Non-Orthogonal Explicit Semantic Analysis

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 92–100, Denver, Colorado, June 4–5, 2015. Non-Orthogonal Explicit Semantic Analysis Nitish Aggarwal Kartik Asooja Georgeta Bordea Paul Buitelaar Insight Centre for Data Analytics National University of Ireland Galway, Ireland firstname.lastname@insight-centre.org Abstract Explicit Semantic Analysis (ESA) utilizes the Wikipedia knowledge base to represent the se- mantics of a word by a vector where every dimension refers to an explicitly defined con- cept like a Wikipedia article. ESA inherently assumes that Wikipedia concepts are orthog- onal to each other, therefore, it considers that two words are related only if they co-occur in the same articles. However, two words can be related to each other even if they appear separately in related articles rather than co- occurring in the same articles. This leads to a need for extending the ESA model to con- sider the relatedness between the explicit con- cepts (i.e. Wikipedia articles in Wikipedia based implementation) for computing textual relatedness. In this paper, we present Non- Orthogonal ESA (NESA) which represents more fine grained semantics of a word as a vector of explicit concept dimensions, where every such concept dimension further consti- tutes a semantic vector built in another vec- tor space. Thus, NESA considers the concept correlations in computing the relatedness be- tween two words. We explore different ap- proaches to compute the concept correlation weights, and compare these approaches with other existing methods. Furthermore, we eval- uate our model NESA on several word related- ness benchmarks showing that it outperforms the state of the art methods. 1 Introduction Significance of quantifying relatedness between two natural language texts has been shown in various tasks which deal with information retrieval (IR), natural language processing (NLP), or other related fields. The semantics of a word can be obtained from existing lexical resources like WordNet and FrameNet. However, such lexical resources require domain expertise for defining the hierarchical structure, which makes their creation very expen- sive. Therefore, distributional semantic models (DSMs) have achieved much attention as they uti- lize available document collections like Wikipedia, and do not depend upon human expertise (Harris, 1954). DSMs represent the semantics of a word by transforming it to a high dimensional distributional vector in a predefined concept space. Many models have been proposed that derive this concept space by using explicit concepts or implicit concepts. Explicit Semantic Analysis (ESA) (Gabrilovich and Markovitch, 2007) utilizes the concepts which are explicitly derived under human cognition like Wikipedia concepts (articles). However, Latent Semantic Analysis (LSA) derives a latent concept space by performing dimensionality reduction (Landauer et al., 1998). Gabrilovich and Markovitch (2007) introduced ESA model in which Wikipedia and Open Directory Project were used to obtain the explicit concepts, however, Wikipedia has been a popular choice in further ESA implementations (Polajnar et al., 2013; Gottron et al., 2011; Aggarwal et al., 2014). ESA represents the semantics of a word with a high di- mensional vector over the Wikipedia concepts. The tf-idf weight of the word with …

Content truncated to 500 words