AgreeSum: Agreement-Oriented Multi-Document Summarization

We aim to renew interest in a particular multi-document summarization (MDS) task which we call AgreeSum: agreement-oriented multi-document summarization. Given a cluster of articles, the goal is to provide abstractive summaries that represent information common and faithful to all input articles. Given the lack of existing datasets, we create a dataset for AgreeSum, and provide annotations on article-summary entailment relations for a subset of the clusters in the dataset. We aim to create strong baselines for the task by applying the top-performing pretrained single-document summarization model PEGASUS onto AgreeSum, leveraging both annotated clusters by supervised losses, and unannotated clusters by T5-based entailment-related and language-related losses. Compared to other baselines, both automatic evaluation and human evaluation show better article-summary and cluster-summary entailment in generated summaries. On a separate note, we hope that our article-summary entailment annotations contribute to the community's effort in improving abstractive summarization faithfulness.


Introduction
Recent works have made great progress in singledocument summarization (SDS) thanks to the encoder-decoder framework and pretraining procedures (Cho et al., 2014;Rush et al., 2015;Narayan et al., 2018;Zhang et al., 2020a). There is a growing interest in multi-document summarization (MDS; Zopf, 2018;Fabbri et al., 2019;Liu and Lapata, 2019;Chu and Liu, 2019;Wang et al., 2020b, et seq.), with applications in search engines, news clustering, timeline generation, and other areas. Past MDS research has primarily focused on § Work completed during internship at Google Research. * Equal contribution.
summarizing articles such that the summary covers an event "comprehensively" while "avoiding redundancy" (Fabbri et al., 2019). We can say that most existing MDS tasks summarize the "union" of the articles.
In this paper, we discuss agreement-oriented multi-document summarization (AgreeSum), in which we aim to abstractively summarize the "intersection" of the articles. More specifically, the input to the task is a cluster of articles, and the expected output is a summary that represents information common and faithful to all input articles in the cluster (Section 3). A few works (discussed in Section 2) have investigated the problem, without using modern neural-network-based methods. The motivation for reviving interest in AgreeSum is twofold. First, given that certain microscopic details are not likely present in all articles in a given cluster, they would be filtered out through AgreeSum. If the source articles reflect different points of view, AgreeSum provides a way of capturing the common ground.
The second motivation for AgreeSum lies in the pursuit of summarization faithfulness. AgreeSum is timely, given that recent works have shown difficulty of producing faithful abstractive summaries (Falke et al., 2019;Maynez et al., 2020;Kryscinski et al., 2020;Durmus et al., 2020;Zhou et al., 2021), though in the SDS setting. AgreeSum could allow practitioners of abstractive summarization systems to carefully study topics related to faithfulness and hallucination.
Given the scarcity of readily available data, we create a dataset 1 based on English Wikipedia current events portal (WCEP) 2 . WCEP contains neutral human-written news summaries with links to news articles (with usually one link per summary). We first extract human-written summaries on WCEP, the linked articles, and semantically similar articles to the linked articles, in order to obtain article clusters associated with summaries. We then annotate the entailment relationship for a subset of the cluster-summary pairs (i.e., whether or not a given article from the cluster semantically entails the summary). From there, we build clustersummary pairs for AgreeSum (Section 3.2).
We build upon previous SDS research, using the top-performing pretrained PEGASUS model (Zhang et al., 2020a) as the starting point for our models. Using our dataset, we first examine a few baseline models and show that several pretraining-based summarization models fail at generating summaries that satisfy AgreeSum's requirements. We also propose an approach that integrates both supervised (using the annotated portion of the dataset) and unsupervised losses (namely, an entailment loss and a language loss to be discussed later, using both the annotated portion and the unannotated portion) while leveraging PEGA-SUS. We show the effectiveness of a simple policy gradient-based algorithm (Sutton et al., 2000) in which the rewards are based on a T5-based articlesummary entailment model (Raffel et al., 2020). To summarize the contributions: • We introduce the AgreeSum task and the WCEP-based dataset ( §3.2).
• A subset of article-summary pairs are annotated with entailment information ( §4). On a separate note, the article-summary-level recognizing textual entailment (RTE) 3 task could stand as a challenging task on its own. The annotations could be of interest to the research in improving abstract summarization faithfulness, in the context of not only AgreeSum but also general single-document summarization tasks.
• We develop simple ways of applying PEGA-SUS to AgreeSum. We provide a few baselines as well as a model that uses unsupervised entailment-related and language-related losses to complement the supervised finetuning of PEGASUS ( §5). Both automatic and human evaluations are performed ( §6). In particular, we show that the T5-based entailment model can be integrated into summarization models to produce summaries that are entailed in source articles.

