# nutribullets_summarizing_health_studies_by_composing_segments__eaf91691.pdf

Nutri-bullets: Summarizing Health Studies by Composing Segments

Darsh J Shah,1 Lili Yu,2 Tao Lei,2 Regina Barzilay1

1Computer Science and Artificial Intelligence Lab, MIT 2ASAPP, Inc. darsh@csail.mit.edu, liliyu@asapp.com, tao@asapp.com, regina@csail.mit.edu

We introduce Nutri-bullets, a multi-document summarization task for health and nutrition. First, we present two datasets of food and health summaries from multiple scientific studies. Furthermore, we propose a novel extract-compose model to solve the problem in the regime of limited parallel data. We explicitly select key spans from several abstracts using a policy network, followed by composing the selected spans to present a summary via a task specific language model. Compared to state-of-the-art methods, our approach leads to more faithful, relevant and diverse summarization properties imperative to this application. For instance, on the Breast Cancer dataset our approach gets a more than 50% improvement on relevance and faithfulness.1

1 Introduction Multi-document summarization is essential in domains like health and nutrition, where new studies are continuously reported (see Figure 1). Websites like Healthline.com2 are critical in making food and nutrition summaries available to web users and subscribers. Such summaries provide a collective view of information a crucial property, especially when the consumption of such knowledge leads to health related decisions. In the current environment, it is critical to provide users with the the latest health related findings. Unfortunately, summarization is time consuming and requires domain experts. Through the introduction of Nutri-bullets, a multidocument summarization task, we aim to automate this summarization and expand the availability and coverage of health and nutrition information. Recently, summarization has been predominantly solved by training sequence-to-sequence (seq2seq) models (Rush, Chopra, and Weston 2015; Vaswani et al. 2017; Hoang et al. 2019; Lewis et al. 2019). In the multi-document setting, popular seq2seq methods either concatenate all input documents as a single source or consider a hierarchical setting (Liu et al. 2018). While such methods generate fluent text, in our case, they fail to produce content faithful to the inputs (examples

Copyright 2021, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. 1Our code and data is submitted and will be made publicly available on acceptance. 2https://healthline.com

Figure 1: Our model s extract-compose summary describing impacts of whole grains on cancer. Segments from multiple scientific studies are composed to present the output.

Figure 2: Examples of unfaithful summaries generated by seq2seq models trained on the Health Line dataset. Such fictitious texts (red) can be extremely misleading for readers.

shown in Figure 2). Two factors make Nutri-bullets particularly challenging for seq2seq models: (i) The concatenation of health documents constitutes long sequences with key information being scattered making composing a good summary extremely difficult; and (ii) While a vast number of scientific abstracts are available in libraries such as Pubmed3, a very limited number of summaries, ranging from several hundreds to a few thousands, are available to train a summarization system. In this paper, we present a novel and practical approach which learns to extract and compose knowledge pieces from scientific abstracts into a summary. We side-step the scarcity of parallel data by focusing on knowledge-extraction from scientific abstracts, made possible by modules trained on such crowd-sourced annotations. Furthermore, our model learns to select a subset of knowledge pieces well-suited for a summary, from a large pool. The selection is performed through a multi-step framework, guided by reinforcement learning and distant supervision from limited available par-

3https://pubmed.ncbi.nlm.nih.gov/

The Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI-21)

allel data. While considering the extracted text as inputs to a seq2seq model already makes the learning-to-generate process easier, we further explore an alternative approach to compose the extracted text into a complete summary using a text infilling language model. Once key spans are selected from different documents, they are fused using a task specific language model (Shen et al. 2020) as illustrated in Figure 3. The fusion of separate yet related spans through generation is critical in producing meaningful and readable summaries.4 We conduct human and empirical evaluation to comprehensively study the applicability and quality of these approaches. While seq2seq models enjoy high fluency scores, the extract-compose method performs much stronger on metrics such as content, relevance, faithfulness and informativeness. Our method is particularly dominant in scenarios with scarce parallel data, since our model requires little summary data for training. For instance, on the Breast Cancer dataset, humans rate the extract-compose model s summaries more than 50% higher for relevance and faithfulness than the next best baseline. Comparison with strong baselines and model ablation variants highlights the necessity of a distant supervision and reinforcement learning based multi-step approach, in selecting key ordered spans amongst several possible combinations, for text composition. Our contributions are threefold: (i) We collect two new nutrition and health related datasets for multi-document summarization. We also collect largescale knowledge extraction annotations, applicable to numerous tasks. (ii) We demonstrate the effectiveness of our modelling approach for generating health summaries given limited parallel data. Our approach strongly outperforms all baselines and variants on human and automatic evaluation. (iii) We conduct comprehensive (human and automatic) evaluation focusing on content relevance, faithfulness and informativeness metrics more relevant to the task. These set new benchmarks to critically evaluate summaries in high impact domains.

