DEANE: Context-Aware Dual-Craft Graph Contrastive Learning for Enhanced Extractive Question Answering
Abstract Extractive Question Answering (EQA) involves extracting accurate answer spans from a background passage in response to a given question. In recent years, there has been significant interest in leveraging Pre-trained Language Models (PLMs) and Graph Convolutional Networks (GCNs) to address E...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-04-01
|
| Series: | International Journal of Computational Intelligence Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44196-025-00801-y |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Extractive Question Answering (EQA) involves extracting accurate answer spans from a background passage in response to a given question. In recent years, there has been significant interest in leveraging Pre-trained Language Models (PLMs) and Graph Convolutional Networks (GCNs) to address EQA tasks. PLMs usually function as context encoders, while GCNs are employed to capture latent semantic relationships between answer spans and the passage/question. This combined approach has shown promise, yielding notable outcomes in EQA performance. However, current graph-based methods encounter a challenge where the graph structure is predefined without sufficient justification. This graph ambiguity can potentially lead to error propagation within the subsequent graph encoder. To alleviate this issue, this paper introduces Dual-craft basEd grAph coNtrastive lEarning (DEANE) for EQA, where the graph structure and node features are context-aware and data-driven. Initially, the passage and question are represented as a connected graph. Subsequently, the adaptive augmentation strategy is introduced to generate two distinct views of the original graph via reparameterization networks, where important graph edges and node features are prioritized. Finally, a multi-view contrastive loss is leveraged to learn latent representations from augmented graphs. Empirically, our method outperforms existing graph-based approaches on six well-established EQA benchmarks. Ablation studies further demonstrate the effectiveness of the proposed approach in mitigating structural ambiguity, enhancing encoder flexibility, and improving model performance through multi-view data integration. |
|---|---|
| ISSN: | 1875-6883 |