An Unsupervised Approach for Semantic Relation Interpretation

AbstractIn this work we propose a hybrid unsupervised approach for semantic relation extraction from Italian and English texts. The system takes as input pairs of ""distributionally similar"" terms, possibly involved in a semantic relation. To validate and label the anonymous relations holding between the terms in input, the candidate pairs of terms are looked for on the Web in the context of reliable lexico-syntactic patterns. This paper focuses on the definition of the patterns, on the measures used to assess the reliability of the suggested specific semantic relation and on the evaluation of the implemented system. So far, the system is able to extract the following types of semantic relations: hyponymy, meronymy, and co-hyponymy. The approach can however be easily extended to manage other relations by defining the appropriate battery of reliable lexico-syntactic patterns. Accuracy of the system was measured with scores of 83.3% for hyponymy, 75% for meronymy and 72.2% for co-hyponymy extraction.