12, 'test': 12, 'train': 69, 'val': 6}, group leaks: [] | | |
| 24 | +| Mock rewrite run (zero cost) 60 passages × 6 generations | 360 rewrites, acceptance 0.7917 (fails = near-verbatim mock outputs) | | |
| 25 | + | |
| 26 | +## Quality audit (manual, 2026-11-09) | |
| 27 | + | |
| 28 | +- Citations: numeric xrefs are normalised to `[n]`; author–year citations are kept as text; both are detected by `protect()` and | |
| 29 | + replaced by `⟦Cn⟧` placeholders before any LLM call. A doubled-bracket bug (`[[16]]`) was found and fixed in this gate. | |
| 30 | +- Section roles: an "Introduction" → "other" bug caused by the numbering-stripping regex (it ate the leading "I") was found by | |
| 31 | + auditing the 40 most frequent titles of "other" passages, and fixed. Remaining "other" passages are untitled sections or | |
| 32 | + topical headings (e.g. "The Follow-up Period"), which is the expected residue. | |
| 33 | +- Historical back-scanned articles (1877–1986, OCR text) are present in PMC OA under CC licences; they pass the quality gate | |
| 34 | + rarely (2 eligible passages from 6 such documents), so OCR noise is not a dominant risk but must be re-checked at scale. | |
| 35 | +- Captions and list items are extracted but never eligible; references, acknowledgements, funding, contributions, | |
| 36 | + ethics, supplementary sections are excluded. | |
| 37 | + | |
| 38 | +## Issues / risks carried forward | |
| 39 | + | |
| 40 | +1. The PMC FTP bulk packages and the OA Web Service were retired in August 2026; the only official bulk channel is the | |
| 41 | + AWS Open Data bucket `pmc-oa-opendata` (per-article JSON + JATS). The id-range sampling prior must be re-validated when | |
| 42 | + scaling (date is verified per document anyway). | |
| 43 | +2. Discipline labels for PMC come from journal/subject keywords — coarse. Non-biomedical strata (economics, finance, | |
| 44 | + mathematics, physics, computer science, books) come from the other adapters (arXiv, Crossref journals, Gutenberg/Bookshelf). | |
| 45 | +3. `TDM` author manuscripts and `CC-BY-ND` articles are ingested as metadata only (31 + 4 of 147 found objects). | |
| 46 | + | |
| 47 | +## Next action | |
| 48 | + | |
| 49 | +Gate B: pilot rewrite of ~1 000 validated passages with Claude (budget ceiling 30 USD, Message Batches), human review of a | |
| 50 | +stratified sample, acceptance-rate report. | |
added
docs/literature/verified_sources.jsonl
+85 −0
@@ -0,0 +1,85 @@ | ||
| 1 | +{"bibkey": "liang2025quantifying", "title": "Quantifying large language model usage in scientific papers", "first_author": "Weixin Liang", "year": 2025, "venue": "Nature Human Behaviour", "doi": "10.1038/s41562-025-02273-8", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1038/s41562-025-02273-8", "verified_at": "2026-11-09T17:12:41Z", "theme": "LLM influence on scientific writing"} | |
| 2 | +{"bibkey": "kobak2025delving", "title": "Delving into LLM-assisted writing in biomedical publications through excess vocabulary", "first_author": "Dmitry Kobak", "year": 2025, "venue": "Science Advances", "doi": "10.1126/sciadv.adt3813", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1126/sciadv.adt3813", "verified_at": "2026-11-09T17:12:41Z", "theme": "LLM influence on scientific writing"} | |
| 3 | +{"bibkey": "geng2024chatgpt", "title": "Is ChatGPT Transforming Academics' Writing Style?", "first_author": "Mingmeng Geng", "year": 2024, "venue": "arXiv preprint arXiv:2404.08627", "doi": null, "arxiv_id": "2404.08627", "verified_via": "http://export.arxiv.org/api/query?id_list=2404.08627", "verified_at": "2026-11-09T17:09:49Z", "theme": "LLM influence on scientific writing"} | |
| 4 | +{"bibkey": "geng2025humanllm", "title": "Human-LLM Coevolution: Evidence from Academic Writing", "first_author": "Mingmeng Geng", "year": 2025, "venue": "arXiv preprint arXiv:2502.09606", "doi": null, "arxiv_id": "2502.09606", "verified_via": "http://export.arxiv.org/api/query?id_list=2502.09606", "verified_at": "2026-11-09T17:09:49Z", "theme": "LLM influence on scientific writing"} | |
| 5 | +{"bibkey": "gray2024chatgpt", "title": "ChatGPT \"contamination\": estimating the prevalence of LLMs in the scholarly literature", "first_author": "Andrew Gray", "year": 2024, "venue": "arXiv preprint arXiv:2403.16887", "doi": null, "arxiv_id": "2403.16887", "verified_via": "http://export.arxiv.org/api/query?id_list=2403.16887", "verified_at": "2026-11-09T17:09:49Z", "theme": "LLM influence on scientific writing"} | |
| 6 | +{"bibkey": "liang2024monitoring", "title": "Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews", "first_author": "Weixin Liang", "year": 2024, "venue": "arXiv preprint arXiv:2403.07183", "doi": null, "arxiv_id": "2403.07183", "verified_via": "http://export.arxiv.org/api/query?id_list=2403.07183", "verified_at": "2026-11-09T17:09:49Z", "theme": "LLM influence on scientific writing"} | |
| 7 | +{"bibkey": "bao2025examining", "title": "Examining Linguistic Shifts in Academic Writing Before and After the Launch of ChatGPT: A Study on Preprint Papers", "first_author": "Tong Bao", "year": 2025, "venue": "Scientometrics,2025", "doi": null, "arxiv_id": "2505.12218", "verified_via": "http://export.arxiv.org/api/query?id_list=2505.12218", "verified_at": "2026-11-09T17:09:49Z", "theme": "LLM influence on scientific writing"} | |
| 8 | +{"bibkey": "juzek2024delve", "title": "Why Does ChatGPT \"Delve\" So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models", "first_author": "Tom S. Juzek", "year": 2024, "venue": "arXiv preprint arXiv:2412.11385", "doi": null, "arxiv_id": "2412.11385", "verified_via": "http://export.arxiv.org/api/query?id_list=2412.11385", "verified_at": "2026-11-09T17:09:56Z", "theme": "LLM influence on scientific writing"} | |
| 9 | +{"bibkey": "reinhart2025llms", "title": "Do LLMs write like humans? Variation in grammatical and rhetorical styles", "first_author": "Alex Reinhart", "year": 2025, "venue": "Proceedings of the National Academy of Sciences", "doi": "10.1073/pnas.2422455122", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1073/pnas.2422455122", "verified_at": "2026-11-09T17:12:41Z", "theme": "LLM influence on scientific writing"} | |
| 10 | +{"bibkey": "munozortiz2023contrasting", "title": "Contrasting Linguistic Patterns in Human and LLM-Generated News Text", "first_author": "Alberto Muñoz-Ortiz", "year": 2023, "venue": "Artificial Intelligence Review 57, 265 (2024)", "doi": "10.1007/s10462-024-10903-2", "arxiv_id": "2308.09067", "verified_via": "http://export.arxiv.org/api/query?id_list=2308.09067", "verified_at": "2026-11-09T17:09:56Z", "theme": "LLM influence on scientific writing"} | |
| 11 | +{"bibkey": "padmakumar2023writing", "title": "Does Writing with Language Models Reduce Content Diversity?", "first_author": "Vishakh Padmakumar", "year": 2023, "venue": "arXiv preprint arXiv:2309.05196", "doi": null, "arxiv_id": "2309.05196", "verified_via": "http://export.arxiv.org/api/query?id_list=2309.05196", "verified_at": "2026-11-09T17:09:49Z", "theme": "LLM influence on scientific writing"} | |
| 12 | +{"bibkey": "jang2026scizoom", "title": "SciZoom: A Large-scale Benchmark for Hierarchical Scientific Summarization across the LLM Era", "first_author": "Han Jang", "year": 2026, "venue": "arXiv preprint arXiv:2603.16131", "doi": null, "arxiv_id": "2603.16131", "verified_via": "http://export.arxiv.org/api/query?id_list=2603.16131", "verified_at": "2026-11-09T17:12:04Z", "theme": "LLM influence on scientific writing"} | |
| 13 | +{"bibkey": "kosch2026chi", "title": "How Did Writing Change At CHI? Analyzing 44 Years of CHI Writing Before and After the Introduction of Large Language Models", "first_author": "Thomas Kosch", "year": 2026, "venue": "arXiv preprint arXiv:2609.23090", "doi": null, "arxiv_id": "2609.23090", "verified_via": "http://export.arxiv.org/api/query?id_list=2609.23090", "verified_at": "2026-11-09T17:12:04Z", "theme": "LLM influence on scientific writing"} | |
| 14 | +{"bibkey": "gao2023comparing", "title": "Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers", "first_author": "Catherine A. Gao", "year": 2023, "venue": "npj Digital Medicine", "doi": "10.1038/s41746-023-00819-6", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1038/s41746-023-00819-6", "verified_at": "2026-11-09T17:12:41Z", "theme": "LLM influence on scientific writing"} | |
| 15 | +{"bibkey": "krishna2023paraphrasing", "title": "Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense", "first_author": "Kalpesh Krishna", "year": 2023, "venue": "arXiv preprint arXiv:2303.13408", "doi": null, "arxiv_id": "2303.13408", "verified_via": "http://export.arxiv.org/api/query?id_list=2303.13408", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 16 | +{"bibkey": "sadasivan2023reliably", "title": "Can AI-Generated Text be Reliably Detected?", "first_author": "Vinu Sankar Sadasivan", "year": 2023, "venue": "arXiv preprint arXiv:2303.11156", "doi": null, "arxiv_id": "2303.11156", "verified_via": "http://export.arxiv.org/api/query?id_list=2303.11156", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 17 | +{"bibkey": "mitchell2023detectgpt", "title": "DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature", "first_author": "Eric Mitchell", "year": 2023, "venue": "arXiv preprint arXiv:2301.11305", "doi": null, "arxiv_id": "2301.11305", "verified_via": "http://export.arxiv.org/api/query?id_list=2301.11305", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 18 | +{"bibkey": "bao2023fastdetectgpt", "title": "Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature", "first_author": "Guangsheng Bao", "year": 2023, "venue": "arXiv preprint arXiv:2310.05130", "doi": null, "arxiv_id": "2310.05130", "verified_via": "http://export.arxiv.org/api/query?id_list=2310.05130", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 19 | +{"bibkey": "hans2024binoculars", "title": "Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text", "first_author": "Abhimanyu Hans", "year": 2024, "venue": "arXiv preprint arXiv:2401.12070", "doi": null, "arxiv_id": "2401.12070", "verified_via": "http://export.arxiv.org/api/query?id_list=2401.12070", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 20 | +{"bibkey": "liang2023detectors", "title": "GPT detectors are biased against non-native English writers", "first_author": "Weixin Liang", "year": 2023, "venue": "arXiv preprint arXiv:2304.02819", "doi": null, "arxiv_id": "2304.02819", "verified_via": "http://export.arxiv.org/api/query?id_list=2304.02819", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 21 | +{"bibkey": "kirchenbauer2023watermark", "title": "A Watermark for Large Language Models", "first_author": "John Kirchenbauer", "year": 2023, "venue": "arXiv preprint arXiv:2301.10226", "doi": null, "arxiv_id": "2301.10226", "verified_via": "http://export.arxiv.org/api/query?id_list=2301.10226", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 22 | +{"bibkey": "dugan2024raid", "title": "RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors", "first_author": "Liam Dugan", "year": 2024, "venue": "arXiv preprint arXiv:2405.07940", "doi": null, "arxiv_id": "2405.07940", "verified_via": "http://export.arxiv.org/api/query?id_list=2405.07940", "verified_at": "2026-11-09T17:09:52Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 23 | +{"bibkey": "saha2025polished", "title": "Almost AI, Almost Human: The Challenge of Detecting AI-Polished Writing", "first_author": "Shoumik Saha", "year": 2025, "venue": "arXiv preprint arXiv:2502.15666", "doi": null, "arxiv_id": "2502.15666", "verified_via": "http://export.arxiv.org/api/query?id_list=2502.15666", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 24 | +{"bibkey": "chen2024imitate", "title": "Imitate Before Detect: Aligning Machine Stylistic Preference for Machine-Revised Text Detection", "first_author": "Jiaqi Chen", "year": 2024, "venue": "arXiv preprint arXiv:2412.10432", "doi": null, "arxiv_id": "2412.10432", "verified_via": "http://export.arxiv.org/api/query?id_list=2412.10432", "verified_at": "2026-11-09T17:12:04Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 25 | +{"bibkey": "artemova2024beemo", "title": "Beemo: Benchmark of Expert-edited Machine-generated Outputs", "first_author": "Ekaterina Artemova", "year": 2024, "venue": "arXiv preprint arXiv:2411.04032", "doi": null, "arxiv_id": "2411.04032", "verified_via": "http://export.arxiv.org/api/query?id_list=2411.04032", "verified_at": "2026-11-09T17:12:04Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 26 | +{"bibkey": "zheng2025thbench", "title": "TH-Bench: Evaluating Evading Attacks via Humanizing AI Text on Machine-Generated Text Detectors", "first_author": "Jingyi Zheng", "year": 2025, "venue": "arXiv preprint arXiv:2503.08708", "doi": null, "arxiv_id": "2503.08708", "verified_via": "http://export.arxiv.org/api/query?id_list=2503.08708", "verified_at": "2026-11-09T17:09:49Z", "theme": "AI-text detection and paraphrase attacks"} | |
