Predicting Word Association Strengths

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1283–1288 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Predicting Word Association Strengths Andrew Cattle Xiaojuan Ma Hong Kong University of Science and Technology Department of Computer Science and Engineering Clear Water Bay, Hong Kong {acattle,mxj}@cse.ust.hk Abstract This paper looks at the task of pre- dicting word association strengths across three datasets; WordNet Evocation (Boyd- Graber et al., 2006), University of South- ern Florida Free Association norms (Nel- son et al., 2004), and Edinburgh Associa- tive Thesaurus (Kiss et al., 1973). We achieve results of r = 0.357 and ρ = 0.379, r = 0.344 and ρ = 0.300, an ρ = 0.292 and ρ = 0.363, respectively. We find Word2Vec (Mikolov et al., 2013) and GloVe (Pennington et al., 2014) co- sine similarities, as well as vector offsets, to be the highest performing features. Fur- thermore, we examine the usefulness of Gaussian embeddings (Vilnis and McCal- lum, 2014) for predicting word association strength, the first work to do so. 1 Introduction Word embeddings such as Word2Vec (Mikolov et al., 2013) or GloVe (Pennington et al., 2014) have received increasing attention in the world of natural language processing and computational linguistics. Under such embeddings, the seman- tic relatedness of two words is generally taken to be the cosine similarity of their word vectors. Al- though this approach performs well for variety of applications, it is not without its limitations. First, it defines “relatedness” quite narrowly as the ex- tent to which the two words appear in similar con- texts. Second, it fails to capture how humans in- ternally represent words (De Deyne et al., 2016b). Word associations offer a more flexible view of semantic relatedness by leveraging “lexical knowledge acquired through world experience” (Nelson et al., 2004). While word embeddings capture distributional relationships, word associ- ations are able to capture more nuanced relation- ships “which are based on human perception and experiences [and] are not reflected in common lan- guage usage.” (Ma, 2013) For example, “yellow” is so closely associated with “banana” that many people would only specify a banana’s colour if it is not yellow. This is backed up by De Deyne et al. (2016b) which found word associations performed better than word embeddings across a variety of semantic relatedness tasks. Furthermore, word associations, unlike cosine similarities, are asymmetric; when presented with the word “beer”, many people think of the word “glass” but when presented with the word “glass”, few people think of the word “beer” (Ma, 2013). This directionality allows for more fine-grained exploration of semantic links, with applications in word similarity (Jabeen et al., 2013) and computa- tional humour (Cattle and Ma, 2016). Although several word association datasets ex- ist, such as the Edinburgh Associative Thesaurus (EAT, Kiss et al., 1973), the University of South Florida Free Association Norms (USF, Nelson et al., 2004), or WordNet Evocation (Evocation, Boyd-Graber et al., 2006), their reliance on human annotations mean they all suffer from …

Content truncated to 500 words