Automated Verb Sense Labelling Based on Linked Lexical Resources

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 68–77, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Automated Verb Sense Labelling Based on Linked Lexical Resources Kostadin Cholakov1 Judith Eckle-Kohler2,3 Iryna Gurevych2,3 1 Humboldt-Universit¨at zu Berlin, kostadin.cholakov@anglistik.hu-berlin.de 2 Ubiquitous Knowledge Processing Lab (UKP-TUDA) Dept. of Computer Science, Technische Universit¨at Darmstadt 3 Ubiquitous Knowledge Processing Lab (UKP-DIPF) German Institute for Educational Research and Educational Information http://www.ukp.tu-darmstadt.de Abstract We present a novel approach for creat- ing sense annotated corpora automatically. Our approach employs shallow syntactico- semantic patterns derived from linked lex- ical resources to automatically identify in- stances of word senses in text corpora. We evaluate our labelling method intrinsically on SemCor and extrinsically by using au- tomatically labelled corpus text to train a classifier for verb sense disambiguation. Testing this classifier on verbs from the English MASC corpus and on verbs from the Senseval-3 all-words disambiguation task shows that it matches the performance of a classifier which has been trained on manually annotated data. 1 Introduction Sense annotated corpora are important resources in NLP as they can be used as training data (e.g., for word sense disambiguation (WSD) or semantic role labelling) or as sources for the acquisition of lexical information (e.g., selectional preference in- formation). Typically, a particular sense inventory from a lexical resource is used to annotate some or all words with word senses from this sense inven- tory. For instance, various sense-annotated cor- pora based on WordNet (WN; (Fellbaum, 1998)) exist, such as the data from the Senseval competi- tions,1 or the SemCor corpus.2 Such corpora are usually created manually which is expensive and time consuming. Furthermore, the corpora are of- ten domain specific (e.g. newspaper texts) which makes statistical systems trained on them strongly biased. We present a novel approach for creating sense annotated corpora automatically. Our approach 1http://www.senseval.org 2http://www.cse.unt.edu/˜rada/ downloads.html#semcor employs shallow syntactico-semantic patterns de- rived from linked lexical resources (LLRs) to auto- matically identify instances of word senses in text corpora. We significantly extend previous work on this task by making two important contributions: (i) we employ a large-scale LLR for automatically creating sense annotated data and (ii) we perform meaningful intrinsic and application-based eval- uations of our method on large sense annotated datasets. LLRs are the result of integrating several lexical-semantic resources by linking them at the word sense level. Examples of large LLRs are the multilingual BabelNet (Navigli and Ponzetto, 2012), an integration of wordnets and Wikipedia3, or UBY, (Gurevych et al., 2012), the resource we employ in our work here. UBY is an integration of multiple resources, such as wordnets, Wikipedia, Wiktionary (WKT)4, FrameNet (FN; (Baker et al., 1998)) and VerbNet (VN; (Kipper et al., 2008)) for English and German. A distinguishing feature of LLRs is the enriched sense representation for word senses that are in- terlinked since different resources provide differ- ent, often complementary information. Annotat- ing corpora with such enriched sense representa- tions turns them into versatile training data for …

Content truncated to 500 words