| 27 | +{"bibkey": "stamatatos2008survey", "title": "A survey of modern authorship attribution methods", "first_author": "Efstathios Stamatatos", "year": 2008, "venue": "Journal of the American Society for Information Science and Technology", "doi": "10.1002/asi.21001", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1002/asi.21001", "verified_at": "2026-11-09T17:12:41Z", "theme": "Stylometry and authorship"} | |
| 28 | +{"bibkey": "juola2008authorship", "title": "Authorship Attribution", "first_author": "Patrick Juola", "year": 2008, "venue": "Foundations and Trends® in Information Retrieval", "doi": "10.1561/1500000005", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1561/1500000005", "verified_at": "2026-11-09T17:12:41Z", "theme": "Stylometry and authorship"} | |
| 29 | +{"bibkey": "koppel2008computational", "title": "Computational methods in authorship attribution", "first_author": "Moshe Koppel", "year": 2008, "venue": "Journal of the American Society for Information Science and Technology", "doi": "10.1002/asi.20961", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1002/asi.20961", "verified_at": "2026-11-09T17:12:41Z", "theme": "Stylometry and authorship"} | |
| 30 | +{"bibkey": "burrows2002delta", "title": "'Delta': a Measure of Stylistic Difference and a Guide to Likely Authorship", "first_author": "J. F. Burrows", "year": 2002, "venue": "Literary and Linguistic Computing", "doi": "10.1093/llc/17.3.267", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1093/llc/17.3.267", "verified_at": "2026-11-09T17:12:41Z", "theme": "Stylometry and authorship"} | |
| 31 | +{"bibkey": "riverasoto2021luar", "title": "Learning Universal Authorship Representations", "first_author": "Rafael Rivera Soto", "year": 2021, "venue": "Conference on Empirical Methods in Natural Language Processing (EMNLP)", "doi": "10.18653/v1/2021.emnlp-main.70", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.18653/v1/2021.emnlp-main.70", "verified_at": "2026-11-09T17:12:41Z", "theme": "Stylometry and authorship"} | |
| 32 | +{"bibkey": "wegmann2022same", "title": "Same Author or Just Same Topic? Towards Content-Independent Style Representations", "first_author": "Anna Wegmann", "year": 2022, "venue": "arXiv preprint arXiv:2204.04907", "doi": null, "arxiv_id": "2204.04907", "verified_via": "http://export.arxiv.org/api/query?id_list=2204.04907", "verified_at": "2026-11-09T17:09:49Z", "theme": "Stylometry and authorship"} | |
| 33 | +{"bibkey": "fisher2024styleremix", "title": "StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements", "first_author": "Jillian Fisher", "year": 2024, "venue": "arXiv preprint arXiv:2408.15666", "doi": null, "arxiv_id": "2408.15666", "verified_via": "http://export.arxiv.org/api/query?id_list=2408.15666", "verified_at": "2026-11-09T17:09:49Z", "theme": "Stylometry and authorship"} | |
| 34 | +{"bibkey": "shen2017style", "title": "Style Transfer from Non-Parallel Text by Cross-Alignment", "first_author": "Tianxiao Shen", "year": 2017, "venue": "arXiv preprint arXiv:1705.09655", "doi": null, "arxiv_id": "1705.09655", "verified_via": "http://export.arxiv.org/api/query?id_list=1705.09655", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 35 | +{"bibkey": "fu2017style", "title": "Style Transfer in Text: Exploration and Evaluation", "first_author": "Zhenxin Fu", "year": 2017, "venue": "arXiv preprint arXiv:1711.06861", "doi": null, "arxiv_id": "1711.06861", "verified_via": "http://export.arxiv.org/api/query?id_list=1711.06861", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 36 | +{"bibkey": "john2018disentangled", "title": "Disentangled Representation Learning for Non-Parallel Text Style Transfer", "first_author": "Vineet John", "year": 2018, "venue": "arXiv preprint arXiv:1808.04339", "doi": null, "arxiv_id": "1808.04339", "verified_via": "http://export.arxiv.org/api/query?id_list=1808.04339", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 37 | +{"bibkey": "krishna2020reformulating", "title": "Reformulating Unsupervised Style Transfer as Paraphrase Generation", "first_author": "Kalpesh Krishna", "year": 2020, "venue": "arXiv preprint arXiv:2010.05700", "doi": null, "arxiv_id": "2010.05700", "verified_via": "http://export.arxiv.org/api/query?id_list=2010.05700", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 38 | +{"bibkey": "riley2020textsettr", "title": "TextSETTR: Few-Shot Text Style Extraction and Tunable Targeted Restyling", "first_author": "Parker Riley", "year": 2020, "venue": "arXiv preprint arXiv:2010.03802", "doi": null, "arxiv_id": "2010.03802", "verified_via": "http://export.arxiv.org/api/query?id_list=2010.03802", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 39 | +{"bibkey": "patel2022lowresource", "title": "Low-Resource Authorship Style Transfer: Can Non-Famous Authors Be Imitated?", "first_author": "Ajay Patel", "year": 2022, "venue": "arXiv preprint arXiv:2212.08986", "doi": null, "arxiv_id": "2212.08986", "verified_via": "http://export.arxiv.org/api/query?id_list=2212.08986", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 40 | +{"bibkey": "mir2019evaluating", "title": "Evaluating Style Transfer for Text", "first_author": "Remi Mir", "year": 2019, "venue": "arXiv preprint arXiv:1904.02295", "doi": null, "arxiv_id": "1904.02295", "verified_via": "http://export.arxiv.org/api/query?id_list=1904.02295", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 41 | +{"bibkey": "jin2020survey", "title": "Deep Learning for Text Style Transfer: A Survey", "first_author": "Di Jin", "year": 2020, "venue": "arXiv preprint arXiv:2011.00416", "doi": null, "arxiv_id": "2011.00416", "verified_via": "http://export.arxiv.org/api/query?id_list=2011.00416", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 42 | +{"bibkey": "rao2018gyafc", "title": "Dear Sir or Madam, May I introduce the GYAFC Dataset: Corpus, Benchmarks and Metrics for Formality Style Transfer", "first_author": "Sudha Rao", "year": 2018, "venue": "arXiv preprint arXiv:1803.06535", "doi": null, "arxiv_id": "1803.06535", "verified_via": "http://export.arxiv.org/api/query?id_list=1803.06535", "verified_at": "2026-11-09T17:09:52Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 43 | +{"bibkey": "riverasoto2024paraphrase", "title": "Mitigating Paraphrase Attacks on Machine-Text Detectors via Paraphrase Inversion", "first_author": "Rafael Rivera Soto", "year": 2024, "venue": "arXiv preprint arXiv:2410.21637", "doi": null, "arxiv_id": "2410.21637", "verified_via": "http://export.arxiv.org/api/query?id_list=2410.21637", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 44 | +{"bibkey": "riverasoto2026unsupervised", "title": "Unsupervised Style Representation Learning for AI-Text Detection via Paraphrase Inversion", "first_author": "Rafael Rivera Soto", "year": 2026, "venue": "arXiv preprint arXiv:2606.10099", "doi": null, "arxiv_id": "2606.10099", "verified_via": "http://export.arxiv.org/api/query?id_list=2606.10099", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 45 | +{"bibkey": "morris2023lminversion", "title": "Language Model Inversion", "first_author": "John X. Morris", "year": 2023, "venue": "arXiv preprint arXiv:2311.13647", "doi": null, "arxiv_id": "2311.13647", "verified_via": "http://export.arxiv.org/api/query?id_list=2311.13647", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 46 | +{"bibkey": "morris2023embeddings", "title": "Text Embeddings Reveal (Almost) As Much As Text", "first_author": "John X. Morris", "year": 2023, "venue": "arXiv preprint arXiv:2310.06816", "doi": null, "arxiv_id": "2310.06816", "verified_via": "http://export.arxiv.org/api/query?id_list=2310.06816", "verified_at": "2026-11-09T17:09:49Z", "theme": "Style transfer, paraphrase and inversion"} | |
| 47 | +{"bibkey": "ganin2014unsupervised", "title": "Unsupervised Domain Adaptation by Backpropagation", "first_author": "Yaroslav Ganin", "year": 2014, "venue": "arXiv preprint arXiv:1409.7495", "doi": null, "arxiv_id": "1409.7495", "verified_via": "http://export.arxiv.org/api/query?id_list=1409.7495", "verified_at": "2026-11-09T17:09:49Z", "theme": "Disentanglement and identifiability"} | |
| 48 | +{"bibkey": "bousmalis2016domain", "title": "Domain Separation Networks", "first_author": "Konstantinos Bousmalis", "year": 2016, "venue": "arXiv preprint arXiv:1608.06019", "doi": null, "arxiv_id": "1608.06019", "verified_via": "http://export.arxiv.org/api/query?id_list=1608.06019", "verified_at": "2026-11-09T17:09:49Z", "theme": "Disentanglement and identifiability"} | |
| 49 | +{"bibkey": "hu2017controlled", "title": "Toward Controlled Generation of Text", "first_author": "Zhiting Hu", "year": 2017, "venue": "arXiv preprint arXiv:1703.00955", "doi": null, "arxiv_id": "1703.00955", "verified_via": "http://export.arxiv.org/api/query?id_list=1703.00955", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 50 | +{"bibkey": "chen2020simclr", "title": "A Simple Framework for Contrastive Learning of Visual Representations", "first_author": "Ting Chen", "year": 2020, "venue": "arXiv preprint arXiv:2002.05709", "doi": null, "arxiv_id": "2002.05709", "verified_via": "http://export.arxiv.org/api/query?id_list=2002.05709", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 51 | +{"bibkey": "oord2018cpc", "title": "Representation Learning with Contrastive Predictive Coding", "first_author": "Aaron van den Oord", "year": 2018, "venue": "arXiv preprint arXiv:1807.03748", "doi": null, "arxiv_id": "1807.03748", "verified_via": "http://export.arxiv.org/api/query?id_list=1807.03748", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 52 | +{"bibkey": "elazar2018adversarial", "title": "Adversarial Removal of Demographic Attributes from Text Data", "first_author": "Yanai Elazar", "year": 2018, "venue": "arXiv preprint arXiv:1808.06640", "doi": null, "arxiv_id": "1808.06640", "verified_via": "http://export.arxiv.org/api/query?id_list=1808.06640", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 53 | +{"bibkey": "locatello2018challenging", "title": "Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations", "first_author": "Francesco Locatello", "year": 2018, "venue": "Proceedings of the 36th International Conference on Machine Learning (ICML 2019)", "doi": null, "arxiv_id": "1811.12359", "verified_via": "http://export.arxiv.org/api/query?id_list=1811.12359", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 54 | +{"bibkey": "hyvarinen1999nonlinear", "title": "Nonlinear independent component analysis: Existence and uniqueness results", "first_author": "Aapo Hyvärinen", "year": 1999, "venue": "Neural Networks", "doi": "10.1016/s0893-6080(98)00140-3", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1016/s0893-6080(98)00140-3", "verified_at": "2026-11-09T17:12:41Z", "theme": "Disentanglement and identifiability"} | |
