Use this URL to cite or link to this record in EThOS:
Title: Aggregating and analysing opinions for argument-based relations
Author: Rajendran, Pavithra
ISNI:       0000 0004 7970 4958
Awarding Body: University of Liverpool
Current Institution: University of Liverpool
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Computational argumentation is a widely studied research area that has developed many formal models of argumentation. Argumentation is itself a multi-disciplinary field that branches from several research domains such as pragmatics, philosophy and logic. The idea of automatically extracting arguments and their relations from social media texts gives rise to a new research domain known as argument mining. The advancement in the field of natural language processing and machine learning has been helpful for extracting argument structures from informal texts. However, work on argument mining itself suffers from several drawbacks with one main problem related to lack of annotated corpora that can deal with argumentative texts. Another main drawback that can be related is the heterogeneous nature of data which prevents the use of a generalised annotation corpus for identifying argument structures. This thesis studies the intersection of computational argumentation and natural language processing to understand and process natural language arguments by developing an argumentation process which consists of identifying arguments and their relations and evaluating these arguments. Arguments are identified as structured or abstract arguments using linguistic attributes such as sentiment, stance and topic. The linguistic expression of the content towards a particular topic as a stance gives a pattern among opinions leading to two types of opinions - explicit opinions and implicit opinions. A binary classification approach is proposed for automatically classifying opinions as explicit or implicit opinions based on the way the stance is expressed. A set of hotel reviews is selected and the opinions are annotated by human annotators. The dataset is developed further by using different semi-supervised and weakly supervised approaches that automatically labels a large unlabeled dataset. This automatically labelled dataset is evaluated for deep learning models with the best performance of an LSTM model on the annotated dataset giving an accuracy of 84%. The second step of the argumentation process uses this classification of opinions to identify different types of relations that occur among these opinions such that it leads to constructing argument structures supporting a particular conclusion. Linguistic attributes such as sentiment and topic, along with the stance classification and domain-based knowledge are used for proposing a distant-supervision based approach that relates opinions as premises leading to a conclusion. The relation among the premises is similar to the entailment relation present in textual entailment and hence it is termed 'support-based entailment' relation. Another relation that is identified is the rephrase relation, in which, two opinions have similar argument meaning and wherein one can replace another without changing the meaning. These relations are useful for constructing argument structures as well as to identify enthymemes from arguments where an enthymeme is an argument with certain information missing. The different steps of the argumentation process for processing arguments in opinionated texts are carried out by considering opinions as abstract arguments. These arguments are built into bipolar argumentation graphs where a set of arguments are related to the support and attack relation. Different existing computational argumentation methods to compute the strength of these arguments are investigated in combination with natural language processing methods. The support relation in these graphs is used to convert them into coalitions of arguments, where a set of arguments support each other directly or indirectly and no attack relation exists within the coalition. These arguments are evaluated by investigating different ways of choosing coalitions, computing their strength and using them to support arguments as a whole. The evaluation process of these arguments is empirically evaluated for an NLP based task, which is to predict the overall sentiment of a review. The thesis explores a series of steps to identify and understand arguments present in opinionated texts by considering how natural language arguments fit within the argumentation process. It is shown that there exist different types of relations among these arguments if they are considered as structured arguments and that such relations help in identifying enthymemes from arguments. These relations are similar to existing relations in natural language processing but the latter has several drawbacks as they were not designed to detect argument-based relations. It is also shown that existing computational argumentation frameworks that are not developed for natural language arguments can be adopted for real-world tasks.
Supervisor: Bollegala, Danushka ; Parsons, Simon Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral