Supervised distributional hypernym discovery via domain adaptation

Comunicació presentada a la Conference on Empirical Methods in Natural Language Processing celebrada els dies 1 a 5 de novembre de 2016 a Austin, Texas.Lexical taxonomies are graph-like hierarchical structures that provide a formal representation of knowledge. Most knowledge graphs to date rely on is-a (hypernymic) relations as the backbone of their semantic structure. In this paper, we propose a supervised distributional framework for hypernym discovery which operates at the sense level, enabling large-scale automatic acquisition of disambiguated taxonomies. By exploiting semantic regularities between hyponyms and hypernyms in embeddings spaces, and integrating a domain clustering algorithm, our model becomes sensitive to the target data. We evaluate several configurations of our approach, training with information derived from a manually created knowledge base, along with hypernymic relations obtained from Open Information Extraction systems. The integration of both sources of knowledge yields the best overall results according to both automatic and manual evaluation on ten different domains.This work is partially funded by the Spanish Ministry of Economy and Competitiveness under the Maria de Maeztu Units of Excellence Programme (MDM-2015-0502) and under the TUNER project (TIN2015-65308-C5-5-R, MINECO/FEDER, UE). The authors also acknowledge support from Dr. Inventor (FP7-ICT-2013.8.1611383). José CamachoCollados is supported by a Google Doctoral Fellowship in Natural Language Processing