Aspectual Flexibility Increases with Agentivity and ConcretenessA Computational Classification Experiment on Polysemous Verbs

Aspectual Flexibility Increases with Agentivity and Concreteness A Computational Classification Experiment on Polysemous Verbs Ingrid Falk and Fabienne Martin Universität Stuttgart first.second@ling.uni-stuttgart.de Abstract We present an experimental study making use of a machine learning approach to identify the factors that affect the aspectual value that characterises polysemous verbs under each of their readings. The study is based on various morpho-syntactic and semantic features collected from a French lexical resource and on a gold standard aspectual classification of verb readings designed by an expert. Our results support the tested hypothesis, namely that agentivity and abstractness influence lexical aspect. Keywords: lexical aspect; aspectual flexibility; French lexicon 1. Introduction It is generally accepted that the event type of a sentence and in particular its aspectual value result from a complex inter- play between lexical features of the predicate on one hand and its linguistic context on the other. In previous work by Siegel and McKeown (2000), Zarcone and Lenci (2008) and Friedrich and Palmer (2014), the event type of corpus utterances is predicted on the basis of syntactic and seman- tic features of the main predicate and the linguistic con- text thereof. In contrast, in this study, we investigate lexi- cal (inner) aspect through a French lexical resource, “Les Verbes Français” (François et al., 2007), a valence lexi- con of French verbs providing a detailed morpho-syntactic and semantic description for each reading (use) of a verb. Since the basic lexical units are the readings of the predi- cate rather than the (ambiguous) verbs, this resource allows one to study lexical aspect on an intermediate level between the coarse-grained predicate level and the very fine-grained corpus utterance level. Thanks to its detailed syntactic and semantic description, it also enables us to investigate the effect of particular complex semantic dimensions, as e.g. agentivity or abstractness, on the lexical aspect of verbs across their different readings. 2. Outline In this work, we focus on the way the lexical aspect of a same verb varies across its different readings, using the lex- ical resource “Les Verbes Français” (François et al. (2007), henceforth LVF). We address this question through a su- pervised classification task. For this, we first extracted ver- bal readings from the LVF for a set of 167 verbs, cho- sen in such a way that each of the four Vendlerian aspec- tual classes are roughly equally represented. We manually annotated each of these readings on the basis of a fine- grained aspectual classification based on eight aspectual classes (see below for a rough description). This manual annotation provides the gold standard for our classification experiments. For each annotated reading, we then auto- matically collected morpho-syntactic and semantic features from the LVF. Based on these features, we trained clas- sifiers to automatically predict the aspectual class of the (manually annotated) readings. We then evaluated the dif- ferent classifiers used by comparing their accuracy in 10- fold cross validation on the basis of the gold annotation. We conducted experiments with three feature sets and their combinations. The …

Content truncated to 500 words