Towards Exploiting Argumentative Context for Argumentative Relation Identification

Argument mining is the task of identifying argument structures in argumentative texts. This task is useful for many applications such as document summarization, opinion mining, and automated essay scoring [10, 13]. In the literature, several subtasks for argument mining have been extensively studied, such as argument component type classification, stance classification, and argumentative relation identification [7, 11, 4, 8, 9, 5]. This paper addresses the task of argumentative relation identification due to its recent popularity in argument mining. Consider the argumentative text in Figure 1, where argument components (ACs), basic units of arguments, are already identified. Argument component type classification aims at classifying ACs into a premise or claim (e.g. classifying AC1 into a claim and AC2 into a premise). Stance classification aims at classifying the stance of ACs towards a claim as either proponent or opponent (e.g. classifying AC1 into a proponent stance and AC2 into a opponent stance). Argumentative relation identification aims to identify an argumentative link between two ACs, and if it exists, classify it into two classes: attack or support (e.g. identifying the attack relation from AC2 to AC1). Conventional approaches have focused on creating features using the local input ACs rather than using macro-level information such as the overall structure of an argument [7, 11, 1]. However, argumentative relations are closely related to each other and they form argument diagrams [6]. Thus, we speculate that the information of surrounding argumentative context (e.g. other argumentative relations) can be useful for predicting a relation. For example, in Figure 1, AC2 attacks AC1 and AC3 attacks AC2. If we were to predict the attack relation from AC3 to AC2, knowing whether AC2 is attacking another AC would be useful information, because a writer frequently uses such macro structure as a tactic for strengthening their argument. For example, in Figure 1, the writer gives a possible counter-argument to their claim (AC2 attacks AC1) and then attacks it immediately (AC3 attacks AC2), which makes it difficult for others to attack.