{
  "document_id": "lb_qasper_8fa5af6a36dd0b6b73900b2ec6f6e43a652a3e7d2b827a58",
  "family": "public_qasper",
  "split": "test",
  "text": "Introduction\nLanguage model pretraining has advanced the state of the art in many NLP tasks ranging from sentiment analysis, to question answering, natural language inference, named entity recognition, and textual similarity. State-of-the-art pretrained models include ELMo BIBREF1, GPT BIBREF2, and more recently Bidirectional Encoder Representations from Transformers (Bert; BIBREF0). Bert combines both word and sentence representations in a single very large Transformer BIBREF3; it is pretrained on vast amounts of text, with an unsupervised objective of masked language modeling and next-sentence prediction and can be fine-tuned with various task-specific objectives.\nIn most cases, pretrained language models have been employed as encoders for sentence- and paragraph-level natural language understanding problems BIBREF0 involving various classification tasks (e.g., predicting whether any two sentences are in an entailment relationship; or determining the completion of a sentence among four alternative sentences). In this paper, we examine the influence of language model pretraining on text summarization. Different from previous tasks, summarization requires wide-coverage natural language understanding going beyond the meaning of individual words and sentences. The aim is to condense a document into a shorter version while preserving most of its meaning. Furthermore, under abstractive modeling formulations, the task requires language generation capabilities in order to create summaries containing novel words and phrases not featured in the source text, while extractive summarization is often defined as a binary classification task with labels indicating whether a text span (typically a sentence) should be included in the summary.\nWe explore the potential of Bert for text summarization under a general framework encompassing both extractive and abstractive modeling paradigms. We propose a novel document-level encoder based on Bert which is able to encode a document and obtain representations for its sentences. Our extractive model is built on top of this encoder by stacking several inter-sentence Transformer layers to capture document-level features for extracting sentences. Our abstractive model adopts an encoder-decoder architecture, combining the same pretrained Bert encoder with a randomly-initialized Transformer decoder BIBREF3. We design a new training schedule which separates the optimizers of the encoder and the decoder in order to accommodate the fact that the former is pretrained while the latter must be trained from scratch. Finally, motivated by previous work showing that the combination of extractive and abstractive objectives can help generate better summaries BIBREF4, we present a two-stage approach where the encoder is fine-tuned twice, first with an extractive objective and subsequently on the abstractive summarization task.\nWe evaluate the proposed approach on three single-document news summarization datasets representative of different writing conventions (e.g., important information is concentrated at the beginning of the document or distributed more evenly throughout) and summary styles (e.g., verbose vs. more telegraphic; extractive vs. abstractive). Across datasets, we experimentally show that the proposed models achieve state-of-the-art results under both extractive and abstractive settings. Our contributions in this work are three-fold: a) we highlight the importance of document encoding for the summarization task; a variety of recently proposed techniques aim to enhance summarization performance via copying mechanisms BIBREF5, BIBREF6, BIBREF7, reinforcement learning BIBREF8, BIBREF9, BIBREF10, and multiple communicating encoders BIBREF11. We achieve better results with a minimum-requirement model without using any of these mechanisms; b) we showcase ways to effectively employ pretrained language models in summarization under both extractive and abstractive settings; we would expect any improvements in model pretraining to translate in better summarization in the future; and c) the proposed models can be used as a stepping stone to further improve summarization performance as well as baselines against which new proposals are tested.\nBackground ::: Pretrained Language Models\nPretrained language models BIBREF1, BIBREF2, BIBREF0, BIBREF12, BIBREF13 have recently emerged as a key technology for achieving impressive gains in a wide variety of natural language tasks. These models extend the idea of word embeddings by learning contextual representations from large-scale corpora using a language modeling objective. Bidirectional Encoder Representations from Transformers (Bert; BIBREF0) is a new language representation model which is trained with a masked language modeling and a “next sentence prediction” task on a corpus of 3,300M words.