spb/wp30_uqo

1{2  "document_id": "lb_qasper_894a0e08b526f2093c854d91c680190c898ae6acbc1ba131",3  "family": "public_qasper",4  "split": "test",5  "text": "Introduction\nThere is a recent spark of interest in the task of Question Answering (QA) over unstructured textual data, also referred to as Machine Reading Comprehension (MRC). This is mostly due to wide-spread success of advances in various facets of deep learning related research, such as novel architectures BIBREF0, BIBREF1 that allow for efficient optimisation of neural networks consisting of multiple layers, hardware designed for deep learning purposes and software frameworks BIBREF2, BIBREF3 that allow efficient development and testing of novel approaches. These factors enable researchers to produce models that are pre-trained on large scale corpora and provide contextualised word representations BIBREF4 that are shown to be a vital component towards solutions for a variety of natural language understanding tasks, including MRC BIBREF5. Another important factor that led to the recent success in MRC-related tasks is the widespread availability of various large datasets, e.g., SQuAD BIBREF6, that provide sufficient examples for optimising statistical models. The combination of these factors yields notable results, even surpassing human performance BIBREF7.\nMRC is a generic task format that can be used to probe for various natural language understanding capabilities BIBREF8. Therefore it is crucially important to establish a rigorous evaluation methodology in order to be able to draw reliable conclusions from conducted experiments. While increasing effort is put into the evaluation of novel architectures, such as keeping the evaluation data from public access to prevent unintentional overfitting to test data, performing ablation and error studies and introducing novel metrics BIBREF9, surprisingly little is done to establish the quality of the data itself. Additionally, recent research arrived at worrisome findings: the data of those gold standards, which is usually gathered involving a crowd-sourcing step, suffers from flaws in design BIBREF10 or contains overly specific keywords BIBREF11. Furthermore, these gold standards contain “annotation artefacts”, cues that lead models into focusing on superficial aspects of text, such as lexical overlap and word order, instead of actual language understanding BIBREF12, BIBREF13. These weaknesses cast some doubt on whether the data can reliably evaluate the reading comprehension performance of the models they evaluate, i.e. if the models are indeed being assessed for their capability to read.\nFigure FIGREF3 shows an example from HotpotQA BIBREF14, a dataset that exhibits the last kind of weakness mentioned above, i.e., the presence of unique keywords in both the question and the passage (in close proximity to the expected answer).\nAn evaluation methodology is vital to the fine-grained understanding of challenges associated with a single gold standard, in order to understand in greater detail which capabilities of MRC models it evaluates. More importantly, it allows to draw comparisons between multiple gold standards and between the results of respective state-of-the-art models that are evaluated on them.\nIn this work, we take a step back and propose a framework to systematically analyse MRC evaluation data, typically a set of questions and expected answers to be derived from accompanying passages. Concretely, we introduce a methodology to categorise the linguistic complexity of the textual data and the reasoning and potential external knowledge required to obtain the expected answer. Additionally we propose to take a closer look at the factual correctness of the expected answers, a quality dimension that appears under-explored in literature.\nWe demonstrate the usefulness of the proposed framework by applying it to precisely describe and compare six contemporary MRC datasets. Our findings reveal concerns about their factual correctness, the presence of lexical cues that simplify the task of reading comprehension and the lack of semantic altering grammatical modifiers. We release the raw data comprised of 300 paragraphs, questions and answers richly annotated under the proposed framework as a resource for researchers developing natural language understanding models and datasets to utilise further.\nTo the best of our knowledge this is the first attempt to introduce a common evaluation methodology for MRC gold standards and the first across-the-board qualitative evaluation of MRC datasets with respect to the proposed categories.\nFramework for MRC Gold Standard Analysis ::: Problem definition\nWe define the task of machine reading comprehension, the target application of the proposed methodology as follows: Given a paragraph $P$ that consists of tokens (words) $p_1, \\ldots , p_{n_P}$ and a question $Q$ that consists of tokens $q_1 \\ldots q_{n_Q}$, the goal is to retrieve an answer $A$ with tokens $a_1 \\ldots a_{n_A}$. $A$ is commonly constrained to be one of the following cases BIBREF15, illustrated in Figure FIGREF9:\nMultiple choice, where the goal is to predict $A$ from a given set of choices $\\mathcal {A}$.\nCloze-style, where $S$ is a sentence, and $A$ and $Q$ are obtained by removing a sequence of words such that $Q = S - A$. The task is to fill in the resulting gap in $Q$ with the expected answer $A$ to form $S$.\nSpan, where is a continuous subsequence of tokens from the paragraph ($A \\subseteq P$). Flavours include multiple spans as the correct answer or $A \\subseteq Q$.