Research map / Graph learning / Knowledge and temporal graphs
Question Calibration and Multi-Hop Modeling for Temporal Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QC-MHM, a temporal KGQA model that calibrates question representations with time-constrained KG facts (SPOs), injects timestamp order via an auxiliary task, and uses multi-hop GNN attention over temporal KG subgraphs.
Problem
Temporal knowledge graph question answering struggles with time-constrained questions because PLM question encoders over-rely on entities and ignore temporal entity shifts, and existing GNN-based methods do not model multi-hop relations explicitly.
Approach
Calibrate the question embedding by multi-view attention fusion with retrieved SPO (subject-predicate-object) embeddings via a gate, add a timestamp-order auxiliary loss on TComplEx embeddings, and perform multi-hop message passing via a powered attention matrix (sum of powers of A) over the question's sub-graph.
Key results
- On CronQuestions, QC-MHM reaches overall Hits@1 0.971 vs 0.920 for the best baseline CTRN, with complex-question Hits@1 0.946 (+7.7 points over CTRN per table).
- On TimeQuestions, QC-MHM achieves Hits@1 0.531 overall vs 0.465 for CTRN.
- Ablation: removing question calibration drops overall Hits@1 from 0.971 to 0.826; removing multi-hop drops it to 0.868.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QC-MHM instead of CTRN | Temporal KGQA · CronQuestions overall | Hits@1: 0.920 -> 0.971 | Helped |
| Use QC-MHM instead of CTRN | Temporal KGQA · CronQuestions complex questions | Hits@1: 0.869 -> 0.946 | Helped |
| Use QC-MHM instead of CTRN | Temporal KGQA · TimeQuestions overall | Hits@1: 0.465 -> 0.531 | Helped |
| Use QC-MHM instead of CTRN | Temporal KGQA · TimeQuestions Temporal | Hits@1: 0.512 -> 0.607 | Helped |
| Add temporal order encoding to timestamp embeddings (ablation) | QC-MHM · CronQuestions overall | Hits@1: 0.902 -> 0.971 | Helped |
| Add multi-hop attention instead of one-hop attention (ablation) | QC-MHM · CronQuestions overall | Hits@1: 0.868 -> 0.971 | Helped |
| Add question calibration (ablation) | QC-MHM · CronQuestions overall | Hits@1: 0.826 -> 0.971 | Helped |
Surprising findings
- PLM-only models (BERT, RoBERTa) perform far below TKG-embedding-based models, with Hits@1 around 0.23-0.24 on CronQuestions.
Limitations
- Relies on pre-trained TComplEx embeddings and entity annotations in questions for subgraph extraction.
- Inconsistencies in reported improvement numbers between abstract and text (e.g., 5.1% and 1.2% vs. 7.7% and 4.9%).
Experimental setup
Datasets: CronQuestions TimeQuestions
Benchmarks: CronQuestions TimeQuestions
Metrics: Hits@1 Hits@10
Baselines: BERT RoBERTa KnowBERT T5-3B EmbedKGQA EaE CronKGQA TMA TSQA TempoQR CTRN
Builds on: TComplEx CronKGQA EmbedKGQA Sentence-BERT Graph Attention Network
Closest papers by meaning
- Knowledge Editing with Dynamic Knowledge Graphs for Multi-Hop Question Answering (AAAI 2025 · similarity 0.89)
- MEQA: A Benchmark for Multi-hop Event-centric Question Answering with Explanations (NeurIPS 2024 · similarity 0.89)
- Diagnosing and Addressing Pitfalls in KG-RAG Datasets: Toward More Reliable Benchmarking (NeurIPS 2025 · similarity 0.89)
- Tree-of-Reasoning Question Decomposition for Complex Question Answering with Large Language Models (AAAI 2024 · similarity 0.88)
- KG-Guard: Graph-Based Hallucination Detection for Knowledge Base Question Answering (NeurIPS 2026 · similarity 0.88)
- Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.