\nThe general architecture of Bert is shown in the left part of Figure FIGREF2. Input text is first preprocessed by inserting two special tokens. [cls] is appended to the beginning of the text; the output representation of this token is used to aggregate information from the whole sequence (e.g., for classification tasks). And token [sep] is inserted after each sentence as an indicator of sentence boundaries. The modified text is then represented as a sequence of tokens $X=[w_1,w_2,\\cdots ,w_n]$. Each token $w_i$ is assigned three kinds of embeddings: token embeddings indicate the meaning of each token, segmentation embeddings are used to discriminate between two sentences (e.g., during a sentence-pair classification task) and position embeddings indicate the position of each token within the text sequence. These three embeddings are summed to a single input vector $x_i$ and fed to a bidirectional Transformer with multiple layers:\nwhere $h^0=x$ are the input vectors; $\\mathrm {LN}$ is the layer normalization operation BIBREF14; $\\mathrm {MHAtt}$ is the multi-head attention operation BIBREF3; superscript $l$ indicates the depth of the stacked layer. On the top layer, Bert will generate an output vector $t_i$ for each token with rich contextual information.\nPretrained language models are usually used to enhance performance in language understanding tasks. Very recently, there have been attempts to apply pretrained models to various generation problems BIBREF15, BIBREF16. When fine-tuning for a specific task, unlike ELMo whose parameters are usually fixed, parameters in Bert are jointly fine-tuned with additional task-specific parameters.\nBackground ::: Extractive Summarization\nExtractive summarization systems create a summary by identifying (and subsequently concatenating) the most important sentences in a document. Neural models consider extractive summarization as a sentence classification problem: a neural encoder creates sentence representations and a classifier predicts which sentences should be selected as summaries. SummaRuNNer BIBREF7 is one of the earliest neural approaches adopting an encoder based on Recurrent Neural Networks. Refresh BIBREF8 is a reinforcement learning-based system trained by globally optimizing the ROUGE metric. More recent work achieves higher performance with more sophisticated model structures. Latent BIBREF17 frames extractive summarization as a latent variable inference problem; instead of maximizing the likelihood of “gold” standard labels, their latent model directly maximizes the likelihood of human summaries given selected sentences. Sumo BIBREF18 capitalizes on the notion of structured attention to induce a multi-root dependency tree representation of the document while predicting the output summary. NeuSum BIBREF19 scores and selects sentences jointly and represents the state of the art in extractive summarization.\nBackground ::: Abstractive Summarization\nNeural approaches to abstractive summarization conceptualize the task as a sequence-to-sequence problem, where an encoder maps a sequence of tokens in the source document $\\mathbf {x} = [x_1, ..., x_n]$ to a sequence of continuous representations $\\mathbf {z} = [z_1, ..., z_n]$, and a decoder then generates the target summary $\\mathbf {y} = [y_1, ..., y_m]$ token-by-token, in an auto-regressive manner, hence modeling the conditional probability: $p(y_1, ..., y_m|x_1, ..., x_n)$.\nBIBREF20 and BIBREF21 were among the first to apply the neural encoder-decoder architecture to text summarization. BIBREF6 enhance this model with a pointer-generator network (PTgen) which allows it to copy words from the source text, and a coverage mechanism (Cov) which keeps track of words that have been summarized. BIBREF11 propose an abstractive system where multiple agents (encoders) represent the document together with a hierarchical attention mechanism (over the agents) for decoding. Their Deep Communicating Agents (DCA) model is trained end-to-end with reinforcement learning. BIBREF9 also present a deep reinforced model (DRM) for abstractive summarization which handles the coverage problem with an intra-attention mechanism where the decoder attends over previously generated words. BIBREF4 follow a bottom-up approach (BottomUp); a content selector first determines which phrases in the source document should be part of the summary, and a copy mechanism is applied only to preselected phrases during decoding. BIBREF22 propose an abstractive model which is particularly suited to extreme summarization (i.e., single sentence summaries), based on convolutional neural networks and additionally conditioned on topic distributions (TConvS2S).\nFine-tuning Bert for Summarization ::: Summarization Encoder\nAlthough Bert has been used to fine-tune various NLP tasks, its application to summarization is not as straightforward. Since Bert is trained as a masked-language model, the output vectors are grounded to tokens instead of sentences, while in extractive summarization, most models manipulate sentence-level representations. Although segmentation embeddings represent different sentences in Bert, they only apply to sentence-pair inputs, while in summarization we must encode and manipulate multi-sentential inputs. Figure FIGREF2 illustrates our proposed Bert architecture for Summarization (which we call BertSum).