Detecting Optional Arguments of Verbs

Detecting optional arguments of verbs Andr´as Kornai, D´avid M´ark Nemeskey G´abor Recski HAS Computer and Automation Research Institute HAS Research Institute for Linguistics H-1111 Kende u 13-17, Budapest H-1068 Benczur u 33, Budapest {kornai,nemeskey}@sztaki.hu recski@mokk.bme.hu Abstract We propose a novel method for detecting optional arguments of Hungarian verbs using only positive data. We introduce a custom variant of collexeme analysis that explicitly models the noise in verb frames. Our method is, for the most part, unsupervised: we use the spectral clustering algorithm described in Brew and Schulte in Walde (2002) to build a noise model from a short, manually verified seed list of verbs. We experimented with both raw count- and context-based clusterings and found their performance almost identical. The code for our algorithm and the frame list are freely available at http://hlt.bme.hu/en/resources/tade. Keywords: optional arguments, spectral clustering, collexeme analysis 1. Introduction One of the classical puzzles in linguistics is to make the distinction between obligatory verbal arguments as in John likes broccoli which cannot be omitted (*John likes) and optional arguments as in John eats broccoli where removal of the argument results in a sentence that is less informative (John eats) but still grammatical. There is nothing in the semantics that would make this distinction obvious: there can be no act of liking that doesn’t involve liking some- thing just as there can be no act of eating that doesn’t in- volve eating something. Note that the lack of bare like is negative evidence, not directly present in primary linguis- tic data. Even in large corpora, the empirical frequency of very rare but grammatical constructions is zero, making them indistinguishable from constructions that are ungram- matical, an observation that has led many linguists (starting with Chomsky 1957) to the conclusion that one must rely on introspection to tell the two apart. Stefanowitsch (2006) uses 2x2 contingency tables and the standard Fisher exact test to show that negative evidence can be meaningfully extracted from large labeled corpora especially for features that are “relatively uncontroversially tagged”. Unfortunately, the recall of this method is greatly limited by the size of the gold data: for example Li and Abe (1999) or G´abor and H´eja (2007) relied on manually tagged gold corpora (the Penn and Szeged Treebanks respectively) to obtain results for 354 English (resp. 150 Hungarian) verbs. Here we extend the less supervised approach pio- neered by Korhonen (1998) for English and Sass (2010) for Hungarian: we take large and unanalyzed corpora, extract patterns by shallow parsing (Briscoe 1997), and postpro- cess the results. The standard approach is to set rather high thresholds, in effect trading in recall for acceptable preci- sion: for example Sass (2010) keeps only patterns that oc- cur over 250 times in the data, yielding 2,200 verbs and 175 frames. Strong thresholding, however, destroys the sensi- tivity to optional arguments, since lack of above-threshold evidence for the intransitive pattern is not evidence for lack of this pattern. In terms of learning an embedding, the unsupervised method would amount to …

Content truncated to 500 words