TAR-KU: Temporal-Aware Risk-Constrained Knowledge Updating for High-Recall Political Text Correction
Main Article Content
Keywords
political text correction, temporal knowledge, knowledge updating, retrieval-augmented generation, entity-relation alignment, risk-constrained editing
Abstract
Political texts contain time-sensitive facts about people, offices, institutions, places, and official terminology. A knowledge base that is merely expanded with newly retrieved material may improve coverage, but it can also introduce entity confusion, outdated facts, conflicting evidence, and unjustified edits. This paper presents TAR-KU, a temporal-aware risk-constrained knowledge updating method for high-recall political text correction. TAR-KU separates knowledge admission from correction authorization. Incremental materials are first converted into candidate records, then checked through entity-relation alignment, temporal validity modeling, source-evidence aggregation, conflict governance, and version snapshots. Only records that satisfy the query-time authorization conditions are allowed to support local edits. This design preserves historical knowledge while making newly verified facts available to downstream correction. Experiments on 688 political-text samples, together with an independent 128-sample source set and a 100-sample clean control set, compare TAR-KU with retrieval, memory, parameter-editing, snapshot, and prompt-update baselines. TAR-KU achieves 83.19% correction precision, 87.59% correction recall, and 85.34% F1, obtaining the highest recall and F1 among the compared methods, at an average latency of 4.07 s. Updating the knowledge state increases recall from 85.63% to 87.59% and correct edits from 566 to 579. The results show that TAR-KU improves dynamic-fact coverage while preserving an explicit boundary between evidence availability and automatic modification.
References
- [1]Lewis P, Perez E, Piktus A, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS, 2020: 9459–9474.
- [2]Ram O, Levine Y, Dalmedigos I, et al. In-Context Retrieval-Augmented Language Models. TACL, 2023, 11: 1316–1331.
- [3]Asai A, Wu Z, Wang Y, et al. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. ICLR, 2024.
- [4]Jiang Z, Xu F, Gao L, et al. Active Retrieval Augmented Generation. EMNLP, 2023: 7969–7992.
- [5]Meng K, Bau D, Andonian A, Belinkov Y. Locating and Editing Factual Associations in GPT. NeurIPS, 2022: 17359–17372.
- [6]Meng K, Sharma A S, Andonian A, et al. Mass-Editing Memory in a Transformer. ICLR, 2023.
- [7]Mitchell E, Lin C, Bosselut A, et al. Memory-Based Model Editing at Scale. ICML, 2022, 162: 15817–15831.
- [8]Wang P, Li Z, Zhang N, et al. WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models. NeurIPS, 2024, 37: 53764–53797.
- [9]Hu C, Cao P, Chen Y, et al. WilKE: Wise-Layer Knowledge Editor for Lifelong Knowledge Editing. Findings of ACL, 2024: 3476–3503.
- [10]Gu J C, Xu H X, Ma J Y, et al. Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue. EMNLP, 2024: 16801–16819.
- [11]Peng B, Zhu Y, Liu Y, et al. Graph Retrieval-Augmented Generation: A Survey. ACM TOIS, 2025, 44(2): 1–52.
- [12]Zhang J, Shao J, Cui B. StreamE: Learning to Update Representations for Temporal Knowledge Graphs in Streaming Scenarios. SIGIR, 2023: 622–631.
- [13]Chen Z, Xu C, Su F, et al. Temporal Extrapolation and Knowledge Transfer for Lifelong Temporal Knowledge Graph Reasoning. Findings of EMNLP, 2023: 6736–6746.
- [14]Zhang F, Zhang Z, Zhuang F, et al. Temporal Knowledge Graph Reasoning With Dynamic Memory Enhancement. IEEE TKDE, 2024, 36(11): 7115–7128.
- [15]Xin H, Chen L. KartGPS: Knowledge Base Update with Temporal Graph Pattern-based Semantic Rules. ICDE, 2024: 5075–5087.
- [16]Zheng C, Li L, Dong Q, et al. Can We Edit Factual Knowledge by In-Context Learning? EMNLP, 2023: 4862–4876.
- [17]Qian X, Zhang Y, Zhao Y, et al. TimeR4: Time-aware Retrieval-Augmented Large Language Models for Temporal Knowledge Graph Question Answering. EMNLP, 2024: 6942-6952. DOI: 10.18653/v1/2024.emnlp-main.394.
- [18]Zhang S, Xue Y, Zhang Y, et al. MRAG: A Modular Retrieval Framework for Time-Sensitive Question Answering. Findings of EMNLP, 2025: 3080-3118. DOI: 10.18653/v1/2025.findings-emnlp.167.
- [19]Cohen R, Biran E, Yoran O, Globerson A, Geva M. Evaluating the Ripple Effects of Knowledge Editing in Language Models. TACL, 2024, 12: 283-298. DOI: 10.1162/tacl_a_00644.
- [20]Hsueh C H, Huang P K M, Lin T H, et al. Editing the Mind of Giants: An In-Depth Exploration of Pitfalls of Knowledge Editing in Large Language Models. Findings of EMNLP, 2024: 9417-9429. DOI: 10.18653/v1/2024.findings-emnlp.550.
- [21]Yin X, Jiang J, Yang L, Wan X. History Matters: Temporal Knowledge Editing in Large Language Model. AAAI, 2024, 38(17): 19413-19421. DOI: 10.1609/aaai.v38i17.29912.
- [22]Zhao S, Pang N, Tan Z, et al. Multi-granularity Temporal Knowledge Editing over Large Language Models. AAAI, 2026, 40(34): 28742-28750. DOI: 10.1609/aaai.v40i34.40107.
- [23]Chen Q, Zhang T, He X, et al. Lifelong Knowledge Editing for LLMs with Retrieval-Augmented Continuous Prompt Learning. EMNLP, 2024: 13565-13580. DOI: 10.18653/v1/2024.emnlp-main.751.
- [24]Qiao S, Liu X, Na S H. COMEM: In-Context Retrieval-Augmented Mass-Editing Memory in Large Language Models. Findings of NAACL, 2024: 2333-2347. DOI: 10.18653/v1/2024.findings-naacl.151.
Article Sidebar
This work is licensed under a Creative Commons Attribution 4.0 International License.