\nIn order to represent individual sentences, we insert external [cls] tokens at the start of each sentence, and each [cls] symbol collects features for the sentence preceding it. We also use interval segment embeddings to distinguish multiple sentences within a document. For $sent_i$ we assign segment embedding $E_A$ or $E_B$ depending on whether $i$ is odd or even. For example, for document $[sent_1, sent_2, sent_3, sent_4, sent_5]$, we would assign embeddings $[E_A, E_B, E_A,E_B, E_A]$. This way, document representations are learned hierarchically where lower Transformer layers represent adjacent sentences, while higher layers, in combination with self-attention, represent multi-sentence discourse.\nPosition embeddings in the original Bert model have a maximum length of 512; we overcome this limitation by adding more position embeddings that are initialized randomly and fine-tuned with other parameters in the encoder.\nFine-tuning Bert for Summarization ::: Extractive Summarization\nLet $d$ denote a document containing sentences $[sent_1, sent_2, \\cdots , sent_m]$, where $sent_i$ is the $i$-th sentence in the document. Extractive summarization can be defined as the task of assigning a label $y_i \\in \\lbrace 0, 1\\rbrace $ to each $sent_i$, indicating whether the sentence should be included in the summary. It is assumed that summary sentences represent the most important content of the document.\nWith BertSum, vector $t_i$ which is the vector of the $i$-th [cls] symbol from the top layer can be used as the representation for $sent_i$. Several inter-sentence Transformer layers are then stacked on top of Bert outputs, to capture document-level features for extracting summaries:\nwhere $h^0=\\mathrm {PosEmb}(T)$; $T$ denotes the sentence vectors output by BertSum, and function $\\mathrm {PosEmb}$ adds sinusoid positional embeddings BIBREF3 to $T$, indicating the position of each sentence.\nThe final output layer is a sigmoid classifier:\nwhere $h^L_i$ is the vector for $sent_i$ from the top layer (the $L$-th layer ) of the Transformer. In experiments, we implemented Transformers with $L=1, 2, 3$ and found that a Transformer with $L=2$ performed best. We name this model BertSumExt.\nThe loss of the model is the binary classification entropy of prediction $\\hat{y}_i$ against gold label $y_i$. Inter-sentence Transformer layers are jointly fine-tuned with BertSum. We use the Adam optimizer with $\\beta _1=0.9$, and $\\beta _2=0.999$). Our learning rate schedule follows BIBREF3 with warming-up ($ \\operatorname{\\operatorname{warmup}}=10,000$):\nFine-tuning Bert for Summarization ::: Abstractive Summarization\nWe use a standard encoder-decoder framework for abstractive summarization BIBREF6. The encoder is the pretrained BertSum and the decoder is a 6-layered Transformer initialized randomly. It is conceivable that there is a mismatch between the encoder and the decoder, since the former is pretrained while the latter must be trained from scratch. This can make fine-tuning unstable; for example, the encoder might overfit the data while the decoder underfits, or vice versa. To circumvent this, we design a new fine-tuning schedule which separates the optimizers of the encoder and the decoder.\nWe use two Adam optimizers with $\\beta _1=0.9$ and $\\beta _2=0.999$ for the encoder and the decoder, respectively, each with different warmup-steps and learning rates:\nwhere $\\tilde{lr}_{\\mathcal {E}}=2e^{-3}$, and $\\operatorname{\\operatorname{warmup}}_{\\mathcal {E}}=20,000$ for the encoder and $\\tilde{lr}_{\\mathcal {D}}=0.1$, and $\\operatorname{\\operatorname{warmup}}_{\\mathcal {D}}=10,000$ for the decoder. This is based on the assumption that the pretrained encoder should be fine-tuned with a smaller learning rate and smoother decay (so that the encoder can be trained with more accurate gradients when the decoder is becoming stable).\nIn addition, we propose a two-stage fine-tuning approach, where we first fine-tune the encoder on the extractive summarization task (Section SECREF8) and then fine-tune it on the abstractive summarization task (Section SECREF13). Previous work BIBREF4, BIBREF23 suggests that using extractive objectives can boost the performance of abstractive summarization. Also notice that this two-stage approach is conceptually very simple, the model can take advantage of information shared between these two tasks, without fundamentally changing its architecture. We name the default abstractive model BertSumAbs and the two-stage fine-tuned model BertSumExtAbs.\nExperimental Setup\nIn this section, we describe the summarization datasets used in our experiments and discuss various implementation details.\nExperimental Setup ::: Summarization Datasets\nWe evaluated our model on three benchmark datasets, namely the CNN/DailyMail news highlights dataset BIBREF24, the New York Times Annotated Corpus (NYT; BIBREF25), and XSum BIBREF22. These datasets represent different summary styles ranging from highlights to very brief one sentence summaries. The summaries also vary with respect to the type of rewriting operations they exemplify (e.g., some showcase more cut and paste operations while others are genuinely abstractive). Table TABREF12 presents statistics on these datasets (test set); example (gold-standard) summaries are provided in the supplementary material.