Automatic Lexical Classification – Balancing between Machine Learning and Linguistics

Automatic Lexical Classification - Balancing between Machine Learning and Linguistics⋆ Anna Korhonen University of Cambridge, Computer Laboratory 15 JJ Thomson Avenue, Cambridge CB3 0GD, UK alk23@cl.cam.ac.uk Abstract. Verb classifications have been used to support a number of practical tasks and ap- plications, such as parsing, information extraction, question-answering, and machine trans- lation. However, large-scale exploitation of verb classes in real-world or domain-sensitive tasks has not been possible because existing manually built classifications are incomprehen- sive. This paper describes recent and on-going research on extending and acquiring lexical classifications automatically. The automatic approach is attractive since it is cost-effective and opens up the opportunity of learning and tuning lexical classifications for the application and domain in question. However, the development of an optimal approach is challenging, and requires not only expertise in machine learning but also a good understanding of the linguistic principles of lexical classification. Keywords: lexical-semantic classification, verb classes, automatic lexical acquisition 1 Introduction Verb classifications have attracted a great deal of interest in both linguistics and natural language processing (NLP). They have proved useful for various important tasks and applications, includ- ing e.g. computational lexicography, parsing, word sense disambiguation, semantic role labeling, information extraction, question-answering, and machine translation (Swier and Stevenson, 2004; Dang, 2004; Shi and Mihalcea, 2005; Kipper et al., 2008; Zapirain et al., 2008). Particularly useful are classes which capture generalizations over a range of (cross-)linguistic properties, such as the ones proposed by Levin (1993). Being defined in terms of similar meaning and (morpho-)syntactic behaviour of words, these classes generally incorporate a wider range of properties than e.g. classes defined solely on semantic grounds (Miller, 1995). For example, verbs which share the meaning component of ‘manner of motion’ (e.g. travel, run, walk), behave similarly in terms of subcategorization (e.g. I travelled/ran/walked, I trav- elled/ran/walked to London, I travelled/ran/walked five miles) and usually have zero-related nom- inals (e.g. a run, a walk) can be grouped to the same lexical class. Such verb classes can be identified across the entire lexicon and they can also apply across languages, since the basic mean- ing components they are comprised of are cross-linguistically applicable or overlapping. While the classes do not provide means for full semantic inferencing, they can offer a powerful tool for generalization, abstraction and prediction which is beneficial for practical tasks. Funda- mentally, the classes are a critical component of any system which needs mapping from surface realization of arguments to predicate-argument structure. As the classes capture higher level ab- stractions they can be used as a principled means to abstract away from individual words when required. For example, they can be utilized to organize a default inheritance hierarchy which ⋆ This work was funded by the Royal Society University Research Fellowship, the EPSRC grant EP/F030061/1 and the British Council Alliance grant. The author would like to thank Lin Sun for his contribution to this paper. Copyright 2009 by Anna Korhonen 19 23rd Pacific Asia Conference on Language, Information and Computation, pages 19–28 effectively captures generalizations over words …

Content truncated to 500 words