2 Related Work Multi-document Summarization Approaches in neural sequence-to-sequence learning (Rush, Chopra, and Weston 2015; Cheng and Lapata 2016; See, Liu, and Manning 2017) for document summarization have shown promise and have been adapted successfully for multi-document summarization (Zhang, Tan, and Wan 2018; Lebanoff, Song, and Liu 2018; Baumel, Eyal, and Elhadad 2018; Amplayo and Lapata 2019; Fabbri et al. 2019). Trained on large amounts of data, these methods have improved upon traditional extractive (Carbonell and Goldstein 1998; Radev and Mc Keown 1998; Haghighi and Vanderwende 2009) and abstractive approaches (Barzilay, Mc Keown, and Elhadad 1999; Mc Keown and Radev 1995; Ganesan, Zhai, and Han 2010). De-

4Our analysis shows that 91% of all produced tokens are from extracted components. The generated words allow fusion and cohesion.

spite producing fluent text, these techniques also tend to generate false information which is not faithful to the original inputs (Puduppully, Dong, and Lapata 2019; Kry sci nski et al. 2019). Side-information, such as citations in scientific domains (Qazvinian and Radev 2008; Qazvinian et al. 2013) or semantic representations (Liu et al. 2015), can be used to improve this (Sharma et al. 2019; Wenbo et al. 2019; Puduppully, Dong, and Lapata 2019; Koncel-Kedziorski et al. 2019a). However, such methods struggle in low resource scenarios. In this work, we are interested in producing faithful and fluent text in a technical domain where few parallel examples are available. Text Fusion Traditionally, sentence fusion approaches (Barzilay and Mc Keown 2005) aid the concatenation of different text fragments for summarization. Recent language modeling approaches like Devlin et al. (2018); Stern et al. (2019) can also be extended for completion and fusion of partial text. These models have more flexibility than those trained on text fusion datasets (Narayan et al. 2017; Geva et al. 2019) that can combine two fragments only. In this work, we modify the Blank Language Model (Shen et al. 2020) to combine fragments coming from different source documents. Deep Reinforcement Learning for Text Summarization The inherent discrete and sequential nature of text generation tasks has made optimizing generation models with reinforcement learning (Sutton and Barto 2018) very popular (Keneshloo et al. 2019). Typically, automatic evaluation metrics, like BLEU (Papineni et al. 2002) or ROUGE (Lin 2004), are used to provide the reward signal to train reinforcement learning algorithms in translation (Wu et al. 2018), summarization (Paulus, Xiong, and Socher 2017; Pasunuru and Bansal 2018; Chen and Bansal 2018; Xu et al. 2019) and question generation (Zhang and Bansal 2019) tasks. In this work, we are interested in using reinforcement learning to iteratively, select an explicit ordered subset of text phrases for downstream fusion.

3 Method In this section, we describe the framework of our extractcompose solution for nutri-bullets. Our goal is to produce a text summary y for a food item from a pool of multiple scientific abstracts X.

3.1 Extract-compose Framework We attain food health entity-entity relations, for both input documents X and the summary y, from entity extraction and relation classification modules trained on corresponding annotations(Table 1). For input documents, we collect {(xp, Gp)N p=1}, where xp is the sentence text and Gp is the set of entity-entity relations, in N sentences and X = {xp} is the raw text. Gp = {(ek i , ek j , rk)K} is composed of K {0, 1, 2, . . . } tuples of two entities ei, ej and their relation r. r represents relations such as the effect of a nutrition entity ei on a condition ej (see Table 1). 5

5We train an entity tagger and relation classifier to predict G and also for computing knowledge based evaluation scores.

Flaxseed extract may improve heart health by lowering cholesterol and triglycerides.

(i) Sentences from multiple Scientific Abstracts

(ii) Knowledge span extraction (iii) Selection for summarization P(C|Z) X P(Z|X)

(iv) Text fusion and generation P(y|C)

Figure 3: Illustrating the flow of our extract-compose system. (i) We begin with multiple scientific abstracts to summarize from; (ii) We extract knowledge spans as possible candidates for generation; (iii) We select key spans ( improve heart health by lowering cholestrol and triglycerides in this case) from all possible candidates and (iv) Present a summary sentence using the spans and a domain specific language model.

Similarly, we denote the summary data as {(yq, Gq)M q=1}, where yq is a concise summary and Gq is the set of entityentity relation tuples, in M summaries. Joint learning of content selection, information aggregation and text generation for multi-document summarization can be challenging. This is exacerbated in our technical domain with few parallel examples. We propose to overcome these challenges by first learning to extract and then composing the informative text pieces. We model two intermediate variables, distilled text spans Z associated with all entity-entity relation information G, and the key content C selected from them. The probability of an output summary y, conditioned on the input X is

C,Z P(y|C)P(C|Z)P(Z|X) (1)

Where: (i) P(Z|X) models a knowledge extraction process gathering all entity-entity relations and corresponding text spans, G; (ii) P(C|Z) models the process of selecting an important subset from all extracted text spans and (iii) P(y|C) models the text fusion process that composes fluent text incorporating the extracted content.