\nExperimental Setup ::: Summarization Datasets ::: CNN/DailyMail\ncontains news articles and associated highlights, i.e., a few bullet points giving a brief overview of the article. We used the standard splits of BIBREF24 for training, validation, and testing (90,266/1,220/1,093 CNN documents and 196,961/12,148/10,397 DailyMail documents). We did not anonymize entities. We first split sentences with the Stanford CoreNLP toolkit BIBREF26 and pre-processed the dataset following BIBREF6. Input documents were truncated to 512 tokens.\nExperimental Setup ::: Summarization Datasets ::: NYT\ncontains 110,540 articles with abstractive summaries. Following BIBREF27, we split these into 100,834/9,706 training/test examples, based on the date of publication (the test set contains all articles published from January 1, 2007 onward). We used 4,000 examples from the training as validation set. We also followed their filtering procedure, documents with summaries less than 50 words were removed from the dataset. The filtered test set (NYT50) includes 3,452 examples. Sentences were split with the Stanford CoreNLP toolkit BIBREF26 and pre-processed following BIBREF27. Input documents were truncated to 800 tokens.\nExperimental Setup ::: Summarization Datasets ::: XSum\ncontains 226,711 news articles accompanied with a one-sentence summary, answering the question “What is this article about?”. We used the splits of BIBREF22 for training, validation, and testing (204,045/11,332/11,334) and followed the pre-processing introduced in their work. Input documents were truncated to 512 tokens.\nAside from various statistics on the three datasets, Table TABREF12 also reports the proportion of novel bi-grams in gold summaries as a measure of their abstractiveness. We would expect models with extractive biases to perform better on datasets with (mostly) extractive summaries, and abstractive models to perform more rewrite operations on datasets with abstractive summaries. CNN/DailyMail and NYT are somewhat abstractive, while XSum is highly abstractive.\nExperimental Setup ::: Implementation Details\nFor both extractive and abstractive settings, we used PyTorch, OpenNMT BIBREF28 and the `bert-base-uncased' version of Bert to implement BertSum. Both source and target texts were tokenized with Bert's subwords tokenizer.\nExperimental Setup ::: Implementation Details ::: Extractive Summarization\nAll extractive models were trained for 50,000 steps on 3 GPUs (GTX 1080 Ti) with gradient accumulation every two steps. Model checkpoints were saved and evaluated on the validation set every 1,000 steps. We selected the top-3 checkpoints based on the evaluation loss on the validation set, and report the averaged results on the test set. We used a greedy algorithm similar to BIBREF7 to obtain an oracle summary for each document to train extractive models. The algorithm generates an oracle consisting of multiple sentences which maximize the ROUGE-2 score against the gold summary.\nWhen predicting summaries for a new document, we first use the model to obtain the score for each sentence. We then rank these sentences by their scores from highest to lowest, and select the top-3 sentences as the summary.\nDuring sentence selection we use Trigram Blocking to reduce redundancy BIBREF9. Given summary $S$ and candidate sentence $c$, we skip $c$ if there exists a trigram overlapping between $c$ and $S$. The intuition is similar to Maximal Marginal Relevance (MMR; BIBREF29); we wish to minimize the similarity between the sentence being considered and sentences which have been already selected as part of the summary.\nExperimental Setup ::: Implementation Details ::: Abstractive Summarization\nIn all abstractive models, we applied dropout (with probability $0.1$) before all linear layers; label smoothing BIBREF30 with smoothing factor $0.1$ was also used. Our Transformer decoder has 768 hidden units and the hidden size for all feed-forward layers is 2,048. All models were trained for 200,000 steps on 4 GPUs (GTX 1080 Ti) with gradient accumulation every five steps. Model checkpoints were saved and evaluated on the validation set every 2,500 steps. We selected the top-3 checkpoints based on their evaluation loss on the validation set, and report the averaged results on the test set.\nDuring decoding we used beam search (size 5), and tuned the $\\alpha $ for the length penalty BIBREF31 between $0.6$ and 1 on the validation set; we decode until an end-of-sequence token is emitted and repeated trigrams are blocked BIBREF9. It is worth noting that our decoder applies neither a copy nor a coverage mechanism BIBREF6, despite their popularity in abstractive summarization. This is mainly because we focus on building a minimum-requirements model and these mechanisms may introduce additional hyper-parameters to tune. Thanks to the subwords tokenizer, we also rarely observe issues with out-of-vocabulary words in the output; moreover, trigram-blocking produces diverse summaries managing to reduce repetitions.