Related Work
Traditional MDS. DUC (Paul and James, 2004;Dang, 2005) and TAC (Owczarzak and Dang, 2011) MDS data are among the first high-quality relevant datasets.  (2020), also based on WCEP, is especially relevant. However, our dataset is different in the following ways. First, all of our articles in the same cluster are about the same event. Next, a large part of our dataset is annotated with article-summary entailment information (i.e., in each of the clusters, for each article in the cluster, whether the article entails the summary; see Section 3). Further, among the annotated article-summary pairs, about half of the articles entail the summary, and half of the articles do not entail the summary. This property makes a realistic and difficult setting for AgreeSum tasks.
In terms of recent MDS neural methods, Chu and Liu (2019) summarize opinions using an autoencoder, in which case the input is much shorter than a typical article. Liu et al. (2019a) improve the model by encoding articles and summaries in the same space. Other novel approaches include using sentence compression in the seq2seq framework (Baziotis et al., 2019), jointly learning sentence fusion and paraphrasing (Nayeem et al., 2018), using graph neural networks to help extraction (Wang et al., 2020b), using spectral methods (Wang et al., 2020c), using transfer learning based on a novel pretraining method called gap-sentence prediction (Zhang et al., 2020a) on a news-specific corpus, among a few others Gu et al., 2020;Mao et al., 2020).
For MDS tasks that summarize the intersection of articles, a few past works have discussed the helpfulness of models that identify common infor-mation "centroids" among multiple related documents, so as to allow internet users to more efficiently understand events (Radev et al., 2000;Barzilay and McKeown, 2005). The attempted non-neural models rely heavily on topic/theme detection and tracking, and are more extractive than abstractive (Radev et al., 2004). The AgreeSum idea has not been fully explored by researchers since much stronger text generation technologies became available. It is timely to revisit the problem also because recent stronger neural abstractive summarization models are prone to hallucination, as are neural text generation in general (Wiseman et al., 2017;Tian et al., 2019;Wang and Sennrich, 2020;Pang and He, 2020).

Summarization hallucination and evaluation.
Non-hallucination is a necessary but not sufficient condition for performing well in AgreeSum: the summary not only needs to be entailed in the union of the articles, but also must be entailed in each of the articles. Unfortunately, recent works have shown the difficulty of identifying and mitigating hallucination (Maynez et al., 2020).
Evaluation-wise, researchers have found that metrics like ROUGE (Lin, 2004) and BERTScore (Zhang et al., 2020b) are only weakly correlated with factuality. Durmus et al. (2020) and Wang et al. (2020a) have therefore proposed using questionanswering systems for evaluating summarizers. Recently, Zhou et al. (2021) have made progress in creating token-level hallucination detectors which rely on negative (i.e., hallucination) data augmentation to train.
In terms of improving faithfulness and factuality, researchers did not find natural language inference (NLI) 4 models trained on standard NLI datasets to be robust enough for summarization-related downstream tasks (Falke et al., 2019). Contemporaneously, entity chains are used to explicitly ground the generations so that they become more faithful (Narayan et al., 2021). More broadly, there have been other recent works striving to develop techniques for high-precision generation (Malmi et al., 2019;Tian et al., 2019;Pang and He, 2020;Parikh et al., 2020;Dušek and Kasner, 2020).