\nFree form, where $A$ is an unconstrained natural language string.\nA gold standard $G$ is composed of $m$ entries $(Q_i, A_i, P_i)_{i\\in \\lbrace 1,\\ldots , m\\rbrace }$.\nThe performance of an approach is established by comparing its answer predictions $A^*_{i}$ on the given input $(Q_i, T_i)$ (and $\\mathcal {A}_i$ for the multiple choice setting) against the expected answer $A_i$ for all $i\\in \\lbrace 1,\\ldots , m\\rbrace $ under a performance metric. Typical performance metrics are exact match (EM) or accuracy, i.e. the percentage of exactly predicted answers, and the F1 score – the harmonic mean between the precision and the recall of the predicted tokens compared to expected answer tokens. The overall F1 score can either be computed by averaging the F1 scores for every instance or by first averaging the precision and recall and then computing the F1 score from those averages (macro F1). Free-text answers, meanwhile, are evaluated by means of text generation and summarisation metrics such as BLEU BIBREF16 or ROUGE-L BIBREF17.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest\nIn this section we describe a methodology to categorise gold standards according to linguistic complexity, required reasoning and background knowledge, and their factual correctness. Specifically, we use those dimensions as high-level categories of a qualitative annotation schema for annotating question, expected answer and the corresponding context. We further enrich the qualitative annotations by a metric based on lexical cues in order to approximate a lower bound for the complexity of the reading comprehension task. By sampling entries from each gold standard and annotating them, we obtain measurable results and thus are able to make observations about the challenges present in that gold standard data.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Problem setting\nWe are interested in different types of the expected answer. We differentiate between Span, where an answer is a continuous span taken from the passage, Paraphrasing, where the answer is a paraphrase of a text span, Unanswerable, where there is no answer present in the context, and Generated, if it does not fall into any of the other categories. It is not sufficient for an answer to restate the question or combine multiple Span or Paraphrasing answers to be annotated as Generated. It is worth mentioning that we focus our investigations on answerable questions. For a complementary qualitative analysis that categorises unanswerable questions, the reader is referred to Yatskar2019.\nFurthermore, we mark a sentence as Supporting Fact if it contains evidence required to produce the expected answer, as they are used further in the complexity analysis.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Factual Correctness\nAn important factor for the quality of a benchmark is its factual correctness, because on the one hand, the presence of factually wrong or debatable examples introduces an upper bound for the achievable performance of models on those gold standards. On the other hand, it is hard to draw conclusions about the correctness of answers produced by a model that is evaluated on partially incorrect data.\nOne way by which developers of modern crowd-sourced gold standards ensure quality is by having the same entry annotated by multiple workers BIBREF18 and keeping only those with high agreement. We investigate whether this method is enough to establish a sound ground truth answer that is unambiguously correct. Concretely we annotate an answer as Debatable when the passage features multiple plausible answers, when multiple expected answers contradict each other, or an answer is not specific enough with respect to the question and a more specific answer is present. We annotate an answer as Wrong when it is factually wrong and a correct answer is present in the context.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Required Reasoning\nIt is important to understand what types of reasoning the benchmark evaluates, in order to be able to accredit various reasoning capabilities to the models it evaluates. Our proposed reasoning categories are inspired by those found in scientific question answering literature BIBREF19, BIBREF20, as research in this area focuses on understanding the required reasoning capabilities. We include reasoning about the Temporal succession of events, Spatial reasoning about directions and environment, and Causal reasoning about the cause-effect relationship between events. We further annotate (multiple-choice) answers that can only be answered By Exclusion of every other alternative.\nWe further extend the reasoning categories by operational logic, similar to those required in semantic parsing tasks BIBREF21, as solving those tasks typically requires “multi-hop” reasoning BIBREF14, BIBREF22. When an answer can only be obtained by combining information from different sentences joined by mentioning a common entity, concept, date, fact or event (from here on called entity), we annotate it as Bridge. We further annotate the cases, when the answer is a concrete entity that satisfies a Constraint specified in the question, when it is required to draw a Comparison of multiple entities' properties or when the expected answer is an Intersection of their properties (e.g. “What do Person A and Person B have in common?”)