REPOSITORY > RESULTS

Abstract

Over the past decade, rapid advancements in natural language processing have opened
up new avenues for tackling complex issues such as news bias analysis. This progress has
empowered researchers to explore innovative approaches to uncovering the complex biases
inherent in news production and coverage processes. News bias, a multifaceted reflection
of the inherent biases in the news production and coverage workflow, has far-reaching
implications for public opinion and decision-making. Analysing it poses a formidable challenge,
given its susceptibility to a variety of influencing factors, including but not limited
to political affiliation, editorial independence, and writing style.
The thesis at hand delves into this complex landscape, aiming not only to identify and
analyse news bias but also to extend these analyses to a multilingual context. By critically
evaluating existing research, proposing future strategies, and establishing comprehensive
methodologies for bias mitigation, the thesis lays the foundation for comprehensive methodologies
for bias mitigation in news reporting. It also presents the design and provides the
implementation of adaptable data generation scripts for generating customised datasets
for related tasks. Furthermore, it introduces a publicly available, novel event-centric news
dataset.
In addition, the thesis introduces a novel learning framework for predicting bias in news
headlines, incorporating inferential commonsense knowledge. It demonstrates how such
knowledge improves comprehension of short headlines that lack proper semantic and syntactic
context. Specifically, it illustrates the crucial role played by commonsense when
coupled with proper selection and refinement techniques in simplifying, interpreting, and
explaining events not explicitly stated in the headlines, thereby significantly enhancing the
task of bias analysis.
This thesis aims to close the gap by broadening its focus to embrace multilingualism in
response to the shortcomings of existing approaches, which primarily focus on languages
with abundant resources. This is accomplished by extending the previously introduced
framework to efficiently harness inferential knowledge in a multilingual context, employing
the translate-retrieve-translate strategy. This strategy is then used to propose a multilingual
bias prediction framework capable of handling low-resource languages under an
imbalanced sample distribution. It assesses the generalisability of the frameworks presented
and investigates the effects of knowledge augmentation and attention mechanisms.
It also provides a qualitative analysis of bias prediction performance across languages and
provides comprehensive explanations.