GRASP: Rich Patterns for Argumentation Mining

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1345–1350 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics GrASP: Rich Patterns for Argumentation Mining Eyal Shnarch Ran Levy∗ Vikas Raykar‡ Noam Slonim IBM Research - Haifa, Israel ‡IBM Research - Bangalore, India {eyals, ranl, noams}@il.ibm.com viraykar@in.ibm.com Abstract We present the GrASP algorithm for au- tomatically extracting patterns that char- acterize subtle linguistic phenomena. To that end, GrASP augments each term of input text with multiple layers of linguis- tic information. These different facets of the text terms are systematically combined to reveal rich patterns. We report highly promising experimental results in several challenging text analysis tasks within the field of Argumentation Mining. We be- lieve that GrASP is general enough to be useful for other domains too. 1 Introduction Many standard text analysis tasks can be ad- dressed relatively well while exploiting simple textual features, e.g., Bag-Of-Words representa- tion and Naive Bayes for document classification (McCallum and Nigam, 1998). However, the iden- tification of more subtle linguistic phenomena, that are further reflected via relatively short texts – as opposed to whole documents – may require a wider spectrum of linguistic features. The main contribution of this work is in out- lining a simple method to automatically extract rich linguistic features in the form of patterns, and demonstrate their utility on tasks related to Ar- gumentation Mining (Mochales Palau and Moens, 2009), although we believe that the proposed ap- proach is not limited to this domain. Argumentation mining involves automatically identifying argumentative structures within a cor- pus – e.g., claims or conclusions, and evidence instances or premises – as well as their inter- relations. For instance, each of the following sen- tences includes a claim for a [topic]. ∗First two authors contributed equally. 1. Opponents often argue that the open primary is unconstitutional. [Open Primaries] 2. Prof. Smith suggested that affirmative action devalues the accomplishments of the chosen. [Af- 3.firmative Action] The majority stated that the First Amend- ment does not guarantee the right to offend others. [Freedom of Speech] These sentences share almost no words in com- mon, however, they are similar at a more abstract level. A human observer may notice the follow- ing underlying common structure, or pattern: [someone][argue/suggest/state][that] [topic term][sentiment term] We present GrASP, standing for GReedy Aug- mented Sequential Patterns, an algorithm that aims to automatically capture such underlying struc- tures of the given data. Table 1 shows the pat- tern that GrASP may find for the above examples, along with its matches in those texts. Such pat- terns can then be used to detect the existence of the phenomenon in new texts. The algorithm starts with augmenting the terms of the input with various layers of attributes, such as hypernyms from WordNet (Fellbaum, 1998), named entity types, and domain knowledge (Sec- tion 3.1). This multi-layered representation en- ables GrASP to consider many facets of each term. Next, it finds the most indicative attributes (Sec- tion 3.2) and iteratively grows …

Content truncated to 500 words