| 55 | +{"bibkey": "vonkugelgen2021selfsupervised", "title": "Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style", "first_author": "Julius von Kügelgen", "year": 2021, "venue": "arXiv preprint arXiv:2106.04619", "doi": null, "arxiv_id": "2106.04619", "verified_via": "http://export.arxiv.org/api/query?id_list=2106.04619", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 56 | +{"bibkey": "zimmermann2021contrastive", "title": "Contrastive Learning Inverts the Data Generating Process", "first_author": "Roland S. Zimmermann", "year": 2021, "venue": "arXiv preprint arXiv:2102.08850", "doi": null, "arxiv_id": "2102.08850", "verified_via": "http://export.arxiv.org/api/query?id_list=2102.08850", "verified_at": "2026-11-09T17:09:52Z", "theme": "Disentanglement and identifiability"} | |
| 57 | +{"bibkey": "vaswani2017attention", "title": "Attention Is All You Need", "first_author": "Ashish Vaswani", "year": 2017, "venue": "arXiv preprint arXiv:1706.03762", "doi": null, "arxiv_id": "1706.03762", "verified_via": "http://export.arxiv.org/api/query?id_list=1706.03762", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 58 | +{"bibkey": "raffel2019t5", "title": "Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer", "first_author": "Colin Raffel", "year": 2019, "venue": "arXiv preprint arXiv:1910.10683", "doi": null, "arxiv_id": "1910.10683", "verified_via": "http://export.arxiv.org/api/query?id_list=1910.10683", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 59 | +{"bibkey": "lewis2019bart", "title": "BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension", "first_author": "Mike Lewis", "year": 2019, "venue": "arXiv preprint arXiv:1910.13461", "doi": null, "arxiv_id": "1910.13461", "verified_via": "http://export.arxiv.org/api/query?id_list=1910.13461", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 60 | +{"bibkey": "see2017pointer", "title": "Get To The Point: Summarization with Pointer-Generator Networks", "first_author": "Abigail See", "year": 2017, "venue": "arXiv preprint arXiv:1704.04368", "doi": null, "arxiv_id": "1704.04368", "verified_via": "http://export.arxiv.org/api/query?id_list=1704.04368", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 61 | +{"bibkey": "gu2016copying", "title": "Incorporating Copying Mechanism in Sequence-to-Sequence Learning", "first_author": "Jiatao Gu", "year": 2016, "venue": "arXiv preprint arXiv:1603.06393", "doi": null, "arxiv_id": "1603.06393", "verified_via": "http://export.arxiv.org/api/query?id_list=1603.06393", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 62 | +{"bibkey": "kudo2018sentencepiece", "title": "SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing", "first_author": "Taku Kudo", "year": 2018, "venue": "arXiv preprint arXiv:1808.06226", "doi": null, "arxiv_id": "1808.06226", "verified_via": "http://export.arxiv.org/api/query?id_list=1808.06226", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 63 | +{"bibkey": "xiong2020layernorm", "title": "On Layer Normalization in the Transformer Architecture", "first_author": "Ruibin Xiong", "year": 2020, "venue": "Published on ICML 2020", "doi": null, "arxiv_id": "2002.04745", "verified_via": "http://export.arxiv.org/api/query?id_list=2002.04745", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 64 | +{"bibkey": "su2021roformer", "title": "RoFormer: Enhanced Transformer with Rotary Position Embedding", "first_author": "Jianlin Su", "year": 2021, "venue": "arXiv preprint arXiv:2104.09864", "doi": null, "arxiv_id": "2104.09864", "verified_via": "http://export.arxiv.org/api/query?id_list=2104.09864", "verified_at": "2026-11-09T17:09:52Z", "theme": "Sequence-to-sequence architectures"} | |
| 65 | +{"bibkey": "popovic2015chrf", "title": "chrF: character n-gram F-score for automatic MT evaluation", "first_author": "Maja Popović", "year": 2015, "venue": "Conference on Machine Translation (WMT)", "doi": "10.18653/v1/w15-3049", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.18653/v1/w15-3049", "verified_at": "2026-11-09T17:12:41Z", "theme": "Evaluation metrics"} | |
| 66 | +{"bibkey": "papineni2002bleu", "title": "BLEU", "first_author": "Kishore Papineni", "year": 2001, "venue": "Proceedings of the conference - Association for Computational Linguistics. Meeting", "doi": "10.3115/1073083.1073135", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.3115/1073083.1073135", "verified_at": "2026-11-09T17:12:41Z", "theme": "Evaluation metrics", "note": "OpenAlex/Crossref title \"BLEU\", issued 2001; proceedings of ACL-02 (bib year 2002)"} | |
| 67 | +{"bibkey": "lin2004rouge", "title": "ROUGE: A Package for Automatic Evaluation of Summaries", "first_author": "Chin-Yew Lin", "year": 2004, "venue": null, "doi": null, "arxiv_id": null, "verified_via": "https://api.openalex.org/works/W2154652894", "verified_at": "2026-11-09T17:12:41Z", "theme": "Evaluation metrics"} | |
| 68 | +{"bibkey": "zhang2019bertscore", "title": "BERTScore: Evaluating Text Generation with BERT", "first_author": "Tianyi Zhang", "year": 2019, "venue": "arXiv preprint arXiv:1904.09675", "doi": null, "arxiv_id": "1904.09675", "verified_via": "http://export.arxiv.org/api/query?id_list=1904.09675", "verified_at": "2026-11-09T17:09:52Z", "theme": "Evaluation metrics"} | |
| 69 | +{"bibkey": "reimers2019sentencebert", "title": "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks", "first_author": "Nils Reimers", "year": 2019, "venue": "arXiv preprint arXiv:1908.10084", "doi": null, "arxiv_id": "1908.10084", "verified_via": "http://export.arxiv.org/api/query?id_list=1908.10084", "verified_at": "2026-11-09T17:09:52Z", "theme": "Evaluation metrics"} | |
| 70 | +{"bibkey": "carlini2020extracting", "title": "Extracting Training Data from Large Language Models", "first_author": "Nicholas Carlini", "year": 2020, "venue": "arXiv preprint arXiv:2012.07805", "doi": null, "arxiv_id": "2012.07805", "verified_via": "http://export.arxiv.org/api/query?id_list=2012.07805", "verified_at": "2026-11-09T17:09:52Z", "theme": "Memorisation and contamination"} | |
| 71 | +{"bibkey": "carlini2022quantifying", "title": "Quantifying Memorization Across Neural Language Models", "first_author": "Nicholas Carlini", "year": 2022, "venue": "arXiv preprint arXiv:2202.07646", "doi": null, "arxiv_id": "2202.07646", "verified_via": "http://export.arxiv.org/api/query?id_list=2202.07646", "verified_at": "2026-11-09T17:09:52Z", "theme": "Memorisation and contamination"} | |
| 72 | +{"bibkey": "lee2021dedup", "title": "Deduplicating Training Data Makes Language Models Better", "first_author": "Katherine Lee", "year": 2021, "venue": "arXiv preprint arXiv:2107.06499", "doi": null, "arxiv_id": "2107.06499", "verified_via": "http://export.arxiv.org/api/query?id_list=2107.06499", "verified_at": "2026-11-09T17:09:52Z", "theme": "Memorisation and contamination"} | |
| 73 | +{"bibkey": "magar2022contamination", "title": "Data Contamination: From Memorization to Exploitation", "first_author": "Inbal Magar", "year": 2022, "venue": "arXiv preprint arXiv:2203.08242", "doi": null, "arxiv_id": "2203.08242", "verified_via": "http://export.arxiv.org/api/query?id_list=2203.08242", "verified_at": "2026-11-09T17:09:52Z", "theme": "Memorisation and contamination"} | |
| 74 | +{"bibkey": "sainz2023contamination", "title": "NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark", "first_author": "Oscar Sainz", "year": 2023, "venue": "arXiv preprint arXiv:2310.18018", "doi": null, "arxiv_id": "2310.18018", "verified_via": "http://export.arxiv.org/api/query?id_list=2310.18018", "verified_at": "2026-11-09T17:09:52Z", "theme": "Memorisation and contamination"} | |
| 75 | +{"bibkey": "shumailov2024collapse", "title": "AI models collapse when trained on recursively generated data", "first_author": "Ilia Shumailov", "year": 2024, "venue": "Nature", "doi": "10.1038/s41586-024-07566-y", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1038/s41586-024-07566-y", "verified_at": "2026-11-09T17:12:41Z", "theme": "Synthetic data and model collapse"} | |
| 76 | +{"bibkey": "alemohammad2023mad", "title": "Self-Consuming Generative Models Go MAD", "first_author": "Sina Alemohammad", "year": 2023, "venue": "arXiv preprint arXiv:2307.01850", "doi": null, "arxiv_id": "2307.01850", "verified_via": "http://export.arxiv.org/api/query?id_list=2307.01850", "verified_at": "2026-11-09T17:09:52Z", "theme": "Synthetic data and model collapse"} | |
| 77 | +{"bibkey": "dohmatob2024tails", "title": "A Tale of Tails: Model Collapse as a Change of Scaling Laws", "first_author": "Elvis Dohmatob", "year": 2024, "venue": "ICML 2024", "doi": null, "arxiv_id": "2402.07043", "verified_via": "http://export.arxiv.org/api/query?id_list=2402.07043", "verified_at": "2026-11-09T17:09:52Z", "theme": "Synthetic data and model collapse"} | |
| 78 | +{"bibkey": "lo2019s2orc", "title": "S2ORC: The Semantic Scholar Open Research Corpus", "first_author": "Kyle Lo", "year": 2019, "venue": "arXiv preprint arXiv:1911.02782", "doi": null, "arxiv_id": "1911.02782", "verified_via": "http://export.arxiv.org/api/query?id_list=1911.02782", "verified_at": "2026-11-09T17:09:52Z", "theme": "Corpora and licensing"} | |
| 79 | +{"bibkey": "saier2023unarxive", "title": "unarXive 2022: All arXiv Publications Pre-Processed for NLP, Including Structured Full-Text and Citation Network", "first_author": "Tarek Saier", "year": 2023, "venue": "arXiv preprint arXiv:2303.14957", "doi": "10.1109/JCDL57899.2023.00020", "arxiv_id": "2303.14957", "verified_via": "http://export.arxiv.org/api/query?id_list=2303.14957", "verified_at": "2026-11-09T17:09:52Z", "theme": "Corpora and licensing"} | |
| 80 | +{"bibkey": "gao2020pile", "title": "The Pile: An 800GB Dataset of Diverse Text for Language Modeling", "first_author": "Leo Gao", "year": 2020, "venue": "arXiv preprint arXiv:2101.00027", "doi": null, "arxiv_id": "2101.00027", "verified_via": "http://export.arxiv.org/api/query?id_list=2101.00027", "verified_at": "2026-11-09T17:09:52Z", "theme": "Corpora and licensing"} | |
| 81 | +{"bibkey": "soldaini2024dolma", "title": "Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research", "first_author": "Luca Soldaini", "year": 2024, "venue": "arXiv preprint arXiv:2402.00159", "doi": null, "arxiv_id": "2402.00159", "verified_via": "http://export.arxiv.org/api/query?id_list=2402.00159", "verified_at": "2026-11-09T17:09:52Z", "theme": "Corpora and licensing"} | |
| 82 | +{"bibkey": "longpre2023provenance", "title": "The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI", "first_author": "Shayne Longpre", "year": 2023, "venue": "arXiv preprint arXiv:2310.16787", "doi": null, "arxiv_id": "2310.16787", "verified_via": "http://export.arxiv.org/api/query?id_list=2310.16787", "verified_at": "2026-11-09T17:09:52Z", "theme": "Corpora and licensing"} | |
