Extracting Argument and Domain Words for Identifying Argument Components in Texts

Proceedings of the 2nd Workshop on Argumentation Mining, pages 22–28, Denver, Colorado, June 4, 2015. c⃝2015 Association for Computational Linguistics Extracting argument and domain words for identifying argument components in texts Huy V. Nguyen Computer Science Department University of Pittsburgh Pittsburgh, PA 15260, USA hvn3@pitt.edu Diane J. Litman Computer Science Department & LRDC University of Pittsburgh Pittsburgh, PA 15260, USA litman@cs.pitt.edu Abstract Argument mining studies in natural language text often use lexical (e.g. n-grams) and syn- tactic (e.g. grammatical production rules) fea- tures with all possible values. In prior work on a corpus of academic essays, we demonstrated that such large and sparse feature spaces can cause difficulty for feature selection and pro- posed a method to design a more compact fea- ture space. The proposed feature design is based on post-processing a topic model to ex- tract argument and domain words. In this pa- per we investigate the generality of this ap- proach, by applying our methodology to a new corpus of persuasive essays. Our exper- iments show that replacing n-grams and syn- tactic rules with features and constraints us- ing extracted argument and domain words sig- nificantly improves argument mining perfor- mance for persuasive essays. 1 Introduction Argument mining in text involves automatically identifying argument components as well as argu- mentative relations between components. Argument mining has been studied in a variety of contexts including essay assessment and feedback (Burstein et al., 2003; Stab and Gurevych, 2014b), visu- alization and search in legal text (Moens et al., 2007), and opinion mining in online reviews and de- bates (Park and Cardie, 2014; Boltuži´c and Šnajder, 2014). Problem formulations of argument mining have ranged from argument detection (e.g. does a sentence contain argumentative content?) to argu- ment component (e.g. claims vs. premise) and/or relation (e.g. support vs. attack) classification. Due to the loosely-organized nature of many types of texts, associated argument mining studies have typically used generic linguistic features, e.g. n- grams and syntactic rules, and counted on feature selection to reduce large and sparse feature spaces. For example, in texts such as student essays and product reviews there are optional titles but typically no section headings, and claims are substantiated by personal experience rather than cited sources. Thus, specialized features as used in scientific ar- ticles (Teufel and Moens, 2002) are not available. While this use of generic linguistic features has been effective, we propose a feature reduction method based on the semi-supervised derivation of lexical signals of argumentative and domain content. Our approach was initially developed to identify ar- gument elements, i.e. hypothesis and findings, in academic essays (written following APA guidelines) of college students (Nguyen and Litman, submitted). In particular, we post-processed a topic model to ex- tract argument words (lexical signals of argumen- tative content) and domain words (terminologies in argument topics) using seeds from the assignment description and essay prompts. The extracted ar- gument and domain words were then used to cre- ate novel features and constraints for argument min- ing, and significantly outperformed features …

Content truncated to 500 words