Task
Short description of AgreeSum. The input is a cluster of around four articles (refer to Section 3.2 for more details) that describe the same event. However, the articles may have different levels of details and/or different levels of neutrality; e.g., one article in a cluster may be an opinion, while other articles may be neutral news. The expected output is an abstractive summary that represents information common and faithful to all input articles in the cluster. Moreover, the summary needs to be informative.

English AgreeSum Dataset Based on WCEP
Step 1. Recall that WCEP contains neutral human-written news summaries, each of which is linked to a news article. The first step is to obtain the summaries and 8 on-topic articles 5 for each summary based on WCEP. Specifically, we collect 5564 human-written summaries {y i } 5564 i=1 from WCEP. For each summary, we have one linked article to each summary on WCEP; we call the set of such articles {x . Given that we want to generate abstractive summaries, and to make the dataset challenging, the set of articles {x will not be used in the final dataset, to prevent excessive textual overlaps between the input articles and the target summaries. For each i, we obtain 8 other news articles x i that are semantically similar to x (0) i , based on a proprietary BERT-based clustering algorithm.
Step 2. The second step is to annotate entailment relations. Annotators are asked to judge if each of the articles in a cluster entails the summary (i.e., "does the article contain all the information presented in the summary?"). 1025 cluster-summary pairs are annotated. We designate ∼10% of the annotated clusters to be in the dev set. This step is discussed further in Section 4.
Step 3. To make AgreeSum moderately but not overly difficult, we set the maximum number of articles per cluster to be four, and the next step is to transform the dataset to have four articles per cluster. To take advantage of entailment annotations, we duplicate each annotated cluster in the training Figure 1: An annotated example (i.e., input-output pair where the input is a cluster of articles and the output is a summary) in the training set. In the figure, two articles in the input cluster entail the summary while the other two do not. An example is either annotated (meaning all article-summary pairs in the cluster-summary pair are annotated) or unannotated (meaning none of the article-summary pairs are annotated). Note that in the dev set, all articles entail the summary given that we would like to compare between generated summaries and the gold summaries.
i } ten times so that there is roughly an equal number of annotated and unannotated clusters. 6 Next, for each cluster, we randomly choose four articles to keep in the cluster, and discard the rest.
Dev and test sets. For the development set, we aim to designate the WCEP summaries as gold summaries. Therefore, for each cluster, we only keep the articles that entail the corresponding summary based on human annotation. In the case that a cluster has > 4 articles that entail the summary, we split the cluster into two, such that each newly formed cluster has ≥ 2 articles that entail the summary.
We use WCEP entries corresponding to dates before August 2019 for the training and development sets; for the test set, we sample 150 clusters of articles containing WCEP linked articles published between August 2019 to August 2020 to ensure that the test set does not have overlapping articles or overlapping publication times with the training and development sets.
For each sampled test cluster, we sample four articles from the cluster. Unlike the development set, these four articles are not annotated and are not guaranteed to entail any particular common summary. As a result, the provided WCEP summaries in the test set are not gold-standard summaries.
The test set is expected to be challenging due to construction methods, as well as the shift in topics of recent events (e.g., the COVID-19 pan-demic). The test set WCEP summaries are not goldstandard summaries, but are nevertheless provided given the potential use to approximate generation informedness (Section 6).
We thus obtain the dataset split shown in Table 1  4 Article-Summary Entailment

The Entailment Dataset
Workers have annotated the entailment relation of 1025 cluster-summary pairs (a subset of the clustersummary pairs obtained in Step 1 in Section 3.2), which correspond to 7750 article-summary pairs. For each article-summary pair, we ask five professionally trained annotators, randomly chosen from a pool of ∼800 raters, whether the summary is semantically entailed in the article; we then take the majority answer. See the appendix for more details. Our AgreeSum models leverage this entailment dataset, as discussed in Section 5. In addition, this dataset could be seen as a challenging RTE-style (or two-class NLI-style) task.

