Toward Improving Robustness of Coreference Resolution for Thai Language

Coreference resolution aims to identify expressions in a text that refer to the same entity and establish connections between them. This paper presents an improved method for Thai coreference resolution, extending the F-coref architecture with two key enhancements. First, to handle the absence of explicit word boundaries in Thai, a pre-tokenization step is implemented before applying the model tokenizer. This ensures accurate alignment between gold coreference labels and resulting tokens. Second, an improved loss function is proposed to overcome a challenge encountered by F-coref during training. This modification prevents the model from solely optimizing coreference to null spans, ensuring a more balanced training trajectory. Empirical evaluations demonstrate the effectiveness of these modifications in boosting the robustness of Thai coreference resolution.