| 83 | +{"bibkey": "plavensigray2017readability", "title": "The readability of scientific texts is decreasing over time", "first_author": "Pontus Plavén‐Sigray", "year": 2017, "venue": "eLife", "doi": "10.7554/elife.27725", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.7554/elife.27725", "verified_at": "2026-11-09T17:12:41Z", "theme": "Scientific style change over time"} | |
| 84 | +{"bibkey": "hyland2016stance", "title": "Change of Attitude? A Diachronic Study of Stance", "first_author": "Ken Hyland", "year": 2016, "venue": "Written Communication", "doi": "10.1177/0741088316650399", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1177/0741088316650399", "verified_at": "2026-11-09T17:12:41Z", "theme": "Scientific style change over time"} | |
| 85 | +{"bibkey": "hyland2016informal", "title": "Is academic writing becoming more informal?", "first_author": "Ken Hyland", "year": 2016, "venue": "English for Specific Purposes", "doi": "10.1016/j.esp.2016.09.001", "arxiv_id": null, "verified_via": "https://api.openalex.org/works?filter=doi:10.1016/j.esp.2016.09.001", "verified_at": "2026-11-09T17:12:41Z", "theme": "Scientific style change over time"} | |
added
paper/references.bib
+956 −0
@@ -0,0 +1,956 @@ | ||
| 1 | +% WP-AI-0002 references -- every entry verified against arXiv API / OpenAlex (see docs/literature/verified_sources.jsonl) | |
| 2 | +% Years are those returned by the API (arXiv: first-version year; OpenAlex: publication_year). | |
| 3 | +% arXiv-only records carry no venue: upgrade to the published venue after a Crossref/OpenAlex check. | |
| 4 | + | |
| 5 | +%% ---- LLM influence on scientific writing ---- | |
| 6 | +@article{liang2025quantifying, | |
| 7 | + author = {Liang, Weixin and Zhang, Yaohui and Wu, Zhengxuan and Lepp, Haley and Ji, Wenlong and Zhao, Xuandong and Cao, Hancheng and Liu, Sheng and He, Siyu and Huang, Zhi and Yang, Diyi and Potts, Christopher and Manning, Christopher D. and Zou, James}, | |
| 8 | + title = {Quantifying large language model usage in scientific papers}, | |
| 9 | + year = {2025}, | |
| 10 | + journal = {Nature {Human} {Behaviour}}, | |
| 11 | + volume = {9}, | |
| 12 | + number = {12}, | |
| 13 | + pages = {2599--2609}, | |
| 14 | + doi = {10.1038/s41562-025-02273-8}, | |
| 15 | + url = {https://doi.org/10.1038/s41562-025-02273-8} | |
| 16 | +} | |
| 17 | + | |
| 18 | +@article{kobak2025delving, | |
| 19 | + author = {Kobak, Dmitry and González-Márquez, Rita and Horvát, Emőke-Ágnes and Lause, Jan}, | |
| 20 | + title = {Delving into {LLM}-assisted writing in biomedical publications through excess vocabulary}, | |
| 21 | + year = {2025}, | |
| 22 | + journal = {Science {Advances}}, | |
| 23 | + volume = {11}, | |
| 24 | + number = {27}, | |
| 25 | + pages = {eadt3813}, | |
| 26 | + doi = {10.1126/sciadv.adt3813}, | |
| 27 | + url = {https://doi.org/10.1126/sciadv.adt3813} | |
| 28 | +} | |
| 29 | + | |
| 30 | +@misc{geng2024chatgpt, | |
| 31 | + author = {Geng, Mingmeng and Trotta, Roberto}, | |
| 32 | + title = {Is {ChatGPT} {Transforming} {Academics}' {Writing} {Style}?}, | |
| 33 | + year = {2024}, | |
| 34 | + howpublished = {arXiv preprint arXiv:2404.08627}, | |
| 35 | + eprint = {2404.08627}, | |
| 36 | + archivePrefix = {arXiv}, | |
| 37 | + primaryClass = {cs.CL}, | |
| 38 | + url = {https://arxiv.org/abs/2404.08627} | |
| 39 | +} | |
| 40 | + | |
| 41 | +@misc{geng2025humanllm, | |
| 42 | + author = {Geng, Mingmeng and Trotta, Roberto}, | |
| 43 | + title = {Human-{LLM} {Coevolution}: {Evidence} from {Academic} {Writing}}, | |
| 44 | + year = {2025}, | |
| 45 | + howpublished = {arXiv preprint arXiv:2502.09606}, | |
| 46 | + eprint = {2502.09606}, | |
| 47 | + archivePrefix = {arXiv}, | |
| 48 | + primaryClass = {cs.CL}, | |
| 49 | + url = {https://arxiv.org/abs/2502.09606} | |
| 50 | +} | |
| 51 | + | |
| 52 | +@misc{gray2024chatgpt, | |
| 53 | + author = {Gray, Andrew}, | |
| 54 | + title = {{ChatGPT} "contamination": estimating the prevalence of {LLMs} in the scholarly literature}, | |
| 55 | + year = {2024}, | |
| 56 | + howpublished = {arXiv preprint arXiv:2403.16887}, | |
| 57 | + eprint = {2403.16887}, | |
| 58 | + archivePrefix = {arXiv}, | |
| 59 | + primaryClass = {cs.DL}, | |
| 60 | + url = {https://arxiv.org/abs/2403.16887} | |
| 61 | +} | |
| 62 | + | |
| 63 | +@misc{liang2024monitoring, | |
| 64 | + author = {Liang, Weixin and Izzo, Zachary and Zhang, Yaohui and Lepp, Haley and Cao, Hancheng and Zhao, Xuandong and Chen, Lingjiao and Ye, Haotian and Liu, Sheng and Huang, Zhi and McFarland, Daniel A. and Zou, James Y.}, | |
| 65 | + title = {Monitoring {AI}-{Modified} {Content} at {Scale}: {A} {Case} {Study} on the {Impact} of {ChatGPT} on {AI} {Conference} {Peer} {Reviews}}, | |
| 66 | + year = {2024}, | |
| 67 | + howpublished = {arXiv preprint arXiv:2403.07183}, | |
| 68 | + eprint = {2403.07183}, | |
| 69 | + archivePrefix = {arXiv}, | |
| 70 | + primaryClass = {cs.CL}, | |
| 71 | + url = {https://arxiv.org/abs/2403.07183} | |
| 72 | +} | |
| 73 | + | |
| 74 | +@article{bao2025examining, | |
| 75 | + author = {Bao, Tong and Zhao, Yi and Mao, Jin and Zhang, Chengzhi}, | |
| 76 | + title = {Examining {Linguistic} {Shifts} in {Academic} {Writing} {Before} and {After} the {Launch} of {ChatGPT}: {A} {Study} on {Preprint} {Papers}}, | |
| 77 | + year = {2025}, | |
| 78 | + journal = {Scientometrics,2025}, | |
| 79 | + eprint = {2505.12218}, | |
| 80 | + archivePrefix = {arXiv}, | |
| 81 | + primaryClass = {cs.CL}, | |
| 82 | + url = {https://arxiv.org/abs/2505.12218} | |
| 83 | +} | |
| 84 | + | |
| 85 | +@misc{juzek2024delve, | |
| 86 | + author = {Juzek, Tom S. and Ward, Zina B.}, | |
| 87 | + title = {Why {Does} {ChatGPT} "{Delve}" {So} {Much}? {Exploring} the {Sources} of {Lexical} {Overrepresentation} in {Large} {Language} {Models}}, | |
| 88 | + year = {2024}, | |
| 89 | + howpublished = {arXiv preprint arXiv:2412.11385}, | |
| 90 | + eprint = {2412.11385}, | |
| 91 | + archivePrefix = {arXiv}, | |
| 92 | + primaryClass = {cs.CL}, | |
| 93 | + url = {https://arxiv.org/abs/2412.11385} | |
| 94 | +} | |
| 95 | + | |
| 96 | +@inproceedings{reinhart2025llms, | |
| 97 | + author = {Reinhart, Alex and Markey, Ben and Laudenbach, Michael and Pantusen, Kachatad and Yurko, Ronald and Weinberg, Gordon and Brown, David West}, | |
| 98 | + title = {Do {LLMs} write like humans? {Variation} in grammatical and rhetorical styles}, | |
| 99 | + year = {2025}, | |
| 100 | + booktitle = {Proceedings of the {National} {Academy} of {Sciences}}, | |
| 101 | + volume = {122}, | |
| 102 | + number = {8}, | |
| 103 | + doi = {10.1073/pnas.2422455122}, | |
| 104 | + url = {https://doi.org/10.1073/pnas.2422455122} | |
| 105 | +} | |
| 106 | + | |
| 107 | +@article{munozortiz2023contrasting, | |
| 108 | + author = {Mu{\~n}oz-Ortiz, Alberto and Gómez-Rodríguez, Carlos and Vilares, David}, | |
| 109 | + title = {Contrasting {Linguistic} {Patterns} in {Human} and {LLM}-{Generated} {News} {Text}}, | |
| 110 | + year = {2023}, | |
| 111 | + journal = {Artificial {Intelligence} {Review} 57, 265 (2024)}, | |
| 112 | + eprint = {2308.09067}, | |
| 113 | + archivePrefix = {arXiv}, | |
| 114 | + primaryClass = {cs.CL}, | |
| 115 | + doi = {10.1007/s10462-024-10903-2}, | |
| 116 | + url = {https://arxiv.org/abs/2308.09067} | |
| 117 | +} | |
| 118 | + | |
| 119 | +@misc{padmakumar2023writing, | |
| 120 | + author = {Padmakumar, Vishakh and He, He}, | |
| 121 | + title = {Does {Writing} with {Language} {Models} {Reduce} {Content} {Diversity}?}, | |
| 122 | + year = {2023}, | |
| 123 | + howpublished = {arXiv preprint arXiv:2309.05196}, | |
| 124 | + eprint = {2309.05196}, | |
| 125 | + archivePrefix = {arXiv}, | |
| 126 | + primaryClass = {cs.CL}, | |
| 127 | + url = {https://arxiv.org/abs/2309.05196} | |
| 128 | +} | |
| 129 | + | |
| 130 | +@misc{jang2026scizoom, | |
| 131 | + author = {Jang, Han and Lee, Junhyeok and Choi, Kyu Sung}, | |
| 132 | + title = {{SciZoom}: {A} {Large}-scale {Benchmark} for {Hierarchical} {Scientific} {Summarization} across the {LLM} {Era}}, | |
| 133 | + year = {2026}, | |
| 134 | + howpublished = {arXiv preprint arXiv:2603.16131}, | |
| 135 | + eprint = {2603.16131}, | |
| 136 | + archivePrefix = {arXiv}, | |
| 137 | + primaryClass = {cs.CL}, | |
| 138 | + url = {https://arxiv.org/abs/2603.16131} | |
| 139 | +} | |
| 140 | + | |
| 141 | +@misc{kosch2026chi, | |
| 142 | + author = {Kosch, Thomas and Welsch, Robin and Hedderich, Michael and Katins, Christopher}, | |
| 143 | + title = {How {Did} {Writing} {Change} {At} {CHI}? {Analyzing} 44 {Years} of {CHI} {Writing} {Before} and {After} the {Introduction} of {Large} {Language} {Models}}, | |
| 144 | + year = {2026}, | |
| 145 | + howpublished = {arXiv preprint arXiv:2609.23090}, | |
| 146 | + eprint = {2609.23090}, | |
| 147 | + archivePrefix = {arXiv}, | |
| 148 | + primaryClass = {cs.HC}, | |
| 149 | + url = {https://arxiv.org/abs/2609.23090} | |
| 150 | +} | |
| 151 | + | |
| 152 | +@article{gao2023comparing, | |
| 153 | + author = {Gao, Catherine A. and Howard, Frederick Matthew and Markov, Nikolay S. and Dyer, Emma and Ramesh, Siddhi and Luo, Yuan and Pearson, Alexander T.}, | |
| 154 | + title = {Comparing scientific abstracts generated by {ChatGPT} to real abstracts with detectors and blinded human reviewers}, | |
| 155 | + year = {2023}, | |
| 156 | + journal = {npj {Digital} {Medicine}}, | |
| 157 | + volume = {6}, | |
| 158 | + number = {1}, | |
| 159 | + pages = {75}, | |
| 160 | + doi = {10.1038/s41746-023-00819-6}, | |
| 161 | + url = {https://doi.org/10.1038/s41746-023-00819-6} | |
| 162 | +} | |
| 163 | + | |
| 164 | +%% ---- AI-text detection and paraphrase attacks ---- | |
| 165 | +@misc{krishna2023paraphrasing, | |
| 166 | + author = {Krishna, Kalpesh and Song, Yixiao and Karpinska, Marzena and Wieting, John and Iyyer, Mohit}, | |
| 167 | + title = {Paraphrasing evades detectors of {AI}-generated text, but retrieval is an effective defense}, | |
| 168 | + year = {2023}, | |
| 169 | + howpublished = {arXiv preprint arXiv:2303.13408}, | |
| 170 | + eprint = {2303.13408}, | |
| 171 | + archivePrefix = {arXiv}, | |
| 172 | + primaryClass = {cs.CL}, | |
| 173 | + url = {https://arxiv.org/abs/2303.13408} | |
| 174 | +} | |
| 175 | + | |
| 176 | +@misc{sadasivan2023reliably, | |
| 177 | + author = {Sadasivan, Vinu Sankar and Kumar, Aounon and Balasubramanian, Sriram and Wang, Wenxiao and Feizi, Soheil}, | |
| 178 | + title = {Can {AI}-{Generated} {Text} be {Reliably} {Detected}?}, | |
| 179 | + year = {2023}, | |
| 180 | + howpublished = {arXiv preprint arXiv:2303.11156}, | |
| 181 | + eprint = {2303.11156}, | |
| 182 | + archivePrefix = {arXiv}, | |
| 183 | + primaryClass = {cs.CL}, | |
| 184 | + url = {https://arxiv.org/abs/2303.11156} | |
| 185 | +} | |
| 186 | + | |
| 187 | +@misc{mitchell2023detectgpt, | |
| 188 | + author = {Mitchell, Eric and Lee, Yoonho and Khazatsky, Alexander and Manning, Christopher D. and Finn, Chelsea}, | |
| 189 | + title = {{DetectGPT}: {Zero}-{Shot} {Machine}-{Generated} {Text} {Detection} using {Probability} {Curvature}}, | |
| 190 | + year = {2023}, | |
| 191 | + howpublished = {arXiv preprint arXiv:2301.11305}, | |
| 192 | + eprint = {2301.11305}, | |
| 193 | + archivePrefix = {arXiv}, | |
| 194 | + primaryClass = {cs.CL}, | |
