ARGUMENT MINING FROM BIOMEDICAL LITERATURE WITH STRUCTURAL FEATURES

Student thesis: Phd

Abstract

Argument mining (AM), the process of automatically extracting argument structure from an argumentative document, has emerged as a crucial area of research within natural language processing (NLP) recently. It has a high impact on real-world applications such as decision-making, information retrieval, fact-checking and fake news detection. Correctly understanding an argument is challenging because it requires understanding not only each sentence, but also the global structural features of the document that these sentences make up. However, most previous AM models only pay attention to a single argument component (AC) to classify the type of AC or a pair of ACs to identify and classify the argument relation (AR) between the two ACs. These models tend to focus more on local features rather than global structural features and ignore the fact that argumentative texts of the same genre such as biomedical literature have global similarities in text organisation and argument structure due to the implicit rule of scientific paper writing. We refer to these similarities as genre-specific structural features and argument-specific structural features, respectively, in this thesis. Therefore, this thesis aims to explore the impact of structural features in argument mining. Firstly, we explore the impact of text zoning labels as the representation of genrespecific structural features on AM. One type of text zoning schemes, argumentative zoning, considered as a forerunner of AM, can be regarded as a coarse-grained argument structure at the sentence level. It shows the common organisation of an abstract in biomedical literature. To leverage such structural features, we propose a method based on multi-head attention for argument component identification and classification subtasks. The results on two biomedical argument mining datasets demonstrate the positive impact of the genre structural features on AM. Further, with the successful trend of treating other NLP tasks as machine reading comprehension (MRC) tasks to mimic the logical process that how humans do recently, which is similar to the reasoning logic of an argument, we propose a multi-turn MRC model that generates the argument structure incrementally to exploit graph-level argument-specific structural features. Specifically, at each turn, all ACs related to the query AC are generated simultaneously, such that the interaction that reveals the structural features of a group of ACs between the answer ACs is considered. In addition, the partially constructed graph is used as sub-graph level argument-specific structural features through a graph convolutional network to support the extension of the graph with additional ACs. Experiments performed on two biomedical argument mining corpus demonstrate the effectiveness of our method in terms of improving the model performance. Finally, since the multi-turn MRC model suffers from the problem of error propagation, we propose a generative multi-hop MRC model to alleviate error propagation while exploiting argument-specific structural features. This multi-hop MRC model learns the path-level structural features and extracts the argument structure at the same time. We validate the proposed method on the same two datasets, showing that even the path information is enough as a representation for the structural features to improve the performance. Overall, we illustrate that exploring the use of global structural features is an important step towards improving argument mining. The use of structural features is a promising way for argument mining and should receive more attention from researchers.
Date of Award1 Aug 2025
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorViktor Schlegel (Supervisor), Sophia Ananiadou (Supervisor) & Riza Theresa Batista-Navarro (Supervisor)

Keywords

  • Argument Mining
  • Structural Features
  • Natural Language Processing

Cite this

'