3.2 Span Extraction P(Z|X)

We model P(Z|X) as a span extraction task leveraging a rationale extraction system (Lei, Barzilay, and Jaakkola 2016). The extraction model picks the fewest words from the input necessary to make the correct relation label prediction for an entity pair. Let (ei, ej, r) be one entity relationship tuple, x be the associated sentence text and z be the considered rationale text span. P(Z|X) is trained to minimize the loss:

L = L(z, r) + L(z) + L(z, ei, ej) (2)

. Where L(z, r) is the cross entropy loss for predicting r with z as the extracted text. L(z) is a regularization term to select short and coherent text, by minimizing the span lengths and discontinuities among the spans. In addition to prior work, we introduced L(z, ei, ej) to encourage the selection of

phrases that contain entities ei and ej. Specifically, we construct verb phrases (from constituency parse trees) containing the condition entity, and minimize the distance between the selected text span z and the verb phrase. Empirically, this loss stabilizes the span extraction and improves the quality of selected text spans, using r labels as indirect supervision as in Shah, Schuster, and Barzilay (2019). By running the extraction model on every (ei, ej, r) tuple, we distill the input text into a set of text spans Z = {z1, z2, . . . , zm}.

3.3 Content Selection Policy Network P(C|Z)

P(C|Z) is a policy network, that takes a large set of text spans Z as input, and outputs C, an ordered set of key text spans. We model our content selection as a finite Markov decision process (MDP). Where the state is represented as st = (t, {c1, . . . , ct}, {z1, z2, ..., zm t}) for step t, content selected so far {c1, . . . , ct} and remaining text spans {z1, z2, ..., zm t}. Our action space is all the remaining text spans plus one special token, Z {STOP}. The number of actions is |m t| + 1. We parameterize the policy πθ(a|st) with a neural network to map the state s to a probability distribution over all available actions. At each step, the probability that the policy selects zi as a candidate is:

πθ(a=zi|st) = exp(f(t, ˆzi, ˆ ci )) Pm t+1 j=1 exp(f(t, ˆzj, ˆ cj )) (3)

where ci = arg maxcj(cos(ˆzi, ˆcj)) is the selected content closest to zi, ˆzi and ˆ ci are the encoded dense vectors, cos(u, v) = u v ||u|| ||v|| and f is a feed-forward neural network that outputs a scalar score. The selection process begins from Z. Our module iteratively samples actions from πθ(a|st). On picking STOP, we end with the selected content C and a corresponding reward. Our rewards guide the selection of informative, diverse and readable phrases:

Re = P c C cos( ˆ eic, ˆ eiy) + cos( ˆ ejc, ˆ ejy) is the cosine similarity of the structures of the selected content C with the structures present in the gold summary y (each gold summary structure accounted with only one c), encouraging the model to select relevant content.

Rd = 1[maxi,j(cos( ˆcj, ˆci)) < δ] computes the similarity between pairs within selected content C, encouraging the selection of diverse spans.

Rs = P c 1[ M P

q=1 cos(ˆc, ˆyq) > M P

q=1 cos(ˆz , ˆyq)], where z

is a text span randomly sampled from scientific abstract corpus, predicts if c is closer to a human summary than a random abstract sentence, encouraging the selection of concise spans and ignoring numerical details. Rm is the Meteor (Denkowski and Lavie 2014) score between the final selection with the gold summary, a signal which can only be provided at the completion of the episode. rp is a small penalty for each action step.

The final multi-objective reward is computed as

R = we Re+wd Rd + ws Rs + wm Rm |C|rp, (4)

where, we, wd, ws, wm and rp are hyper-parameters. During training, the network is updated with these rewards. Our paradigm allows an exploration of different span combinations while incorporating delayed feedback.

3.4 Text Fusion P(y|C) The text fusion module, P(y|C), composes a complete summary using the text spans C selected by the policy network. We propose to utilize the recently developed Blank Language Model (BLM) (Shen et al. 2020), which fills in the blanks by iteratively determining the word to place in a blank or adding a new blank, until all blanks are filled. The model is trained on the Wiki Text-103 dataset (Merity et al. 2016). We extend this model with additional categorical blanks between different text spans in C, according to their relation type. This ensures control over facts and relations captured from the scientific abstracts X to present a semantically valid and fluent summary (details see Appendix). In the Transformer variant of the model, we train a seq2seq P(y|C) using the limited parallel data.

4 Data In this section, we describe the dataset collected for our Nutri-bullet system.

4.1 Corpus Collection Our Healthline6 and Breast Cancer7 datasets consist of scientific abstracts as inputs and human written summaries as outputs. Scientific Abstracts We collect 7750 scientific abstracts from Pubmed and the Science Direct, each averaging 327 words. The studies in these abstracts are cited by domain experts when writing summaries in the Healthline and Breast Cancer datasets. A particular food and its associated abstracts are fed as inputs to our Nutri-bullet systems. We exploit the large scientific abstract corpus when gathering

6https://www.healthline.com/nutrition 7https://foodforbreastcancer.com/

entity, relation and sentiment annotations (see Table 3) to overcome the challenge of limited parallel examples. Modules trained on these annotations can be applied to any food health scientific abstract. Summaries Domain experts curate summaries for a general audience in the Healthline and Breast Cancer datasets. These summaries describe nutrition and health benefits of a specific food (examples shown in Appendix). In the Health Line dataset, each food has multiple bullet summaries, where each bullet typically talks about a different health impact (hydration, anti-diabetic etc). In Breast Cancer, each food has a single summary, describing in great detail its impact on breast and other cancers. Parallel Instances The references in the human written summaries form natural pairings with the scientific abstracts. We harness this to collect 1894 parallel (abstracts, summary) instances in Health Line, and 141 parallel instances in Breast Cancer (see Table 2). Summaries in Health Line average 24.46 words, created using an average of 3 articles. Summaries in Breast Cancer have an average length of 155.71 words referencing an average of 18 articles. Unsurprisingly, Breast Cancer is the more challenging of the two.