| 195 | + url = {https://arxiv.org/abs/2301.11305} | |
| 196 | +} | |
| 197 | + | |
| 198 | +@misc{bao2023fastdetectgpt, | |
| 199 | + author = {Bao, Guangsheng and Zhao, Yanbin and Teng, Zhiyang and Yang, Linyi and Zhang, Yue}, | |
| 200 | + title = {Fast-{DetectGPT}: {Efficient} {Zero}-{Shot} {Detection} of {Machine}-{Generated} {Text} via {Conditional} {Probability} {Curvature}}, | |
| 201 | + year = {2023}, | |
| 202 | + howpublished = {arXiv preprint arXiv:2310.05130}, | |
| 203 | + eprint = {2310.05130}, | |
| 204 | + archivePrefix = {arXiv}, | |
| 205 | + primaryClass = {cs.CL}, | |
| 206 | + url = {https://arxiv.org/abs/2310.05130} | |
| 207 | +} | |
| 208 | + | |
| 209 | +@misc{hans2024binoculars, | |
| 210 | + author = {Hans, Abhimanyu and Schwarzschild, Avi and Cherepanova, Valeriia and Kazemi, Hamid and Saha, Aniruddha and Goldblum, Micah and Geiping, Jonas and Goldstein, Tom}, | |
| 211 | + title = {Spotting {LLMs} {With} {Binoculars}: {Zero}-{Shot} {Detection} of {Machine}-{Generated} {Text}}, | |
| 212 | + year = {2024}, | |
| 213 | + howpublished = {arXiv preprint arXiv:2401.12070}, | |
| 214 | + eprint = {2401.12070}, | |
| 215 | + archivePrefix = {arXiv}, | |
| 216 | + primaryClass = {cs.CL}, | |
| 217 | + url = {https://arxiv.org/abs/2401.12070} | |
| 218 | +} | |
| 219 | + | |
| 220 | +@misc{liang2023detectors, | |
| 221 | + author = {Liang, Weixin and Yuksekgonul, Mert and Mao, Yining and Wu, Eric and Zou, James}, | |
| 222 | + title = {{GPT} detectors are biased against non-native {English} writers}, | |
| 223 | + year = {2023}, | |
| 224 | + howpublished = {arXiv preprint arXiv:2304.02819}, | |
| 225 | + eprint = {2304.02819}, | |
| 226 | + archivePrefix = {arXiv}, | |
| 227 | + primaryClass = {cs.CL}, | |
| 228 | + url = {https://arxiv.org/abs/2304.02819} | |
| 229 | +} | |
| 230 | + | |
| 231 | +@misc{kirchenbauer2023watermark, | |
| 232 | + author = {Kirchenbauer, John and Geiping, Jonas and Wen, Yuxin and Katz, Jonathan and Miers, Ian and Goldstein, Tom}, | |
| 233 | + title = {A {Watermark} for {Large} {Language} {Models}}, | |
| 234 | + year = {2023}, | |
| 235 | + howpublished = {arXiv preprint arXiv:2301.10226}, | |
| 236 | + eprint = {2301.10226}, | |
| 237 | + archivePrefix = {arXiv}, | |
| 238 | + primaryClass = {cs.LG}, | |
| 239 | + url = {https://arxiv.org/abs/2301.10226} | |
| 240 | +} | |
| 241 | + | |
| 242 | +@misc{dugan2024raid, | |
| 243 | + author = {Dugan, Liam and Hwang, Alyssa and Trhlik, Filip and Ludan, Josh Magnus and Zhu, Andrew and Xu, Hainiu and Ippolito, Daphne and Callison-Burch, Chris}, | |
| 244 | + title = {{RAID}: {A} {Shared} {Benchmark} for {Robust} {Evaluation} of {Machine}-{Generated} {Text} {Detectors}}, | |
| 245 | + year = {2024}, | |
| 246 | + howpublished = {arXiv preprint arXiv:2405.07940}, | |
| 247 | + eprint = {2405.07940}, | |
| 248 | + archivePrefix = {arXiv}, | |
| 249 | + primaryClass = {cs.CL}, | |
| 250 | + url = {https://arxiv.org/abs/2405.07940} | |
| 251 | +} | |
| 252 | + | |
| 253 | +@misc{saha2025polished, | |
| 254 | + author = {Saha, Shoumik and Feizi, Soheil}, | |
| 255 | + title = {Almost {AI}, {Almost} {Human}: {The} {Challenge} of {Detecting} {AI}-{Polished} {Writing}}, | |
| 256 | + year = {2025}, | |
| 257 | + howpublished = {arXiv preprint arXiv:2502.15666}, | |
| 258 | + eprint = {2502.15666}, | |
| 259 | + archivePrefix = {arXiv}, | |
| 260 | + primaryClass = {cs.CL}, | |
| 261 | + url = {https://arxiv.org/abs/2502.15666} | |
| 262 | +} | |
| 263 | + | |
| 264 | +@misc{chen2024imitate, | |
| 265 | + author = {Chen, Jiaqi and Zhu, Xiaoye and Liu, Tianyang and Chen, Ying and Chen, Xinhui and Yuan, Yiwen and Leong, Chak Tou and Li, Zuchao and Long, Tang and Zhang, Lei and Yan, Chenyu and Mei, Guanghao and Zhang, Jie and Zhang, Lefei}, | |
| 266 | + title = {Imitate {Before} {Detect}: {Aligning} {Machine} {Stylistic} {Preference} for {Machine}-{Revised} {Text} {Detection}}, | |
| 267 | + year = {2024}, | |
| 268 | + howpublished = {arXiv preprint arXiv:2412.10432}, | |
| 269 | + eprint = {2412.10432}, | |
| 270 | + archivePrefix = {arXiv}, | |
| 271 | + primaryClass = {cs.CL}, | |
| 272 | + url = {https://arxiv.org/abs/2412.10432} | |
| 273 | +} | |
| 274 | + | |
| 275 | +@misc{artemova2024beemo, | |
| 276 | + author = {Artemova, Ekaterina and Lucas, Jason and Venkatraman, Saranya and Lee, Jooyoung and Tilga, Sergei and Uchendu, Adaku and Mikhailov, Vladislav}, | |
| 277 | + title = {Beemo: {Benchmark} of {Expert}-edited {Machine}-generated {Outputs}}, | |
| 278 | + year = {2024}, | |
| 279 | + howpublished = {arXiv preprint arXiv:2411.04032}, | |
| 280 | + eprint = {2411.04032}, | |
| 281 | + archivePrefix = {arXiv}, | |
| 282 | + primaryClass = {cs.CL}, | |
| 283 | + url = {https://arxiv.org/abs/2411.04032} | |
| 284 | +} | |
| 285 | + | |
| 286 | +@misc{zheng2025thbench, | |
| 287 | + author = {Zheng, Jingyi and Wang, Junfeng and Sun, Zhen and Dong, Wenhan and Liu, Yule and He, Xinlei}, | |
| 288 | + title = {{TH}-{Bench}: {Evaluating} {Evading} {Attacks} via {Humanizing} {AI} {Text} on {Machine}-{Generated} {Text} {Detectors}}, | |
| 289 | + year = {2025}, | |
| 290 | + howpublished = {arXiv preprint arXiv:2503.08708}, | |
| 291 | + eprint = {2503.08708}, | |
| 292 | + archivePrefix = {arXiv}, | |
| 293 | + primaryClass = {cs.CR}, | |
| 294 | + url = {https://arxiv.org/abs/2503.08708} | |
| 295 | +} | |
| 296 | + | |
| 297 | +%% ---- Stylometry and authorship ---- | |
| 298 | +@article{stamatatos2008survey, | |
| 299 | + author = {Stamatatos, Efstathios}, | |
| 300 | + title = {A survey of modern authorship attribution methods}, | |
| 301 | + year = {2008}, | |
| 302 | + journal = {Journal of the {American} {Society} for {Information} {Science} and {Technology}}, | |
| 303 | + volume = {60}, | |
| 304 | + number = {3}, | |
| 305 | + pages = {538--556}, | |
| 306 | + doi = {10.1002/asi.21001}, | |
| 307 | + url = {https://doi.org/10.1002/asi.21001} | |
| 308 | +} | |
| 309 | + | |
| 310 | +@article{juola2008authorship, | |
| 311 | + author = {Juola, Patrick}, | |
| 312 | + title = {Authorship {Attribution}}, | |
| 313 | + year = {2008}, | |
| 314 | + journal = {Foundations and {Trends}® in {Information} {Retrieval}}, | |
| 315 | + volume = {1}, | |
| 316 | + number = {3}, | |
| 317 | + pages = {233--334}, | |
| 318 | + doi = {10.1561/1500000005}, | |
| 319 | + url = {https://doi.org/10.1561/1500000005} | |
| 320 | +} | |
| 321 | + | |
| 322 | +@article{koppel2008computational, | |
| 323 | + author = {Koppel, Moshe and Schler, Jonathan and Argamon, Shlomo}, | |
| 324 | + title = {Computational methods in authorship attribution}, | |
| 325 | + year = {2008}, | |
| 326 | + journal = {Journal of the {American} {Society} for {Information} {Science} and {Technology}}, | |
| 327 | + volume = {60}, | |
| 328 | + number = {1}, | |
| 329 | + pages = {9--26}, | |
| 330 | + doi = {10.1002/asi.20961}, | |
| 331 | + url = {https://doi.org/10.1002/asi.20961} | |
| 332 | +} | |
| 333 | + | |
| 334 | +@article{burrows2002delta, | |
| 335 | + author = {Burrows, J. F.}, | |
| 336 | + title = {'{Delta}': a {Measure} of {Stylistic} {Difference} and a {Guide} to {Likely} {Authorship}}, | |
| 337 | + year = {2002}, | |
| 338 | + journal = {Literary and {Linguistic} {Computing}}, | |
| 339 | + volume = {17}, | |
| 340 | + number = {3}, | |
| 341 | + pages = {267--287}, | |
| 342 | + doi = {10.1093/llc/17.3.267}, | |
| 343 | + url = {https://doi.org/10.1093/llc/17.3.267} | |
| 344 | +} | |
| 345 | + | |
| 346 | +@inproceedings{riverasoto2021luar, | |
| 347 | + author = {{Rivera Soto}, Rafael and Miano, Olivia and Ordoñez, Juanita and Chen, Barry Y. and Khan, Aleem and Bishop, Marcus and Andrews, Nicholas}, | |
| 348 | + title = {Learning {Universal} {Authorship} {Representations}}, | |
| 349 | + year = {2021}, | |
| 350 | + booktitle = {Conference on {Empirical} {Methods} in {Natural} {Language} {Processing} ({EMNLP})}, | |
| 351 | + pages = {913--919}, | |
| 352 | + doi = {10.18653/v1/2021.emnlp-main.70}, | |
| 353 | + url = {https://doi.org/10.18653/v1/2021.emnlp-main.70} | |
| 354 | +} | |
| 355 | + | |
| 356 | +@misc{wegmann2022same, | |
| 357 | + author = {Wegmann, Anna and Schraagen, Marijn and Nguyen, Dong}, | |
| 358 | + title = {Same {Author} or {Just} {Same} {Topic}? {Towards} {Content}-{Independent} {Style} {Representations}}, | |
| 359 | + year = {2022}, | |
| 360 | + howpublished = {arXiv preprint arXiv:2204.04907}, | |
| 361 | + eprint = {2204.04907}, | |
| 362 | + archivePrefix = {arXiv}, | |
| 363 | + primaryClass = {cs.CL}, | |
| 364 | + url = {https://arxiv.org/abs/2204.04907} | |
| 365 | +} | |
| 366 | + | |
| 367 | +@misc{fisher2024styleremix, | |
| 368 | + author = {Fisher, Jillian and Hallinan, Skyler and Lu, Ximing and Gordon, Mitchell and Harchaoui, Zaid and Choi, Yejin}, | |
| 369 | + title = {{StyleRemix}: {Interpretable} {Authorship} {Obfuscation} via {Distillation} and {Perturbation} of {Style} {Elements}}, | |
| 370 | + year = {2024}, | |
| 371 | + howpublished = {arXiv preprint arXiv:2408.15666}, | |
| 372 | + eprint = {2408.15666}, | |
| 373 | + archivePrefix = {arXiv}, | |
| 374 | + primaryClass = {cs.CL}, | |
| 375 | + url = {https://arxiv.org/abs/2408.15666} | |
| 376 | +} | |
| 377 | + | |
| 378 | +%% ---- Style transfer, paraphrase and inversion ---- | |
| 379 | +@misc{shen2017style, | |
| 380 | + author = {Shen, Tianxiao and Lei, Tao and Barzilay, Regina and Jaakkola, Tommi}, | |
| 381 | + title = {Style {Transfer} from {Non}-{Parallel} {Text} by {Cross}-{Alignment}}, | |
| 382 | + year = {2017}, | |
| 383 | + howpublished = {arXiv preprint arXiv:1705.09655}, | |
| 384 | + eprint = {1705.09655}, | |
| 385 | + archivePrefix = {arXiv}, | |
| 386 | + primaryClass = {cs.CL}, | |
| 387 | + url = {https://arxiv.org/abs/1705.09655} | |
| 388 | +} | |
| 389 | + | |
| 390 | +@misc{fu2017style, | |
| 391 | + author = {Fu, Zhenxin and Tan, Xiaoye and Peng, Nanyun and Zhao, Dongyan and Yan, Rui}, | |
| 392 | + title = {Style {Transfer} in {Text}: {Exploration} and {Evaluation}}, | |
| 393 | + year = {2017}, | |
| 394 | + howpublished = {arXiv preprint arXiv:1711.06861}, | |
| 395 | + eprint = {1711.06861}, | |
| 396 | + archivePrefix = {arXiv}, | |
| 397 | + primaryClass = {cs.CL}, | |
| 398 | + url = {https://arxiv.org/abs/1711.06861} | |
| 399 | +} | |
| 400 | + | |
| 401 | +@misc{john2018disentangled, | |
| 402 | + author = {John, Vineet and Mou, Lili and Bahuleyan, Hareesh and Vechtomova, Olga}, | |
| 403 | + title = {Disentangled {Representation} {Learning} for {Non}-{Parallel} {Text} {Style} {Transfer}}, | |
| 404 | + year = {2018}, | |
| 405 | + howpublished = {arXiv preprint arXiv:1808.04339}, | |
| 406 | + eprint = {1808.04339}, | |
| 407 | + archivePrefix = {arXiv}, | |
| 408 | + primaryClass = {cs.CL}, | |
| 409 | + url = {https://arxiv.org/abs/1808.04339} | |
| 410 | +} | |
| 411 | + | |
| 412 | +@misc{krishna2020reformulating, | |
| 413 | + author = {Krishna, Kalpesh and Wieting, John and Iyyer, Mohit}, | |
| 414 | + title = {Reformulating {Unsupervised} {Style} {Transfer} as {Paraphrase} {Generation}}, | |
| 415 | + year = {2020}, | |
| 416 | + howpublished = {arXiv preprint arXiv:2010.05700}, | |
| 417 | + eprint = {2010.05700}, | |
| 418 | + archivePrefix = {arXiv}, | |
| 419 | + primaryClass = {cs.CL}, | |
| 420 | + url = {https://arxiv.org/abs/2010.05700} | |
| 421 | +} | |
| 422 | + | |
| 423 | +@misc{riley2020textsettr, | |
