Acquiring Verb Classes Through Bottom-Up Semantic Verb Clustering

Acquiring Verb Classes Through Bottom-Up Semantic Verb Clustering Olga Majewska, Diana McCarthy, Ivan Vuli´c and Anna Korhonen University of Cambridge Language Technology Lab Dept. of Theoretical & Applied Linguistics Cambridge CB3 9DA, United Kingdom om304@cam.ac.uk, diana@dianamccarthy.co.uk, iv250@cam.ac.uk, alk23@cam.ac.uk Abstract In this paper, we present the first analysis of bottom-up manual semantic clustering of verbs in three languages, English, Polish and Croatian. Verb classes including syntactic and semantic information have been shown to support many NLP tasks by allowing abstraction from individual words and thereby alleviating data sparseness. The availability of such classifications is however still non-existent or limited in most languages. While a range of automatic verb classification approaches have been proposed, high-quality resources and gold standards are needed for evaluation and to improve the performance of NLP systems. We investigate whether semantic verb classes in three different languages can be reliably obtained from native speakers without linguistics training. The analysis of inter-annotator agreement shows an encouraging degree of overlap in the classifications produced for each language individually, as well as across all three languages. Comparative examination of the resultant classifications provides interesting insights into cross-linguistic semantic commonalities and patterns of ambiguity. Keywords: verb classes, semantic clustering, multilingual NLP 1. Introduction With the recent advances in automatic lexical acquisition, the need for high-quality evaluation resources is ever grow- ing. Due to the pivotal role played by verbs in sentence structure, the problem of creation of verbal classifications has attracted a lot of attention in natural language process- ing (NLP). Different approaches to creation of verbal clas- sifications have been proposed, varying with regard to the guiding criteria by which the class architecture is organ- ised, prioritising semantic (WordNet (Miller, 1995; Fell- baum, 1998), FrameNet (Baker et al., 1998), PropBank (Palmer et al., 2005)) or syntactic information (COMLEX (Grishman et al., 1994), VALEX (Korhonen et al., 2006)), or combining the two (Levin, 1993; Kipper et al., 2000; Kipper Schuler, 2005). Kipper Schuler’s (2005) Verb- Net, grouping English verbs into classes defined by shared meaning components and syntactic behaviour, is one of the richest lexical verb resources currently available, and its utility in various NLP applications has been repeatedly demonstrated (Rios et al., 2011; Windisch Brown et al., 2011; Schmitz et al., 2012; Lippincott et al., 2013; Bailey et al., 2015). However, creation of a similar resource from scratch, draw- ing simultaneously on semantic and syntactic criteria, is a challenging and time-consuming task when attempted by annotators without theoretical linguistics background (Ma- jewska et al., 2017). A number of approaches to automatic verb classification have been proposed (Joanis et al., 2008; Sun et al., 2010; Falk et al., 2012; Kawahara et al., 2014; Scarton et al., 2014; Peterson et al., 2016; Vuli´c et al., 2017), allowing to minimise the time required and elimi- nate the need to employ trained lexicographers. However, evaluation of such systems relies on the availability of gold standard classes, and these are still lacking for a great ma- jority of languages. In light of these challenges and the …

Content truncated to 500 words