Opinion Holder and Target Extraction based on the Induction of Verbal Categories

Proceedings of the 19th Conference on Computational Language Learning, pages 215–225, Beijing, China, July 30-31, 2015. c⃝2015 Association for Computational Linguistics Opinion Holder and Target Extraction based on the Induction of Verbal Categories Michael Wiegand Spoken Language Systems Saarland University D-66123 Saarbr¨ucken, Germany michael.wiegand@lsv.uni-saarland.de Josef Ruppenhofer Dept. of Information Science and Language Technology Hildesheim University D-31141 Hildesheim, Germany ruppenho@uni-hildesheim.de Abstract We present an approach for opinion role induction for verbal predicates. Our model rests on the assumption that opinion verbs can be divided into three different types where each type is associated with a char- acteristic mapping between semantic roles and opinion holders and targets. In sev- eral experiments, we demonstrate the rel- evance of those three categories for the task. We show that verbs can easily be categorized with semi-supervised graph- based clustering and some appropriate similarity metric. The seeds are obtained through linguistic diagnostics. We evalu- ate our approach against a new manually- compiled opinion role lexicon and perform in-context classification. 1 Introduction While there has been much research in senti- ment analysis on subjectivity detection and po- larity classification, there has been less work on the extraction of opinion roles, i.e. entities that express an opinion (opinion holders), and enti- ties or propositions at which sentiment is directed (opinion targets). Previous research relies on large amounts of labeled training data or leverages gen- eral semantic resources which are expensive to construct, e.g. FrameNet (Baker et al., 1998). In this paper, we present an approach to induce opinion roles of verbal predicates. The input is a set of opinion verbs that can be found in a com- mon sentiment lexicon. Our model rests on the assumption that those verbs can be divided into three different types. Each type has a character- istic mapping between semantic roles and opinion holders and targets. Thus, the problem of opinion role induction is reduced to automatically catego- rizing opinion verbs. We frame the task of opinion role extraction as a triple (pred, const, role) where pred is a predi- cate evoking an opinion (we exclusively focus on opinion verbs), const is some constituent bearing a semantic role assigned by pred, and role is the opinion role that is assigned to const. Our work assumes the knowledge of opinion words. We do not cover polarity classification. Many lexicons with that kind of information al- ready exist. Our sole interest is the assignment of opinion holder and target given some opinion verb. There does not exist any publicly available lexical resource specially designed for this task. For the induction of opinion verb types, we con- sider semi-supervised graph clustering with some appropriate similarity metric. We also propose an effective method for deriving seeds automatically by applying some linguistic diagnostics. Our approach is evaluated in a supervised learn- ing scenario on a set of sentences with annotated opinion holders and targets. We employ differ- ent kinds of features, including features derived from a semantic parser based on FrameNet. We also compare our proposed model based …

Content truncated to 500 words