Unknown word sense detection as outlier detection

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 128–135, New York, June 2006. c⃝2006 Association for Computational Linguistics Unknown word sense detection as outlier detection Katrin Erk Computational Linguistics Saarland University Saarbr¨ucken, Germany erk@coli.uni-sb.de Abstract We address the problem of unknown word sense detection: the identification of cor- pus occurrences that are not covered by a given sense inventory. We model this as an instance of outlier detection, using a simple nearest neighbor-based approach to measuring the resemblance of a new item to a training set. In combination with a method that alleviates data sparseness by sharing training data across lemmas, the approach achieves a precision of 0.77 and recall of 0.82. 1 Introduction If a system has seen only positive examples, how does it recognize a negative example? This is the problem addressed by outlier detection, also called novelty detection1 (Markou and Singh, 2003a; Markou and Singh, 2003b; Marsland, 2003): to de- tect novel or unknown items that differ from all the seen training data. Outlier detection approaches typ- ically derive some model of “normal” objects from the training set and use a distance measure and a threshold to detect abnormal items. In this paper, we apply outlier detection tech- niques to the task of unknown sense detection: the identification of corpus occurrences that are not cov- ered by a given sense inventory. The training set 1The term novelty detection is also used for the distinction of novel and repeated information in information retrieval, a different if related topic. Figure 1: Wrong assignment due to missing sense: from the Hound of the Baskervilles, Ch. 14 against which new occurrences are compared will consist of sense-annotated text. Unknown sense detection is related to word sense disambiguation (WSD) and to word sense discrim- ination (Sch¨utze, 1998), but differs from both. In WSD all senses are assumed known, and the task is to select one of them, while in unknown sense detec- tion the task is to decide whether a given occurrence matches any of the known senses or none of them, and all training instances, regardless of the sense to which they belong, are modeled as one group of known data. Unknown sense detection also differs from word sense discrimination, where no sense in- ventory is given and the task is to group occurrences into senses. In unknown sense detection the model respects the given word senses. The main motivation for this study comes from shallow semantic parsing, by which we mean a com- bination of WSD and the automatic assignment of 128 semantic roles to free text. In cases where a sense is missing from the inventory, WSD will wrongly assign one of the existing senses. Figure 1 shows an example, a sentence from the Hound of the Baskervilles, analyzed by the SHALMANESER (Erk and Pado, 2006) shallow semantic parser. The anal- ysis is based on FrameNet (Baker et al., 1998), a resource that lists senses and semantic roles for En- glish …

Content truncated to 500 words