Bridging Languages by SuperSense Entity Tagging

Abstract

This paper explores a very basic linguistic phenomenon in multilingualism: the lexicalizations of entities are very often identical within different languages while concepts are usually lexicalized differently. Since entities are commonly referred to by proper names in natural language, we measured their distribution in the lexical overlap of the terminologies extracted from comparable corpora. Results show that the lexical overlap is mostly composed by unambiguous words, which can be regarded as anchors to bridge languages: most of terms having the same spelling refer exactly to the same entities. Thanks to this important feature of Named Entities, we developed a multilingual super sense tagging system capable to distinguish between concepts and individuals. Individuals adopted for training have been extracted both by YAGO and by a heuristic procedure. The general F1 of the English tagger is over 76%, which is in line with the state of the art on super sense tagging while augmenting the number of classes. Performances for Italian are slightly lower, while ensuring a reasonable accuracy level which is capable to show effective results for knowledge acquisition.