Semantic Language Model Improves Temporal Knowledge Graph Completion on Benchmark Datasets

A semantic-driven framework named Temporal Semantic Language Model has demonstrated state-of-the-art performance on temporal knowledge graph completion benchmarks, outperforming existing structural baselines across multiple event datasets.

Temporal knowledge graph completion centers on predicting missing facts in knowledge graphs where entities and relations change over time, supporting applications such as intelligent question answering and event forecasting. Traditional completion methods depend on graph structures and process temporal quadruples in isolation, an approach that restricts their capacity to capture long-range temporal dependencies and rich semantic interactions across events.

To address these structural constraints, researchers designed the Temporal Semantic Language Model, or TSemLM, to operate without explicit graph construction or graph neural networks. The framework transforms temporal quadruples directly into natural language text representations to capture temporal dynamics through language modeling rather than graph topology.

The system constructs time-ordered event sequences by implementing a sampling strategy that jointly evaluates global frequency and temporal sensitivity. This sequence construction enables the model to effectively represent contextual temporal semantics across evolving relational facts.

Using a pretrained language model, the framework extracts context-aware semantic embeddings at the event level to predict missing entities or relations. By prioritizing semantic representations over structural encodings, the architecture aims to achieve greater adaptability and generalization across different temporal knowledge graphs.

Evaluation on standard benchmarks showed that TSemLM raised the Mean Reciprocal Rank by 2.57 percentage points on the ICEWS14 dataset and by 2.51 percentage points on ICEWS18, representing relative improvements of 6.4 percent and 9.1 percent over the strongest baseline. Further evaluations on the ICEWS05-15 dataset showed consistent performance gains across a broader time range and a substantially larger scale of events, accompanied by component-wise ablations, prompt template robustness tests, and efficiency analysis.

Subject of Research: Temporal knowledge graph completion using pretrained language models and semantic event sequences

Article Title: Pretrained language models with semantic embeddings for temporal knowledge graph completion

Article References: Lv, S., Dong, J., & Wang, X. (2026). Pretrained language models with semantic embeddings for temporal knowledge graph completion. Knowledge and Information Systems, 68(1), Article 314. https://doi.org/10.1007/s10115-026-02932-w

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02932-w

Keywords: temporal knowledge graph completion, TSemLM, pretrained language models, semantic embeddings, temporal quadruples, event forecasting, intelligent question answering, ICEWS14, ICEWS18, ICEWS05-15, mean reciprocal rank, time-ordered event sequences, temporal sensitivity, global frequency sampling, contextual temporal semantics, entity prediction, relation prediction, prompt template robustness