| 424 | + author = {Riley, Parker and Constant, Noah and Guo, Mandy and Kumar, Girish and Uthus, David and Parekh, Zarana}, | |
| 425 | + title = {{TextSETTR}: {Few}-{Shot} {Text} {Style} {Extraction} and {Tunable} {Targeted} {Restyling}}, | |
| 426 | + year = {2020}, | |
| 427 | + howpublished = {arXiv preprint arXiv:2010.03802}, | |
| 428 | + eprint = {2010.03802}, | |
| 429 | + archivePrefix = {arXiv}, | |
| 430 | + primaryClass = {cs.CL}, | |
| 431 | + url = {https://arxiv.org/abs/2010.03802} | |
| 432 | +} | |
| 433 | + | |
| 434 | +@misc{patel2022lowresource, | |
| 435 | + author = {Patel, Ajay and Andrews, Nicholas and Callison-Burch, Chris}, | |
| 436 | + title = {Low-{Resource} {Authorship} {Style} {Transfer}: {Can} {Non}-{Famous} {Authors} {Be} {Imitated}?}, | |
| 437 | + year = {2022}, | |
| 438 | + howpublished = {arXiv preprint arXiv:2212.08986}, | |
| 439 | + eprint = {2212.08986}, | |
| 440 | + archivePrefix = {arXiv}, | |
| 441 | + primaryClass = {cs.CL}, | |
| 442 | + url = {https://arxiv.org/abs/2212.08986} | |
| 443 | +} | |
| 444 | + | |
| 445 | +@misc{mir2019evaluating, | |
| 446 | + author = {Mir, Remi and Felbo, Bjarke and Obradovich, Nick and Rahwan, Iyad}, | |
| 447 | + title = {Evaluating {Style} {Transfer} for {Text}}, | |
| 448 | + year = {2019}, | |
| 449 | + howpublished = {arXiv preprint arXiv:1904.02295}, | |
| 450 | + eprint = {1904.02295}, | |
| 451 | + archivePrefix = {arXiv}, | |
| 452 | + primaryClass = {cs.CL}, | |
| 453 | + url = {https://arxiv.org/abs/1904.02295} | |
| 454 | +} | |
| 455 | + | |
| 456 | +@misc{jin2020survey, | |
| 457 | + author = {Jin, Di and Jin, Zhijing and Hu, Zhiting and Vechtomova, Olga and Mihalcea, Rada}, | |
| 458 | + title = {Deep {Learning} for {Text} {Style} {Transfer}: {A} {Survey}}, | |
| 459 | + year = {2020}, | |
| 460 | + howpublished = {arXiv preprint arXiv:2011.00416}, | |
| 461 | + eprint = {2011.00416}, | |
| 462 | + archivePrefix = {arXiv}, | |
| 463 | + primaryClass = {cs.CL}, | |
| 464 | + url = {https://arxiv.org/abs/2011.00416} | |
| 465 | +} | |
| 466 | + | |
| 467 | +@misc{rao2018gyafc, | |
| 468 | + author = {Rao, Sudha and Tetreault, Joel}, | |
| 469 | + title = {Dear {Sir} or {Madam}, {May} {I} introduce the {GYAFC} {Dataset}: {Corpus}, {Benchmarks} and {Metrics} for {Formality} {Style} {Transfer}}, | |
| 470 | + year = {2018}, | |
| 471 | + howpublished = {arXiv preprint arXiv:1803.06535}, | |
| 472 | + eprint = {1803.06535}, | |
| 473 | + archivePrefix = {arXiv}, | |
| 474 | + primaryClass = {cs.CL}, | |
| 475 | + url = {https://arxiv.org/abs/1803.06535} | |
| 476 | +} | |
| 477 | + | |
| 478 | +@misc{riverasoto2024paraphrase, | |
| 479 | + author = {{Rivera Soto}, Rafael and Chen, Barry and Andrews, Nicholas}, | |
| 480 | + title = {Mitigating {Paraphrase} {Attacks} on {Machine}-{Text} {Detectors} via {Paraphrase} {Inversion}}, | |
| 481 | + year = {2024}, | |
| 482 | + howpublished = {arXiv preprint arXiv:2410.21637}, | |
| 483 | + eprint = {2410.21637}, | |
| 484 | + archivePrefix = {arXiv}, | |
| 485 | + primaryClass = {cs.CL}, | |
| 486 | + url = {https://arxiv.org/abs/2410.21637} | |
| 487 | +} | |
| 488 | + | |
| 489 | +@misc{riverasoto2026unsupervised, | |
| 490 | + author = {{Rivera Soto}, Rafael and Chen, Barry and Andrews, Nicholas}, | |
| 491 | + title = {Unsupervised {Style} {Representation} {Learning} for {AI}-{Text} {Detection} via {Paraphrase} {Inversion}}, | |
| 492 | + year = {2026}, | |
| 493 | + howpublished = {arXiv preprint arXiv:2606.10099}, | |
| 494 | + eprint = {2606.10099}, | |
| 495 | + archivePrefix = {arXiv}, | |
| 496 | + primaryClass = {cs.LG}, | |
| 497 | + url = {https://arxiv.org/abs/2606.10099} | |
| 498 | +} | |
| 499 | + | |
| 500 | +@misc{morris2023lminversion, | |
| 501 | + author = {Morris, John X. and Zhao, Wenting and Chiu, Justin T. and Shmatikov, Vitaly and Rush, Alexander M.}, | |
| 502 | + title = {Language {Model} {Inversion}}, | |
| 503 | + year = {2023}, | |
| 504 | + howpublished = {arXiv preprint arXiv:2311.13647}, | |
| 505 | + eprint = {2311.13647}, | |
| 506 | + archivePrefix = {arXiv}, | |
| 507 | + primaryClass = {cs.CL}, | |
| 508 | + url = {https://arxiv.org/abs/2311.13647} | |
| 509 | +} | |
| 510 | + | |
| 511 | +@misc{morris2023embeddings, | |
| 512 | + author = {Morris, John X. and Kuleshov, Volodymyr and Shmatikov, Vitaly and Rush, Alexander M.}, | |
| 513 | + title = {Text {Embeddings} {Reveal} ({Almost}) {As} {Much} {As} {Text}}, | |
| 514 | + year = {2023}, | |
| 515 | + howpublished = {arXiv preprint arXiv:2310.06816}, | |
| 516 | + eprint = {2310.06816}, | |
| 517 | + archivePrefix = {arXiv}, | |
| 518 | + primaryClass = {cs.CL}, | |
| 519 | + url = {https://arxiv.org/abs/2310.06816} | |
| 520 | +} | |
| 521 | + | |
| 522 | +%% ---- Disentanglement and identifiability ---- | |
| 523 | +@misc{ganin2014unsupervised, | |
| 524 | + author = {Ganin, Yaroslav and Lempitsky, Victor}, | |
| 525 | + title = {Unsupervised {Domain} {Adaptation} by {Backpropagation}}, | |
| 526 | + year = {2014}, | |
| 527 | + howpublished = {arXiv preprint arXiv:1409.7495}, | |
| 528 | + eprint = {1409.7495}, | |
| 529 | + archivePrefix = {arXiv}, | |
| 530 | + primaryClass = {stat.ML}, | |
| 531 | + url = {https://arxiv.org/abs/1409.7495} | |
| 532 | +} | |
| 533 | + | |
| 534 | +@misc{bousmalis2016domain, | |
| 535 | + author = {Bousmalis, Konstantinos and Trigeorgis, George and Silberman, Nathan and Krishnan, Dilip and Erhan, Dumitru}, | |
| 536 | + title = {Domain {Separation} {Networks}}, | |
| 537 | + year = {2016}, | |
| 538 | + howpublished = {arXiv preprint arXiv:1608.06019}, | |
| 539 | + eprint = {1608.06019}, | |
| 540 | + archivePrefix = {arXiv}, | |
| 541 | + primaryClass = {cs.CV}, | |
| 542 | + url = {https://arxiv.org/abs/1608.06019} | |
| 543 | +} | |
| 544 | + | |
| 545 | +@misc{hu2017controlled, | |
| 546 | + author = {Hu, Zhiting and Yang, Zichao and Liang, Xiaodan and Salakhutdinov, Ruslan and Xing, Eric P.}, | |
| 547 | + title = {Toward {Controlled} {Generation} of {Text}}, | |
| 548 | + year = {2017}, | |
| 549 | + howpublished = {arXiv preprint arXiv:1703.00955}, | |
| 550 | + eprint = {1703.00955}, | |
| 551 | + archivePrefix = {arXiv}, | |
| 552 | + primaryClass = {cs.LG}, | |
| 553 | + url = {https://arxiv.org/abs/1703.00955} | |
| 554 | +} | |
| 555 | + | |
| 556 | +@misc{chen2020simclr, | |
| 557 | + author = {Chen, Ting and Kornblith, Simon and Norouzi, Mohammad and Hinton, Geoffrey}, | |
| 558 | + title = {A {Simple} {Framework} for {Contrastive} {Learning} of {Visual} {Representations}}, | |
| 559 | + year = {2020}, | |
| 560 | + howpublished = {arXiv preprint arXiv:2002.05709}, | |
| 561 | + eprint = {2002.05709}, | |
| 562 | + archivePrefix = {arXiv}, | |
| 563 | + primaryClass = {cs.LG}, | |
| 564 | + url = {https://arxiv.org/abs/2002.05709} | |
| 565 | +} | |
| 566 | + | |
| 567 | +@misc{oord2018cpc, | |
| 568 | + author = {{van den Oord}, Aaron and Li, Yazhe and Vinyals, Oriol}, | |
| 569 | + title = {Representation {Learning} with {Contrastive} {Predictive} {Coding}}, | |
| 570 | + year = {2018}, | |
| 571 | + howpublished = {arXiv preprint arXiv:1807.03748}, | |
| 572 | + eprint = {1807.03748}, | |
| 573 | + archivePrefix = {arXiv}, | |
| 574 | + primaryClass = {cs.LG}, | |
| 575 | + url = {https://arxiv.org/abs/1807.03748} | |
| 576 | +} | |
| 577 | + | |
| 578 | +@misc{elazar2018adversarial, | |
| 579 | + author = {Elazar, Yanai and Goldberg, Yoav}, | |
| 580 | + title = {Adversarial {Removal} of {Demographic} {Attributes} from {Text} {Data}}, | |
| 581 | + year = {2018}, | |
| 582 | + howpublished = {arXiv preprint arXiv:1808.06640}, | |
| 583 | + eprint = {1808.06640}, | |
| 584 | + archivePrefix = {arXiv}, | |
| 585 | + primaryClass = {cs.CL}, | |
| 586 | + url = {https://arxiv.org/abs/1808.06640} | |
| 587 | +} | |
| 588 | + | |
| 589 | +@article{locatello2018challenging, | |
| 590 | + author = {Locatello, Francesco and Bauer, Stefan and Lucic, Mario and Rätsch, Gunnar and Gelly, Sylvain and Schölkopf, Bernhard and Bachem, Olivier}, | |
| 591 | + title = {Challenging {Common} {Assumptions} in the {Unsupervised} {Learning} of {Disentangled} {Representations}}, | |
| 592 | + year = {2018}, | |
| 593 | + journal = {Proceedings of the 36th {International} {Conference} on {Machine} {Learning} ({ICML} 2019)}, | |
| 594 | + eprint = {1811.12359}, | |
| 595 | + archivePrefix = {arXiv}, | |
| 596 | + primaryClass = {cs.LG}, | |
| 597 | + url = {https://arxiv.org/abs/1811.12359} | |
| 598 | +} | |
| 599 | + | |
| 600 | +@article{hyvarinen1999nonlinear, | |
| 601 | + author = {Hyvärinen, Aapo and Pajunen, Petteri}, | |
| 602 | + title = {Nonlinear independent component analysis: {Existence} and uniqueness results}, | |
| 603 | + year = {1999}, | |
| 604 | + journal = {Neural {Networks}}, | |
| 605 | + volume = {12}, | |
| 606 | + number = {3}, | |
| 607 | + pages = {429--439}, | |
| 608 | + doi = {10.1016/s0893-6080(98)00140-3}, | |
| 609 | + url = {https://doi.org/10.1016/s0893-6080(98)00140-3} | |
| 610 | +} | |
| 611 | + | |
| 612 | +@misc{vonkugelgen2021selfsupervised, | |
| 613 | + author = {{von K\"{u}gelgen}, Julius and Sharma, Yash and Gresele, Luigi and Brendel, Wieland and Schölkopf, Bernhard and Besserve, Michel and Locatello, Francesco}, | |
| 614 | + title = {Self-{Supervised} {Learning} with {Data} {Augmentations} {Provably} {Isolates} {Content} from {Style}}, | |
| 615 | + year = {2021}, | |
| 616 | + howpublished = {arXiv preprint arXiv:2106.04619}, | |
| 617 | + eprint = {2106.04619}, | |
| 618 | + archivePrefix = {arXiv}, | |
| 619 | + primaryClass = {stat.ML}, | |
| 620 | + url = {https://arxiv.org/abs/2106.04619} | |
| 621 | +} | |
| 622 | + | |
| 623 | +@misc{zimmermann2021contrastive, | |
| 624 | + author = {Zimmermann, Roland S. and Sharma, Yash and Schneider, Steffen and Bethge, Matthias and Brendel, Wieland}, | |
| 625 | + title = {Contrastive {Learning} {Inverts} the {Data} {Generating} {Process}}, | |
| 626 | + year = {2021}, | |
| 627 | + howpublished = {arXiv preprint arXiv:2102.08850}, | |
| 628 | + eprint = {2102.08850}, | |
| 629 | + archivePrefix = {arXiv}, | |
| 630 | + primaryClass = {cs.LG}, | |
| 631 | + url = {https://arxiv.org/abs/2102.08850} | |
| 632 | +} | |
| 633 | + | |
| 634 | +%% ---- Sequence-to-sequence architectures ---- | |
| 635 | +@misc{vaswani2017attention, | |
| 636 | + author = {Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia}, | |
| 637 | + title = {Attention {Is} {All} {You} {Need}}, | |
| 638 | + year = {2017}, | |
| 639 | + howpublished = {arXiv preprint arXiv:1706.03762}, | |
| 640 | + eprint = {1706.03762}, | |
| 641 | + archivePrefix = {arXiv}, | |
| 642 | + primaryClass = {cs.CL}, | |
| 643 | + url = {https://arxiv.org/abs/1706.03762} | |
| 644 | +} | |
| 645 | + | |
| 646 | +@misc{raffel2019t5, | |
| 647 | + author = {Raffel, Colin and Shazeer, Noam and Roberts, Adam and Lee, Katherine and Narang, Sharan and Matena, Michael and Zhou, Yanqi and Li, Wei and Liu, Peter J.}, | |
| 648 | + title = {Exploring the {Limits} of {Transfer} {Learning} with a {Unified} {Text}-to-{Text} {Transformer}}, | |
| 649 | + year = {2019}, | |
| 650 | + howpublished = {arXiv preprint arXiv:1910.10683}, | |
| 651 | + eprint = {1910.10683}, | |
| 652 | + archivePrefix = {arXiv}, | |
| 653 | + primaryClass = {cs.LG}, | |
| 654 | + url = {https://arxiv.org/abs/1910.10683} | |
| 655 | +} | |
| 656 | + | |
| 657 | +@misc{lewis2019bart, | |
