UNITOR: Combining Semantic Text Similarity functions through SV Regression

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 597–602, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics UNITOR: Combining Semantic Text Similarity functions through SV Regression Danilo Croce, Paolo Annesi, Valerio Storch and Roberto Basili Department of Enterprise Engineering University of Roma, Tor Vergata 00133 Roma, Italy {croce,annesi,storch,basili}@info.uniroma2.it Abstract This paper presents the UNITOR system that participated to the SemEval 2012 Task 6: Se- mantic Textual Similarity (STS). The task is here modeled as a Support Vector (SV) regres- sion problem, where a similarity scoring func- tion between text pairs is acquired from exam- ples. The semantic relatedness between sen- tences is modeled in an unsupervised fashion through different similarity functions, each capturing a specific semantic aspect of the STS, e.g. syntactic vs. lexical or topical vs. paradigmatic similarity. The SV regressor ef- fectively combines the different models, learn- ing a scoring function that weights individual scores in a unique resulting STS. It provides a highly portable method as it does not depend on any manually built resource (e.g. WordNet) nor controlled, e.g. aligned, corpus. 1 Introduction Semantic Textual Similarity (STS) measures the de- gree of semantic equivalence between two phrases or texts. An effective method to compute similar- ity between short texts or sentences has many appli- cations in Natural Language Processing (Mihalcea et al., 2006) and related areas such as Information Retrieval, e.g. to improve the effectiveness of a se- mantic search engine (Sahami and Heilman, 2006), or databases, where text similarity can be used in schema matching to solve semantic heterogeneity (Islam and Inkpen, 2008). STS is here modeled as a Support Vector (SV) re- gression problem, where a SV regressor learns the similarity function over text pairs. Regression learn- ing has been already applied to different NLP tasks. In (Pang and Lee, 2005) it is applied to Opinion Mining, in particular to the rating-inference prob- lem, wherein one must determine an author evalua- tion with respect to a multi-point scale. In (Albrecht and Hwa, 2007) a method is proposed for develop- ing sentence-level MT evaluation metrics using re- gression learning without directly relying on human reference translations. In (Biadsy et al., 2008) it has been used to rank candidate sentences for the task of producing biographies from Wikipedia. Finally, in (Becker et al., 2011) SV regressor has been used to rank questions within their context in the multi- modal tutorial dialogue problem. In this paper, the semantic relatedness between two sentences is modeled as a combination of dif- ferent similarity functions, each describing the anal- ogy between the two texts according to a specific semantic perspective: in this way, we aim at captur- ing syntactic and lexical equivalences between sen- tences and exploiting either topical relatedness or paradigmatic similarity between individual words. The variety of semantic evidences that a system can employ here grows quickly, according to the genre and complexity of the targeted sentences. We thus propose to combine such a body of evidence to learn a comprehensive scoring function y …

Content truncated to 500 words