DeInStance: Creating and Evaluating a German Corpus for Fine-Grained Inferred Stance Detection

DeInStance: Creating and Evaluating a German Corpus for Fine-Grained Inferred Stance Detection Anne G¨ohring, Manfred Klenner, Sophia Conrad Department of Computational Linguistics University of Zurich, Switzerland goehring,klenner,conrad@cl.uzh.ch Abstract We introduce deInStance, a corpus of 1000 politicians’ answers in German (de) contain- ing sentences labeled with explicitly expressed and inferred stances - pro and con relations - by 3 annotators. They achieved an accept- able inter-rater agreement given the inherent subjective nature of the task. A first base- line, a fine-tuned BERT-based token classifier, achieved F1-scores of around 70% . Our focus is on the difficult subclass of sentences com- prising only non-polar words, but still with an (implicit) pro or con perspective of the writer. 1 Introduction When people are asked about their position with regard to a certain topic, they typically answer by elaborating an argumentation in favor or against this topic. Argument mining is concerned with the structure of such arguments and the classification of each part. This happens at the clause level: a clause might be a claim, the support of a claim, etc. But what about the entities and events within the different parts of a clause? What can the reader infer from the writer’s perspective on these differ- ent subtopics contained in the author’s text? In this work, we present a new resource together with a first neural sequence labeling model of such in- ferred fine-grained stances. The goal is to find all those entities (called targets) in a text that the writer approves (pro) or disapproves (con), explicitly or implicitly. These targets might be aspects of the overall topic of the text, but also any entity men- tioned in the text towards which the writer seems to bear a positive or negative attitude. Among these non-aspect targets are entities reflecting the writer’s moral convictions, political views, and all sorts of other preferences. 2 Corpus Annotation As a starting point, we took the German part of the freely available xstance1 corpus. The original corpus contains politicians’ stances consisting of an explicit position (from strongly/weakly against to weakly/strongly in favor) together with a comment as answer to given questions from different topics. We annotated a subcorpus2 of 1000 answers where each word receives a label: pro (in favour of), con (against) or none (neutral). A pro relation indicates that the writer approves (i.e. is in favor of) the denoted entity or event; correspondingly for con.3 2.1 Objective of the Annotation In order to clarify the annotation task and to show the differences to aspect-based sentiment analy- sis, take the following question: Do you support the introduction of minimum wage for employees? One answer is: Another unhelpful blanket proposal from the mothballs of socialism, which would fur- ther weaken our country’s competitiveness. Social- ism (text author is against it) and the competitive- ness (author is in favour) are somehow related to the question, but they are not aspects in the sense of aspect-based sentiment analysis. Aspects are strongly correlated categories of an item (e.g. the price of …

Content truncated to 500 words