Learning to Generate Coherent Summary with Discriminative Hidden Semi-Markov Model

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1648–1659, Dublin, Ireland, August 23-29 2014. Learning to Generate Coherent Summary with Discriminative Hidden Semi-Markov Model Hitoshi Nishikawa1, Kazuho Arita1, Katsumi Tanaka1, Tsutomu Hirao2, Toshiro Makino1 and Yoshihiro Matsuo1 Nippon Telegraph and Telephone Corporation 1 1-1 Hikari-no-oka, Yokosuka-shi, Kanagawa, 239-0847 Japan 2 2-4 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0237 Japan { nishikawa.hitoshi, arita.kazuho, tanaka.katsumi hirao.tsutomu, makino.toshiro, matsuo.yoshihiro } @lab.ntt.co.jp Abstract In this paper we introduce a novel single-document summarization method based on a hidden semi-Markov model. This model can naturally model single-document summarization as the optimization problem of selecting the best sequence from among the sentences in the input doc- ument under the given objective function and knapsack constraint. This advantage makes it possible for sentence selection to take the coherence of the summary into account. In addition our model can also incorporate sentence compression into the summarization process. To demon- strate the effectiveness of our method, we conduct an experimental evaluation with a large-scale corpus consisting of 12,748 pairs of a document and its reference. The results show that our method significantly outperforms the competitive baselines in terms of ROUGE evaluation, and the linguistic quality of summaries is also improved. Our method successfully mimicked the reference summaries, about 20 percent of the summaries generated by our method were com- pletely identical to their references. Moreover, we show that large-scale training samples are quite effective for training a summarizer. 1 Introduction Single-document summarization is attracting much more attention as a key technology in providing better information access in a commercial context. The Financial Times and CNN have been providing summaries of articles in their websites to attract users, and Summly, which has been acquired by Yahoo!, provided the service of automatically summarizing articles on the Internet. Given the cost of manual summarization, we can greatly improve the information access of Internet users by creating an automatic summarizer that can approach the summarization quality of humans. To mimic manually-written summaries, one important aspect is coherence (Nenkova and McKeown, 2011). Although coherence has been studied widely in a field of multi-document summarization (Kara- manis et al., 2004; Barzilay and Lapata, 2005; Nishikawa et al., 2010; Christensen et al., 2013), it has not been studied enough in the context of single-document summarization. In this paper, we revisit the prob- lem of coherence and employ it to produce both informative and linguistically high-quality summaries. To obtain such summaries, we introduce a novel summarization method based on a hidden semi- Markov model. The method has the properties of both the popular single-document summarization model, the knapsack problem, which packs the sentences into the given length and the hidden Markov model, which takes summary coherence into account by determining sentence context when selecting sentences. By leveraging this, we can build a summarizer that naturally achieves coherence. We state the novelty and contributions of this paper as follows: • We regard single-document summarization as a combinatorial optimization problem modeled by a hidden semi-Markov model …

Content truncated to 500 words