Model Performance
To explore how models perform on the articlesummary entailment dataset, and to see whether we can have a good entailment-classification model to guide the learning of our summarizer that encourages faithfulness to all input articles in the same cluster, we treat the dataset as a two-class NLI task where the label is either "entailed" or "not entailed," and examine the classification accuracy. Specifically, we evaluated the following three models. RoBERTa-large (Liu et al., 2019b) fine-tuned on MNLI fails on our entailment data 7 , likely in part because the premises in MNLI are sentences but are articles in our article-summary entailment task. It is also worth noting that unlike T5 (Raffel et al., 2020), RoBERTa-large is not pretrained on a multi-task mixture of many tasks, with each task converted into the text-to-text format.
We also attempt a model that integrates the PEGASUS-encoder which is pretrained in the news domain using the HugeNews corpus which contains around 1.5 billion articles. Given that pretrained PEGASUS does not include [CLS] tokens and it is very expensive to re-pretrain, we use a CNNbased classifier (Kim, 2014) whose input is the PEGASUS-encoder-outputs; more specifically, the convolutional layers pool over the sequence of encoder outputs. Based on the architecture by Kim (2014), our CNNs use filter n-gram sizes of 2, 3, 4, and 5, with 256 filters each. The resulting classifier achieves ∼68% accuracy on our entailment dataset. 8 In comparison, T5 (Raffel et al., 2020) shows encouraging results. The multi-task-trained T5large fine-tuned on our training set achieves 81.3% accuracy (79.1% for vanilla T5-large). T5-small fine-tuned on our training set achieves 79.5% accuracy (76.4% for vanilla T5-small).
We aim to see if using an entailment signal from a moderately good article-summary entailment classification model would help produce summaries that satisfy the AgreeSum criteria.

Notations and Baselines
We first provide some notations so as to allow easier discussion. Suppose the clusters of articles in the training set are denoted by {c i } i∈N . Each cluster contains at most four articles: For cluster i, let e i be the set of indices that correspond to articles that entail the summary. For example, in Figure 1 which corresponds to the ith cluster, we have e i = {1, 3}. Let E and D denote the encoder and the decoder of PEGASUS, respectively.
The following baselines are based on PEGA-SUS pretrained using gap-sentence prediction on HugeNews with 1.5 billion articles (3.8 terabytes). Note that despite being an SDS model, PEGASUS also achieves near-SoTA results on the Multi-News MDS dataset (Fabbri et al., 2019), so it is a competitive baseline for MDS as well.
B1: finetuning on (x (0) i , y i ) pairs. Recall that for each summary y i , we drop the WCEP-linked article x (0) i to prevent excessive textual overlaps between the input cluster and the output summary. In this baseline, however, we use y i 's in the training set and its corresponding x i as gold targets? On WCEP, it is reasonable to assume that the summaries are directly connected to the linked articles, and therefore, x i entails y i . Moreover, these summaries also inform the model of the style of the WCEP summaries. However, the downside is that the model could potentially prioritize extractions from the articles over entailment (i.e., the property that the generated summary is entailed in each article), as we would see in Section 6.
B2: concatenating truncated inputs. We finetune PEGASUS on the following: for each cluster c i , we truncate {x (j) i } j for each j ∈ e i and concatenate them such that the concatenated sequence has length ≤ 1024, given hardware constraints. We use a special symbol to delineate article boundaries.
B4: merging encodings and decode. Inspired by Chu and Liu (2019), we first encode x (j) i separately for each j ∈ e i , and pass the average of the encodings to the decoder (i.e., 1 Next, we do supervised learning based on the WCEP summaries. B5: best lead-1 sentence by entailment score. We first extract the first sentence of each article in the cluster. Next, we rerank the sentences using an entailment score, which is the mean of the binary entailment labels, predicted by fine-tuned T5-large (Section 4), between each article in the cluster and the given sentence. We choose the summary corresponding to the highest score, breaking ties randomly.

