Subcategorisation Acquisition from Raw Text for a Free Word-Order Language

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 298–307, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Subcategorisation Acquisition from Raw Text for a Free Word-Order Language Will Roberts and Markus Egg and Valia Kordoni Institute f¨ur Anglistik und Amerikanistik, Humboldt University 10099 Berlin, Germany {will.roberts,markus.egg,evangelia.kordoni}@anglistik.hu-berlin.de Abstract We describe a state-of-the-art automatic system that can acquire subcategorisation frames from raw text for a free word-order language. We use it to construct a subcate- gorisation lexicon of German verbs from a large Web page corpus. With an automatic verb classification paradigm we evaluate our subcategorisation lexicon against a pre- vious classification of German verbs; the lexicon produced by our system performs better than the best previous results. 1 Introduction We introduce a state-of-the-art system for the ac- quisition of subcategorisation frames (SCFs) from large corpora, which can deal with languages with very free word order. The concrete language we treat is German; its word order variability is illus- trated in (1)–(4), all of which express the sentence The man gave the old dog a chop: (1) Dem alten Hund gab der Mann ein Schnitzel. (2) Ein Schnitzel gab dem alten Hund der Mann. (3) Ein Schnitzel gab der Mann dem alten Hund. (4) Der Mann gab dem alten Hund ein Schnitzel. On the basis of raw text, the system can be used to build extensive SCF lexicons for German verbs. Subcategorisation means that lexical items require specific obligatory concomitants or argu- ments; we focus on verb subcategorisation. E.g., the verb geben ‘give’ requires three arguments, the nominative subject der Mann ‘the man’, the dative indirect object dem alten Hund ‘the old dog’, and the accusative direct object ein Schnitzel ‘a chop’. Other syntactic items may be subcategorised for, too, e.g. both stellen and its English translation put subcategorise for subject, direct object, and a prepositional phrase (PP) like on the shelf: (5) [NP Al] put [NP the book] [PP on the shelf]. Subcategorisation frames describe a combina- tion of arguments required by a specific verb. The set of SCFs for a verb is called its subcategori- sation preference. Our system follows much pre- vious work by counting PPs that accompany the verb among its complements, even though they are not obligatory (so-called ‘adjuncts’), because PP adjuncts are excellent clues to a verb’s semantics (Sun et al., 2008). However, nominal and clausal adjuncts do not count as verbal complements. SCF information can benefit all applications that need information on predicate-argument struc- ture, e.g., parsing, verb clustering, semantic role la- belling, or machine translation. Automatic acquisi- tion of SCF information with minimal supervision is also crucial to construct useful resources quickly. The main innovation of the presented new sys- tem is to address two challenges simultaneously, viz., SCF acquisition from raw text and the focus on languages with a very free word order. With this system, we create an SCF lexicon for German verbs and evaluate this lexicon against a previously published manual …

Content truncated to 500 words