Information Spreading in Expanding Wordnet Hypernymy Structure

Proceedings of Recent Advances in Natural Language Processing, pages 553–561, Hissar, Bulgaria, 7-13 September 2013. Information Spreading in Expanding Wordnet Hypernymy Structure Maciej Piasecki Institute of Informatics Wrocław Univ. of Technology maciej.piasecki@pwr.wroc.pl Radosław Ramocki Institute of Informatics Wrocław Univ. of Technology rramocki@gmail.com Michał Kali´nski Institute of Informatics Wrocław Univ. of Technology 168023@student.pwr.wroc.pl Abstract The paper presents a wordnet expan- sion algorithm, which is based on lexico- semantic relations extracted from large text corpora. We do not assume that the extracted relation instances (i.e. word pairs) are described by probabilities. Thus, results produced by any method, includ- ing pattern-based and Distributional Se- mantics approaches can be used. The algo- rithm is based on a general spreading acti- vation model. Support for word-to-word semantic associations is first mapped on the existing wordnet structure. Next, the support is spread over the wordnet net- work in order to find attachment areas for a new word. Evaluation and comparison with other approaches in experiments on Princeton WordNet 3.0 is presented. 1 Introduction Wordnets became important large scale language resources providing relational description of lex- ical meanings. e.g. WordNet (Fellbaum, 1998), GermaNet (Hamp and Feldweg, 1997) or plWord- Net (Maziarz et al., 2012). The required large amount of work on wordnet construction can be lessened by supporting manual work with auto- mated tools for the extraction of lexico-semantic relations and wordnet expansion. A scheme for lexico-semantic network extraction from corpus includes, e.g. (Yang and Callan, 2009; Navigli et al., 2011): term extraction, extraction of term as- sociations and taxonomy induction. A taxonomy structure is mostly a subset of the whole word- net hyper/hyponynymy structure. Thus, a more general task, for the last phase, is extraction of lexico-semantic relations (sensu stricto), called re- lation formation in (Yang and Callan, 2009). In our work, we focus on the automated expansion of a such wordnet hypernymy structure. Upper levels of a wordnet hypernymy describe more general, often highly abstract lexical units (i.e. pairs: lemma and its sense). Such fine grained distinctions are hard to trace in a corpus, but mostly it is this part of a wordnet that is cre- ated first. Thus, we assumed that upper hyper- nymic levels are already built manually, and what is needed is to expand the wordnet structure to- wards the lower levels. Our goal is to develop a method of automated expansion of wordnet hy- pernymy structure based on both lexico-semantic associations extracted automatically from a large text corpus and the prior partial wordnet structure. We do not assume the existence of any kind of se- mantic annotation or document structure, to make the proposed method general. Most taxonomy induction methods use only the existing hypernymy structure as a basis for the in- cremental wordnet expansion, e.g. (Snow et al., 2006; Piasecki et al., 2009b). We explore all dif- ferent types of wordnet links to identify the appro- priate location for a new lemma sense. 2 Related works (Alfonseca and Manandhar, 2002) and (Witschel, 2005) treat wordnet hypernymy as a kind of …

Content truncated to 500 words