| dc.description.abstract | In the era of AI technology breakthroughs, natural language processing (NLP) has
become a key field with the goal of bridging the gap between machine understanding and
human language. Text classification is a significant challenge in the field of natural language
processing (NLP) and has various practical uses, such as sentiment analysis, spam detection,
and subject categorization. The use of attention mechanisms in deep learning has recently made
substantial progress in text categorization, leading to the development of more advanced and
precise models.
This thesis focuses on a critical application of text classification: the categorization of
claims based on available evidence from documents. This task is essential for fact-checking and
evidence-based decision-making, requiring the ability to determine whether a claim is
supported by evidence, refuted by it, or if there is insufficient evidence to make a decision. The
accuracy of this categorization is vital for ensuring the credibility and reliability of information
in various domains, including journalism, legal analysis, and scientific research.
The challenge of categorizing claims involves the efficient selection of relevant features
that capture the nuanced meanings and implications of textual evidence. Deep learning models,
particularly those based on graph neural networks, have shown great promise in learning these
features from large datasets. However, the reliability of evidence sources remains a significant
concern, as the precision of classification depends heavily on the qualified evidence.
Additionally, context-aware classification is crucial, as the meaning of a statement can
vary significantly depending on its context. This necessitates models capable of dynamically
interpreting context to make accurate classifications. Furthermore, handling lengthy documents
presents scalability challenges, requiring systems that can efficiently process and analyze large
volumes of text without compromising performance. | en_US |
