Semantic Role Labeling for Sentiment Inference: A Case Study

Semantic Role Labeling for Sentiment Inference: A Case Study Manfred Klenner, Anne Göhring Department of Computational Linguistics University of Zurich {klenner,goehring}@cl.uzh.ch Abstract In this paper, we evaluate in a case study whether semantic role labelling (SRL) can be reliably used for verb-based sentiment infer- ence (SI). SI strives to identify polar relations (against, in-favour-of) between discourse enti- ties. We took 300 sentences with 10 different verbs that show verb alternations or are am- biguous in order to find out if current SRL sys- tems actually can assign the correct semantic roles and find the correct underlying predicates. Since in SI each verb reading comes with a par- ticular polar profile, SRL is useful only if its analyses are consistent and reliable. We found that this is not (yet) given for German. 1 Introduction Sentiment Inference (SI) is the task of predicting opponents and proponents given a text. SI reveals how the writer conceptualises the world and how she perceives the discourse entities she refers to. Take for instance the sentence This government cheats the world. The writer tries to convey that the government is against the world and that it is - in the perspective of the writer - a negative actor and the world is the victim, which means that there is a negative effect on the world. We, thus, can talk about positive and negative actors, positive and negative effects, about negative (opponents) and positive (proponents) relations. We call these specifications the polar profile of a verb. In (Klenner et al., 2017), we introduced a verb- based SI system that uses dependency labels in order to express such polar profiles. For instance, the subject of the verb cheat - if used in a factual sentence - is identified as indicating a negative ac- tor, the filler of the direct object receives a negative effect, and a negative relation (against) between the two is casted. Even after normalization of de- pendency trees, e.g. by resolving passive voice, some problems remain, namely verb alternations and verb ambiguity. It certainly will lead to false analyses. Verb alternation, among others, is given if a semantic role changes its syntactic host. As an example of an instrument-subject verb alternation, compare The police man killed the aggressor with a knife versus The knife killed the aggressor. For a dependency-based approach the police man and the knife are both the subjects although the police man is the agent and the knife is the instrument. There should be a negative polar relation between police man and aggressor, but not between knife and ag- gressor (a knife cannot be against somebody). If SRL was used instead of dependency parsing, the agent role would indicate the against relation while the instrument role would block such an inference1 and thus might be a means to provide a general solution to this problem. SRL could also be useful for verb sense disam- biguation. Part of SRL is a step called predicate identification (Conia et al., 2021b), where a verb is …

Content truncated to 500 words