Event Graph based Sentence Fusion

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4075–4084 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4075 Event Graph based Sentence Fusion Ruifeng Yuan ∗ The Hong Kong Polytechnic University csryuan@comp.polyu.edu.hk Zili Wang ∗ Xidian University ziliwang.do@gmail.com Wenjie Li The Hong Kong Polytechnic University cswjli@comp.polyu.edu.hk Abstract Sentence fusion is a conditional generation task that merges several related sentences into a coherent one, which can be deemed as a sum- mary sentence. The importance of sentence fu- sion has long been recognized by communities in natural language generation, especially in text summarization. It remains challenging for a state-of-the-art neural abstractive summariza- tion model to generate a well-integrated sum- mary sentence. In this paper, we explore the ef- fective sentence fusion method in the context of text summarization. We propose to build an event graph from the input sentences to ef- fectively capture and organize related events in a structured way and use the constructed event graph to guide sentence fusion. In ad- dition to make use of the attention over the content of sentences and graph nodes, we fur- ther develop a graph flow attention mechanism to control the fusion process via the graph structure. When evaluated on sentence fusion data built from two summarization datasets, CNN/DaliyMail and Multi-News, our model shows to achieve state-of-the-art performance in terms of Rouge and other metrics like fusion rate and faithfulness. 1 Introduction Sentence fusion aims to combine several related sentences into a single coherent text. It is impor- tant in many NLP tasks such as text summariza- tion, question answering and retrieval-based dia- logue. In text summarization, it is a common prac- tice for a proficient editor to fuse the information from several related sentences, however, it remains challenging for a state-of-the-art neural abstrac- tive summarization model to achieve effective sen- tence fusion. As pointed out in (Lebanoff et al., 2019a), the human-written summaries contain 32% fusion sentences on the CNN/DailyMail dataset, 1These authors contributed equally to this work. Fused Sentence: Johnny Kemp is “believed to have drowned at a beach in Montego Bay,” police say. Source Sentences: (A) Bahamian R&B singer Johnny Kemp, best known for the 1988 party anthem “Just Got Paid,” died this week in Jamaica. (B) The singer is believed to have drowned at a beach in Montego Bay on Thursday, the Jamaica Constabulatory Force said in a press release. Disparate Sentence Fusion Similar Sentence Fusion Source Sentences: (A) Meng Wanzhou, Huawei’s chief fifinancial offificer and deputy chair, was arrested in Vancouver. (B) Canadian officials have arrested Meng Wanzhou on Dec. 1 Fused Sentence: Meng was arrested in Vancouver on Dec. 1 by Canadian officials. Figure 1: Examples of two types of sentence fusion in text summarization. while only 6% of the summary sentences gener- ated by the Pointer-Generator model (See et al., 2017) are shown to fuse the information spread over sentences. Besides, without proper guidance, many sentences generated by fusion contain fac- tual errors. Therefore, it is worthwhile to explore effective …

Content truncated to 500 words