Association-Augmented Retrieval Improves Multi-Hop Question Answering
Researchers have unveiled Association-Augmented Retrieval (AAR), an efficient transductive reranking technique that enhances multi-hop passage retrieval by identifying associative links among passages. AAR employs a compact MLP with 4.2M parameters, utilizing contrastive learning based on co-occurrence annotations to evaluate bi-directional relationships in the embedding space. In tests on HotpotQA, AAR elevates passage Recall@5 from 0.831 to 0.916, marking an increase of 8.6 points without the need for tuning on the evaluation set, and achieves a remarkable gain of 28.5 points on challenging questions. Additionally, AAR records a +10.1 point improvement in the transductive setting on MuSiQue. However, an inductive model trained on training-split associations does not show notable enhancements on unseen validation associations.
Key facts
- AAR is a transductive reranking method for multi-hop retrieval
- Uses a small MLP with 4.2M parameters
- Trained via contrastive learning on co-occurrence annotations
- Improves HotpotQA Recall@5 from 0.831 to 0.916
- Gains of +28.5 points on hard questions
- Achieves +10.1 points on MuSiQue in transductive setting
- Inductive model shows no significant improvement on unseen associations
- Method does not require evaluation-set tuning
Entities
—