Learning to Rank Semantic Coherence for Topic Segmentation

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1340–1344 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Learning to Rank Semantic Coherence for Topic Segmentation Liang Wang1 Sujian Li1,2 Yajuan Lyu3 Houfeng Wang1,2 1Key Laboratory of Computational Linguistics, Peking University, MOE, China 2Collaborative Innovation Center for Language Ability, Xuzhou, Jiangsu, China 3Baidu Inc., Beijing, China {intfloat,lisujian,wanghf}@pku.edu.cn lvyajuan@baidu.com Abstract Topic segmentation plays an important role for discourse parsing and information retrieval. Due to the absence of train- ing data, previous work mainly adopts un- supervised methods to rank semantic co- herence between paragraphs for topic seg- mentation. In this paper, we present an intuitive and simple idea to automatically create a “quasi” training dataset, which in- cludes a large amount of text pairs from the same or different documents with dif- ferent semantic coherence. With the train- ing corpus, we design a symmetric CNN neural network to model text pairs and rank the semantic coherence within the learning to rank framework. Experiments show that our algorithm is able to achieve competitive performance over strong base- lines on several real-world datasets. 1 Introduction The goal of topic segmentation is to segment a document into several topically coherent parts, with different parts corresponding to different top- ics. Topic segmentation enables better understand- ing of document structure, and makes long doc- ument much easier to navigate. It also provides helpful information for tasks such as information retrieval, topic tracking etc (Purver, 2011). Due to the lack of large scale annotated topic segmentation dataset, previous work mainly focus on unsupervised models to measure the coherence between two textual segments. The intuition be- hind unsupervised models is that two adjacent seg- ments from the same topic are more coherent than those from different topics. Under this intuition, one direction of research attempts to measure co- herence by computing text similarity. The typi- cal methods include TextTiling (Hearst, 1997) and its variants, such as C99 (Choi, 2000), TopicTil- ing (Riedl and Biemann, 2012b) etc. The other di- rection of research develops topic modeling tech- niques to explore topic representation of text and topic change between textual segments (Yam- ron et al., 1998; Eisenstein and Barzilay, 2008; Riedl and Biemann, 2012a; Du et al., 2013; Jameel and Lam, 2013). With carefully designed gen- erative process and efficient inference algorithm, topic models are able to model coherence as latent variables and outperform lexical similarity based models. Though unsupervised models make progress in modeling text coherence, they mostly suffer from one of the following two limitations. First, it is not precise to measure coherence with text sim- ilarity, since text similarity is just one aspect to influence coherence. Second, many assumptions and manually set parameters are usually involved in the complex modeling techniques, due to the absence of supervised information. To overcome aforementioned limitations, we prefer to directly model the text coherence by exploring possible su- pervised information. Then, we can learn a func- tion f(s1, s2) which takes two textual segments s1 …

Content truncated to 500 words