| 658 | + author = {Lewis, Mike and Liu, Yinhan and Goyal, Naman and Ghazvininejad, Marjan and Mohamed, Abdelrahman and Levy, Omer and Stoyanov, Ves and Zettlemoyer, Luke}, | |
| 659 | + title = {{BART}: {Denoising} {Sequence}-to-{Sequence} {Pre}-training for {Natural} {Language} {Generation}, {Translation}, and {Comprehension}}, | |
| 660 | + year = {2019}, | |
| 661 | + howpublished = {arXiv preprint arXiv:1910.13461}, | |
| 662 | + eprint = {1910.13461}, | |
| 663 | + archivePrefix = {arXiv}, | |
| 664 | + primaryClass = {cs.CL}, | |
| 665 | + url = {https://arxiv.org/abs/1910.13461} | |
| 666 | +} | |
| 667 | + | |
| 668 | +@misc{see2017pointer, | |
| 669 | + author = {See, Abigail and Liu, Peter J. and Manning, Christopher D.}, | |
| 670 | + title = {Get {To} {The} {Point}: {Summarization} with {Pointer}-{Generator} {Networks}}, | |
| 671 | + year = {2017}, | |
| 672 | + howpublished = {arXiv preprint arXiv:1704.04368}, | |
| 673 | + eprint = {1704.04368}, | |
| 674 | + archivePrefix = {arXiv}, | |
| 675 | + primaryClass = {cs.CL}, | |
| 676 | + url = {https://arxiv.org/abs/1704.04368} | |
| 677 | +} | |
| 678 | + | |
| 679 | +@misc{gu2016copying, | |
| 680 | + author = {Gu, Jiatao and Lu, Zhengdong and Li, Hang and Li, Victor O. K.}, | |
| 681 | + title = {Incorporating {Copying} {Mechanism} in {Sequence}-to-{Sequence} {Learning}}, | |
| 682 | + year = {2016}, | |
| 683 | + howpublished = {arXiv preprint arXiv:1603.06393}, | |
| 684 | + eprint = {1603.06393}, | |
| 685 | + archivePrefix = {arXiv}, | |
| 686 | + primaryClass = {cs.CL}, | |
| 687 | + url = {https://arxiv.org/abs/1603.06393} | |
| 688 | +} | |
| 689 | + | |
| 690 | +@misc{kudo2018sentencepiece, | |
| 691 | + author = {Kudo, Taku and Richardson, John}, | |
| 692 | + title = {{SentencePiece}: {A} simple and language independent subword tokenizer and detokenizer for {Neural} {Text} {Processing}}, | |
| 693 | + year = {2018}, | |
| 694 | + howpublished = {arXiv preprint arXiv:1808.06226}, | |
| 695 | + eprint = {1808.06226}, | |
| 696 | + archivePrefix = {arXiv}, | |
| 697 | + primaryClass = {cs.CL}, | |
| 698 | + url = {https://arxiv.org/abs/1808.06226} | |
| 699 | +} | |
| 700 | + | |
| 701 | +@article{xiong2020layernorm, | |
| 702 | + author = {Xiong, Ruibin and Yang, Yunchang and He, Di and Zheng, Kai and Zheng, Shuxin and Xing, Chen and Zhang, Huishuai and Lan, Yanyan and Wang, Liwei and Liu, Tie-Yan}, | |
| 703 | + title = {On {Layer} {Normalization} in the {Transformer} {Architecture}}, | |
| 704 | + year = {2020}, | |
| 705 | + journal = {Published on {ICML} 2020}, | |
| 706 | + eprint = {2002.04745}, | |
| 707 | + archivePrefix = {arXiv}, | |
| 708 | + primaryClass = {cs.LG}, | |
| 709 | + url = {https://arxiv.org/abs/2002.04745} | |
| 710 | +} | |
| 711 | + | |
| 712 | +@misc{su2021roformer, | |
| 713 | + author = {Su, Jianlin and Lu, Yu and Pan, Shengfeng and Murtadha, Ahmed and Wen, Bo and Liu, Yunfeng}, | |
| 714 | + title = {{RoFormer}: {Enhanced} {Transformer} with {Rotary} {Position} {Embedding}}, | |
| 715 | + year = {2021}, | |
| 716 | + howpublished = {arXiv preprint arXiv:2104.09864}, | |
| 717 | + eprint = {2104.09864}, | |
| 718 | + archivePrefix = {arXiv}, | |
| 719 | + primaryClass = {cs.CL}, | |
| 720 | + url = {https://arxiv.org/abs/2104.09864} | |
| 721 | +} | |
| 722 | + | |
| 723 | +%% ---- Evaluation metrics ---- | |
| 724 | +@inproceedings{popovic2015chrf, | |
| 725 | + author = {Popović, Maja}, | |
| 726 | + title = {chrF: character n-gram {F}-score for automatic {MT} evaluation}, | |
| 727 | + year = {2015}, | |
| 728 | + booktitle = {Conference on {Machine} {Translation} ({WMT})}, | |
| 729 | + pages = {392--395}, | |
| 730 | + doi = {10.18653/v1/w15-3049}, | |
| 731 | + url = {https://doi.org/10.18653/v1/w15-3049} | |
| 732 | +} | |
| 733 | + | |
| 734 | +@inproceedings{papineni2002bleu, | |
| 735 | + author = {Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing}, | |
| 736 | + title = {{BLEU}: a Method for Automatic Evaluation of Machine Translation}, | |
| 737 | + year = {2002}, | |
| 738 | + note = {OpenAlex/Crossref return the title as ``BLEU'' (issued 2001); subtitle and 2002 follow the ACL-02 proceedings named in the Crossref container-title}, | |
| 739 | + booktitle = {Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics}, | |
| 740 | + pages = {311}, | |
| 741 | + doi = {10.3115/1073083.1073135}, | |
| 742 | + url = {https://doi.org/10.3115/1073083.1073135} | |
| 743 | +} | |
| 744 | + | |
| 745 | +@inproceedings{lin2004rouge, | |
| 746 | + author = {Lin, Chin-Yew}, | |
| 747 | + title = {{ROUGE}: {A} {Package} for {Automatic} {Evaluation} of {Summaries}}, | |
| 748 | + year = {2004}, | |
| 749 | + pages = {74--81}, | |
| 750 | + booktitle = {Text Summarization Branches Out}, | |
| 751 | + url = {https://aclanthology.org/W04-1013/} | |
| 752 | +} | |
| 753 | + | |
| 754 | +@misc{zhang2019bertscore, | |
| 755 | + author = {Zhang, Tianyi and Kishore, Varsha and Wu, Felix and Weinberger, Kilian Q. and Artzi, Yoav}, | |
| 756 | + title = {{BERTScore}: {Evaluating} {Text} {Generation} with {BERT}}, | |
| 757 | + year = {2019}, | |
| 758 | + howpublished = {arXiv preprint arXiv:1904.09675}, | |
| 759 | + eprint = {1904.09675}, | |
| 760 | + archivePrefix = {arXiv}, | |
| 761 | + primaryClass = {cs.CL}, | |
| 762 | + url = {https://arxiv.org/abs/1904.09675} | |
| 763 | +} | |
| 764 | + | |
| 765 | +@misc{reimers2019sentencebert, | |
| 766 | + author = {Reimers, Nils and Gurevych, Iryna}, | |
| 767 | + title = {Sentence-{BERT}: {Sentence} {Embeddings} using {Siamese} {BERT}-{Networks}}, | |
| 768 | + year = {2019}, | |
| 769 | + howpublished = {arXiv preprint arXiv:1908.10084}, | |
| 770 | + eprint = {1908.10084}, | |
| 771 | + archivePrefix = {arXiv}, | |
| 772 | + primaryClass = {cs.CL}, | |
| 773 | + url = {https://arxiv.org/abs/1908.10084} | |
| 774 | +} | |
| 775 | + | |
| 776 | +%% ---- Memorisation and contamination ---- | |
| 777 | +@misc{carlini2020extracting, | |
| 778 | + author = {Carlini, Nicholas and Tramer, Florian and Wallace, Eric and Jagielski, Matthew and Herbert-Voss, Ariel and Lee, Katherine and Roberts, Adam and Brown, Tom and Song, Dawn and Erlingsson, Ulfar and Oprea, Alina and Raffel, Colin}, | |
| 779 | + title = {Extracting {Training} {Data} from {Large} {Language} {Models}}, | |
| 780 | + year = {2020}, | |
| 781 | + howpublished = {arXiv preprint arXiv:2012.07805}, | |
| 782 | + eprint = {2012.07805}, | |
| 783 | + archivePrefix = {arXiv}, | |
| 784 | + primaryClass = {cs.CR}, | |
| 785 | + url = {https://arxiv.org/abs/2012.07805} | |
| 786 | +} | |
| 787 | + | |
| 788 | +@misc{carlini2022quantifying, | |
| 789 | + author = {Carlini, Nicholas and Ippolito, Daphne and Jagielski, Matthew and Lee, Katherine and Tramer, Florian and Zhang, Chiyuan}, | |
| 790 | + title = {Quantifying {Memorization} {Across} {Neural} {Language} {Models}}, | |
| 791 | + year = {2022}, | |
| 792 | + howpublished = {arXiv preprint arXiv:2202.07646}, | |
| 793 | + eprint = {2202.07646}, | |
| 794 | + archivePrefix = {arXiv}, | |
| 795 | + primaryClass = {cs.LG}, | |
| 796 | + url = {https://arxiv.org/abs/2202.07646} | |
| 797 | +} | |
| 798 | + | |
| 799 | +@misc{lee2021dedup, | |
| 800 | + author = {Lee, Katherine and Ippolito, Daphne and Nystrom, Andrew and Zhang, Chiyuan and Eck, Douglas and Callison-Burch, Chris and Carlini, Nicholas}, | |
| 801 | + title = {Deduplicating {Training} {Data} {Makes} {Language} {Models} {Better}}, | |
| 802 | + year = {2021}, | |
| 803 | + howpublished = {arXiv preprint arXiv:2107.06499}, | |
| 804 | + eprint = {2107.06499}, | |
| 805 | + archivePrefix = {arXiv}, | |
| 806 | + primaryClass = {cs.CL}, | |
| 807 | + url = {https://arxiv.org/abs/2107.06499} | |
| 808 | +} | |
| 809 | + | |
| 810 | +@misc{magar2022contamination, | |
| 811 | + author = {Magar, Inbal and Schwartz, Roy}, | |
| 812 | + title = {Data {Contamination}: {From} {Memorization} to {Exploitation}}, | |
| 813 | + year = {2022}, | |
| 814 | + howpublished = {arXiv preprint arXiv:2203.08242}, | |
| 815 | + eprint = {2203.08242}, | |
| 816 | + archivePrefix = {arXiv}, | |
| 817 | + primaryClass = {cs.CL}, | |
| 818 | + url = {https://arxiv.org/abs/2203.08242} | |
| 819 | +} | |
| 820 | + | |
| 821 | +@misc{sainz2023contamination, | |
| 822 | + author = {Sainz, Oscar and Campos, Jon Ander and García-Ferrero, Iker and Etxaniz, Julen and Lacalle, Oier Lopez de and Agirre, Eneko}, | |
| 823 | + title = {{NLP} {Evaluation} in trouble: {On} the {Need} to {Measure} {LLM} {Data} {Contamination} for each {Benchmark}}, | |
| 824 | + year = {2023}, | |
| 825 | + howpublished = {arXiv preprint arXiv:2310.18018}, | |
| 826 | + eprint = {2310.18018}, | |
| 827 | + archivePrefix = {arXiv}, | |
| 828 | + primaryClass = {cs.CL}, | |
| 829 | + url = {https://arxiv.org/abs/2310.18018} | |
| 830 | +} | |
| 831 | + | |
| 832 | +%% ---- Synthetic data and model collapse ---- | |
| 833 | +@article{shumailov2024collapse, | |
| 834 | + author = {Shumailov, Ilia and Shumaylov, Zakhar and Zhao, Yiren and Papernot, Nicolas and Anderson, Ross and Gal, Yarin}, | |
| 835 | + title = {{AI} models collapse when trained on recursively generated data}, | |
| 836 | + year = {2024}, | |
| 837 | + journal = {Nature}, | |
| 838 | + volume = {631}, | |
| 839 | + number = {8022}, | |
| 840 | + pages = {755--759}, | |
| 841 | + doi = {10.1038/s41586-024-07566-y}, | |
| 842 | + url = {https://doi.org/10.1038/s41586-024-07566-y} | |
| 843 | +} | |
| 844 | + | |
| 845 | +@misc{alemohammad2023mad, | |
| 846 | + author = {Alemohammad, Sina and Casco-Rodriguez, Josue and Luzi, Lorenzo and Humayun, Ahmed Imtiaz and Babaei, Hossein and LeJeune, Daniel and Siahkoohi, Ali and Baraniuk, Richard G.}, | |
| 847 | + title = {Self-{Consuming} {Generative} {Models} {Go} {MAD}}, | |
| 848 | + year = {2023}, | |
| 849 | + howpublished = {arXiv preprint arXiv:2307.01850}, | |
| 850 | + eprint = {2307.01850}, | |
| 851 | + archivePrefix = {arXiv}, | |
| 852 | + primaryClass = {cs.LG}, | |
| 853 | + url = {https://arxiv.org/abs/2307.01850} | |
| 854 | +} | |
| 855 | + | |
| 856 | +@article{dohmatob2024tails, | |
| 857 | + author = {Dohmatob, Elvis and Feng, Yunzhen and Yang, Pu and Charton, Francois and Kempe, Julia}, | |
| 858 | + title = {A {Tale} of {Tails}: {Model} {Collapse} as a {Change} of {Scaling} {Laws}}, | |
| 859 | + year = {2024}, | |
| 860 | + journal = {{ICML} 2024}, | |
| 861 | + eprint = {2402.07043}, | |
| 862 | + archivePrefix = {arXiv}, | |
| 863 | + primaryClass = {cs.LG}, | |
| 864 | + url = {https://arxiv.org/abs/2402.07043} | |
| 865 | +} | |
| 866 | + | |
| 867 | +%% ---- Corpora and licensing ---- | |
| 868 | +@misc{lo2019s2orc, | |
| 869 | + author = {Lo, Kyle and Wang, Lucy Lu and Neumann, Mark and Kinney, Rodney and Weld, Dan S.}, | |
| 870 | + title = {{S2ORC}: {The} {Semantic} {Scholar} {Open} {Research} {Corpus}}, | |
| 871 | + year = {2019}, | |
| 872 | + howpublished = {arXiv preprint arXiv:1911.02782}, | |
| 873 | + eprint = {1911.02782}, | |
| 874 | + archivePrefix = {arXiv}, | |
| 875 | + primaryClass = {cs.CL}, | |
| 876 | + url = {https://arxiv.org/abs/1911.02782} | |
| 877 | +} | |
| 878 | + | |
| 879 | +@misc{saier2023unarxive, | |
| 880 | + author = {Saier, Tarek and Krause, Johan and Färber, Michael}, | |
| 881 | + title = {unarXive 2022: {All} {arXiv} {Publications} {Pre}-{Processed} for {NLP}, {Including} {Structured} {Full}-{Text} and {Citation} {Network}}, | |
| 882 | + year = {2023}, | |
| 883 | + howpublished = {arXiv preprint arXiv:2303.14957}, | |