\nWe are interested in the linguistic reasoning capabilities probed by a gold standard, therefore we include the appropriate category used by Wang2019. Specifically, we annotate occurrences that require understanding of Negation, Quantifiers (such as “every”, “some”, or “all”), Conditional (“if ...then”) statements and the logical implications of Con-/Disjunction (i.e. “and” and “or”) in order to derive the expected answer.\nFinally, we investigate whether arithmetic reasoning requirements emerge in MRC gold standards as this can probe for reasoning that is not evaluated by simple answer retrieval BIBREF23. To this end, we annotate the presence of of Addition and Subtraction, answers that require Ordering of numerical values, Counting and Other occurrences of simple mathematical operations.\nAn example can exhibit multiple forms of reasoning. Notably, we do not annotate any of the categories mentioned above if the expected answer is directly stated in the passage. For example, if the question asks “How many total points were scored in the game?” and the passage contains a sentence similar to “The total score of the game was 51 points”, it does not require any reasoning, in which case we annotate it as Retrieval.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Knowledge\nWorthwhile knowing is whether the information presented in the context is sufficient to answer the question, as there is an increase of benchmarks deliberately designed to probe a model's reliance on some sort of background knowledge BIBREF24. We seek to categorise the type of knowledge required. Similar to Wang2019, on the one hand we annotate the reliance on factual knowledge, that is (Geo)political/Legal, Cultural/Historic, Technical/Scientific and Other Domain Specific knowledge about the world that can be expressed as a set of facts. On the other hand, we denote Intuitive knowledge requirements, which is challenging to express as a set of facts, such as the knowledge that a parenthetic numerical expression next to a person's name in a biography usually denotes his life span.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Linguistic Complexity\nAnother dimension of interest is the evaluation of various linguistic capabilities of MRC models BIBREF25, BIBREF26, BIBREF27. We aim to establish which linguistic phenomena are probed by gold standards and to which degree. To that end, we draw inspiration from the annotation schema used by Wang2019, and adapt it around lexical semantics and syntax.\nMore specifically, we annotate features that introduce variance between the supporting facts and the question. With regard to lexical semantics, we focus on the use of redundant words that do not alter the meaning of a sentence for the task of retrieving the expected answer (Redundancy), requirements on the understanding of words' semantic fields (Lexical Entailment) and the use of Synonyms and Paraphrases with respect to the question wording. Furthermore we annotate cases where supporting facts contain Abbreviations of concepts introduced in the question (and vice versa) and when a Dative case substitutes the use of a preposition (e.g. “I bought her a gift” vs “I bought a gift for her”). Regarding syntax, we annotate changes from passive to active Voice, the substitution of a Genitive case with a preposition (e.g. “of”) and changes from nominal to verbal style and vice versa (Nominalisation).\nWe recognise features that add ambiguity to the supporting facts, for example when information is only expressed implicitly by using an Ellipsis. As opposed to redundant words, we annotate Restrictivity and Factivity modifiers, words and phrases whose presence does change the meaning of a sentence with regard to the expected answer, and occurrences of intra- or inter-sentence Coreference in supporting facts (that is relevant to the question). Lastly, we mark ambiguous syntactic features, when their resolution is required in order to obtain the answer. Concretely, we mark argument collection with con- and disjunctions (Listing) and ambiguous Prepositions, Coordination Scope and Relative clauses/Adverbial phrases/Appositions.\nFramework for MRC Gold Standard Analysis ::: Dimensions of Interest ::: Complexity\nFinally, we want to approximate the presence of lexical cues that might simplify the reading required in order to arrive at the answer. Quantifying this allows for more reliable statements about and comparison of the complexity of gold standards, particularly regarding the evaluation of comprehension that goes beyond simple lexical matching. We propose the use of coarse metrics based on lexical overlap between question and context sentences. Intuitively, we aim to quantify how much supporting facts “stand out” from their surrounding passage context. This can be used as proxy for the capability to retrieve the answer BIBREF10. Specifically, we measure (i) the number of words jointly occurring in a question and a sentence, (ii) the length of the longest n-gram shared by question and sentence and (iii) whether a word or n-gram from the question uniquely appears in a sentence.\nThe resulting taxonomy of the framework is shown in Figure FIGREF10. The full catalogue of features, their description, detailed annotation guideline as well as illustrating examples can be found in Appendix .