Persuasive argumentation is the medium of communication used to shift public opinion, sell goods through advertisements, influence legal decisions, and motivate policy changes. With widespread adaptation of social media outlets, news channels, and mass-advertisement platforms, the task of predicting argument persuasiveness has become increasingly important.
Discourse is the level of linguistic analysis characterizing a text’s structure at the sentence or clause level. However, current transformer language models tend to neglect discourse-oriented tasks during pre-training. Since generalizable morphological, syntactic, and semantic linguistic properties are learned during the pre-training phase, linguistic properties related to discourse should be learned by the model at this phase as well.
Structural Inductive Biases
While language models might fail to learn notions of discourse, graph neural nets pose a modeling approach through which discourse structure is fundamental to the learning process. By framing independent ideas within text as nodes, and the relations between them as edges, we consider a graphical representation of text. From here, we use graph neural networks to model how the structure of individual components in a text impact the perception of readers on its entirety.