Joint prediction in MST-style discourse parsing for argumentation mining

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 938–948, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Joint prediction in MST-style discourse parsing for argumentation mining Andreas Peldszus Applied Computational Linguistics UFS Cognitive Science University of Potsdam peldszus@uni-potsdam.de Manfred Stede Applied Computational Linguistics UFS Cognitive Science University of Potsdam stede@uni-potsdam.de Abstract We introduce a new approach to argumen- tation mining that we applied to a parallel German/English corpus of short texts an- notated with argumentation structure. We focus on structure prediction, which we break into a number of subtasks: relation identification, central claim identification, role classification, and function classifica- tion. Our new model jointly predicts dif- ferent aspects of the structure by combin- ing the different subtask predictions in the edge weights of an evidence graph; we then apply a standard MST decoding algo- rithm. This model not only outperforms two reasonable baselines and two data- driven models of global argument struc- ture for the difficult subtask of relation identification, but also improves the results for central claim identification and func- tion classification and it compares favor- ably to a complex mstparser pipeline. 1 Introduction Argumentation mining is a task that has drawn increased interest in the last years. In its full- fledged version, it seeks to automatically recog- nize the structure of argumentation in a text by identifying and connecting the central claim of the text, supporting premises, possible objections, and counter-objections to these objections.1 A variety of applications can profit from ac- cess to the argumentative structure of text, includ- ing the retrieval of relevant court decisions from legal databases (Palau and Moens, 2011), auto- matic document summarization systems (Teufel and Moens, 2002), the analysis of scientific papers in biomedical text mining (Teufel, 2010; Liakata 1A comprehensive overview of the research field is given in (Peldszus and Stede, 2013). et al., 2012), or essay scoring. Importantly, argu- ment analysis can also be an extension of opinion mining applications. To make argumentation structures available for these applications, their robust automatic recogni- tion is required, a task that is very challenging: argumentative strategies and styles vary across text genres and languages; classifying arguments might require domain knowledge; furthermore, ar- gumentation can often rely on implicitly conveyed messages. The full-fledged task can be decomposed into several subtasks: • Segmentation: splitting the text into elemen- tary discourse units (EDUs as used in gen- eral kinds of discourse parsing, typically sen- tences or clauses) • Identification of argumentative discourse units (ADUs): discarding argumentatively ir- relevant EDUs, joining adjacent EDUs to form larger ADUs • ADU type classification: determining the type of argumentative unit; different schemes have been proposed, involving stance, evi- dence types, rhetorical status, argumentative function • Relation identification: building a connected tree- or graph-structure to represent argumen- tative relations between the ADUs • Relation type classification: determining the type of argumentative relation (e.g. support- ing versus attacking relations or more fine- grained types) In this paper, we address the last three subtasks: Given a …

Content truncated to 500 words