Electronic Thesis and Dissertation Repository

Thesis Format



Master of Science


Computer Science

Collaborative Specialization

Artificial Intelligence


Mercer, Robert E.


We develop a novel unified representation for the argumentation mining task facilitating the extracting from text and the labelling of the non-argumentative units and argumentation components—premises, claims, and major claims—and the argumentative relations—premise to claim or premise in a support or attack relation, and claim to major claim in a for or against relation—in an end-to-end machine learning pipeline. This tightly integrated representation combines the
component and relation identification sub-problems and enables a unitary solution for detecting argumentation structures. This new representation together with a new deep learning architecture composed of a mixed embedding method, a multi-head attention layer, two biLSTM layers, and a final linear layer obtains state-of-the-art accuracy and F1 scores on the Persuasive Essays (PE) dataset. Furthermore, the augmentation of the PE corpus by including copies of major claims substituting the n-gram tokens that occur right before the major claim tokens with other major claim-introducing n-grams has aided in this increased performance on the PE dataset.

Summary for Lay Audience

Argumentation Mining is a research area in the field of Natural Language Processing. To detect, extract and identify argumentative structures from natural text we have to consider several sub-tasks which define the whole argumentation mining problem. Previous research works consider the argumentation detection problem as a set of many different sub-tasks. The subtasks are: 1) Separating non-argumentative units (sentences, words) from the argumentative units. 2) Predicting different types of argumentative components. 3) Identifying relations (support or attack) between the argumentative components. 4) Predicting the distance (textual distance measured by the number of sentences before or after) between the detected argumentative components. For example, the following sentence contains argumentative elements: ``We should not get a long-haired cat (Claim) because cats with long hair shed all over the house (Premise supporting the Claim).'' Previous research works have solved the argumentation detection task in a de-coupled way. They first detect the argumentative components, then identify stance or other relational attributes between each of the detected components. Some of the previous research works have worked with fewer sub-tasks and have obtained significant improvement regarding the detection of argumentation components. They assume that the spans of the argumentative elements (sentences, words) have been given and try to predict the correct type of argumentative components and relations from them. We have introduced a novel representation to jointly solve all four sub-tasks mentioned above. We have developed a novel neural network architecture to detect and solve all the sub-tasks related to argumentation mining. With our novel representation of the argumentation problem and the deep learning architecture, we have achieved state-of-the-art results on the Persuasive Essays (PE) Corpus.