4.2 Entity, Relation and Sentiment Annotations

Despite having a small parallel data compared to Hermann et al. (2015); Narayan, Cohen, and Lapata (2018), we conduct large-scale crowd-sourcing tasks to collect entity, relation and sentiment annotations on Amazon Mechanical Turk. The annotations (see Table 3) are designed to capture the rich technical information ingrained in such domains, alleviating the difficulty of multi-document summarization and are broadly applicable to different systems (Koncel Kedziorski et al. 2019b). Entity and Relation Annotations Workers identify food, nutrition, condition and population entities by highlighting the corresponding text spans. Given the annotated entities in text, workers are asked to enumerate all the valid relation tuples (ei, ej, r). Table 1 lists possible combinations of ei, ej and r for each relation type, along with some examples. The technical information present in our domain can make annotating challenging. To collect reliable annotations, we set up several rounds of qualification tasks 8, offer direct communication channels to answer annotators questions and take majority vote among 3 annotators for each data point. As shown in Table 3, we collected 91K entities, 34K pairs of relations, and 7K sentiments9. The high value of mean Cohen s κ highlights high annotator agreement for all the tasks, despite various challenges. Food Health Sentiment Additionally, we also collect food sentiment annotations for evaluating our system. Food health sentiment (positive, negative, neutral) indicates the

8To set up the qualification, the authors first annotate tens of examples which serve as gold answers. We leverage Mturk APIs to grade the annotation by comparing with the gold answers. 9For Breast Cancer, we use the entity tagger and relation classifier finetuned on scientific abstracts and Health Line datasets to extract the entities and relations.

Relation Type ei ej r Example

Containing Food, Nutrition Nutrition Contain (apple, fiber, contain)

Causing Food, Nutrition, Condition Condition Increase, Decrease, Satisfy, Control

(bananas, metabolism, increase), (orange juice, hydration, satisfy)

Table 1: Details of entity-entity relationships that we study and some tuple examples.

Data Train Dev Test Input Scientific Abstracts 6110 750 866 Average words 327.7 323.3 332.3 Health Line summaries 1522 179 193 Average words 24.7 23.3 23.9 Abstracts Per Instance 3.19 2.82 3.43 Breast Cancer summaries 104 18 19 Average words 161.4 131.6 148.1 Abstracts Per Instance 18.90 17.21 17.67

Table 2: Statistics for scientific abstracts, Health Line and Breast Cancer datasets.

Data Task # annotations mean κ

Scientific entity 83543 0.75 Abstracts relation 28088 0.79, 0.81 sentiment 5000 0.65

Health Line entity 7860 0.86 relation 5974 0.73, 0.90 sentiment 2000 0.89

Table 3: Entity, relation and sentiment annotation statistics. Each annotation is from three annotators. Mean κ is the mean pairwise Cohen s κ score.

kind of impact the food has on human health. The annotation is performed at a sentence level, and modules trained on this data are used to assess the contrastiveness in a food s summary bullets. Annotation interfaces, instructions as well as more data details can be found in Appendix.

5 Experimental Setup Datasets We randomly split both Health Line and Breast Cancer datasets into training, development and testing sets(see Table 2). Evaluation The subjective nature of summarization demands human judgements to comprehensively evaluate model performance. We have human annotators score our models on faithfulness, relevance, fluency and informativeness. Given input scientific abstracts, faithfulness characterizes if the output text is consistent with the input. Given a reference summary, relevance indicates if the output text shares similar information. Fluency represents if the output text is grammatically correct and written in well-formed English. Informativeness characterizes the degree of health related knowledge conveyed by the summaries.10

10On the Health Line datasets, each food article contains multiple bullet summaries. We group these bullet summaries per food for annotator comparison. For Breast Cancer, a single summary output

Annotators rate faithfulness, relevance and fluency on a 1-4 scale likert score, which is commonly used as a standard psychometric scale to measure responses in social science and attitude research projects (Croasmun and Ostrom 2011; Li 2013; Sullivan and Artino Jr 2013). For rating informativeness, we perform a one-vs-one evaluation between our model and the strong Transformer baseline (Hoang et al. 2019). We have 3 annotators score every data point and take an average across the scores. We further evaluate our systems using the following automatic metrics. Meteor is an automatic metric used to compare the model output with the gold reference. KG(G) computes the number of entity-entity pairs with a relation in the gold reference, that are present in the output.11 This captures relevance in context of the reference. KG(I), similarly, computes the number of entity-entity pairs in the output that are present in the input scientific abstracts. This measures faithfulness with respect to the input documents. Diversity calculates the proportion of unique trigrams in the outputs (Li et al. 2016; Rao and Daum e III 2019). Contrastiveness calculates the proportion of article summaries belonging to the same food that contain both positive and negative/neutral sentiment about a food s health impact. Baselines We compare our method against several stateof-the-art methods.

