Data-driven relation discovery from unstructured texts

This work proposes a data driven methodology for the extraction of subject-verb-object triplets from a text corpus. Previous works on the field solved the problem by means of complex learning algorithms requiring hand-crafted examples; our proposal completely avoids learning triplets from a dataset and is built on top of a well-known baseline algorithm designed by Delia Rusu et al.. The baseline algorithm uses only syntactic information for generating triplets and is characterized by a very low precision i.e., very few triplets are meaningful. Our idea is to integrate the semantics of the words with the aim of filtering out the wrong triplets, thus increasing the overall precision of the system. The algorithm has been tested over the Reuters Corpus and has it as shown good performance with respect to the baseline algorithm for triplet extraction.