UNITOR-CORE_TYPED: Combining Text Similarity and Semantic Filters through SV Regression

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 59–65, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics UNITOR-CORE TYPED: Combining Text Similarity and Semantic Filters through SV Regression Danilo Croce, Valerio Storch and Roberto Basili Department of Enterprise Engineering University of Roma, Tor Vergata 00133 Roma, Italy {croce,storch,basili}@info.uniroma2.it Abstract This paper presents the UNITOR system that participated in the *SEM 2013 shared task on Semantic Textual Similarity (STS). The task is modeled as a Support Vector (SV) regression problem, where a similarity scoring function between text pairs is acquired from examples. The proposed approach has been implemented in a system that aims at providing high ap- plicability and robustness, in order to reduce the risk of over-fitting over a specific datasets. Moreover, the approach does not require any manually coded resource (e.g. WordNet), but mainly exploits distributional analysis of un- labeled corpora. A good level of accuracy is achieved over the shared task: in the Typed STS task the proposed system ranks in 1st and 2nd position. 1 Introduction Semantic Textual Similarity (STS) measures the de- gree of semantic equivalence between two phrases or texts. An effective method to compute similarity between sentences or semi-structured material has many applications in Natural Language Processing (Mihalcea et al., 2006) and related areas such as Information Retrieval, improving the effectiveness of semantic search engines (Sahami and Heilman, 2006), or databases, using text similarity in schema matching to solve semantic heterogeneity (Islam and Inkpen, 2008). This paper describes the UNITOR system partic- ipating in both tasks of the *SEM 2013 shared task on Semantic Textual Similarity (STS), described in (Agirre et al., 2013): • the Core STS tasks: given two sentences, s1 and s2, participants are asked to provide a score reflecting the corresponding text similarity. It is the same task proposed in (Agirre et al., 2012). • the Typed-similarity STS task: given two semi-structured records t1 and t2, containing several typed fields with textual values, partic- ipants are asked to provide multiple similarity scores: the types of similarity to be studied in- clude location, author, people involved, time, events or actions, subject and description. In line with several participants of the STS 2012 challenge, such as (Banea et al., 2012; Croce et al., 2012a; ˇSari´c et al., 2012), STS is here modeled as a Support Vector (SV) regression problem, where a SV regressor learns the similarity function over text pairs. The semantic relatedness between two sen- tences is first modeled in an unsupervised fashion by several similarity functions, each describing the analogy between the two texts according to a spe- cific semantic perspective. We aim at capturing sep- arately syntactic and lexical equivalences between sentences and exploiting either topical relatedness or paradigmatic similarity between individual words. Such information is then combined in a supervised schema through a scoring function y = f(⃗x) over individual measures from labeled data through SV regression: y is the gold similarity score (provided …

Content truncated to 500 words