Cross-Domain Mining of Argumentative Text through Distant Supervision

Proceedings of NAACL-HLT 2016, pages 1395–1404, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Cross-Domain Mining of Argumentative Text through Distant Supervision Khalid Al-Khatib Henning Wachsmuth Matthias Hagen Jonas Köhler Benno Stein Faculty of Media, Bauhaus-Universität Weimar, Germany <firstname>.<lastname>@uni-weimar.de Abstract Argumentation mining is considered as a key technology for future search engines and au- tomated decision making. In such applica- tions, argumentative text segments have to be mined from large and diverse document col- lections. However, most existing argumenta- tion mining approaches tackle the classifica- tion of argumentativeness only for a few man- ually annotated documents from narrow do- mains and registers. This limits their practi- cal applicability. We hence propose a distant supervision approach that acquires argumen- tative text segments automatically from online debate portals. Experiments across domains and registers show that training on such a cor- pus improves the effectiveness and robustness of mining argumentative text. We freely pro- vide the underlying corpus for research. 1 Introduction Argumentation mining attracts much attention re- cently: it is an important building block of applica- tions like automated decision making (Bench-Capon et al., 2009) or pro-and-con search engines (Cabrio and Villata, 2012c). In such applications, argumen- tation mining usually consists of solving three tasks for each document: (1) Identifying all argumenta- tive text segments in the document, (2) classifying the type of each segment, and (3) classifying rela- tions between the segments. In this paper we focus on the first task taking on the retrieval perspective of a search engine: Given a large-scale collection of documents (e.g., the web) and a query on some topic, return all argumenta- tive text segments relevant to the topic. Among oth- ers, a classifier is needed for this task that can dis- tinguish argumentative from non-argumentative seg- ments. Since we cannot presuppose a specific do- main or register within a general retrieval scenario, the classifier needs to robustly deal with documents from diverse domains and registers. In this regard the following two key issues arise. First, existing approaches to classifying argu- mentativeness usually focus on specific text do- mains (e.g., education) and registers (e.g., student essays). Therefore, many used features capture not only local linguistic properties of a text segment, but also global document properties (e.g., that a segment is part of the introduction). Such kinds of features tend to be effective only within a certain domain or a particular register while often failing for others. Second, all major existing approaches follow a su- pervised learning scheme based on manual annota- tion of argumentative text segments. However, the annotation of arguments is particularly intricate and thus expensive due to the complex linguistic struc- ture and the partly subjective interpretation of ar- gumentativeness. Different types of argumentative and non-argumentative segments may come in any order, segment boundaries are not always unam- biguous, and parts of an argument may be implicit. Studies reveal that annotators need multiple train- ing sessions to identify and classify argumentative segments with moderate inter-annotator agreement, and crowdsourcing-based annotation does …

Content truncated to 500 words