Relational Features in Fine-Grained Opinion Analysis

摘要: Fine-grained opinion analysis methods often make use of linguistic features but typically do not take the interaction between opinions into account. This article describes a set experiments that demonstrate relational features, mainly derived from dependency-syntactic and semantic role structures, can significantly improve performance automatic systems for number fine-grained tasks: marking up expressions, finding holders, determining polarities expressions. These it possible to model way expressed in natural-language discourse interact sentence over arbitrary distances. The relations requires us consider multiple simultaneously, which makes search optimal intractable. However, reranker be used as sufficiently accurate efficient approximation. A feature sets machine learning approaches rerankers are evaluated. For task expression extraction, best shows 10-point absolute improvement soft recall on MPQA corpus conventional sequence labeler based local contextual while precision decreases only slightly. Significant improvements also seen extended tasks where holders considered: 10 7 points recall, respectively. In addition, outperform previously published results unlabeled (6 F-measure points) polarity-labeled (10–15 extraction. Finally, an extrinsic evaluation, extracted MPQA-style expressions practical mining tasks. all scenarios considered, lead statistically significant improvements.