Unsupervised Argumentation Mining in Student Essays

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6795–6803 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6795 Unsupervised Argumentation Mining in Student Essays Isaac Persing and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688, USA {persingq,vince}@hlt.utdallas.edu Abstract State-of-the-art systems for argumentation mining are supervised, thus relying on training data containing manually annotated argument components and the relationships between them. To eliminate the reliance on annotated data, we present a novel approach to unsupervised argument mining. The key idea is to bootstrap from a small set of argument components automatically identified using simple heuristics in combination with reliable contextual cues. Results on a Stab and Gurevych’s corpus of 402 essays show that our unsupervised ap- proach rivals two supervised baselines in performance and achieves 73.5−83.7% of the performance of a state-of-the-art neural approach. Keywords: Opinion Mining / Sentiment Analysis, Discourse Annotation, Representation and Processing, Information Extraction, Information Retrieval 1. Introduction Recent years have seen a surge of interest in argumentation mining (see Cabrio and Villata (2018) and Lawrence and Reed (2019) for comprehensive surveys of this area of re- search). Argumentation mining typically involves address- ing two subtasks: (1) argument component identification (ACI), which consists of identifying the locations and types of the components that make up the arguments (i.e., major claims, claims, and premises), and (2) relation identifica- tion (RI), which involves identifying the type of relation that holds between two argument components (i.e., support, attack, none). As an example, consider the following text segment taken from a corpus of student essays that Stab and Gurevych (S&G) annotated with argument components and their relations (Stab and Gurevych, 2017): In my view point, (1) I would agree to idea of taking a break before starting higher education. (2) Students who take break benefit by traveling around the places or by working. (3) They tend to contribute more due to their real world experiences which makes them more mature. In this example, premise (3) supports claim (2), and (1) is a major claim. State-of-the-art argument mining systems are supervised, adopting a feature-rich approach that typically include large variety of structural, lexical, and syntactic features. The plethora of features seem to suggest that argument- annotated training data are indispensable for training these systems, as they provide a supervised learner guidance on which features are useful and how features should be com- bined. Moreover, world knowledge and domain-specific knowledge can potentially be learned from annotated data, as can be seen from the aforementioned example. A natural question, then, is: in the absence of argument- annotated training data, is argument mining doomed to fail? In particular, are argument-annotated training data as in- dispensable as they seem in enabling state-of-the-art argu- ment mining systems to achieve their current level of per- formance? By blindly applying supervised approaches with a rich feature set, previous work has largely failed to answer this question. To shed light on this question, we …

Content truncated to 500 words