Veracity Assessment: A semantic approach

Financiador

Organização

Publicações

Explorations into neural argument mining

Publication . Rodrigues, João; Branco, António Manuel Horta

Automatic argument mining seeks to endow artificial devices with the ability to identify structured arguments from plain unstructured text in natural language. A major motivation for empowering artificial devices with such skill is the fundamental role of argumentation in human cognition, by allowing the transmission and acquisition of knowledge, a hallmark of human interactions. While inheriting the complexities of defining a structure for arguments that are faced by argumentation theory, in general, argument mining also addresses challenges that are specific to the automatic extraction of structured arguments, as for example, the need for a representation of natural language semantics for an argument to be handled computationally. Current natural language processing techniques have successfully captured, to some extent, the meaning of words and sentences with distributional semantic representations and different neural network architectures. The performance of these techniques has improved with ever larger volumes of data, particularly annotated data. However, annotating data is time-consuming and laborious, and several natural language processing tasks have an inherent scarcity of annotated data, including argument mining. In this dissertation, we explored several neural network solutions to argument mining inspired by distributional semantics and transfer learning to mitigate the scarcity of annotated data in argument mining and improve its performance. We make contributions regarding several argument mining sub-tasks, namely in the so-called argument component identification, clausal classification and relational classification. These solutions experiment and resort to cross-lingual transfer from the most resourced language, English, to a low-resourced language, Portuguese, as also diverse knowledge bases from different lexical semantic families and over forty confluent language processing tasks. Additionally, we assess the impact of data artifacts in annotated data through a reproduction study, followed by a study of Transformer-based architectures on a most demanding argument mining sub-task concerned with argument reasoning comprehension. Throughout these explorations into neural argument mining, we set new state-of-the-art results across several argument mining sub-tasks, providing empirical evidence that our proposed solutions outperform previous computational models with alternative approaches.

2023-07Tese de doutoramento

Acesso aberto

Ver mais

Entidade financiadora

Fundação para a Ciência e a Tecnologia

Número da atribuição

SFRH/BD/129824/2017