| 884 | + eprint = {2303.14957}, | |
| 885 | + archivePrefix = {arXiv}, | |
| 886 | + primaryClass = {cs.DL}, | |
| 887 | + doi = {10.1109/JCDL57899.2023.00020}, | |
| 888 | + url = {https://arxiv.org/abs/2303.14957} | |
| 889 | +} | |
| 890 | + | |
| 891 | +@misc{gao2020pile, | |
| 892 | + author = {Gao, Leo and Biderman, Stella and Black, Sid and Golding, Laurence and Hoppe, Travis and Foster, Charles and Phang, Jason and He, Horace and Thite, Anish and Nabeshima, Noa and Presser, Shawn and Leahy, Connor}, | |
| 893 | + title = {The {Pile}: {An} {800GB} {Dataset} of {Diverse} {Text} for {Language} {Modeling}}, | |
| 894 | + year = {2020}, | |
| 895 | + howpublished = {arXiv preprint arXiv:2101.00027}, | |
| 896 | + eprint = {2101.00027}, | |
| 897 | + archivePrefix = {arXiv}, | |
| 898 | + primaryClass = {cs.CL}, | |
| 899 | + url = {https://arxiv.org/abs/2101.00027} | |
| 900 | +} | |
| 901 | + | |
| 902 | +@misc{soldaini2024dolma, | |
| 903 | + author = {Soldaini, Luca and Kinney, Rodney and Bhagia, Akshita and Schwenk, Dustin and Atkinson, David and Authur, Russell and Bogin, Ben and Chandu, Khyathi and Dumas, Jennifer and Elazar, Yanai and Hofmann, Valentin and Jha, Ananya Harsh and Kumar, Sachin and Lucy, Li and Lyu, Xinxi and Lambert, Nathan and Magnusson, Ian and Morrison, Jacob and Muennighoff, Niklas and Naik, Aakanksha and Nam, Crystal and Peters, Matthew E. and Ravichander, Abhilasha and Richardson, Kyle and Shen, Zejiang and Strubell, Emma and Subramani, Nishant and Tafjord, Oyvind and Walsh, Pete and Zettlemoyer, Luke and Smith, Noah A. and Hajishirzi, Hannaneh and Beltagy, Iz and Groeneveld, Dirk and Dodge, Jesse and Lo, Kyle}, | |
| 904 | + title = {Dolma: an {Open} {Corpus} of {Three} {Trillion} {Tokens} for {Language} {Model} {Pretraining} {Research}}, | |
| 905 | + year = {2024}, | |
| 906 | + howpublished = {arXiv preprint arXiv:2402.00159}, | |
| 907 | + eprint = {2402.00159}, | |
| 908 | + archivePrefix = {arXiv}, | |
| 909 | + primaryClass = {cs.CL}, | |
| 910 | + url = {https://arxiv.org/abs/2402.00159} | |
| 911 | +} | |
| 912 | + | |
| 913 | +@misc{longpre2023provenance, | |
| 914 | + author = {Longpre, Shayne and Mahari, Robert and Chen, Anthony and Obeng-Marnu, Naana and Sileo, Damien and Brannon, William and Muennighoff, Niklas and Khazam, Nathan and Kabbara, Jad and Perisetla, Kartik and Wu, Xinyi and Shippole, Enrico and Bollacker, Kurt and Wu, Tongshuang and Villa, Luis and Pentland, Sandy and Hooker, Sara}, | |
| 915 | + title = {The {Data} {Provenance} {Initiative}: {A} {Large} {Scale} {Audit} of {Dataset} {Licensing} \& {Attribution} in {AI}}, | |
| 916 | + year = {2023}, | |
| 917 | + howpublished = {arXiv preprint arXiv:2310.16787}, | |
| 918 | + eprint = {2310.16787}, | |
| 919 | + archivePrefix = {arXiv}, | |
| 920 | + primaryClass = {cs.CL}, | |
| 921 | + url = {https://arxiv.org/abs/2310.16787} | |
| 922 | +} | |
| 923 | + | |
| 924 | +%% ---- Scientific style change over time ---- | |
| 925 | +@article{plavensigray2017readability, | |
| 926 | + author = {Plavén‐Sigray, Pontus and Matheson, Granville James and Schiffler, Björn Christian and Thompson, William Hedley}, | |
| 927 | + title = {The readability of scientific texts is decreasing over time}, | |
| 928 | + year = {2017}, | |
| 929 | + journal = {eLife}, | |
| 930 | + volume = {6}, | |
| 931 | + doi = {10.7554/elife.27725}, | |
| 932 | + url = {https://doi.org/10.7554/elife.27725} | |
| 933 | +} | |
| 934 | + | |
| 935 | +@article{hyland2016stance, | |
| 936 | + author = {Hyland, Ken and Jiang, Feng Kevin}, | |
| 937 | + title = {Change of {Attitude}? {A} {Diachronic} {Study} of {Stance}}, | |
| 938 | + year = {2016}, | |
| 939 | + journal = {Written {Communication}}, | |
| 940 | + volume = {33}, | |
| 941 | + number = {3}, | |
| 942 | + pages = {251--274}, | |
| 943 | + doi = {10.1177/0741088316650399}, | |
| 944 | + url = {https://doi.org/10.1177/0741088316650399} | |
| 945 | +} | |
| 946 | + | |
| 947 | +@article{hyland2016informal, | |
| 948 | + author = {Hyland, Ken and Jiang, Feng Kevin}, | |
| 949 | + title = {Is academic writing becoming more informal?}, | |
| 950 | + year = {2016}, | |
| 951 | + journal = {English for {Specific} {Purposes}}, | |
| 952 | + volume = {45}, | |
| 953 | + pages = {40--51}, | |
| 954 | + doi = {10.1016/j.esp.2016.09.001}, | |
| 955 | + url = {https://doi.org/10.1016/j.esp.2016.09.001} | |
| 956 | +} | |
added
pyproject.toml
+49 −0
Large diff (49 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/build_corpus.py
+35 −0
Large diff (35 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/discover_arxiv.py
+60 −0
Large diff (60 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/discover_books.py
+83 −0
Large diff (83 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/discover_pmc.py
+40 −0
Large diff (40 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/evaluate.py
+32 −0
Large diff (32 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/make_splits.py
+39 −0
Large diff (39 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/predict.py
+60 −0
Large diff (60 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/rewrite.py
+101 −0
Large diff (101 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/train.py
+103 −0
Large diff (103 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/train_tokenizer.py
+46 −0
Large diff (46 lines) is not loaded by default to keep the page fast.Load diff
added
scripts/validate_pairs.py
+30 −0
Large diff (30 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/__init__.py
+8 −0
Large diff (8 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/data/__init__.py
+2 −0
Large diff (2 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/data/schema.py
+143 −0
Large diff (143 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/data/splits.py
+89 −0
Large diff (89 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/data/taxonomy.py
+46 −0
Large diff (46 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/dedup/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/dedup/minhash.py
+67 −0
Large diff (67 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/evaluation/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/evaluation/baselines.py
+23 −0
Large diff (23 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/evaluation/bootstrap.py
+40 −0
Large diff (40 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/evaluation/metrics.py
+54 −0
Large diff (54 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/evaluation/runner.py
+89 −0
Large diff (89 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/evaluation/stylometry.py
+124 −0
Large diff (124 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/inference/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/inference/pipeline.py
+103 −0
Large diff (103 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/ingest/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/ingest/arxiv_oai.py
+486 −0
Large diff (486 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/ingest/books.py
+773 −0
Large diff (773 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/ingest/corpus.py
+198 −0
Large diff (198 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/ingest/pmc_s3.py
+187 −0
Large diff (187 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/licenses/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/licenses/policy.py
+146 −0
Large diff (146 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/models/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/models/config.py
+53 −0
Large diff (53 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/models/losses.py
+114 −0
Large diff (114 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/models/tokenizer.py
+76 −0
Large diff (76 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/models/transformer.py
+314 −0
Large diff (314 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/parse/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/parse/books.py
+627 −0
Large diff (627 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/parse/jats.py
+347 −0
Large diff (347 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/parse/latex.py
+740 −0
Large diff (740 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/parse/protect.py
+86 −0
Large diff (86 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/parse/quality.py
+81 −0
Large diff (81 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/release/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/rewrite/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/rewrite/anthropic_provider.py
+112 −0
Large diff (112 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/rewrite/engine.py
+178 −0
Large diff (178 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/rewrite/mock_provider.py
+83 −0
Large diff (83 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/rewrite/prompts.py
+62 −0
Large diff (62 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/rewrite/provider.py
+40 −0
Large diff (40 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/segment/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/segment/passages.py
+70 −0
Large diff (70 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/training/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/training/data.py
+137 −0
Large diff (137 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/training/trainer.py
+230 −0
Large diff (230 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/utils/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/utils/hashing.py
+29 −0
Large diff (29 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/utils/http.py
+102 −0
Large diff (102 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/utils/io.py
+112 −0
Large diff (112 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/utils/provenance.py
+76 −0
Large diff (76 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/utils/text.py
+49 −0
Large diff (49 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/validate/__init__.py
+1 −0
Large diff (1 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/validate/checks.py
+166 −0
Large diff (166 lines) is not loaded by default to keep the page fast.Load diff
added
src/htr/validate/pairs.py
+63 −0
Large diff (63 lines) is not loaded by default to keep the page fast.Load diff
added
tests/conftest.py
+5 −0
Large diff (5 lines) is not loaded by default to keep the page fast.Load diff
added
tests/integration/test_arxiv_live.py
+39 −0
Large diff (39 lines) is not loaded by default to keep the page fast.Load diff
added
tests/integration/test_books_live.py
+27 −0
Large diff (27 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_books_parser.py
+182 −0
Large diff (182 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_jats.py
+43 −0
Large diff (43 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_latex_parser.py
+206 −0
Large diff (206 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_licenses.py
+20 −0
Large diff (20 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_models.py
+81 −0
Large diff (81 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_protect_and_text.py
+21 −0
Large diff (21 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_splits_and_dedup.py
+29 −0
Large diff (29 lines) is not loaded by default to keep the page fast.Load diff
added
tests/unit/test_validate_and_mock.py
+30 −0
Large diff (31 lines) is not loaded by default to keep the page fast.Load diff