A Multi-Task and Multi-Label Classification Model for Implicit Discourse Relation Recognition

2024-08-16

arXiv.org

AI Summary

This paper addresses the multi-label classification challenge in implicit discourse relation recognition (IDRR), arising from semantic ambiguity. We introduce the first multi-label IDRR benchmark aligned with PDTB 3.0’s three-layer semantic hierarchy, along with a unified single- and multi-label joint learning framework. Methodologically, we propose a context encoder–based multi-task deep classifier that jointly optimizes multi-label classification loss (e.g., binary cross-entropy) and single-label cross-entropy loss, trained exclusively on the DiscoGeM corpus. Our contributions are threefold: (1) establishing the first multi-label IDRR benchmark; (2) proposing a joint single-/multi-label learning paradigm that naturally derives optimal single-label predictions from multi-label outputs; and (3) demonstrating, for the first time, effective cross-corpus transfer from DiscoGeM to PDTB 3.0. Experiments show our approach achieves state-of-the-art performance on the DiscoGeM single-label IDRR task.

Technology Category

• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Natural Language Processing: Discourse, Pragmatics & Argument Mining
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

We address the inherent ambiguity in Implicit Discourse Relation Recognition (IDRR) by introducing a novel multi-task classification model capable of learning both multi-label and single-label representations of discourse relations. Our model is trained exclusively on the DiscoGeM corpus and evaluated both on the DiscoGeM and the PDTB 3.0 corpus. We establish the first benchmark on multi-label IDRR classification and achieve SOTA results on single-label IDRR classification using the DiscoGeM corpus. Finally, we present the first evaluation on the potential of transfer learning between the DiscoGeM and the PDTB 3.0 corpus on single-label IDRR classification.

Problem

• Develop multi-label classification for implicit discourse relation recognition
• Jointly learn multi-label representations across PDTB 3.0 sense levels
• Achieve state-of-the-art results in single-label IDRR using DiscoGeM

Innovation

• Multi-task model for multi-label discourse relations
• Adaptable to single-label IDRR via probability selection
• SOTA results using DiscoGeM corpus