\nResults ::: Automatic Evaluation\nWe evaluated summarization quality automatically using ROUGE BIBREF32. We report unigram and bigram overlap (ROUGE-1 and ROUGE-2) as a means of assessing informativeness and the longest common subsequence (ROUGE-L) as a means of assessing fluency. Table TABREF23 summarizes our results on the CNN/DailyMail dataset. The first block in the table includes the results of an extractive Oracle system as an upper bound. We also present the Lead-3 baseline (which simply selects the first three sentences in a document). The second block in the table includes various extractive models trained on the CNN/DailyMail dataset (see Section SECREF5 for an overview). For comparison to our own model, we also implemented a non-pretrained Transformer baseline (TransformerExt) which uses the same architecture as BertSumExt, but with fewer parameters. It is randomly initialized and only trained on the summarization task. TransformerExt has 6 layers, the hidden size is 512, and the feed-forward filter size is 2,048. The model was trained with same settings as in BIBREF3. The third block in Table TABREF23 highlights the performance of several abstractive models on the CNN/DailyMail dataset (see Section SECREF6 for an overview). We also include an abstractive Transformer baseline (TransformerAbs) which has the same decoder as our abstractive BertSum models; the encoder is a 6-layer Transformer with 768 hidden size and 2,048 feed-forward filter size. The fourth block reports results with fine-tuned Bert models: BertSumExt and its two variants (one without interval embeddings, and one with the large version of Bert), BertSumAbs, and BertSumExtAbs. Bert-based models outperform the Lead-3 baseline which is not a strawman; on the CNN/DailyMail corpus it is indeed superior to several extractive BIBREF7, BIBREF8, BIBREF19 and abstractive models BIBREF6. Bert models collectively outperform all previously proposed extractive and abstractive systems, only falling behind the Oracle upper bound. Among Bert variants, BertSumExt performs best which is not entirely surprising; CNN/DailyMail summaries are somewhat extractive and even abstractive models are prone to copying sentences from the source document when trained on this dataset BIBREF6. Perhaps unsurprisingly we observe that larger versions of Bert lead to performance improvements and that interval embeddings bring only slight gains. Table TABREF24 presents results on the NYT dataset. Following the evaluation protocol in BIBREF27, we use limited-length ROUGE Recall, where predicted summaries are truncated to the length of the gold summaries. Again, we report the performance of the Oracle upper bound and Lead-3 baseline. The second block in the table contains previously proposed extractive models as well as our own Transformer baseline. Compress BIBREF27 is an ILP-based model which combines compression and anaphoricity constraints. The third block includes abstractive models from the literature, and our Transformer baseline. Bert-based models are shown in the fourth block. Again, we observe that they outperform previously proposed approaches. On this dataset, abstractive Bert models generally perform better compared to BertSumExt, almost approaching Oracle performance.\nTable TABREF26 summarizes our results on the XSum dataset. Recall that summaries in this dataset are highly abstractive (see Table TABREF12) consisting of a single sentence conveying the gist of the document. Extractive models here perform poorly as corroborated by the low performance of the Lead baseline (which simply selects the leading sentence from the document), and the Oracle (which selects a single-best sentence in each document) in Table TABREF26. As a result, we do not report results for extractive models on this dataset. The second block in Table TABREF26 presents the results of various abstractive models taken from BIBREF22 and also includes our own abstractive Transformer baseline. In the third block we show the results of our Bert summarizers which again are superior to all previously reported models (by a wide margin).\nResults ::: Model Analysis ::: Learning Rates\nRecall that our abstractive model uses separate optimizers for the encoder and decoder. In Table TABREF27 we examine whether the combination of different learning rates ($\\tilde{lr}_{\\mathcal {E}}$ and $\\tilde{lr}_{\\mathcal {D}}$) is indeed beneficial. Specifically, we report model perplexity on the CNN/DailyMail validation set for varying encoder/decoder learning rates. We can see that the model performs best with $\\tilde{lr}_{\\mathcal {E}}=2e-3$ and $\\tilde{lr}_{\\mathcal {D}}=0.1$.