Algoritmi e Risorse per l'analisi del Linguaggio Concreto e Figurato: il Caso delle Metafore

A metaphor is a conceptual device used to quickly deliver some abstract concept in term of something else; the basic underlying mechanism is that a difficult concept is explained by resorting to a simpler one; one that typ- ically belongs to a more physical experience of the real world. For example in \My car drinks gasoline", the operation of a car, of which one as only an indirect experience, is expressed like eating, more understandable by a person. Metaphors are a challenging obstacle in many sectors of natural language processing; being so widespread and important for the compre- hension of concepts and their communication, however, their identification and interpretation is of the utmost relevance for any automated system that has to cope with Natural Language. Abstractness of the concepts involved in metaphors seems to play a major role in metaphors creation and under- standing. By contrast, past work on word abstractness has been limited by the necessity of involving a great number of human annotators even for restricted word sets; moreover, most of the previous experiments in this sense did not attempt to disambiguate among the various senses that words can possibly have. Given the recent creation of a large lexical resource for capturing common-sense knowledge, COVER, we propose an algo- rithm for the automatic abstractness annotation of word senses. A twofold evaluation has been designed: an intrinsic one, aimed at investigating how our abstractness scores correlate with human judgement; and an extrinsic one, where the abstractness scores added to the concepts in the COVER lexical resource are used by a novel algorithm for metaphor detection. Both steps are conducted on international data sets, well established in literature.