OPI: Semeval-2014 Task 3 System Description

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 454–458, Dublin, Ireland, August 23-24, 2014. OPI: Semeval-2014 Task 3 System Description Marek Kozlowski National Information Processing Institute mkozlowski@opi.org.pl Abstract In this paper, we describe the OPI system participating in the Semeval-2014 task 3 Cross-Level Semantic Similarity. Our ap- proach is knowledge-poor, there is no ex- ploitation of any structured knowledge re- sources as Wikipedia, WordNet or Babel- Net. The method is also fully unsuper- vised, the training set is only used in order to tune the system. System measures the semantic similarity of texts using corpus- based measures of termsets similarity. 1 Introduction The task Cross-Level Semantic Similarity of SemEval-2014 aims at an evaluation for seman- tic similarity across different sizes of text (lexi- cal levels). Unlike prior SemEval tasks on textual similarity that have focused on comparing similar- sized texts, the mentioned task evaluates the case where larger text must be compared to smaller text, namely there are covered four semantic sim- ilarity comparisons: paragraph to sentence, sen- tence to phrase, phrase to word and word to sense. We present the method for measuring the se- mantic similarity of texts using a corpus-based measure of termsets (set of words) similarity. We start from preprocessing texts, identifying bound- ary values, computing termsets similarities and de- rive from them the final score, which is normal- ized. The input of the task consists of two text seg- ments of different level. We want to determine a score indicating their semantic similarity of the smaller item to the larger item. Similarity is scored from 0 to 4, when 0 means no semantic intersec- This work is licensed under a Creative Commons Attribution 4.0 International Licence. Page numbers and proceedings footer are added by the organisers. Licence details: http: //creativecommons.org/licenses/by/4.0/ tion, 4 means that two items have very similar meanings. 2 Related Work There are lots of papers about measuring the similarity between documents and single words. Document-level similarity works are based on Vector Space Models (Salton and Lesk, 1971; Salton and McGill, 1983). A significant effort has also been put into measuring similarity at the word level, namely by approaches that use distributional semantics (Turney and Pantel, 2010). Related work can be classified into four ma- jor categories: vector-based document mod- els methods, corpus-based methods, knowledge- based methods and hybrid methods (Islam and Inkpen, 2008). Vector-based document models represent docu- ment as a vector of words and the similarity eval- uation is based on the number of words that oc- cur in both texts. Lexical similarity methods have problems with different words sharing common sense. Next approaches, such as corpus-based and knowledge-based methods, overcome the above issues. Corpus based methods apply scores provided by Pointwise Mutual Information (PMI) and Latent Semantic Analysis (LSA). The Pointwise Mutual Information (PMI) (Tur- ney, 2001) between two words wi and wj is: PMI(wi, wj) = log2 p(wi, wj) p(wi)p(wj) The Latent Semantic Analysis (LSA) (Landauer and Dumais, 1997; Landauer et al., 2007) is a …

Content truncated to 500 words