Learning to Summarise Related Sentences

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1636–1647, Dublin, Ireland, August 23-29 2014. Learning to Summarise Related Sentences Emmanouil Tzouridis† † Dep. of Computer Science TU Darmstadt Germany {tzouridis,brefeld}@kma. informatik.tu-darmstadt.de Jamal Abdul Nasir Dep. of Computer Science LUMS Lahore Pakistan jamaln@lums.edu.pk Ulf Brefeld†‡ ‡ Inform. Center f. Education DIPF Frankfurt/Main Germany brefeld@dipf.de Abstract We cast multi-sentence compression as a structured prediction problem. Related sentences are represented by a word graph so that summaries constitute paths in the graph (Filippova, 2010). We devise a parameterised shortest path algorithm that can be written as a generalised linear model in a joint space of word graphs and compressions. We use a large-margin approach to adapt parameterised edge weights to the data such that the shortest path is identical to the desired summary. Decoding during training is performed in polynomial time using loss augmented infer- ence. Empirically, we compare our approach to the state-of-the-art in graph-based multi-sentence compression and observe significant improvements of about 7% in ROUGE F-measure and 8% in BLEU score, respectively. 1 Introduction Automatic text summarisation is one of oldest forms of natural language processing (Luhn, 1958; Bax- endale, 1958). The goal is to extract the most important part of the content from either a single document or a collection of documents (Mani, 2001; Roussinov and Chen, 2001; McKeown et al., 2005). Frequently, the information of interest is contained in only a part of a sentence or may be distributed across parts of several sentences. Identifying the content carrying part(s) constitutes an essential tech- nique not only for single- and multi-document extractive summarisation but also text simplification in general. Generating a simplified version of a text traditionally has many applications in question answer- ing (Hermjakob et al., 2002) and speech synthesis (Kaji et al., 2004). Due to limited display sizes of mobile devices, recent applications also deal with summarising/simplifying news articles, social media, emails, or websites (Corston-Oliver, 2001). Multi-sentence compression (MSC) unifies many of the mentioned characteristics and challenges and can be seen as a key to text summarisation and simplification (Jing and McKeown, 2000). The task in multi-sentence compression is to map a collection of related sentences to a grammatical short sentence that preserves the most important part of the content. Sentence compression methods have been devised using manually crafted rules (Dorr et al., 2003), language models (Hori et al., 2003; Clarke and Lapata, 2008), or syntactical representations (Barzilay and Lee, 2003; Galley and McKeown, 2007; Filippova and Strube, 2008). Filippova (2010) introduces an elegant graph-based approach to multi-sentence compres- sion that simply relies on the words of the sentences and efficient dynamic programming. Her approach implements the observation that the frequency of words influences their appearance in human summaries (Nenkova et al., 2006). Although being an intuitive rule that does work well in practice, frequency-based strategies often remain heuristic. In this paper we propose a structured learning-based approach to multi-sentence compression. In analogy to Filippova (2010), related sentences are represented by a …

Content truncated to 500 words