\nApplication of the Framework ::: Candidate Datasets\nWe select contemporary MRC benchmarks to represent all four commonly used problem definitions BIBREF15. In selecting relevant datasets, we do not consider those that are considered “solved”, i.e. where the state of the art performance surpasses human performance, as is the case with SQuAD BIBREF28, BIBREF7. Concretely, we selected gold standards that fit our problem definition and were published in the years 2016 to 2019, have at least $(2019 - publication\\ year) \\times 20$ citations, and bucket them according to the answer selection styles as described in Section SECREF4 We randomly draw one from each bucket and add two randomly drawn datasets from the candidate pool. This leaves us with the datasets described in Table TABREF19. For a more detailed description, we refer to Appendix .\nApplication of the Framework ::: Annotation Task\nWe randomly select 50 distinct question, answer and passage triples from the publicly available development sets of the described datasets. Training, development and the (hidden) test set are drawn from the same distribution defined by the data collection method of the respective dataset. For those collections that contain multiple questions over a single passage, we ensure that we are sampling unique paragraphs in order to increase the variety of investigated texts.\nThe samples were annotated by the first author of this paper, using the proposed schema. In order to validate our findings, we further take 20% of the annotated samples and present them to a second annotator (second author). Since at its core, the annotation is a multi-label task, we report the inter-annotator agreement by computing the (micro-averaged) F1 score, where we treat the first annotator's labels as gold. Table TABREF21 reports the agreement scores, the overall (micro) average F1 score of the annotations is 0.82, which means that on average, more than two thirds of the overall annotated labels were agreed on by both annotators. We deem this satisfactory, given the complexity of the annotation schema.\nApplication of the Framework ::: Qualitative Analysis\nWe present a concise view of the annotation results in Figure FIGREF23. The full annotation results can be found in Appendix . We centre our discussion around the following main points:\nApplication of the Framework ::: Qualitative Analysis ::: Linguistic Features\nAs observed in Figure FIGREF23 the gold standards feature a high degree of Redundancy, peaking at 76% of the annotated HotpotQA samples and synonyms and paraphrases (labelled Synonym), with ReCoRd samples containing 58% of them, likely to be attributed to the elaborating type of discourse of the dataset sources (encyclopedia and newswire). This is, however, not surprising, as it is fairly well understood in the literature that current state-of-the-art models perform well on distinguishing relevant words and phrases from redundant ones BIBREF32. Additionally, the representational capability of synonym relationships of word embeddings has been investigated and is well known BIBREF33. Finally, we observe the presence of syntactic features, such as ambiguous relative clauses, appositions and adverbial phrases, (RelAdvApp 40% in HotpotQA and ReCoRd) and those introducing variance, concretely switching between verbal and nominal styles (e.g. Nominalisation 10% in HotpotQA) and from passive to active voice (Voice, 8% in HotpotQA).\nSyntactic features contributing to variety and ambiguity that we did not observe in our samples are the exploitation of verb symmetry, the use of dative and genitive cases or ambiguous prepositions and coordination scope (respectively Symmetry, Dative, Genitive, Prepositions, Scope). Therefore we cannot establish whether models are capable of dealing with those features by evaluating them on those gold standards.\nApplication of the Framework ::: Qualitative Analysis ::: Factual Correctness\nWe identify three common sources that surface in different problems regarding an answer's factual correctness, as reported in Figure FIGREF23 and illustrate their instantiations in Table TABREF31:\nDesign Constraints: Choosing the task design and the data collection method introduces some constraints that lead to factually debatable examples. For example, a span might have been arbitrarily selected from multiple spans that potentially answer a question, but only a single continuous answer span per question is allowed by design, as observed in the NewsQA and MsMarco samples (32% and 34% examples annotated as Debatable with 16% and 53% thereof exhibiting arbitrary selection, respectively). Sometimes, when additional passages are added after the annotation step, they can by chance contain passages that answer the question more precisely than the original span, as seen in HotpotQA (16% Debatable samples, 25% of them due to arbitrary selection). In the case of MultiRC it appears to be inconsistent, whether multiple correct answer choices are expected to be correct in isolation or in conjunction (28% Debatable with 29% of them exhibiting this problem). This might provide an explanation to its relatively weak human baseline performance of 84% F1 score BIBREF31.\nWeak Quality assurance: When the (typically crowd-sourced) annotations are not appropriately validated, incorrect examples will find their way into the gold standards. This typically results in factually wrong expected answers (i.e. when a more correct answer is present in the context) or a question is expected to be Unanswerable, but is actually answerable from the provided context. The latter is observed in MsMarco (83% of examples annotated as Wrong) and NewsQA, where 60% of the examples annotated as Wrong are Unanswerable with an answer present.