Inferring Knowledge with Word Refinements in a Crowdsourced Lexical-Semantic Network

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 365–375, Dublin, Ireland, August 23-29 2014. Inferring Knowledge with Word Refinements in a Crowdsourced Lexical-Semantic Network Manel Zarrouk UM2-LIRMM 161 rue Ada 34095 Montpellier, FRANCE manel.zarrouk@lirmm.fr Mathieu Lafourcade UM2-LIRMM 161 rue Ada 34095 Montpellier, FRANCE mathieu.lafourcade@lirmm.fr Abstract Automatically inferring new relations from already existing ones is a way to improve the qual- ity and coverage of a lexical network and to perform error detection. In this paper, we devise such an approach for the crowdsourced JeuxDeMots lexical network and we focus especially on word refinements. We first present deduction (generic to specific) and induction (specific to generic) which are two inference schemes ontologically founded and then propose a trans- fer schema devoted to infer relations with and for word refinements. 1 Introduction Efficiently building useful resources for Computational Linguistics (CL) is of a crucial interest. Most of existing lexical-semantic networks have been built by hand (like for instance WordNet (Miller et al., 1990)) and, despite that assisting tools are generally designed for consistency checking, the task remains time consuming and costly. Fully automated approaches are generally limited to term co- occurrences as extracting precise semantic relations between terms from corpora remains at best difficult. Crowdsourcing approaches are flowering in CL especially with the advent of Amazon Me- chanical Turk or in a broader scope Wikipedia, to cite the most well known examples. WordNet is such a lexical network, constructed at great cost, based on synsets which can be roughly considered as con- cepts (Fellbaum, 1988). EuroWordnet (Vossen., 1998) a multilingual version of WordNet and WOLF (Sagot., 2008) a French version of WordNet, were built by automated crossing of the original Princeton WordNet and other lexical resources along with some more or less manual checking. Navigli (2010) constructed automatically BabelNet a large multilingual lexical network from term co-occurrences in Wikipedia. Although being very large and multilingually connected (which is tremendously usefull for machine translation, for instance) it contains few various lexical-semantic relations. An ideal lexical-semantic network contains interconnected lemmas, word forms and multi-word expressions as entry points (nodes) along with word meanings and concepts. The idea itself of word senses as forwarded in the lexicographic tradition may be debatable in the context of resources for se- mantic analysis, and we generally prefer to consider the psycholinguistic idea of word usages. A given polysemous word, as identified by locutors, has several usages that might differ substantially from word senses as classically defined. A given usage can also in turn have several deeper refinements and the whole set of usages can take the form of a decision tree. For a very classical example, bank can be related to money or river : bank ⋗ ’bank>money’ and bank ⋗ ’bank>river’. A ’bank>money’ can be distinguished as the financial institution or the actual building. In the context of a collaborative construction, such a lexical resource should be considered as being constantly evolving and a general pragmatic rule of thumb is to …

Content truncated to 500 words