A Unified Method for Extracting Simple and Multiword Verbs with Valence Information and Application for Hungarian

International Conference RANLP 2009 - Borovets, Bulgaria, pages 399–403 A Uni ed Method for Extracting Simple and Multiword Verbs with Valence Information and Application for Hungarian Bálint Sass Research Institute for Linguistics, Hungarian Academy of Sciences, Budapest, Hungary sass.balint@nytud.hu Abstract We present a method for extracting verb- centered constructions (VCCs) from corpora. In our framework, simple and multiword verbs, with or without valence are all VCCs. They are treated uniformly, from e.g. to breathe till e.g. to take something into consideration. In order to extract VCCs we represent the corpus as a sequence of clauses that contain a verb together with all its NP dependents. The method is a gen- eralization of a former subcategorization frame extraction method. It is based on cumulative counting of frequent subframes: small frequency counts are inherited to one of the longest avail- able subframes using random selection. The method nds out automatically the number of elements in VCCs; and it detects automatically whether a content word is integral part of the VCC (forming a multiword verb), or just the verb-dependent relation is important (forming a valence slot of the verb). Signi cance of our method lies in its capability to deal with multi- word verbs and (their) valence simultaneously. The paper includes evaluation for Hungarian, we obtain precision values above 80% using n- best lists evaluation. The representation and the method is in essence language independent, it could be applied to other languages as well. 1 Introduction Multiword expressions (MWEs) consist of several words, but semantically act as one unit, having non- compositional (idiomatic) meaning [12, 9]. Their meaning cannot be deduced, although the meaning of each part is known. Nevertheless, it is necessary to know their meanings if we want to deal with seman- tics in any eld of natural language processing. Be- ing a borderline case between grammar and lexicon, importance of MWES was underestimated until quite recently [12]. In fact, number of MWEs is large, one fth of all verbs can be part of a MWE in runnig text [6]. In NLP applications MWEs are usually stored in a lexical resource together with their meaning, thus the main task (called lexical acquisition) is to build up such a lexicon. The traditional collocation-based approach of collecting/extracting MWEs is based on the fact that words in MWEs appear more frequently together than expected. The strength of association between these words can be measured using particular statistical association measures [3]. Most of them are worked out to handle exactly two words (bigrams), but this is too limiting, because there are longer MWEs, obviously, and there are cases when we do not even know the number of words in the MWEs beforehand. Conventional classes ot MWEs [12, 9, 6], which can be located along a scale from most idiomatic to most literal meaning are shown below, with examples: 1. fully rigid expressions  ad hoc; 2. idioms  kick the bucket; 3. verb particle constructions (VPCs)  hand in; 4. support verb constructions (SVCs)  …

Content truncated to 500 words