Improving Supervised Sense Disambiguation with Web-Scale Selectors

Proceedings of COLING 2012: Technical Papers, pages 2423–2440, COLING 2012, Mumbai, December 2012. Improving Supervised Sense Disambiguation with Web-Scale Selectors H. Andrew Schwartz1 Fernando Gomez2 Lyle H. Ungar1 (1) University of Pennsylvania, Philadelphia, PA USA (2) University of Central Florida, Orlando, FL USA hansens@seas.upenn.edu, gomez@eecs.ucf.edu, ungar@cis.upenn.edu ABSTRACT This paper introduces a method to improve supervised word sense disambiguation perfor- mance by including a new class of features which leverage contextual information from large unannotated corpora. This new feature class, selectors, contains words that appear in other corpora with the same local context as a given lexical instance. We show that support vector sense classifiers trained with selectors achieve higher accuracy than those trained only with standard features, producing error reductions of 15.4% and 6.9% on standard coarse-grained and fine-grained disambiguation tasks respectively. Furthermore, we find an error reduction of 9.3% when including selectors for the classification step of named-entity recognition over a representative sample of OntoNotes. These significant improvements come free of any human annotation cost, only requiring unlabeled Web-Scale corpora. KEYWORDS: word sense disambiguation, lexical semantics, semi-supervised learning. 2423 1 Introduction Supervised word sense disambiguation (WSD) systems often rely directly on the local contexts in which target words appear. For example, the state-of-the-art system of Zhong et al. (2008) uses features based on collocations centered on the target word. Models relying on such features do well with copious amounts of training data, but they are prone to errors when the local context of a test instance differs from local context observed during training. Consider the sentences below. 1. The workers loaded the port onto the ship this morning. 2. She purchased a couple of bottles of port from the store. 3. The couple enjoyed their richly-flavored port. Though referring to the same sense of ‘port’, “a sweet dark-red dessert wine” (Miller et al., 1993), it is difficult to connect any two instances based on local context; the parts-of-speech even differ substantially. Models for disambiguation can benefit from the addition of a feature that does not rely directly on the local context. We present a new class of features which encodes an abstraction of a word’s context, rather than encoding contents of the local context itself. We refer to this new feature class as selectors, borrowing the term from an approach to knowledge-based (unsupervised) word sense disambiguation which uses the idea of searching for words that share the same context (Lin, 1997; Schwartz and Gomez, 2008). More precisely, selectors are words that show up in the same local context as a given instance of another word. For example, selectors for ‘port’ in sentence 1 might be ‘bottles’, ‘crates’, ‘passengers’, ‘wine’, ‘luggage’, etc. Considering that the other sentences may share some of the same selectors such as ‘bottles’ or ‘wine’, one can see how this abstraction of context to selectors can be beneficial. Figure 1 demonstrates mapping the context from one instance to selectors, which match the selectors of another instance. In this sense, it is the contexts (or word instances) …

Content truncated to 500 words