Copy-gen: See, Liu, and Manning (2017) is a top performing technique for summarization, which can copy words from the input or generate new words. Transformer: Hoang et al. (2019) is a system that utilizes a pretrained Transformer for summarization. Graph Writer: Koncel-Kedziorski et al. (2019b) is a graph transformer based model, which generates text using a seed title and a knowledge graph. This model utilizes all the extraction capabilities used by our P(Z|X) module.

In addition, we study variants of our model (1) replacing BLM with a transformer text fusing model Ours (Transformer), (2) replacing the policy gradient selector with a fully supervised selector trained on knowledge structure information (Select w/ Sup.), and (3) different reward combinations. Further, we compare our system with a BERT-based Extractive Summarization system. Implementation Details We adapt the implementation of span extraction from Bao et al. (2018). Our policy network

is a food s complete summary. 11We run entity tagging plus relation classification on top of the model output and gold summaries. We match the gold (eg i , eg j, rg) tuples using word embedding based cosine similarity with the corresponding entities in the output structures (eo i , eo j, ro). If the cosine score exceeds a threshold of 0.7, a match is found.

Breast Cancer Health Line

Faithfulness Relevance Fluency Faithfulness Relevance Fluency

Ours (Transformer)

Transformer

Graph Writer

Figure 4: Human evaluation Gantt charts showing ratings on faithfulness, relevance and fluency for both Breast Cancer and Health Line datasets, where Ours (BLM) extract-compose model strongly outperforms all others. Each example is rated by three judges and the mean is reported.

Breast Cancer Health Line

Better 94 % 75 % Worse 6 % 25 %

Table 4: Human evaluation on informativeness of Ours (BLM) when comparing to Transformer.

is a three layer feedforward neural network. We set we, wd and ws to 1 and wm to 0.75. rp is 0.02 and δ is 0.99. We use a large, 6 layer BLM (Shen et al. 2020) for fusion. Additional experimentation details can be found in the Appendix . We train a Neural CRF tagger (Yang and Zhang 2018) for the food, condition and nutrition entity tagging. We use BERT (Devlin et al. 2018) text classifiers to predict the relation between two entities, and to predict food sentiments of a summary, trained using annotations from Table 3.

6 Results In this section, we describe the performance of Nutri-bullet systems (baselines and models) on the Health Line and Breast Cancer datasets. We also report ablation studies, highlighting the importance of our policy network s ability to explore different span combinations, while incorporating long term rewards from distant supervision. Human Evaluation Figure 4 and Table 4 report the human evaluation results of our model and baselines on the test set of both datasets. Our method outperforms all other models by a large margin on faithfulness, owing to the extraction and controlled text generation process. Our method is also rated the highest on relevance (3.11 & 3.19), demonstrating its ability to select the important and relevant content consistently, even with little supervision. Our model produces fluent text. On the contrary, despite being fluent, transformer models fail to generate faithful content. They achieve decent averaged relevance scores ( 2.9) on the Health Line dataset. However, annotators rate them a score 1 far more often,12 due to hallucinated content and factually incorrect outputs. Copy-gen and Graph Writer struggles to score high on either of our metrics, showing the challenges of learning meaningful summaries in such a low-resource and

12Annotators rate transformer with score 1 6 times and ours(transformer) 3 times more often than BLM model.

knowledge-rich domain. Finally, we evaluate if our output text can benefit the reader and provide useful nutritional information about the food being described. In our human evaluation of informativeness, we outperform 94%-6% and 75%-25% against the Transformer baseline on Breast Cancer and Health Line respectively. Automatic Evaluation Table 6 presents the automatic evaluation results on Breast Cancer and Health Line datasets. High KG(I) (52% & 40%) and KG(G) (96% & 88%) scores for our method highlight that our produced text is faithful and relevant, consistent with human evaluation. Additionally, high diversity (91% & 90%) and contrastiveness (94% & 83%) scores indicate that our model is also able to present distinct (across sentiment and content) information for the same food. In contrast, the sequence-to-sequence based methods tend to get a higher Meteor score with a lower diversity, suggesting that they repeatedly generate similar outputs, regardless of the input text. Low KG scores show that they fail to capture the relevant or faithful text, which is crucial in this domain. Among these, our transformer variation performs strong, especially on the Breast Cancer dataset. This was also observed in Liu and Lapata (2019), where abstractive summarization fed with only a subset of the input sentences outperformed vanilla abstractive summarization. Case Study Table 5 shows examples of various system outputs on Health Line. The food name and titles of corresponding scientific abstracts are also presented. We observe that Transformer, Graph Writer and Copy-gen fail to generate meaningful summaries that are faithful to the original input text. With carefully selected knowledge, our (transformer) and our (BLM) both produce relevant and useful information. Ours (transformer) generates concise and easier language. Ours (BLM) composes faithful text with rich details, for example, it captures anti-aging properties , which is relevant to free radicals in the gold summary. Detailed analysis shows that 91% of all produced tokens by our model are from the extracted segments critical for faithful summarization. The words generated in between allow cohesive fusion. Examples from the Breast Cancer dataset can be found in the Appendix. Additional model outputs with human evaluation scores can be found in Appendix. P(C|Z) Variants To further understand our model, we implement an alternative content selection method, using

Food banana Pubmed Titles (1) High content of dopamine a strong antioxidant, in Cavendish banana; (2) Flavonoid intake and risk of CVD: a systematic review and meta-analysis of prospective cohort studies; (3) Antioxidants in health, disease and aging; (4) Banana(Musa spp) from peel to pulp: ethnopharmacology, source of bioactive compounds and its relevance for human health.

