Here’s My Point: Joint Pointer Architecture for Argument Mining

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1364–1373 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Here’s My Point: Joint Pointer Architecture for Argument Mining Peter Potash, Alexey Romanov, Anna Rumshisky Department of Computer Science University of Massachusetts Lowell {ppotash,aromanov,arum}@cs.uml.edu Abstract In order to determine argument structure in text, one must understand how indi- vidual components of the overall argu- ment are linked. This work presents the first neural network-based approach to link extraction in argument mining. Specif- ically, we propose a novel architecture that applies Pointer Network sequence-to- sequence attention modeling to structural prediction in discourse parsing tasks. We then develop a joint model that extends this architecture to simultaneously address the link extraction task and the classifica- tion of argument components. The pro- posed joint model achieves state-of-the-art results on two separate evaluation corpora, showing far superior performance than the previously proposed corpus-specific and heavily feature-engineered models. Fur- thermore, our results demonstrate that jointly optimizing for both tasks is crucial for high performance. 1 Introduction An important goal in argument mining is to un- derstand the structure in argumentative text (Pers- ing and Ng, 2016; Peldszus and Stede, 2015; Stab and Gurevych, 2016; Nguyen and Litman, 2016). One fundamental assumption when working with argumentative text is the presence of Arguments Components (ACs). The types of ACs are gener- ally characterized as a claim or a premise (Govier, 2013), with premises acting as support (or possi- bly attack) units for claims (though some corpora have further AC types, such as major claim (Stab and Gurevych, 2016, 2014b)). The task of processing argument structure en- capsulates four distinct subtasks (our work fo- cuses on subtasks 2 and 3): (1) Given a sequence of tokens that represents an entire argumentative text, determine the token subsequences that con- stitute non-intersecting ACs; (2) Given an AC, determine the type of AC (claim, premise, etc.); (3) Given a set/list of ACs, determine which ACs have directed links that encapsulate overall argu- ment structure; (4) Given two linked ACs, deter- mine whether the link is a supporting or attack- ing relation. This can be labeled as a ‘micro’ ap- proach to argument mining (Stab and Gurevych, 2016). In contrast, there have been a number of efforts to identify argument structure at a higher level (Boltuzic and ˇSnajder, 2014; Ghosh et al., 2014; Cabrio and Villata, 2012), as well as slightly re-ordering the pipeline with respect to AC types (Rinott et al., 2015)). There are two key assumptions our work makes going forward. First, we assume subtask 1 has been completed, i.e. ACs have already been iden- tified. Second, we follow previous work that as- sumes a tree structure for the linking of ACs (Palau and Moens, 2009; Cohen, 1987; Peldszus and Stede, 2015; Stab and Gurevych, 2016). Specifi- cally, a given AC can only have a single outgoing link, but can have numerous incoming links. Fur- thermore, there is a ‘head’ component that has no outgoing link …

Content truncated to 500 words