Self Verifying Temporal Knowledge Evolution for Safe Contradiction Detection and Information Repair
Abstract Temporal knowledge graph completion has become an effective way to represent evolving facts, but accurate prediction of a missing temporal link is not equivalent to deciding whether two pieces of evidence are compatible, contradictory, obsolete, or simply insufficient for a safe update. This distinction becomes critical when a knowledge system is expected to change its stored information rather than merely rank candidate entities. This paper introduces TRACE-IT, a selfverifying temporal knowledge evolution architecture that couples temporal reasoning, contradiction detection, evidence sufficiency assessment, provenance-aware verification, calibrated confidence, selective abstention, and constrained knowledge repair. The study also introduces T-CONFLICT, a 7,600-pair benchmark with a hierarchical two-task formulation: Temporal Conflict Resolution (TCR) and Evidence Sufficiency Classification (ESC). The benchmark uses entity-disjoint splits and blind two-annotator adjudication, obtaining Cohen’s kappa of 0.813 and Krippendorff’s alpha of 0.809. Across six temporal knowledge graph benchmarks, TRACE-IT improves a same-backbone TimePlex implementation by a mean of 6.65 MRR points with a 95% t-based confidence interval of [5.69, 7.61] and an exact Wilcoxon p-value of 0.031. On T-CONFLICT, TRACE-IT obtains 0.864 TCR Macro-F1, 0.872 ESC Macro-F1, and 0.852 binary contradiction F1. More importantly for trustworthy evolution, the system reaches 0.919 repair precision, 0.812 conditional repair recall, 0.081 unsafe-repair rate, and 0.157 abstention rate on the ICEWS14 evaluation. Calibration yields ECE = 0.0183 over 7,556 non-abstained decisions, while selective prediction reaches 70.1% coverage at 4.8% unsafe-repair risk. Robustness experiments show that the performance margin over TimePlex widens as timestamps, facts, or evidence are corrupted. The results support a view of temporal knowledge systems in which prediction, verification, explanation, and controlled change are treated as one operational problem rather than isolated objectives.
Authors
- Phat Hai Tien Nguyen
Institutions
- Ton Duc Thang University (VN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11295059/v1
- Primary Topic
- Advanced Graph Neural Networks
- Type
- preprint