\nResults ::: Model Analysis ::: Position of Extracted Sentences\nIn addition to the evaluation based on ROUGE, we also analyzed in more detail the summaries produced by our model. For the extractive setting, we looked at the position (in the source document) of the sentences which were selected to appear in the summary. Figure FIGREF31 shows the proportion of selected summary sentences which appear in the source document at positions 1, 2, and so on. The analysis was conducted on the CNN/DailyMail dataset for Oracle summaries, and those produced by BertSumExt and the TransformerExt. We can see that Oracle summary sentences are fairly smoothly distributed across documents, while summaries created by TransformerExt mostly concentrate on the first document sentences. BertSumExt outputs are more similar to Oracle summaries, indicating that with the pretrained encoder, the model relies less on shallow position features, and learns deeper document representations.\nResults ::: Model Analysis ::: Novel N-grams\nWe also analyzed the output of abstractive systems by calculating the proportion of novel n-grams that appear in the summaries but not in the source texts. The results are shown in Figure FIGREF33. In the CNN/DailyMail dataset, the proportion of novel n-grams in automatically generated summaries is much lower compared to reference summaries, but in XSum, this gap is much smaller. We also observe that on CNN/DailyMail, BertExtAbs produces less novel n-ngrams than BertAbs, which is not surprising. BertExtAbs is more biased towards selecting sentences from the source document since it is initially trained as an extractive model. The supplementary material includes examples of system output and additional ablation studies.\nResults ::: Human Evaluation\nIn addition to automatic evaluation, we also evaluated system output by eliciting human judgments. We report experiments following a question-answering (QA) paradigm BIBREF33, BIBREF8 which quantifies the degree to which summarization models retain key information from the document. Under this paradigm, a set of questions is created based on the gold summary under the assumption that it highlights the most important document content. Participants are then asked to answer these questions by reading system summaries alone without access to the article. The more questions a system can answer, the better it is at summarizing the document as a whole. Moreover, we also assessed the overall quality of the summaries produced by abstractive systems which due to their ability to rewrite content may produce disfluent or ungrammatical output. Specifically, we followed the Best-Worst Scaling BIBREF34 method where participants were presented with the output of two systems (and the original document) and asked to decide which one was better according to the criteria of Informativeness, Fluency, and Succinctness.\nBoth types of evaluation were conducted on the Amazon Mechanical Turk platform. For the CNN/DailyMail and NYT datasets we used the same documents (20 in total) and questions from previous work BIBREF8, BIBREF18. For XSum, we randomly selected 20 documents (and their questions) from the release of BIBREF22. We elicited 3 responses per HIT. With regard to QA evaluation, we adopted the scoring mechanism from BIBREF33; correct answers were marked with a score of one, partially correct answers with 0.5, and zero otherwise. For quality-based evaluation, the rating of each system was computed as the percentage of times it was chosen as better minus the times it was selected as worse. Ratings thus range from -1 (worst) to 1 (best).\nResults for extractive and abstractive systems are shown in Tables TABREF37 and TABREF38, respectively. We compared the best performing BertSum model in each setting (extractive or abstractive) against various state-of-the-art systems (whose output is publicly available), the Lead baseline, and the Gold standard as an upper bound. As shown in both tables participants overwhelmingly prefer the output of our model against comparison systems across datasets and evaluation paradigms. All differences between BertSum and comparison models are statistically significant ($p<0.05$), with the exception of TConvS2S (see Table TABREF38; XSum) in the QA evaluation setting.\nConclusions\nIn this paper, we showcased how pretrained Bert can be usefully applied in text summarization. We introduced a novel document-level encoder and proposed a general framework for both abstractive and extractive summarization. Experimental results across three datasets show that our model achieves state-of-the-art results across the board under automatic and human-based evaluation protocols. Although we mainly focused on document encoding for summarization, in the future, we would like to take advantage the capabilities of Bert for language generation.\nAcknowledgments\nThis research is supported by a Google PhD Fellowship to the first author. We gratefully acknowledge the support of the European Research Council (Lapata, award number 681760, “Translating Multiple Modalities into Text”). We would also like to thank Shashi Narayan for providing us with the XSum dataset.",
  "char_len": 30391,
  "document_hash": "8bd512af0c4c3757e43fbef60070ed3c2d2bba018fa9c9b2313bc12c15f526ed",
  "questions": [
    {
      "question_id": "q00",
      "qtype": "qasper",
      "question": "What are the datasets used for evaluation?"
    }
  ]
}