Finding Opinion Targets in News Comments and Book Reviews

摘要

In this paper we approach the extraction of opinion targets in user-generated texts written in Portuguese in two different domains: political news articles and books. The former was represented by SentiCorpus-PT, and the latter by ReLi. We implemented and tested five prototypes for extracting explicit opinion targets from the corpora: one based on Centering, one based on morphosyntactic patterns, and three based on syntactic and morphosyntactic heuristics, which were used as baselines. Experimental results proved target extraction on ReLi to be a more difficult task than on SentiCorpus-PT, probably due to its low proportion of targets per sentence. Also, the baseline based on proper nouns performed better on SentiCorpus-PT, showing that superficial heuristics may yield good results in this domain.