DeDisCo at the DISRPT 2025 Shared Task: A System for Discourse Relation Classification

2025-09-14

AI Summary

This work addresses cross-lingual discourse relation classification in the DISRPT 2025 shared task, where limited annotated data and language-specific annotation biases hinder robust generalization. Method: We propose a dual-path encoder–decoder architecture: mT5 serves as the multilingual encoder to extract semantic representations, while Qwen acts as the reasoning-aware decoder for relation inference. We integrate machine-translated augmented data with explicit linguistic features—including dependency paths and discourse connectives—and introduce a progressive fine-tuning strategy tailored for low-resource settings. Contribution/Results: The model achieves 71.28% macro-accuracy on the DISRPT 2025 test set—ranking among the top systems—demonstrating strong generalization under unsupervised and few-shot conditions. Error analysis confirms its effectiveness in mitigating cross-lingual annotation bias and improving long-distance dependency modeling, establishing a scalable paradigm for low-resource discourse parsing.

Technology Category

• Natural Language Processing: Discourse, Pragmatics & Argument Mining
• Reasoning under Uncertainty: Relational Probabilistic Models
• Knowledge Representation and Reasoning: Qualitative Reasoning

Application Category

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

Abstract

This paper presents DeDisCo, Georgetown University's entry in the DISRPT 2025 shared task on discourse relation classification. We test two approaches, using an mt5-based encoder and a decoder based approach using the openly available Qwen model. We also experiment on training with augmented dataset for low-resource languages using matched data translated automatically from English, as well as using some additional linguistic features inspired by entries in previous editions of the Shared Task. Our system achieves a macro-accuracy score of 71.28, and we provide some interpretation and error analysis for our results.

Problem

• Classifying discourse relations in multilingual text
• Improving performance for low-resource language processing
• Testing transformer-based approaches for relation classification

Innovation

• mt5-based encoder for classification
• Qwen decoder approach for relations
• augmented dataset with automatic translation