\nArbitrary Precision: There appears to be no clear guideline on how precise the answer is expected to be, when the passage expresses the answer in varying granularities. We annotated instances as Debatable when the expected answer was not the most precise given the context (44% and 29% of Debatable instances in NewsQA and MultiRC, respectively).\nApplication of the Framework ::: Qualitative Analysis ::: Semantics-altering grammatical modifiers\nWe took interest in whether any of the benchmarks contain what we call distracting lexical features (or distractors): grammatical modifiers that alter the semantics of a sentence for the final task of answering the given question while preserving a similar lexical form. An example of such features are cues for (double) Negation (e.g., “no”, “not”), which when introduced in a sentence, reverse its meaning. Other examples include modifiers denoting Restrictivity, Factivity and Reasoning (such as Monotonicity and Conditional cues). Examples of question-answer pairs containing a distractor are shown in Table FIGREF37.\nWe posit that the presence of such distractors would allow for evaluating reading comprehension beyond potential simple word matching. However, we observe no presence of such features in the benchmarks (beyond Negation in DROP, ReCoRd and HotpotQA, with 4%, 4% and 2% respectively). This results in gold standards that clearly express the evidence required to obtain the answer, lacking more challenging, i.e., distracting, sentences that can assess whether a model can truly understand meaning.\nApplication of the Framework ::: Qualitative Analysis ::: Other\nIn the Figure FIGREF23 we observe that Operational and Arithmetic reasoning moderately (6% to 8% combined) appears “in the wild”, i.e. when not enforced by the data design as is the case with HotpotQA (80% Operations combined) or DROP (68% Arithmetic combined). Causal reasoning is (exclusively) present in MultiRC (32%), whereas Temporal and Spatial reasoning requirements seem to not naturally emerge in gold standards. In ReCoRd, a fraction of 38% questions can only be answered By Exclusion of every other candidate, due to the design choice of allowing questions where the required information to answer them is not fully expressed in the accompanying paragraph.\nTherefore, it is also a little surprising to observe that ReCoRd requires external resources with regard to knowledge, as seen in Figure FIGREF23. MultiRC requires technical or more precisely basic scientific knowledge (6% Technical/Scientific), as a portion of paragraphs is extracted from elementary school science textbooks BIBREF31. Other benchmarks moderately probe for factual knowledge (0% to 4% across all categories), while Intuitive knowledge is required to derive answers in each gold standard.\nIt is also worth pointing out, as done in Figure FIGREF23, that although MultiRC and MsMarco are not modelled as a span selection problem, their samples still contain 50% and 66% of answers that are directly taken from the context. DROP contains the biggest fraction of generated answers (60%), due to the requirement of arithmetic operations.\nTo conclude our analysis, we observe similar distributions of linguistic features and reasoning patterns, except where there are constraints enforced by dataset design, annotation guidelines or source text choice. Furthermore, careful consideration of design choices (such as single-span answers) is required, to avoid impairing the factual correctness of datasets, as pure crowd-worker agreement seems not sufficient in multiple cases.\nApplication of the Framework ::: Quantitative Results ::: Lexical overlap\nWe used the scores assigned by our proposed set of metrics (discussed in Section SECREF11 Dimensions of Interest: Complexity) to predict the supporting facts in the gold standard samples (that we included in our manual annotation). Concretely, we used the following five features capturing lexical overlap: (i) the number of words occurring in sentence and question, (ii) the length of the longest n-gram shared by sentence and question, whether a (iii) uni- and (iv) bigram from the question is unique to a sentence, and (v) the sentence index, as input to a logistic regression classifier. We optimised on each sample leaving one example for evaluation. We compute the average Precision, Recall and F1 score by means of leave-one-out validation with every sample entry. The averaged results after 5 runs are reported in Table TABREF41.\nWe observe that even by using only our five features based lexical overlap, the simple logistic regression baseline is able to separate out the supporting facts from the context to a varying degree. This is in line with the lack of semantics-altering grammatical modifiers discussed in the qualitative analysis section above. The classifier performs best on DROP (66% F1) and MultiRC (40% F1), which means that lexical cues can considerably facilitate the search for the answer in those gold standards. On MultiRC, yadav2019quick come to a similar conclusion, by using a more sophisticated approach based on overlap between question, sentence and answer choices.\nSurprisingly, the classifier is able to pick up a signal from supporting facts even on data that has been pruned against lexical overlap heuristics by populating the context with additional documents that have high overlap scores with the question. This results in significantly higher scores than when guessing randomly (HotpotQA 26% F1, and MsMarco 11% F1). We observe similar results in the case the length of the question leaves few candidates to compute overlap with $6.3$ and $7.3$ tokens on average for MsMarco and NewsQA (26% F1), compared to $16.9$ tokens on average for the remaining four dataset samples.