Unsupervised Word Usage Similarity in Social Media Texts

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 248–253, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics Unsupervised Word Usage Similarity in Social Media Texts Spandana Gella,♣ Paul Cook,♣ and Bo Han♠♣ ♠ NICTA Victoria Research Laboratory ♣ Department of Computing and Information Systems, The University of Melbourne sgella@student.unimelb.edu.au, paulcook@unimelb.edu.au, hanb@student.unimelb.edu.au Abstract We propose an unsupervised method for au- tomatically calculating word usage similar- ity in social media data based on topic mod- elling, which we contrast with a baseline dis- tributional method and Weighted Textual Ma- trix Factorization. We evaluate these meth- ods against a novel dataset made up of human ratings over 550 Twitter message pairs anno- tated for usage similarity for a set of 10 nouns. The results show that our topic modelling ap- proach outperforms the other two methods. 1 Introduction In recent years, with the growing popularity of so- cial media applications, there has been a steep rise in the amount of “post”-based user-generated text (including microblog posts, status updates and com- ments) (Bennett, 2012). This data has been iden- tified as having potential for applications ranging from trend analysis (Lau et al., 2012a) and event de- tection (Osborne et al., 2012) to election outcome prediction (O’Connor et al., 2010). However, given that posts are generally very short, noisy and lack- ing in context, traditional NLP approaches tend to perform poorly over social media data (Hong and Davison, 2010; Ritter et al., 2011; Han et al., 2012). This is the first paper to address the task of lexi- cal semantic interpretation in microblog data based on word usage similarity. Word usage similar- ity (USIM: Erk et al. (2009)) is a relatively new paradigm for capturing similarity in the usages of a given word independently of any lexicon or sense inventory. The task is to rate on an ordinal scale the similarity in usage between two different usages of the same word. In doing so, it avoids common issues in conventional word sense disambiguation, relating to sense underspecification, the appropriateness of a static sense inventory to a given domain, and the in- ability to capture similarities/overlaps between word senses. As an example of USIM, consider the fol- lowing pairing of Twitter posts containing the target word paper: 1. Deportation of Afghan Asylum Seekers from Australia : This paper aims to critically evalu- ate a newly signed agree. 2. @USER has his number on a piece of paper and I walkd off! The task is to predict a real-valued number in the range [1, 5] for the similarity in the respective us- ages of paper, where 1 indicates the usages are com- pletely different and 5 indicates they are identical. In this paper we develop a new USIM dataset based on Twitter data. In experiments on this dataset we demonstrate that an LDA-based topic modelling approach outperforms a baseline distributional se- mantic approach and Weighted Textual Matrix Fac- torization (WTMF: Guo and …

Content truncated to 500 words