Supervised Sentence Fusion with Single-Stage Inference

International Joint Conference on Natural Language Processing, pages 1410–1418, Nagoya, Japan, 14-18 October 2013. Supervised Sentence Fusion with Single-Stage Inference Kapil Thadani and Kathleen McKeown Department of Computer Science Columbia University New York, NY 10025, USA {kapil,kathy}@cs.columbia.edu Abstract Sentence fusion—the merging of sen- tences containing similar information— has been shown to be useful in an abstrac- tive summarization context. We present a new dataset of sentence fusion in- stances obtained from evaluation datasets in summarization shared tasks and use this dataset to explore supervised approaches to sentence fusion. Our proposed infer- ence approach recovers the highest scor- ing output fusion under an n-gram fac- torization using a compact integer linear programming formulation that avoids cy- cles and disconnected structures. In addi- tion, we introduce simple fusion-specific features and constraints that outperform a compression-inspired baseline as well as a variant that relies on human-identified concept spans for perfect content selec- tion. 1 Introduction Abstractive text summarization has long been a high-level goal of natural language processing. Although progress in text-to-text (T2T) genera- tion tasks such as sentence compression and para- phrase generation has been steady, the fusion of multiple sentences offers a particularly formidable challenge. Sentence fusion refers to the task of combining two or more sentences which overlap in information content, avoiding extraneous de- tails and preserving common information. This procedure has been observed in human summa- rization (Jing and McKeown, 2000) and has been shown to be a valuable component of automated summarization systems (Barzilay and McKeown, 2005). However, research in sentence fusion has long been hampered by the absence of datasets for the task, and the difficulty of generating one has cast doubt on the viability of automated fu- sion (Daum´e III and Marcu, 2004). This paper presents a new fusion dataset gener- ated from existing human annotations and also in- troduces a discriminative T2T system that general- izes the single sentence compression approach of Thadani and McKeown (2013) to n-way sentence fusion. Our fusion dataset is constructed from evaluation data for summarization shared tasks in the Document Understanding Conference (DUC)1 and the Text Analysis Conference (TAC).2 Specif- ically, we use human-generated annotations pro- duced for the pyramid method (Nenkova et al., 2007) for summarization evaluation to produce a dataset of natural human fusions with quan- tifiable agreement. This offers advantages over previous datasets used for standalone English sentence fusion which contain annotator-induced noise (McKeown et al., 2010) or cannot be dis- tributed (Elsner and Santhanam, 2011). In ad- dition, both these datasets contain approximately 300 instances of fusion while the new dataset pre- sented here contains 1858 instances. Crucially, this larger corpus encourages super- vised approaches to sentence fusion and we lever- age this to explore new strategies for the task. Previous approaches to fusion have generally re- lied on variations of dependency graph combina- tion (Barzilay and McKeown, 2005; Filippova and Strube, 2008b; Elsner and Santhanam, 2011) for content selection with a separate step for lineariza- tion that is usually based on a language model (LM). …

Content truncated to 500 words