A Cascaded Classification Approach to Semantic Head Recognition

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 793–803, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics A Cascaded Classification Approach to Semantic Head Recognition Lukas Michelbacher Alok Kothari Martin Forst† Christina Lioma Hinrich Sch¨utze Institute for NLP University of Stuttgart {michells,kotharak,liomaca}@ims.uni-stuttgart.de †Microsoft martin.forst@microsoft.com Abstract Most NLP systems use tokenization as part of preprocessing. Generally, tokenizers are based on simple heuristics and do not recog- nize multi-word units (MWUs) like hot dog or black hole unless a precompiled list of MWUs is available. In this paper, we propose a new cascaded model for detecting MWUs of arbitrary length for tokenization, focusing on noun phrases in the physics domain. We adopt a classification approach because – un- like other work on MWUs – tokenization re- quires a completely automatic approach. We achieve an accuracy of 68% for recognizing non-compositional MWUs and show that our MWU recognizer improves retrieval perfor- mance when used as part of an information re- trieval system. 1 Introduction Most NLP systems use tokenization as part of pre- processing. Generally, tokenizers are based on sim- ple heuristics and do not recognize multi-word units (MWUs) like hot dog or black hole. Our long-term goal is to build MWU-aware tokenizers that are used as part of the standard toolkit for NLP preprocessing alongside part-of-speech and named-entity tagging. We define an MWU as a sequence of words that has properties that cannot be inferred from the com- ponent words (cf. e.g. Manning and Sch¨utze (1999, Ch. 5), Sag et al. (2002)). The most important of these properties is non-compositionality, the fact that the meaning of a phrase cannot be predicted from the meanings of its component words. For ex- ample, a hot dog is not a hot animal but a sausage in a bun and a black hole in astrophysics is a region of space with special properties, not a dark cavity. The correct recognition of MWUs is an important building block of many NLP tasks. For example, in information retrieval (IR) the query hot dog should not retrieve documents that only contain the words hot and dog individually, outside of the phrase hot dog. In this study, we focus on noun phrases in the physics domain. For specialized domains such as physics, adaptable and reliable MWU recognition is of particular importance because comprehensive and up-to-date lists of MWUs are not available and would have to be created by hand. We chose noun phrases because domain-specific terminology is commonly encoded in noun phrase MWUs; other types of phrases – e.g., verb constructions – rarely give rise to fixed domain-specific multi-word se- quences that should be treated as a unit. We cast the task of MWU tokenization as seman- tic head recognition in this paper. The importance of syntactic heads for many NLP tasks is generally ac- cepted. For example, in coreference resolution iden- tity of syntactic heads is predictive of coreference; in parse disambiguation, the syntactic head of a noun phrase is a …

Content truncated to 500 words