Inductive and deductive inferences in a Crowdsourced Lexical-Semantic Network (Inférences déductives et réconciliation dans un réseau lexico-sémantique) [in French])

In Computational Linguistics, building lexical-semantic networks and validating contained relations are paramount issues as well as adding some reasoning skills in order to enrich these knowledge bases. In this paper we devise an inference engine which aims at producing new potential relations from already existing ones in the JeuxDeMots network. This network is constructed with the help of a GWAP (game with a purpose) thanks to thousands of players. It handles terms and weighted relations between these terms, and currently contains over 2 million relation occurences. Polysemous terms may be refined in several senses (bank may be a bank>financial institution or a bank>river) but as the network is indefinitely under construction (in the context of a Never Ending Learning approach) some senses may be missing at a given time. The approach we proposed here is founded on the triangulation method through two kinds of inference schemes: deduction (topdown from generic to specific terms) and induction (bottom-up from specific to generic terms). A blocking mechanism, whose purpose is to avoid proposing highly dubious new relations, is based on logical and statistical constraints. Automatically inferred relations are then proposed to human contributors to be validated. In case of invalidation, a reconciliation dialog is undertaken to identify the cause of the wrong inference: an exception, an error in the premises or a previously undetected confusion due to polysemy on the central term common to both premises.