Unsupervised Estimation of Word Usage Similarity

Unsupervised Estimation of Word Usage Similarity Marco Lui,♠♥ Timothy Baldwin,♠♥ and Diana McCarthy ♣ ♠ NICTA Victoria Research Laboratory ♥ Dept of Computing and Information Systems, The University of Melbourne ♣ Dept of Theoretical and Applied Linguistics, University of Cambridge mhlui@unimelb.edu.au, tb@ldwin.net, diana@dianamccarthy.co.uk Abstract We present a method to estimate word use similarity independent of an external sense inventory. This method utilizes a topic- modelling approach to compute the similar- ity in usage of a single word across a pair of sentences, and we evaluate our method in terms of its ability to reproduce a human- annotated ranking over sentence pairs. We find that our method outperforms a bag-of- words baseline, and that for certain words there is very strong correlation between our method and human annotators. We also find that lemma-specific models do not out- perform general topic models, despite the fact that results with the general model vary substantially by lemma. We provide a de- tailed analysis of the result, and identify open issues for future research. 1 Introduction Automated Word Usage Similarity (Usim) is the task of determining the similarity in use of a par- ticular word across a pair of sentences. It is re- lated to the tasks of word sense disambiguation (WSD) and word sense induction (WSI), but dif- fers in that Usim does not pre-suppose a pre- defined sense inventory. It also captures the fact that word senses may not always be distinct, and that the applicability of word senses is not neces- sarily mutually exclusive. In Usim, we consider pairs of sentences at a time, and quantify the sim- ilarity of the sense of the target word being used in each sentence. An example of a sentence pair (SPAIR) using similar but not identical senses of the word dry is given in Figure 1. Usim is a relatively new NLP task, partly due to the lack of resources for its evaluation. Erk et al. (2009) recently produced a corpus of sentence Part c) All this has been a little dry so far: now for some fun. For people who knew him, it was typical of his dry hu- mor, but some in the audience thought he was tipsy. Figure 1: Example of an SPAIR judged by annotators to use similar but not identical senses of the word dry. pairs annotated for usage similarity judgments, al- lowing Usim to be formulated as a distinct task from the related tasks of word sense disambigua- tion and word sense induction. In this work, we propose a method to estimate word usage similarity in an entirely unsupervised fashion through the use of a topic model. We make use of the well-known Latent Dirichlet Al- location (LDA) model (Blei et al., 2003) to model the distribution of topics in a sentence, then exam- ine the similarity between sentences on the basis of the similarity between their topic distributions. Our main contributions in this work are: (1) we introduce a method to compute word usage sim- ilarity in an …

Content truncated to 500 words