Subgraph Retrieval Enhanced by Graph-Text Alignment for Commonsense Question Answering (SEPTA Framework)

ECML PKDD (European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases) is a major European academic conference focusing on machine learning and knowledge discovery. The conference aims to promote exchange and collaboration among researchers, practitioners, and industry experts worldwide.

A paper co‑authored by Ant Group, Peking University, and Beijing Normal University titled “Subgraph Retrieval Enhanced by Graph-Text Alignment for Commonsense Question Answering” was accepted at ECML PKDD. The paper proposes a method that enhances subgraph information extraction through graph‑text alignment to address commonsense QA tasks.

The approach transforms a knowledge graph into a subgraph vector database and adopts a BFS‑style subgraph sampling strategy to avoid information loss. It also introduces a bidirectional contrastive learning method for graph‑text alignment, which improves subgraph retrieval and knowledge fusion. Finally, a prediction module integrates the retrieved information for reasoning.

Extensive experiments on five CSQA datasets demonstrate the framework’s effectiveness and robustness, showing superior performance and notable advantages in low‑resource scenarios.

A live streaming event titled “Paper Show Live #7” will feature a presentation by Peng Bo‑Ci, a PhD student at Peking University and research intern at Ant Group’s Graph Computing team, on October 17, 2024, from 18:00 to 19:00, providing deeper insight into the research.

Live Viewing Guide: Time: 2024‑10‑17 18:00‑19:00 Platforms: WeChat Channels (Ant Technology Research Institute, Ant Technology AntTech), Bilibili (Ant Technology Research Institute) – please follow and reserve.