Gold Bananas are high in several antioxidants, which may help reduce damage from free radicals and lower your risk of some diseases. Copy-gen Seaweed contain several compounds that can reduce feelings of fullness and prevent fullness hormones, that are needed to health health benefits. Graph Writer The type of insoluble fiber found in banana banana may help reduce blood sugar levels and help lower blood pressure. Transformer The antioxidants in walnuts may help lower blood pressure in people with type 2 diabetes. Ours (Trans.) Bananas are a rich source of antioxidants, which may help reduce the risk of many chronic diseases. Ours (BLM) Bananas are used to help overcome or prevent a substantial number of illnesses , as depression and bananas containing antioxidants may lower the incidence of disease , such as certain cancers , cardiovascular and neurodegenerative diseases , DNA damage , or even have anti-aging properties .

Table 5: Example outputs for a Health Line input.

Breast Cancer Health Line

Model Me KG(G) KG(I) Di Co Me KG(G) KG(I) Di Co

Copy-gen 14.0 21 23 70 0 7.4 21 51 82 43 Graph Writer 10.1 0 0 16 0 7.6 3 69 31 25 Transformer 13.0 31 11 76 66 10.2 21 67 53 28

Ours (Tranformer) 15.0 49 13 81 50 10.3 23 64 55 28 Ours (BLM) 13.8 52 96 91 94 8.7 40 88 90 83

Table 6: Meteor score (Me), KG in gold(G), KG in input(I), Diversity (Di) and Contrastiveness (Co) in our models and various baselines, on both Breast Cancer and Health Line datasets. denotes cases which model generates meaningless results due to small training size. The best results are in bold and Ours(BLM) extract-compose is the most dominant.

a supervised classification module (implementation details described in Appendix . Table 7 reports the results. Being an extract-compose variant, the supervised model (first row) produces faithful summaries (KG(I)). However, our Policy Network s joint selection and ability to explore span combinations with guidance from gold structure rewards and the Meteor score, lead to an improved performance on KG(G), Diversity and Meteor. Additionally, on human evaluated relevance, the Policy Network approach scores a higher 3.19 while the supervised extraction variant scores 2.5. We also observe the importance of Rm for KG(I) and Meteor, and Rd for Diversity.

Model Me KG(G) KG(I) Di Co

Select w/ Sup. 7.6 28 84 75 88 Policy (w/o Rm, Rd) 8.4 36 83 70 88 Policy (w/o Rm) 8.2 33 89 91 88 Policy (full R) 8.7 40 88 90 83

Table 7: Automatic evaluation of extract-compose model s variants on Health Line. The best results are in bold.

Comparison with Extractive Summarization To understand the benefits of fine-span extraction followed by fusion, we compare our model with an extractive summarization system (implementation details in Appendix . Our model achieves 40% KG(G), while the extractive summarization

system achieves 26%. Our extract-compose approach performs strongly since: (1) We extract knowledge pieces more precisely by selecting key spans instead of complex complete sentences from scientific abstracts; (2) RL model s content selection jointly through multi-steps, and (3) The Text Fusion module consolidates knowledge pieces, which may otherwise remain incomplete due to linguistic phenomena, such as coreference and discourse. Even unsupervised methods like Text Rank (Mihalcea and Tarau 2004) are not particularly applicable when we need to select key spans amongst multiple candidates. Instead our model is able to capture relevant, readable and coherent pieces of text by utilizing guidance from distant supervision and the use of a domain specific language model for fusion.

7 Conclusion High impact datasets, content selection, faithful decoding and evaluation are open challenges in building multidocument health summarization systems. First, we propose two new datasets for Nutri-bullets. Next, we tackle this problem by exploiting annotations on the source side and formulating an extraction and composition method. Comprehensive human evaluation demonstrates the efficacy of our method in producing faithful, informative and relevant summaries.

References Amplayo, R. K.; and Lapata, M. 2019. Informative and controllable opinion summarization. ar Xiv preprint ar Xiv:1909.02322 .

Bao, Y.; Chang, S.; Yu, M.; and Barzilay, R. 2018. Deriving machine attention from human rationales. ar Xiv preprint ar Xiv:1808.09367 .

Barzilay, R.; Mc Keown, K.; and Elhadad, M. 1999. Information fusion in the context of multi-document summarization. In ACL, 550 557.

Barzilay, R.; and Mc Keown, K. R. 2005. Sentence fusion for multidocument news summarization. Computational Linguistics 31(3): 297 328.

Baumel, T.; Eyal, M.; and Elhadad, M. 2018. Query focused abstractive summarization: Incorporating query relevance, multi-document coverage, and summary length constraints into seq2seq models. ar Xiv preprint ar Xiv:1801.07704 .

Carbonell, J.; and Goldstein, J. 1998. The use of MMR, diversity-based reranking for reordering documents and producing summaries. In Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, 335 336.

