Abstractive Multi-document Summarization by Partial Tree Extraction, Recombination and Linearization

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 812–821, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Abstractive Multi-document Summarization by Partial Tree Extraction, Recombination and Linearization Litton J Kurisinkel IIIT-H, Hyderabad litton.jKurisinkel@research.iiit.ac.in Yue Zhang SUTD, Singapore yue zhang@sutd.edu.sg Vasudeva Varma IIIT-H, Hyderabad vv@iiit.ac.in Abstract Existing work for abstractive multi- document summarization utilise existing phrase structures directly extracted from input documents to generate summary sentences. These methods can suf- fer from lack of consistence and coher- ence in merging phrases. We introduce a novel approach for abstractive multi- document summarization through partial dependency tree extraction, recombination and linearization. The method entrusts the summarizer to generate its own topi- cally coherent sequential structures from scratch for effective communication. Re- sults on TAC 2011, DUC- 2004 and 2005 show that our system gives competitive results compared with state-of-the-art ab- stractive summarization approaches in the literature. We also achieve competitive re- sults in linguistic quality assessed by hu- man evaluators. 1 Introduction Multi-document summarization generates a tex- tual summary from a corpus of documents deal- ing with a set of related topics. An optimum generated summary should encompass the most relevant and topically diverse content, which can represent the input corpus in stipulated summary space. Extractive multi-document summarization approaches pick out a subset of sentences to con- stitute the summary, which can be noisy and inco- herent, as all the portions of a sentence may not be relevant for summary generation (Lin and Bilmes, 2011). An abstractive multi-document summarizer, in contrast, infers the most relevant information and generates summary sentences exhibiting coher- ence and fluency. There has been relatively little existing work on abstractive multi-document sum- marization. Bing et al. (2015) merge phrases that are extracted from input documents into a coher- ent summary. Banerjee et al. (2015) utilize multi- sentence compression for summarization. Both of the above approaches rely on sequential arrange- ment of phrasal or subsentential structures existing in the input corpus to generate summary. How- ever, an ideal abstractive multi-document summa- rizer should enjoy the freedom to exhibit its own writing style and to generate the sentences from scratch. We build a model to this end by leveraging syn- tactic dependencies. Input for our model is the set of syntactic dependency trees obtained by parsing sentences in the corpus to be summarized. Rel- evant and noise pruned partial tree structures are extracted from the set of dependency trees and dif- ferent subsets of maximally relevant partial depen- dency structures are identified. Partial trees in dif- ferent subsets are linearized to generate individ- ual summary sentences. In this work, we utilize transition-based syntactic linearization approach proposed by Puduppully et al. (2016) to linearize a combination of partial trees and to generate a noise free summary sentence. The combinability of a set of partial trees to form a full dependency tree of a valid sentence is estimated using a generative model of syntactic dependency trees (Zhang et al., 2016). As a result, the model is …

Content truncated to 500 words