Learning to Differentiate Pairwise-Argument Representations for Implicit Discourse Relation Recognition

Implicit Discourse Relation Recognition: A Joint Learning Framework

Implicit Discourse Relation Recognition (IDRR) is a challenging task that involves recognizing relationships between texts without explicit connectives. This paper proposes a joint learning framework that combines prototypical learning, adversarial learning, and hub-migration based redistribution to enhance the performance of Pre-trained Language Models (PLMs) for IDRR. The framework is designed to produce more distinguishable argument representations, which is crucial for connective-free relation determination.

Key Contributions:

• A joint learning framework that combines three learning methods to enhance PLMs for IDRR Prototypical learning that uses contrastive learning to guide the encoder to produce more typical representations
• Adversarial learning that generates highly distracting examples to improve the robustness of the model
• Hub-migration based redistribution that disperses the centers of all classes in the feature space and drives in-class samples to converge towards the centers

Experimental Results:

• The proposed framework achieves substantial improvements compared to BERT, RoBERTa, and DeBERTa baselines on PDTB 2.0, PDTB 3.0, and CoNLL-2016 datasets
• The framework outperforms previous work in the scenario of connective-free IDRR and obtains comparable performance to some of the connective-exposed IDRR models

Conclusion:

This paper proposes a joint learning framework that enhances PLMs for IDRR by producing more distinguishable argument representations. The framework combines prototypical learning, adversarial learning, and hub-migration based redistribution to improve the performance of IDRR models. Experimental results demonstrate the effectiveness of the proposed framework, which achieves state-of-the-art performance on several benchmark datasets.

Download Paper