AgreeSum Model (ASM): Leveraging Unannotated Data
We introduce another baseline model called ASM. The distinguishing feature is that to target clustersummary entailment (meaning the summary is entailed in each article in the cluster), we attempt to use the T5-entailment-classification results (discussed in Section 4) as a training signal.
Supervised pretraining and training. Given pretrained PEGASUS, we first fine-tune according to B1. Next, for each cluster, we concatenate all elements of c i similar to B2, and mask out the articles that do not entail y i by the padding symbol; effectively, we use {x (j) i } j∈e i as input to the transformer. Then, we use the standard cross-entropy loss L ce to train the summarizer. However, only a small number of clusters are annotated, providing a very limited supervised signal.
Unsupervised entailment loss. We complement L ce with the entailment loss L e to learn entailment behavior. We fine-tune T5-small on our training dataset to predict entailment, and use this as our entailment classifier F e . In practice, we obtain T5outputs using remote procedure calls (RPCs).
F e takes in an article x and a summary y as inputs, and outputs 1 if x entails y and −1 otherwise. The loss L e is based on policy gradients (Williams, 1992;Sutton et al., 2000); we aim to maximize a sequence-level metric of a summary decoded from the model during training: where φ stands for the parameters of the encoder E and the decoder D. We thus have the following gradient: is the sequence-level return. Intuitively, during training, we sample a summaryỹ from our model, and run it through the T5 entailment classifier (which is separate from our summarizer) to obtain Q(ỹ). We then weight the MLE gradient (takingỹ as the target) byQ(ỹ). L e thus aims to guide our summarizer to generate summaries that entail in all or most of the articles.
Unsupervised language loss. Only using the entailment loss L e may result in degenerations. One way to encourage the model to generate fluent summaries and summaries that look like WCEP summaries is by using GAN-style (Goodfellow et al., 2014) objectives, which have achieved good performance in some conditional generation tasks like textual style transfer and machine translation (Shen et al., 2017;Wu et al., 2018;Pang and Gimpel, 2019). We thus use F l , a "language classifier" (i.e., discriminator) that distinguishes model generations from real dataset summaries (no matter annotated or not), to force our summarizer to generate sentences that look like the human-written summaries. F l is based on the CNN settings by Kim (2014) identical to the setting introduced in Section 4, while the inputs to CNNs are sentence representations obtained using PEGASUS.
Specifically, suppose there are k examples in a minibatch, then where h i is the decoder hidden states of WCEP summaries, andh i is the decoder hidden states of model generations, by professor forcing (Lamb et al., 2016). In the professor forcing algorithm, the input to F l is the hidden states instead of hard tokens so as to address the mismatch between training-time sequence prefix and test-time sequence prefix.

Summary.
We alternate updates among the following: (1) supervised training: updating E, D to minimize the loss L ce ; (2) unsupervised training: updating E, D to minimize the loss L e − λL l ; (3) language classifier training: updating F l to minimize the loss L l .

Implementation
We implement our model as a fork of the opensource 568M-parameter PEGASUS model (Zhang et al., 2020a). 9 We initialize PEGASUS from the "mixed & stochastic" checkpoint, which was pretrained for 1.5M steps. All baseline models are fine-tuned with a learning rate of 1e-4. All of the ASM models use 5e-5  Table 2: Results (on test set if not specified). ¶ : the first number corresponds to T5-large-evaluated results, and the second number corresponds to human-evaluated results. For each row, human raters annotated the entire test set; each article-generation pair is annotated by three raters, and we take the majority answer. The best result in each column is in blue; the worst in red. n-gram overlap: the proportion of generation n-grams that are also in the source. ‡ : the test set WCEP summaries should not be treated as references, given that there is no guarantee that the WCEP summary is entailed in each of the articles; the test set WCEP summaries are provided as an approximate measure of informedness. Note: B4 results in extensive hallucinations and very low ROUGE (∼10), and the ablation study of ASM without L l produces heavy degenerations, so they are omitted.
(tuned in {1e-5, 5e-5, 1e-4, 2e-4}). All models use beam search for decoding (beam size 8, beam alpha 0.8). Given hardware constraints, all models use a max input length of 1024. The max output decoding length is set to be 128. In addition, we tune λ ∈ {0.05, 0.1, 0.3, 0.5, 1, 2} and choose λ = 0.1 for all reported experiments that include the L l . We do not change any other default hyperparameter settings adopted from PEGASUS. Please refer to the appendix for more details.

