Inducing Fine-Grained Semantic Classes via Hierarchical and Collective Classification

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 931–939, Beijing, August 2010 Inducing Fine-Grained Semantic Classes via Hierarchical and Collective Classification Altaf Rahman and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas {altaf,vince}@hlt.utdallas.edu Abstract Research in named entity recognition and mention detection has typically involved a fairly small number of semantic classes, which may not be adequate if seman- tic class information is intended to sup- port natural language applications. Moti- vated by this observation, we examine the under-studied problem of semantic sub- type induction, where the goal is to au- tomatically determine which of a set of 92 fine-grained semantic classes a noun phrase belongs to. We seek to improve the standard supervised approach to this prob- lem using two techniques: hierarchical classification and collective classification. Experimental results demonstrate the ef- fectiveness of these techniques, whether or not they are applied in isolation or in combination with the standard approach. 1 Introduction Semantic class determination refers to the task of classifying a noun phrase (NP), be it a name or a nominal, as one of a set of pre-defined se- mantic classes. A semantic class classifier is a basic text-processing component in many high- level natural language processing (NLP) applica- tions, including information-extraction (IE) sys- tems and question-answering (QA) systems. In recent years, supervised semantic class determi- nation has been tackled primarily in the context of (1) coreference resolution (e.g., Ng (2007), Huang et al. (2009)), where semantic classes are induced and subsequently used to disallow coreference be- tween semantically incompatible NPs, and (2) the mention detection task in the ACE evaluations (e.g., Florian et al. (2004; 2006)), where the goal is to identify the boundary of a mention (i.e., a noun phrase that belongs to one of the pre-defined ACE semantic classes), its mention type (e.g., pro- noun, name), and its semantic class. The output of a mention detector is then used by downstream IE components, which typically include a coref- erence resolution system and a relation extraction system. Owing in part to its potentially large in- fluence on downstream IE components, accurate semantic class determination is crucial. Over the years, NLP researchers have focused on a relatively small number of semantic classes in both NE recognition and mention detection: seven classes in the MUC-6 and MUC-7 NE recognition task, four classes in the CoNLL 2002 and 2003 NE recognition shared task, and seven classes in the ACE 2005 mention detection task. Given that one of the uses of semantic class information is to support NLP applications, it is questionable whether this purpose can be adequately served by such a small number of semantic classes. For ex- ample, given the question “Which city was the first Olympic Games held in?”, it would be help- ful for a QA system to know which NEs are cities. However, virtually all of the existing NE recog- nizers and mention detectors can only determine whether an NE is a location or not. Our goal …

Content truncated to 500 words