Exploring ESA to Improve Word Relatedness

Proceedings of the Third Joint Conference on Lexical and Computational Semantics (*SEM 2014), pages 51–56, Dublin, Ireland, August 23-24 2014. Exploring ESA to Improve Word Relatedness Nitish Aggarwal Kartik Asooja Paul Buitelaar Insight Centre for Data Analytics National University of Ireland Galway, Ireland firstname.lastname@deri.org Abstract Explicit Semantic Analysis (ESA) is an ap- proach to calculate the semantic relatedness between two words or natural language texts with the help of concepts grounded in human cognition. ESA usage has received much at- tention in the field of natural language pro- cessing, information retrieval and text analy- sis, however, performance of the approach de- pends on several parameters that are included in the model, and also on the text data type used for evaluation. In this paper, we investi- gate the behavior of using different number of Wikipedia articles in building ESA model, for calculating the semantic relatedness for differ- ent types of text pairs: word-word, phrase- phrase and document-document. With our findings, we further propose an approach to improve the ESA semantic relatedness scores for words by enriching the words with their explicit context such as synonyms, glosses and Wikipedia definitions. 1 Introduction Explicit Semantic Analysis (ESA) is a distributional semantic model (Harris, 1954) that computes the relatedness scores between natural language texts by using high dimensional vectors. ESA builds the high dimensional vectors by using the explicit concepts defined in human cognition. Gabrilovich and Markovitch (2007) introduced the ESA model in which Wikipedia and Open Directory Project1 was used to obtain the explicit concepts. ESA con- siders every Wikipedia article as a unique explicit 1http://www.dmoz.org topic. It also assumes that the articles are topically orthogonal. However, recent work (Gottron et al., 2011) has shown that by using the documents from Reuters corpus instead of Wikipedia articles can also achieve comparable results. ESA model includes various parameters (Sorg and Cimiano, 2010) that play important roles on its performance. Therefore, the model requires further investigation in order to better tune the parameters. ESA model has been adapted very quickly in different fields related to text analysis, due to the simplicity of its implementation and the availability of Wikipedia corpus. Gabrilovich and Markovitch (2007) evaluated the ESA against word relatedness dataset WN353 (Finkelstein et al., 2001) and doc- ument relatedness dataset Lee50 (Lee et al., 2005) by using all the articles from Wikipedia snapshot of 11 Nov, 2005. However, the results reported using different implementations (Polajnar et al., 2013) (Hassan and Mihalcea, 2011) of ESA on same datasets (WN353 and Lee50) vary a lot, due the specificity of ESA implementation. For instance, Hassan and Mihalcea (2011) found a significant difference between the scores obtained from their own implementation and the scores reported in the original article (Gabrilovich and Markovitch, 2007). In this paper, first, we investigate the behavior of ESA model in calculating the semantic related- ness for different types of text pairs: word-word, phrase-phrase and document-document by using different number of Wikipedia articles for building the model. Second, we propose an approach 51 …

Content truncated to 500 words