A Deep Dive into QC‑MHM: Boosting Accuracy in Temporal Knowledge Graph Question Answering

Temporal KGQA vs. Traditional KGQA

Temporal KGQA extends the classic (subject, relation, object) triple to a quadruple (subject, relation, object, time interval), turning the answer space from pure entities to entities or timestamps and requiring time‑ordered and cross‑time reasoning, which makes it a more challenging and research‑valuable subfield.

Why Existing Methods Fall Short

Most prior models encode questions with PLMs such as BERT, which focus on entity nouns and ignore temporal expressions like "before", "after", or "during", leading to missed entity state changes. Additionally, they treat the knowledge graph merely as an index, failing to exploit multi‑hop relational structures, so the reasoning process becomes a black box.

QC‑MHM Framework

The proposed QC‑MHM (Question Calibration and Multi‑Hop Modeling) strengthens three layers—embedding, question representation, and graph reasoning—through four tightly coupled modules.

Module 1: Time‑aware KG Embedding

QC‑MHM adopts TComplEx as the base embedding algorithm. To give timestamps a sense of order, sinusoidal positional encodings are added to each timestamp, and an auxiliary binary classification task (“is m earlier than n?”) is introduced with binary cross‑entropy, enabling the embedding space to perceive temporal order.

Module 2: Question Calibration

This module first recalls candidate SPO triples by encoding both the question and SPOs with SentenceBERT and ranking them by cosine similarity, keeping the top‑10 candidates. Then three attention mechanisms—Concat, Dot, and Minus—compare question words with each SPO. A gated adaptive fusion layer lets the model decide how much of the original question representation versus the retrieved temporal information should contribute, producing a calibrated question vector that is time‑aware.

Module 3: Multi‑hop Modeling

To leverage graph structure, a multi‑hop attention message‑passing mechanism is built on a GNN. It starts with subgraph cropping based on entities extracted from the question, reducing the search space. Path scoring aggregates information across arbitrary‑length paths in a single GNN layer, and the final representation is obtained by average pooling. This design allows the model to see long‑range dependencies without iterative hops.

Module 4: Dual‑Channel Answer Prediction

Semantic vectors and graph vectors are concatenated and passed through a Transformer‑style fusion layer. The fused representation is projected separately into entity space and timestamp space, reusing the TComplEx scoring function, and the whole system is trained with cross‑entropy loss.

Experimental Results

On the CronQuestions benchmark, QC‑MHM achieves Hits@1 0.971 (an absolute gain of 5.1 %) and Hits@10 +1.2 %, outperforming all baselines on both entity‑type and time‑type answer tracks. On the TimeQuestions benchmark, it similarly attains top performance, demonstrating the method’s transferability beyond a single dataset.

Fine‑grained analysis shows especially large improvements on "Before‑After" questions, where temporal ordering is critical.

Ablation Study

Removing the question‑calibration module causes a sharp drop in overall Hits@1.

Removing the multi‑hop modeling module collapses performance on complex multi‑hop queries.

Removing the temporal auxiliary task eliminates sensitivity to purely temporal questions.

These results confirm that the three components are mutually reinforcing and indispensable.

Interpretability Visualizations

Path‑visualization figures reveal attention matrices that highlight which graph edges receive high weights, turning the previously black‑box reasoning into a white‑box process. Gradient‑based SPO attribution further shows that the calibrated question vector indeed captures key temporal facts.

Title: Question Calibration and Multi‑Hop Modeling for Temporal Question Answering
Link: https://arxiv.org/abs/2402.13188
Conference: The Thirty‑Eighth AAAI Conference on Artificial Intelligence (AAAI)
Authors: Beijing University of Aeronautics and Astronautics, Fudan University, Zhejiang University