Title Generation with Quasi-Synchronous Grammar

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 513–523, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Title Generation with Quasi-Synchronous Grammar Kristian Woodsend, Yansong Feng and Mirella Lapata School of Informatics, University of Edinburgh Edinburgh EH8 9AB, United Kingdom k.woodsend@ed.ac.uk, Y.Feng-4@sms.ed.ac.uk, mlap@inf.ed.ac.uk Abstract The task of selecting information and render- ing it appropriately appears in multiple con- texts in summarization. In this paper we present a model that simultaneously optimizes selection and rendering preferences. The model operates over a phrase-based represen- tation of the source document which we ob- tain by merging PCFG parse trees and depen- dency graphs. Selection preferences for in- dividual phrases are learned discriminatively, while a quasi-synchronous grammar (Smith and Eisner, 2006) captures rendering prefer- ences such as paraphrases and compressions. Based on an integer linear programming for- mulation, the model learns to generate sum- maries that satisfy both types of preferences, while ensuring that length, topic coverage and grammar constraints are met. Experiments on headline and image caption generation show that our method obtains state-of-the-art per- formance using essentially the same model for both tasks without any major modifications. 1 Introduction Summarization is the process of condensing a source text into a shorter version while preserving its infor- mation content. Humans summarize on a daily ba- sis and effortlessly, yet the automatic production of high-quality summaries remains a challenge. Most work today focuses on extractive summa- rization, where a summary is created by identifying and subsequently concatenating the most important sentences in a document. The advantage of this ap- proach is that it does not require a great deal of lin- guistic analysis to generate grammatical sentences, assuming the source document was well written. Unfortunately, extracts generated this way are often documents of low readability and text quality, and contain much redundant information. The concise- ness can be improved when sentence extraction is interfaced with sentence compression, where words and clauses are deleted based on rules typically op- erating over parsed input (Jing, 2000; Daum´e III and Marcu, 2002; Lin, 2003; Daum´e III, 2006; Zajic et al., 2007; Martins and Smith, 2009). An alternative abstractive or “bottom-up” ap- proach involves identifying high-interest words and phrases in the source text, and combining them into new sentences guided by a language model (Banko et al., 2000; Soricut and Marcu, 2007). This ap- proach has the potential to work well, breaking out of the single-sentence paradigm. Unfortunately, the resulting summaries are not always coherent — indi- vidual constituent phrases are often combined with- out any semantic constraints — or grammatical be- yond the n-gram horizon imposed by the language model. Constituent deletion and recombination are merely two of the many rewrite operations profes- sional editors and abstractors employ when creating summaries (Jing, 2002). Additional operations in- clude truncating sentences, aggregating them, and paraphrasing at word or syntax level. Furthermore, professionals write summaries in a task-specific style. News headlines for example are typically short (three to six words), written in …

Content truncated to 500 words