Chen, Y.-C.; and Bansal, M. 2018. Fast Abstractive Summarization with Reinforce-Selected Sentence Rewriting. In ACL, 675 686. Melbourne, Australia: ACL. doi:10.18653/ v1/P18-1063.

Cheng, J.; and Lapata, M. 2016. Neural Summarization by Extracting Sentences and Words. In ACL, 484 494. Berlin, Germany: ACL. doi:10.18653/v1/P16-1046.

Croasmun, J. T.; and Ostrom, L. 2011. Using Likert-Type Scales in the Social Sciences. Journal of Adult Education 40(1): 19 22.

Denkowski, M.; and Lavie, A. 2014. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, 376 380.

Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. ar Xiv preprint ar Xiv:1810.04805 .

Fabbri, A.; Li, I.; She, T.; Li, S.; and Radev, D. 2019. Multi-News: A Large-Scale Multi-Document Summarization Dataset and Abstractive Hierarchical Model. 1074 1084. Florence, Italy: ACL. doi:10.18653/v1/P19-1102.

Ganesan, K.; Zhai, C.; and Han, J. 2010. Opinosis: A Graph Based Approach to Abstractive Summarization of Highly Redundant Opinions. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), 340 348. Beijing, China: Coling 2010 Organizing Committee. URL https://www.aclweb.org/anthology/C101039.

Geva, M.; Malmi, E.; Szpektor, I.; and Berant, J. 2019. Disco Fuse: A Large-Scale Dataset for Discourse-Based Sentence Fusion. NAACL HLT 3443 3455. doi:10.18653/v1/ N19-1348.

Haghighi, A.; and Vanderwende, L. 2009. Exploring content models for multi-document summarization. In NAACL HLT, 362 370.

Hermann, K. M.; Kocisky, T.; Grefenstette, E.; Espeholt, L.; Kay, W.; Suleyman, M.; and Blunsom, P. 2015. Teaching machines to read and comprehend. In Advances in neural information processing systems, 1693 1701.

Hoang, A.; Bosselut, A.; Celikyilmaz, A.; and Choi, Y. 2019. Efficient adaptation of pretrained transformers for abstractive summarization. ar Xiv preprint ar Xiv:1906.00138 .

Keneshloo, Y.; Shi, T.; Ramakrishnan, N.; and Reddy, C. K. 2019. Deep reinforcement learning for sequence-tosequence models. IEEE Transactions on Neural Networks and Learning Systems .

Koncel-Kedziorski, R.; Bekal, D.; Luan, Y.; Lapata, M.; and Hajishirzi, H. 2019a. Text generation from knowledge graphs with graph transformers. ar Xiv preprint ar Xiv:1904.02342 .

Koncel-Kedziorski, R.; Bekal, D.; Luan, Y.; Lapata, M.; and Hajishirzi, H. 2019b. Text Generation from Knowledge Graphs with Graph Transformers. In NAACL HLT, 2284 2293. Minneapolis, Minnesota: ACL. doi:10.18653/ v1/N19-1238.

Kry sci nski, W.; Mc Cann, B.; Xiong, C.; and Socher, R. 2019. Evaluating the factual consistency of abstractive text summarization. ar Xiv ar Xiv 1910.

Lebanoff, L.; Song, K.; and Liu, F. 2018. Adapting the neural encoder-decoder framework from single to multidocument summarization. ar Xiv preprint ar Xiv:1808.06218 .

Lei, T.; Barzilay, R.; and Jaakkola, T. 2016. Rationalizing Neural Predictions. In EMNLP, 107 117. Austin, Texas: ACL. doi:10.18653/v1/D16-1011.

Lewis, M.; Liu, Y.; Goyal, N.; Ghazvininejad, M.; Mohamed, A.; Levy, O.; Stoyanov, V.; and Zettlemoyer, L. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. ar Xiv preprint ar Xiv:1910.13461 .

Li, J.; Monroe, W.; Ritter, A.; Galley, M.; Gao, J.; and Jurafsky, D. 2016. Deep reinforcement learning for dialogue generation. ar Xiv preprint ar Xiv:1606.01541 .

Li, Q. 2013. A novel Likert scale based on fuzzy sets theory. Expert Systems with Applications 40(5): 1609 1618.

Lin, C.-Y. 2004. Rouge: A package for automatic evaluation of summaries. Text Summarization Branches Out .

Liu, F.; Flanigan, J.; Thomson, S.; Sadeh, N.; and Smith, N. A. 2015. Toward Abstractive Summarization Using Semantic Representations. In NAACL HLT, 1077 1086. Denver, Colorado: ACL. doi:10.3115/v1/N15-1114.

Liu, P. J.; Saleh, M.; Pot, E.; Goodrich, B.; Sepassi, R.; Kaiser, L.; and Shazeer, N. 2018. Generating wikipedia by summarizing long sequences. ar Xiv preprint ar Xiv:1801.10198 .

Liu, Y.; and Lapata, M. 2019. Text summarization with pretrained encoders. ar Xiv preprint ar Xiv:1908.08345 .

Mc Keown, K.; and Radev, D. R. 1995. Generating summaries of multiple news articles. In Proceedings of the 18th annual international ACM SIGIR conference on Research and development in information retrieval, 74 82.

