AbstractAs the volume of data we produce continues to grow, manual techniques increasingly struggle to keep up with the pace at which it is being generated, and greater emphasis is being placed on the automatic extraction of meaning from this data. Opinion mining and sentiment analysis provide valuable information on the views expressed in a text, however, they tell us only what opinions are being put forth and not why people hold the opinions they do. This is the task addressed by argument mining. The majority of argument mining techniques explored to date have focused on applying existing computational linguistic techniques to identify specific facets of the argumentative structure (for example, classifying premise/conclusion or argument/non-argument). The techniques presented in this thesis complement and extend these existing approaches by taking as a starting point the rich heritage of philosophical research in the analysis and understanding of argumentation, and drawing inspiration from the ways in which humans understand the structure of an argument.
The argument mining techniques presented here cover: a study of explicit linguistic expressions of the relationship between statements (e.g. "because", "therefore" or "however"); contextual knowledge in the form of premise-conclusion topic models which capture common patterns of statements matching one topic being used to support or attack statements matching another topic; relating similarity and topical changes to underlying argumentative structure; properties of large scale argument networks such as how central a proposition is to the text, offering a clue to the argumentative structure often intuitively employed by a human annotator, who will naturally connect a range of supporting arguments to a central conclusion; and argumentation schemes, common patterns of human reasoning which have been detailed extensively in philosophy and psychology.
Whilst each of these approaches produces reliable results, illuminating a facet of the full argumentative structure, it is in their combination that these techniques find their greatest strength. The final part of the work presented here looks at combining the output from these individual approaches whilst maintaining explainability of where the structure comes from. Allowing us, for example, to say that there is an inference relation between x and y because they form an instance of a particular argument scheme, or between y and z because of the presence of a discourse indicator. By leveraging the strengths of each, this combined explainable approach is shown to achieve an identification of the argumentative structure that is both more detailed and more accurate than existing argument mining techniques when tested on a corpus of debate from the US 2016 Presidential election, and comparable results to state of the art techniques when tested on widely used third-party corpora.
The work presented in this thesis offers two principal contributions, the development of a range of argument mining techniques grounded in argumentation theory, and, the introduction of Explainable Argument Mining (XAM).
|Date of Award||2021|
|Supervisor||Chris Reed (Supervisor) & Katarzyna Budzynska (Supervisor)|