Results
The following two models are not included in the table given their poor performance.
(1) The ablation study of ASM without language loss does not produce meaningful outputs. In this case, given that the T5 entailment classifier does not encourage language quality, the summaries in fact degenerate heavily.
(2) Model B4 (merging encodings and decode) is omitted from the table given very poor ROUGE performance (∼10) and extensive hallucination. Our conjecture is that the mean of the encodings of articles does not correspond to a meaningful encoding in the case of PEGASUS.

Agreement and Hallucination
First, we claim that ASM achieves better agreement. There are two types of agreement: articlesummary agreement which is the proportion of summaries entailed in the articles, and cluster-summary agreement which is the proportion of clusters in which all articles entail the summary. Table 2 reports both the T5-automatic evaluation and the human evaluation results on agreement.
(Preliminary) automatic metrics. Given an article-summary pair, we use T5-large fine-tuned on our entailment dataset (Section 4) to predict whether the summary is entailed in the article. For article-summary agreement, we compute the percentage of "entailed" classifications. For clustersummary agreement, we compute the percentage of clusters where all article-summary pairs lead to "entailed" classifications. We see that the T5evaluation results for ASM models perform better than the respective results for baseline models.
Human evaluation. For each row of Table 2, human raters annotated the entire test set (150 clusters, which corresponds to 600 article-generation pairs), on whether the generated summaries are entailed in the article. Workers were asked: Does the article contain all the information presented in the summary?
The full prompt is available in the appendix. One design choice is that we merge all articlesummary pairs for each cluster together into one task/HIT. Therefore, each task/HIT corresponds to four article-summary annotations. To reduce the inherent variance in human evaluation, each article-generation pair is annotated by three different raters, and we take the majority answer. Please see the appendix for more details. For article-summary agreement, we see that the human evaluation for the ASM models performs a little better than the extractive results (B5). ASM models perform more than 10 points better than the best abstractive baseline results (B1), which is in turn ∼5 points better than the other abstractive baseline results (B2, B3). For cluster-summary agreement, this improvement is even clearer, with ASM models performing more than 10 points better than any baseline.
In addition, we see that ASM reduces hallucination compared to B2 and B3. One way to approximate hallucination is by the number of clusters in which none of the articles entail the generated summary. Using both automatic and human evaluation results, we see that our model does better than B2 and B3, but a little worse than B1 which copies extensively from the source articles.

Discussion on ROUGE
In SDS tasks, Durmus et al. (2020) and Wang et al. (2020a) observe that ROUGE and BERTScore have a small correlation with summary factuality.
A hallucination or non-entailment 10 can have major text-span overlaps with the reference, thereby having a large ROUGE score. Given the nature of AgreeSum, and by Table 2, we confirm that high ROUGE does not imply entailment and should not be considered heavily in evaluation.
On the other hand, intuitively, we do recognize that an overly small ROUGE may indicate bad generations (e.g., extremely short generations, offtopic generations, and other degenerations like repetitions), which is the case for B4 generations as well as ASM-minus-L l generations.
Thus, practitioners need to rely on and determine the desired tradeoff between the following two automatic metrics: (1) ROUGE as a coarse proxy for summary quality and informedness, and (2) entailment-related and hallucination-related metrics (Section 6.1).
As a reminder, the development set summaries can be treated as gold-standard summaries. However, the test set summaries are not gold-standard summaries; they are only provided so as to allow one way to measure generation quality and informedness. Unlike the development set, none of the test-set input articles are filtered out through the summary-article entailment annotation procedure, given that we do not want to introduce potential bias through too much manipulation and filtering on raw test clusters.

