Coreference resolution in natural language processing

Coreference Resolution (CR), a key task in Natural Language Processing, involves identifying expressions that refer to the same entity or event. This thesis presents a detailed survey of CR models from the past decade, intro- duces a novel Multi-task Learning (MTL) model for the Gold-two-mention CR task, and explores CR’s application in financial sentiment analysis. The models surveyed are classified into feature-based, neural networks (includ- ing multilayer perceptrons and recurrent neural networks), knowledge-based, and transformer-based frameworks. The new MTL model simultaneously ad- dresses mention identification and linking in gold-two-mention style CR, us- ing a dynamic weight balancing mechanism to optimize task-specific weights during training. This model achieves state-of-the-art results on three bench- mark datasets. Furthermore, the thesis integrates advanced CR techniques with domain-specific models such as FinBERT to improve sentiment anal- ysis in financial texts. Rigorous empirical methods, including Propensity Score Matching and multivariate analysis, assess the impact of sentiments on financial outcomes, validating the framework’s predictive capabilities in financial sentiment analysis. This research not only improves CR method- ologies but also demonstrates their practical relevance in financial analytics, encouraging further interdisciplinary studies.