Weiwei: A Simple Unsupervised Latent Semantics based Approach for Sentence Similarity

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 586–590, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Weiwei: A Simple Unsupervised Latent Semantics based Approach for Sentence Similarity Weiwei Guo Department of Computer Science, Columbia University, weiwei@cs.columbia.edu Mona Diab Center for Computational Learning Systems, Columbia University, mdiab@ccls.columbia.edu Abstract The Semantic Textual Similarity (STS) shared task (Agirre et al., 2012) computes the degree of semantic equivalence between two sen- tences.1 We show that a simple unsupervised latent semantics based approach, Weighted Textual Matrix Factorization that only exploits bag-of-words features, can outperform most systems for this task. The key to the approach is to carefully handle missing words that are not in the sentence, and thus rendering it su- perior to Latent Semantic Analysis (LSA) and Latent Dirichlet Allocation (LDA). Our sys- tem ranks 20 out of 89 systems according to the official evaluation metric for the task, Pear- son correlation, and it ranks 10/89 and 19/89 in the other two evaluation metrics employed by the organizers. 1 Introduction Identifying the degree of semantic similarity [SS] between two sentences is helpful for many NLP top- ics. In Machine Translation (Kauchak and Barzi- lay, 2006) and Text Summarization (Zhou et al., 2006), results are automatically evaluated based on sentence comparison. In Text Coherence Detection (Lapata and Barzilay, 2005), sentences are linked to- gether by similar or related words. For Word Sense Disambiguation, researchers (Banerjee and Peder- sen, 2003; Guo and Diab, 2012a) construct a sense similarity measure from the sentence similarity of the sense definitions. Almost all SS approaches decompose the task into word pairwise similarity problems. For example, Is- 1Mona Diab, co-author of this paper, is one of the task orga- nizers lam and Inkpen (2008) create a matrix for each sen- tence pair, where columns are the words in the first sentence and rows are the words in the second sen- tence, and each cell stores the distributional similar- ity of the two words. Then they create an alignment between words in two sentences, and sentence simi- larity is calculated based on the sum of the similarity of aligned word pairs. There are two disadvantages with word similarity based approaches: 1. lexical ambiguity as the word pairwise similarity ignores the semantic interaction between the word and sen- tence/context. 2. word co-occurrence information is not as sufficiently exploited as they are in latent variable models such as Latent Semantic Analysis (LSA) (Landauer et al., 1998) and Latent Dirichilet Allocation (LDA) (Blei et al., 2003). On the other hand, latent variable models can solve the two issues naturally by modeling the semantics of words and sentences simultaneously in the low-dimensional la- tent space. However, attempts at addressing SS using LSA perform significantly below word similarity based models (Mihalcea et al., 2006; O’Shea et al., 2008). We believe the reason is that the observed words in a sentence are too few for latent variable mod- els to learn robust semantics. For example, given the two sentences of WordNet sense definitions …

Content truncated to 500 words