Merity, S.; Xiong, C.; Bradbury, J.; and Socher, R. 2016. Pointer sentinel mixture models. ar Xiv preprint ar Xiv:1609.07843 .

Mihalcea, R.; and Tarau, P. 2004. Textrank: Bringing order into text. In EMNLP, 404 411.

Narayan, S.; Cohen, S. B.; and Lapata, M. 2018. Don t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. ar Xiv preprint ar Xiv:1808.08745 .

Narayan, S.; Gardent, C.; Cohen, S. B.; and Shimorina, A. 2017. Split and Rephrase. EMNLP 606 616. doi:10.18653/ v1/D17-1064.

Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002. BLEU: a method for automatic evaluation of machine translation. In ACL, 311 318. ACL.

Pasunuru, R.; and Bansal, M. 2018. Multi-Reward Reinforced Summarization with Saliency and Entailment. In NAACL HLT, 646 653. New Orleans, Louisiana: ACL. doi: 10.18653/v1/N18-2102.

Paulus, R.; Xiong, C.; and Socher, R. 2017. A deep reinforced model for abstractive summarization. ar Xiv preprint ar Xiv:1705.04304 .

Puduppully, R.; Dong, L.; and Lapata, M. 2019. Datato-text generation with entity modeling. ar Xiv preprint ar Xiv:1906.03221 .

Qazvinian, V.; and Radev, D. R. 2008. Scientific paper summarization using citation summary networks. ar Xiv preprint ar Xiv:0807.1560 .

Qazvinian, V.; Radev, D. R.; Mohammad, S. M.; Dorr, B.; Zajic, D.; Whidby, M.; and Moon, T. 2013. Generating extractive summaries of scientific paradigms. Journal of Artificial Intelligence Research 46: 165 201.

Radev, D. R.; and Mc Keown, K. R. 1998. Generating Natural Language Summaries from Multiple On-Line Sources. Computational Linguistics 24(3): 469 500.

Rao, S.; and Daum e III, H. 2019. Answer-based adversarial training for generating clarification questions. ar Xiv preprint ar Xiv:1904.02281 .

Rush, A. M.; Chopra, S.; and Weston, J. 2015. A Neural Attention Model for Abstractive Sentence Summarization. In EMNLP, 379 389. Lisbon, Portugal: ACL. doi:10.18653/ v1/D15-1044.

See, A.; Liu, P. J.; and Manning, C. D. 2017. Get To The Point: Summarization with Pointer-Generator Networks. In ACL, 1073 1083. Vancouver, Canada: ACL. doi:10.18653/ v1/P17-1099.

Shah, D. J.; Schuster, T.; and Barzilay, R. 2019. Automatic Fact-guided Sentence Modification. ar Xiv preprint ar Xiv:1909.13838 . Sharma, E.; Huang, L.; Hu, Z.; and Wang, L. 2019. An Entity-Driven Framework for Abstractive Summarization. In EMNLP-IJCNLP, 3280 3291. Hong Kong, China: ACL. doi:10.18653/v1/D19-1323. Shen, T.; Quach, V.; Barzilay, R.; and Jaakkola, T. 2020. Blank Language Models. ar Xiv preprint ar Xiv:2002.03079 . Stern, M.; Chan, W.; Kiros, J.; and Uszkoreit, J. 2019. Insertion transformer: Flexible sequence generation via insertion operations. ar Xiv preprint ar Xiv:1902.03249 . Sullivan, G. M.; and Artino Jr, A. R. 2013. Analyzing and interpreting data from Likert-type scales. Journal of graduate medical education 5(4): 541 542. Sutton, R. S.; and Barto, A. G. 2018. Reinforcement learning: An introduction. MIT press. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, Ł.; and Polosukhin, I. 2017. Attention is all you need. In Advances in neural information processing systems, 5998 6008. Wenbo, W.; Yang, G.; Heyan, H.; and Yuxiang, Z. 2019. Concept Pointer Network for Abstractive Summarization. ar Xiv preprint ar Xiv:1910.08486 . Wu, L.; Tian, F.; Qin, T.; Lai, J.; and Liu, T.-Y. 2018. A study of reinforcement learning for neural machine translation. ar Xiv preprint ar Xiv:1808.08866 . Xu, P.; Wu, C.-S.; Madotto, A.; and Fung, P. 2019. Clickbait? Sensational Headline Generation with Auto-tuned Reinforcement Learning. ar Xiv preprint ar Xiv:1909.03582 . Yang, J.; and Zhang, Y. 2018. Ncrf++: An opensource neural sequence labeling toolkit. ar Xiv preprint ar Xiv:1806.05626 . Zhang, J.; Tan, J.; and Wan, X. 2018. Adapting Neural Single-Document Summarization Model for Abstractive Multi-Document Summarization: A Pilot Study. In Proceedings of the 11th International Conference on Natural Language Generation, 381 390. Tilburg University, The Netherlands: ACL. doi:10.18653/v1/W18-6545. Zhang, S.; and Bansal, M. 2019. Addressing Semantic Drift in Question Generation for Semi-Supervised Question Answering. ar Xiv preprint ar Xiv:1909.06356 .