Improving Update Summarization via Supervised ILP and Sentence Reranking

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1317–1322, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Improving Update Summarization via Supervised ILP and Sentence Reranking Chen Li1, Yang Liu1, Lin Zhao2 1 Computer Science Department, The University of Texas at Dallas Richardson, Texas 75080, USA 2 Research and Technology Center, Robert Bosch LLC Palo Alto, California 94304, USA {chenli,yangl@hlt.utdallas.edu} {lin.zhao@us.bosch.com} Abstract Integer Linear Programming (ILP) based sum- marization methods have been widely adopted recently because of their state-of-the-art per- formance. This paper proposes two new mod- ifications in this framework for update sum- marization. Our key idea is to use discrimi- native models with a set of features to mea- sure both the salience and the novelty of words and sentences. First, these features are used in a supervised model to predict the weights of the concepts used in the ILP model. Sec- ond, we generate preliminary sentence candi- dates in the ILP model and then rerank them using sentence level features. We evaluate our method on different TAC update summariza- tion data sets, and the results show that our system performs competitively compared to the best TAC systems based on the ROUGE evaluation metric. 1 Introduction Update summarization has attracted significant re- search focus recently. Different from generic extrac- tive summarization, update summarization assumes that users already have some information about a given topic from an old data set, and thus for a new data set the system aims to generate a summary that contains as much novel information as possi- ble. This task was first introduced at DUC 2007 and then continued until TAC 2011. It is very useful to chronological events in real applications. Most basic update summarization methods are variants of multi-document summarization methods, with some consideration of the difference between the earlier and later document sets (Boudin et al., 2008; Fisher and Roark, 2008; Long et al., 2010; Bysani, 2010). One important line is to use graph- based co-ranking. They rank the sentences in the earlier and later document sets simultaneously by considering the sentence relationship. For example, Li et al. (2008) was inspired by the intuition that “a sentence receives a positive influence from the sentences that correlate to it in the same collection, whereas receives a negative influence from the sen- tences that correlates to it in the different (or previ- ously read) collection’, and proposed a graph based sentence ranking algorithm for update summariza- tion. Wan (2012) integrated two co-ranking pro- cesses by adding some strict constraints, which led to more accurate computation of sentences’ scores for update summarization. A similar method was also applied earlier by (Wan et al., 2011) for mul- tilingual news summarization. In addition, genera- tive models, such as topic models, have also been adopted for this task. For example, Delort and Alfonseca (2012) proposed a novel nonparametric Bayesian approach, a variant of Latent Dirichlet Al- location (LDA), aiming to distinguish between com- mon …

Content truncated to 500 words