Work

Detecting and Mitigating Factual Errors and Biases in Natural Language Processing

Year: 2024
Type: dissertation
Abstract: The importance of fact checking has been emphasized recently due to the increasing harm to society from misleading information and fake news propagated online. This issue is even more pronounced in recent times with the widespread adoption of natural language generation systems such as machine translation and AI chatbots, which have been shown to generate factually incorrect information and exhibit biases learned from data. In this dissertation, I aim to improve the trustworthiness of the Web via developing methods that improve information integrity online on three different aspects. First, I develop a neural model capable of unbiased fact checking for verifying information on the web. Specifically, automatic claim and evidence generation methods are developed to counter the annotation artifacts in the training corpus, and a claim verification model is trained on the augmented data. Second, I develop an incongruent news headline detection model for maintaining information integrity on the web. Factual incongruency between the news headline and the article body is detected by employing a novel graph-based hierarchical encoder architecture. Finally, with the aim of improving trustworthiness of AI-generated texts on the web, I develop gender bias mitigation methods for neural machine translation systems. Specifically, a gender-aware contrastive learning approach based on gender pseudo-labels is employed as an additional fine-tuning step for neural machine translation models.
Author: 이민우
Language: English
Open Access status: green
OpenAlex ID: W7036833116
Cites: 122
Cited by:
Related to: 10