Graph-Based Induction of Word Senses in Croatian

Graph-Based Induction of Word Senses in Croatian Marko Bekavac and Jan ˇSnajder University of Zagreb, Faculty of Electrical Engineering and Computing Text Analysis and Knowledge Engineering Lab Unska 3, 10000 Zagreb, Croatia {jan.snajder, marko.bekavac}@fer.hr Abstract Word sense induction (WSI) seeks to induce senses of words from unannotated corpora. In this paper, we address the WSI task for the Croatian language. We adopt the word clustering approach based on co-occurrence graphs, in which senses are taken to correspond to strongly inter-connected components of co-occurring words. We experiment with a number of graph construction techniques and clustering algorithms, and evaluate the sense inventories both as a clustering problem and extrinsically on a word sense disambiguation (WSD) task. In the cluster-based evaluation, Chinese Whispers algorithm outperformed Markov Clustering, yielding a normalized mutual information score of 64.3. In contrast, in WSD evaluation Markov Clustering performed better, yielding an accuracy of about 75%. We are making available two induced sense inventories of 10,000 most frequent Croatian words: one coarse-grained and one fine-grained inventory, both obtained using the Markov Clustering algorithm. Keywords: word sense induction, co-occurrence graph, graph-based clustering, Croatian language 1. Introduction Word sense disambiguation (WSD) – the task of automati- cally determining the sense of a polysemous word in con- text – is considered one of the fundamental problems of natural language processing (Navigli, 2009). WSD can be approached as either sense labeling or sense discrimina- tion. In the former, words occurrences are assigned sense labels from a predefined sense inventory. In contrast, sense discrimination (Sch¨utze, 1998) addresses a simpler task of differentiating among the different uses of a word, without a reference to a sense inventory. Related to WSD is the task of word sense induction (WSI). WSI seeks to induce senses of words from unannotated corpora. There are two main approaches to WSI: con- text clustering and word clustering. Context clustering ex- ploits the distributional hypothesis (Harris, 1954) to group together the similar usages of a word. In contrast, word clustering groups together different but semantically simi- lar words that pertain to a specific sense (e.g., money, loan, and finance for the financial sense of bank). In particular, word clustering based on co-occurrence graphs posits that senses of a polysemous word correspond to strongly inter- connected components of its co-occurring words. In this paper, we address the WSI task for Croatian, a resource-scarce South Slavic language. Our aim is to build general-domain sense-induced inventories (in the form of word clusters) that can be readily used for WSD and re- lated tasks. We adopt a word clustering approach using co-occurrence graphs and experiment with a number of graph construction techniques and clustering algorithms. The contribution of our work is twofold: (1) we describe and make publicly available induced sense inventories for Croatian and (2) we evaluate the sense inventories both in- trinsically (as a clustering problem) and extrinsically on a WSD task. To the best of our knowledge, this is the first work on WSI for Croatian, and also the first that makes …

Content truncated to 500 words