Structural Techniques for Argument Mining
Argument mining — the generalised, automated recognition of the structure of human reasoning expressed linguistically — is an enormously challenging task. The large and active philosophical research field of argumentation theory demonstrates the breadth of the issues that are apposite.
Our work in this area aims to combine a range of approaches, looking at the kinds of features common in natural language that enable the audience to understand the points being made and the relations between them. These methods cover linguistic features, changes in the topic being discussed and the identification of argumentation schemes, patterns of human reasoning which have been detailed extensively in philosophy and psychology. We have demonstrated that the structure of such schemes can provide rich information to the task of automatically identify complex argumentative structures in natural language text. By training a range of classifiers to identify the individual proposition types which occur in these schemes, it is possible not only to determine where a scheme is being used, but also the roles played by its component parts.
Whilst these methods focus mainly on the logical structure, this can be strengthened by also working with the rich and extensive domain of rhetorical figures. It seems likely that at least some rhetorical figures can be easily recognised automatically. If the occurrence of rhetorical figures can be correlated with some aspects of argument structure, then a new and potentially extremely powerful set of techniques can be brought on stream for substantially improving the performance of argument mining techniques.
We have, so far, focused on a small set of rhetorical figures that can be easily identified, compared instances of these figures with pre-annotated argument structures, and determined that in many cases connections can be found. Although this work is still at an early stage, requiring expansions of both the datasets used for comparison and the figures considered, it is a promising early sign that any one of the types of rhetorical figure could be interestingly challenging to identify and highly correlated with some aspect of argumentative structure. With over 700 to choose from, we could be at the outset of an extremely rich seam for argument mining.