Pattern and semantic analysis to improve unsupervised techniques for opinion target identification

摘要

This research employs patterns and semantic analysis to improve the existingunsupervised opinion targets extraction technique. Two steps are employed to identifyopinion targets: candidate selection and opinion targets selection. For candidateselection; a combined lexical based syntactic pattern is identified. For opinion targetsselection, a hybrid approach that combines the existing likelihood ratio test techniquewith semantic based relatedness is proposed. The existing approach basically extractsfrequently observed targets in text. However, analysis shows that not all target featuresoccur frequently in the texts. Hence the hybrid technique is proposed to extractboth frequent and infrequent targets. The proposed algorithm employs incrementalapproach to improve the performance of existing unsupervised mining of featuresby extracting infrequent features through semantic relatedness with frequent featuresbased on lexical dictionary. Empirical results show that the hybrid technique withcombined patterns outperforms the existing techniques.