More Observations
Language. We also asked workers to judge the language of the generations: Is this summary coherent and wellwritten with no self-contradictions or capitalization, spelling, punctuation, or grammar errors?
The full prompt is in the appendix. We see that the ASM-generated summaries are marginally better at the above. However, ASM without language loss results in heavy degenerations. On a separate note, we see that ASM generations tend to copy more than B2 and B3, but less than B1.
Examples. The appendix contains a few examples that compare generations from different models. For example, in Table 5, given four articles, we see the different generations that the systems produce. Article 4 is an opinion piece. ASM model correctly abstractively summarizes Article 2 and 4 in a way that agrees with 1 and 3. B1 copies from Article 1, while B2 and B3 have hallucinations. Given space constraints, please refer to the appendix.

Extension: Post-hoc Entailment Reranking after Decoding a Beam
To generate summaries that achieve better agreement, we also attempted a decoding trick which we name as entailment-oriented decoding, denoted by entdec in Table 3. We first define an entailment score used in this case. Given a cluster of articles and a generated summary, the entailment score is the mean of the T5-large-predicted binary labels (1 corresponds to "entailed" and 0 corresponds to "not entailed").
A model is suffixed as X-entdeck if we decode from X using beam search with beam size k; and after obtaining the size-k beam, we select the generation that corresponds to the largest T5 entailment score. We pick the beam with the largest score, using the original beam probabilities to break ties. Intuitively, this trick picks the best-entailment summary locally given that the generations in the same beam are usually similar. Table 3 shows that the entdec-trick generations indeed achieve higher article-summary agreement and human-summary agreement. We see that ASM still maintains the advantage in agreement, even if compared to entdec-decoded generations from other baselines.  standard NLI datasets do not offer robust benefits to improving summarization factuality. Maynez et al. (2020), on the other hand, rerank four summaries generated by four different models using BERTbased MNLI models, and find small improvements in faithfulness and factuality. However, these works rank different summaries after decoding assuming an existing summarizer (similar to our entdec trick), instead of updating the model parameters directly. Our contribution lies in the fact that we successfully use entailment models to improve the model during training time.
The major feature of our T5-based NLI model is that (1) our NLI model is based on multi-taskpretrained T5, implying that pretrained T5 can already handle article-length inputs well in certain tasks, and (2) our model is obtained after finetuning on our article-summary entailment dataset. Therefore, our T5-based NLI model is much better adjusted to the length of the premises (given that traditional NLI tasks correspond to sentence-level entailment, but our case corresponds to article-summary entailment). We thus see that using simple T5based binary signals can successfully improve entailment. However, more complicated modeling may be necessary if the AgreeSum cluster size becomes much larger.

Conclusion
We discuss the AgreeSum task with its dataset, and a range of baseline models. AgreeSum is timely given the recent focus on summarization faithfulness. In fact, we show that the summaries produced by several powerful pretraining-based baseline models are not able to follow AgreeSum's requirements satisfactorily. We welcome the community to contribute more advanced methods that work well on AgreeSum, especially when only a small subset of the dataset is labeled with articlesummary entailment information.
Within the AgreeSum dataset, we also provide article-summary entailment annotations on a subset of clusters, which we hope can contribute to the recent effort in improving abstractive summarization faithfulness.
Moreover, while there is contemporaneous development of complex approaches to encourage generated abstractive summaries to be entailed in the source articles, we show that it is feasible to improve the entailment behavior of generated summaries based on a binary article-summary entailment classifier.