Sentiframes: A Resource for Verb-centered German Sentiment Inference

Sentiframes: A Resource for Verb-centered German Sentiment Inference Manfred Klenner, Michael Amsler Institute of Computational Linguistics University of Zurich Binzm¨uhlestrasse 14, 8050 Zurich, Switzerland klenner@cl.uzh.ch, mamsler@cl.uzh.ch Abstract In this paper, a German verb resource for verb-centered sentiment inference is introduced and evaluated. Our model specifies verb polarity frames that capture the polarity effects on the fillers of the verb’s arguments given a sentence with that verb frame. Verb signatures and selectional restrictions are also part of the model. An algorithm to apply the verb resource to treebank sentences and the results of our first evaluation are discussed. Keywords: German Sentiment Analysis, verb-centered, Sentiframes 1. Introduction Some verbs impose a positive or negative polarity on the fillers of some of their semantic roles. Often, an opinion holder is present in such a sentence as well. Then, the opin- ion holder is understood to have a positive or negative at- titude towards the opinion target - as in Europe [holder] criticizes Russia [target]. Sometimes, we do not know the opinion holder, e.g., in passive constructions as he was crit- icized. Moreover, it could be argued that some verbs such to win or to survive not even implicitly refer to an opinion holder and nevertheless imply a target polarity. The relation between holder and target, thus, is not the most general per- spective on these verbs - it is a special case. We have cre- ated Sentiframes1, a resource of verb effects that does not specify holder-target relations, but verb-specific polar con- notations (from which such a relational view could easily be derived). We have created 471 so-called verb polarity frames for 251 German verbs. A verb polarity frame is a subcategoriza- tion frame that specifies for each verb role whether a polar- ity effect is triggered given that the appropriate filler object occupies the role. For instance, the direct object of to criti- cize receives a negative effect, while the direct object of to admire gets a positive effect. Also, subclauses might get an effect. Again, to criticize that assigns a negative effect to the subclause. This (in principle) supports inferences in the style of Deng and Wiebe (2014), where the relation be- tween an opinion holder of the matrix verb and an actor of a subclause can be induced. We have not created such a system, yet. But the resource is meant to serve as the lex- ical basis of such an approach. A matrix verb might form an intensional context, e.g., in I hope that my paper gets accepted nothing positive or negative can be derived: my paper should not receive a positive effect in such a context. In order to prevent such invalid inferences, we have spec- ified for each verb that subcategorizes for a subclause its verb signature in the sense from Nairn et al. (2006). This includes the behaviour under negation as well. Finally, we found that verb sense disambiguation - although in princi- ple needed - often could be avoided by means of shallow 1It was …

Content truncated to 500 words