\nFinally, it is worth mentioning that although the queries in ReCoRd are explicitly independent from the passage, the linear classifier is still capable of achieving 34% F1 score in predicting the supporting facts.\nHowever, neural networks perform significantly better than our admittedly crude baseline (e.g. 66% F1 for supporting facts classification on HotpotQA BIBREF14), albeit utilising more training examples, and a richer sentence representation. This facts implies that those neural models are capable of solving more challenging problems than simple “text matching” as performed by the logistic regression baseline. However, they still circumvent actual reading comprehension as the respective gold standards are of limited suitability to evaluate this BIBREF34, BIBREF35. This suggests an exciting future research direction, that is categorising the scale between text matching and reading comprehension more precisely and respectively positioning state-of-the-art models thereon.\nRelated Work\nAlthough not as prominent as the research on novel architecture, there has been steady progress in critically investigating the data and evaluation aspects of NLP and machine learning in general and MRC in particular.\nRelated Work ::: Adversarial Evaluation\nThe authors of the AddSent algorithm BIBREF11 show that MRC models trained and evaluated on the SQuAD dataset pay too little attention to details that might change the semantics of a sentence, and propose a crowd-sourcing based method to generate adversary examples to exploit that weakness. This method was further adapted to be fully automated BIBREF36 and applied to different gold standards BIBREF35. Our proposed approach differs in that we aim to provide qualitative justifications for those quantitatively measured issues.\nRelated Work ::: Sanity Baselines\nAnother line of research establishes sane baselines to provide more meaningful context to the raw performance scores of evaluated models. When removing integral parts of the task formulation such as question, the textual passage or parts thereof BIBREF37 or restricting model complexity by design in order to suppress some required form of reasoning BIBREF38, models are still able to perform comparably to the state-of-the-art. This raises concerns about the perceived benchmark complexity and is related to our work in a broader sense as one of our goals is to estimate the complexity of benchmarks.\nRelated Work ::: Benchmark evaluation in NLP\nBeyond MRC, efforts similar to ours that pursue the goal of analysing the evaluation of established datasets exist in Natural Language Inference BIBREF13, BIBREF12. Their analyses reveal the existence of biases in training and evaluation data that can be approximated with simple majority-based heuristics. Because of these biases, trained models fail to extract the semantics that are required for the correct inference. Furthermore, a fair share of work was done to reveal gender bias in coreference resolution datasets and models BIBREF39, BIBREF40, BIBREF41.\nRelated Work ::: Annotation Taxonomies\nFinally, related to our framework are works that introduce annotation categories for gold standards evaluation. Concretely, we build our annotation framework around linguistic features that were introduced in the GLUE suite BIBREF42 and the reasoning categories introduced in the WorldTree dataset BIBREF19. A qualitative analysis complementary to ours, with focus on the unanswerability patterns in datasets that feature unanswerable questions was done by Yatskar2019.\nConclusion\nIn this paper, we introduce a novel framework to characterise machine reading comprehension gold standards. This framework has potential applications when comparing different gold standards, considering the design choices for a new gold standard and performing qualitative error analyses for a proposed approach.\nFurthermore we applied the framework to analyse popular state-of-the-art gold standards for machine reading comprehension: We reveal issues with their factual correctness, show the presence of lexical cues and we observe that semantics-altering grammatical modifiers are missing in all of the investigated gold standards. Studying how to introduce those modifiers into gold standards and observing whether state-of-the-art MRC models are capable of performing reading comprehension on text containing them, is a future research goal.\nA future line of research is to extend the framework to be able to identify the different types of exploitable cues such as question or entity typing and concrete overlap patterns. This will allow the framework to serve as an interpretable estimate of reading comprehension complexity of gold standards. Finally, investigating gold standards under this framework where MRC models outperform the human baseline (e.g. SQuAD) will contribute to a deeper understanding of the seemingly superb performance of deep learning approaches on them.",6  "char_len": 32718,7  "document_hash": "1b3ecc85293454ed3b8e1d98977820282d31940306b20a816030d376eaf02955",8  "questions": [9    {10      "question_id": "q00",11      "qtype": "qasper",12      "question": "How does proposed qualitative annotation schema looks like?"13    }14  ]15}