Sentence Similarity Learning by Lexical Decomposition and Composition

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1340–1349, Osaka, Japan, December 11-17 2016. Sentence Similarity Learning by Lexical Decomposition and Composition Zhiguo Wang and Haitao Mi and Abraham Ittycheriah IBM T.J. Watson Research Center Yorktown Heights, NY, USA {zhigwang, hmi, abei}@us.ibm.com Abstract Most conventional sentence similarity methods only focus on similar parts of two input sen- tences, and simply ignore the dissimilar parts, which usually give us some clues and semantic meanings about the sentences. In this work, we propose a model to take into account both the similarities and dissimilarities by decomposing and composing lexical semantics over sentences. The model represents each word as a vector, and calculates a semantic matching vector for each word based on all words in the other sentence. Then, each word vector is decomposed into a sim- ilar component and a dissimilar component based on the semantic matching vector. After this, a two-channel CNN model is employed to capture features by composing the similar and dis- similar components. Finally, a similarity score is estimated over the composed feature vectors. Experimental results show that our model gets the state-of-the-art performance on the answer sentence selection task, and achieves a comparable result on the paraphrase identification task. 1 Introduction Sentence similarity is a fundamental metric to measure the degree of likelihood between a pair of sen- tences. It plays an important role for a variety of tasks in both NLP and IR communities. For example, in paraphrase identification task, sentence similarity is used to determine whether two sentences are para- phrases or not (Yin and Sch¨utze, 2015; He et al., 2015). For question answering and information retrieval tasks, sentence similarities between query-answer pairs are used for assessing the relevance and ranking all the candidate answers (Severyn and Moschitti, 2015; Wang and Ittycheriah, 2015). However, sentence similarity learning has following challenges: 1. There is a lexical gap between semantically equivalent sentences. Take the E1 and E2 in Table 1 for example, they have the similar meaning but with different lexicons. 2. Semantic similarity should be measured at different levels of granularity (word-level, phrase-level and syntax-level). E.g., “not related” in E2 is an indivisible phrase when matching with “irrelevant” in E1 (shown in square brackets). 3. The dissimilarity (shown in angle brackets) between two sentences is also a significant clue (Qiu et al., 2006). For example, by judging the dissimilar parts, we can easily identify that E3 and E5 share the similar meaning “The study is about salmon”, because “sockeye” belongs to the salmon family, and “flounder” does not. Whereas the meaning of E4 is quite different from E3, which emphasizes “The study is about red (a special kind of) salmon”, because both “sockeye” and “coho” are in the salmon family. How we can extract and utilize those information becomes another challenge. In order to handle the above challenges, researchers have been working on sentence similarity al- gorithms for a long time. To bridge the lexical gap (challenge 1), some word …

Content truncated to 500 words