TY - GEN
T1 - AEGI
T2 - 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2026
AU - Fu, Yuqing
AU - Deng, Yimin
AU - Wang, Yejing
AU - Zhu, Li
AU - Qian, Xueming
AU - Zhao, Guoshuai
AU - Zhao, Xiangyu
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.
PY - 2026
Y1 - 2026
N2 - Temporal knowledge graph question answering (TKGQA) aims to answer temporal questions by leveraging dynamic knowledge, which is crucial for various downstream tasks. In this process, aligning the temporal questions and TKGs presents a significant challenge. Current methods either perform implicit similarity ranking through time-aware representations or concentrate on explicit semantic parsing and LLM-based interpretable reasoning. The first approach often struggles with a lack of semantic understanding when dealing with question-rich content, making it challenging to manage complex temporal information. The second approach typically relies on external knowledge retrieval, which can be adversely affected by noise from inaccuracies in the retrieval process that occurs before answering the questions. To overcome these limitations, we introduce an Anchor Event Guided Inference (AEGI) framework. This framework comprises a coarse-grained answer retrieval process and a TKG fact retrieval chain, providing both implicit and explicit reasoning perspectives for LLM-based TKGQA. By aligning temporal evidence with anchor events, AEGI improves knowledge consistency and reduces hallucinations in LLMs, enabling interpretable and accurate reasoning. Extensive experiments on two TKGQA datasets, MultiTQ, TimeQuestions demonstrate the effectiveness of our approach. The code is available at: https://github.com/yqingFU1007/AEGI-public.
AB - Temporal knowledge graph question answering (TKGQA) aims to answer temporal questions by leveraging dynamic knowledge, which is crucial for various downstream tasks. In this process, aligning the temporal questions and TKGs presents a significant challenge. Current methods either perform implicit similarity ranking through time-aware representations or concentrate on explicit semantic parsing and LLM-based interpretable reasoning. The first approach often struggles with a lack of semantic understanding when dealing with question-rich content, making it challenging to manage complex temporal information. The second approach typically relies on external knowledge retrieval, which can be adversely affected by noise from inaccuracies in the retrieval process that occurs before answering the questions. To overcome these limitations, we introduce an Anchor Event Guided Inference (AEGI) framework. This framework comprises a coarse-grained answer retrieval process and a TKG fact retrieval chain, providing both implicit and explicit reasoning perspectives for LLM-based TKGQA. By aligning temporal evidence with anchor events, AEGI improves knowledge consistency and reduces hallucinations in LLMs, enabling interpretable and accurate reasoning. Extensive experiments on two TKGQA datasets, MultiTQ, TimeQuestions demonstrate the effectiveness of our approach. The code is available at: https://github.com/yqingFU1007/AEGI-public.
UR - https://www.scopus.com/pages/publications/105041842820
U2 - 10.1007/978-981-92-1465-5_20
DO - 10.1007/978-981-92-1465-5_20
M3 - 会议稿件
AN - SCOPUS:105041842820
SN - 9789819214648
T3 - Lecture Notes in Computer Science
SP - 253
EP - 264
BT - Advances in Knowledge Discovery and Data Mining - 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2026, Proceedings
A2 - Wong, Raymond Chi-Wing
A2 - Kwok, James
A2 - Tong, Hanghang
A2 - Lu, Hua
A2 - Salim, Flora
A2 - Song, Yuanfeng
A2 - Yiu, Man Lung
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 9 June 2026 through 12 June 2026
ER -