Automatic Extraction of Fixed Multiword Expressions

Automatic Extraction of Fixed Multiword Expressions Campbell Hore⋆, Masayuki Asahara, and Y¯uji Matsumoto Graduate School of Information Science, Nara Institute of Science and Technology, 8916-5 Takayama, Ikoma, Nara 630-0192, Japan {campbe-h, masayu-a, matsu}@is.naist.jp http://cl.naist.jp Abstract. Fixed multiword expressions are strings of words which to- gether behave like a single word. This research establishes a method for the automatic extraction of such expressions. Our method involves three stages. In the first, a statistical measure is used to extract candidate bi- grams. In the second, we use this list to select occurrences of candidate expressions in a corpus, together with their surrounding contexts. These examples are used as training data for supervised machine learning, re- sulting in a classifier which can identify target multiword expressions. The final stage is the estimation of the part of speech of each extracted expression based on its context of occurence. Evaluation demonstrated that collocation measures alone are not effective in identifying target ex- pressions. However, when trained on one million examples, the classifier identified target multiword expressions with precision greater than 90%. Part of speech estimation had precision and recall of over 95%. 1 Introduction 1.1 Multiword Expressions For natural language processing purposes, a naive definition of a word in English is ‘a sequence of letters delimited by spaces’. By this definition, the expression ad hoc, which originally came from Latin, consists of two “words”, ad and hoc. However, in isolation hoc is not a meaningful English word. It is always preceded by ad. This suggests that treating these two words as if they together form a single “word with spaces” more closely models their behaviour in text. A sequence of words which for one reason or another is more sensibly treated as a single lexical item, rather than as individual words, is known as a multiword expression (MWE). In other words, an MWE is a sequence of words which together behave as though they were a single word. MWEs are not limited to imported foreign phrases such as ad hoc. They cover a large range of expression types including proper nouns such as New York, verb- particle constructions such as to call up (i.e. to telephone someone), and light ⋆ Supported by the Japanese government’s MEXT scholarship programme. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 565–575, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 566 C. Hore, M. Asahara, and Y. Matsumoto verbs such as to make a mistake. The justification for treating such expressions as MWEs is that their linguistic properties are odd in some way as compared to “normal” expressions: either their part of speech or their meaning is unpre- dictable despite full knowledge about the parts of speech or meanings of their constituent words. 1.2 Fixed Multiword Expressions By fixed, we mean that this particular type of MWE consists of a contiguous se- quence of words. Other MWE types can consist of discontinuous word sequences. For example, the verb-particle construction to call up takes an indirect object, the person who …

Content truncated to 500 words