Saarland: Vector-based models of semantic textual similarity

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 603–607, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Saarland: Vector-based models of semantic textual similarity Georgiana Dinu Center of Mind/Brain Sciences University of Trento georgiana.dinu@unitn.it Stefan Thater Dept. of Computational Linguistics Universit¨at des Saarlandes stth@coli.uni-saarland.de Abstract This paper describes our system for the Se- meval 2012 Sentence Textual Similarity task. The system is based on a combination of few simple vector space-based methods for word meaning similarity. Evaluation results show that a simple combination of these unsuper- vised data-driven methods can be quite suc- cessful. The simple vector space components achieve high performance on short sentences; on longer, more complex sentences, they are outperformed by a surprisingly competitive word overlap baseline, but they still bring im- provements over this baseline when incorpo- rated into a mixture model. 1 Introduction Vector space models are widely-used methods for word meaning similarity which exploit the so-called distributional hypothesis, stating that semantically similar words tend to occur in similar contexts. Word meaning is represented by the contexts in which a word occurs, and similarity is computed by compar- ing these contexts in a high-dimensional vector space (Turney and Pantel, 2010). Distributional models of word meaning are attractive because they are sim- ple, have wide coverage, and can be easily acquired at virtually no cost in an unsupervised way. Fur- thermore, recent research has shown that, at least to some extent, these models can be generalized to capture similarity beyond the (isolated) word level, either as lexical meaning modulated by context, or as vectorial meaning representations for phrases and sentences. In this paper we evaluate the use of some of these models for the Semantic Textual Similarity (STS) task, which measures the degree of semantic equivalence between two sentences. In recent work Mitchell and Lapata (2008) has drawn the attention to the question of building vecto- rial meaning representations for sentences by combin- ing individual word vectors. They propose a family of simple “compositional” models that compute a vector for a phrase or a sentence by combining vectors of the constituent words, using different operations such as vector addition or component-wise multiplication. More refined models have been proposed recently by Baroni and Zamparelli (2010) and Grefenstette and Sadrzadeh (2011). Thater et al. (2011) and others take a slightly dif- ferent perspective on the problem: Instead of com- puting a vector representation for a complete phrase or sentence, they focus on the problem of “disam- biguating” the vector representation of a target word based on distributional information about the words in the target’s context. While this approach is not “compositional” in the sense described above, it still captures some meaning of the complete phrase in which a target word occurs. In this paper, we report on the system we used in the Semeval 2012 Sentence Textual Similarity shared task and describe an approach that uses a combina- tion of few simple vector-based components. We extend the model of Thater et al. (2011), which …

Content truncated to 500 words