In this paper, we present our approach for Multilingual Open Information Extraction. Our sequence labeling based approach builds only on Universal Dependency representation to capture OpenIE’s regularities and to perform Cross-lingual Multilingual OpenIE. We propose a new two-stage pipeline model for sequence labeling, that first identifies all the arguments of the relation and only then classifies them according to their most likely label. This paper also introduces a new benchmark evaluation for French. Experimental Evaluation shows that our approach achieves the best results in the available Benchmarks (English, French, Spanish and Portuguese).

The JeuxDeMots project aims at building a very large knowledge base in French, both common sense and specialized, using games, contributory approaches, and inference mechanisms. A dozen games have been designed as part of this project, each one allowing to collect specific information, or to consolidate the information acquired through the other games. With this presentation, the data collected and constructed since the launch of the project in the summer of 2007 will be analyzed both qualitatively and quantitatively. In particular, the following aspects will be detailed: the structure of the lexical and semantic network, some types of relations (semantic, ontological, subjective, semantic roles, associations of ideas), annotation of relations (meta-information), semantic refinements (management of polysemy), the creation of clusters allowing the representation of richer knowledge (n-argument relations) that make an implicit neural network. Finally, I will describe some complementary acquisition methods and applications such as a bot for endogenous contributions, a chatbot making inferences and semantic extraction from texts.