Exploiting Dialogue Acts and Context to Identify Argumentative Relations in Online Debates

Argumentative Relation Classification is the task of determining the relationship between two contributions in the context of an argumentative dialogue. Existing models in the literature rely on a combination of lexical features and pre-trained language models to tackle this task; while this approach is somewhat effective, it fails to take into account the importance of pragmatic features such as the illocutionary force of the argument or the structure of previous utterances in the discussion; relying solely on lexical features also produces models that over-fit their initial training set and do not scale to unseen domains. In this work, we introduce ArguNet, a new model for Argumentative Relation Classification which relies on a combination of Dialogue Acts and Dialogue Context to improve the representation of argument structures in opinionated dialogues.

Argumentative Dialogues are discussions between two or more parties involving an opinionated topic, i.e. any topic which may divide the interlocutors into a number of conflicting opinions. These discussions are usually different from ordinary conversations, in that the speakers’ goal is usually to convince their interlocutors of their own point of view by defending their own stance and attacking their opponents’ arguments. Figure 1 shows an example of a debate from the Kialo online debate platform. A key aspect in the study of Argumentative Dialogues is identifying the relationship between an argument step in the discussion and preceding argument steps introduced by other speakers; this task is commonly referred to as Argumentative Relation Classification (Stab and Gurevych, 2014), or sometimes Argument Polarity Prediction

we will use the term Argumentative Relation Classification, to avoid any confusion

with similar tasks such as Sentiment Analysis or Stance Classification.

First, they often ignore any non-lexical aspect of the dialogue, which hinders their capability to correctly understand the conversation. Second, they have limited understanding of the surrounding context of the argument contributions, and struggle to take long-term dependencies into account. Finally, they are often tested in a domain-specific scenario in which the system learns to predict relations between argument contributions that belong in the same dataset it was trained on; this makes it hard to correctly assess their capability to adapt to unseen conversations

The formal study of argumentative discussions is known in the literature as Argumentation Theory (van Eemeren et al., 1996). Walton (2009) divides argumentative study into four separate tasks: identification, which involves identifying Argumentative Dialogue Units (ADUs) in a dialogue and inserting them into a pre-determined argumentation scheme; analysis, which deals with identifying premises and conclusion of each argument; evaluation, which involves assessing an argument’s quality and persuasive power; and invention, which involves the creation of novel arguments for the debate. In this work we will focus on the task of identification of pre-constructed ADUs in an argumentation scheme.

Defeasible Logic, a formalism in which conclusions are supported by premises that may no longer be justified when additional premises are introduced.

The identification of a logical structure for reasoning goes back to the seminal works by Pollock (1987) and (Nute, 1988), which introduced Defeasible Logic, a formalism in which conclusions are supported by premises that may no longer be justified when additional premises are introduced. Dung (1995) introduced an abstract theory of Acceptability of Arguments in which arguments are seen as a set of logical statements, and each argument can be accepted or defeated depending on whether it clashes with other arguments. Prakken (2010) elaborated on this theory and presented a framework for structured arguments in which arguments can be supported with premises that justify their validity, and other arguments can attack the speaker’s viewpoint by either attacking the argument directly, or one of its premises. Cabrio and Villata (2012) combine textual entailment and argumentation graph into a unified framework that aims at automatically detecting accepted and defeated arguments based on the entailment between them. Lenz et al. (2020) adopted this scheme in their study on Argumentative Relation Classification on the Kialo corpus, and defined Default Inference and Default Conflict relations between arguments that support and attack each other respectively. The scheme was adopted by Fabbri et al. (2021), who use Natural Language Inference models to directly compute Argumentative Relations. This approach, however, does not distinguish between the semantic problem of determining logical relations between argument steps and the pragmatic problem of determining dialogue moves in a sequence of contributions.

Bipolar Argumentation Graph (BAG), in which claims are represented as nodes in a weighted graph, and can be supported by other claims or premises that can either Support or Attack each other.

Dimension Communicative Function

Task PropQuestion, SetQuestion,

ChoiceQuestion, Inform, Agree,

Disagree, Answer, Directive,

Commissive

Social Greeting, Goodbye, Thanking,

AcceptThanking, Apology, AcceptApology

Feedback AlloFeedback

Table 1: The DASHNet tagging scheme. Tags, also

known as Communicative Functions, are grouped in

Semantic Dimensions which represent different aspects

of utterance functions

supported