Four Novel Ways Toward Better Factuality of Abstractive Text Summarization
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Abstractive text summarization is deep learning-based generative modeling research in natural language processing. Its goal is to develop models and methods that automatically condense long documents into fluent, coherent, relevant, and consistent summaries. It is an important research in the Big Data era, which demands advanced models and methods to turn voluminous and often long text data into concise but informative and factual summaries or abstracts for efficient human consumption. Although considerable progress has been made attributed to the advancement of encoder-decoder modeling, language modeling, and numerous task-specific methods, abstractive text summarization still faces challenges, particularly hallucinations in the model-generated summaries. Given its importance, the research recorded in this thesis was aimed at improving the factuality of abstractive text summarization by investigating a range of related problems, including undesirable (and ungrammatical) word repeats, distorted sub-phrasal hallucinations, endophoric reference errors, intrinsic named entity-related hallucinations, and factual informativeness issues. To do this, four novel methods, including determinantal point process-based sampling with self-critical reinforcement learning, syntactic structure-aware semantic learning, entity alignment learning facilitated by adaptive margin ranking loss, and optimal transport-based informative attention guided by named entity salience, were explored. Extensive experiments and subsequent result analyses have shown the efficacy of the proposed methods.
Please use this identifier to cite or link to this item: