TCC-RAG: Temporal Causal Chain-Guided RAG for Temporal Knowledge Graph Question Answering

What it claims

TCC-RAG is a training-free, plug-and-play retrieval layer for temporal knowledge graph QA. It replaces semantic-similarity retrieval with structured lookups driven by temporal cues parsed from the question ("before", "after", "first", "last"). The authors report Hits@1 gains of 10.72% and 8.40% on two TKGQA benchmarks against more than 10 baselines, and a 62.97% gain on multi-step temporal questions. They also report over 57% lower token consumption than the strongest baseline. "Causal" is used operationally: it means query-relevant temporal ordering between events, not interventional causal inference.

How it works

Facts are quadruples (head, relation, tail, timestamp). They are reorganised into three views. A Temporal Event Graph (TEG) serves queries naming a known entity pair. Two temporally sorted, trie-structured indices serve multi-step queries anchored on a single entity: the Subject Causal Cue Index (SCCI) and the Object Causal Cue Index (OCCI).

At query time, NER and LLM extraction each propose entities, and a consistency check reconciles them. The LLM then parses the relation, the entity's role (subject or object), and the cues. Cues come in two kinds:

• Ordering cues (before, after, same) filter events against a boundary timestamp.
• Selection cues (first, last) pick the earliest or latest remaining event.

For "Before Mexico, which country did China visit last?", the boundary is China's visit to Mexico, and the cues are (before, last).

The causal evidence filter is deterministic. It first removes causally equivalent duplicates, then keeps a minimal subset of facts that is sufficient and consistent with the query. When structured retrieval returns nothing, semantic retrieval supplies fallback candidates, and these pass through the same filter.

flowchart TD
  Q[Question] --> P[Parse entities, relation, role, cues]
  P --> R{Route by query shape}
  R -->|entity pair| TEG[Temporal Event Graph]
  R -->|single anchor, subject| S[SCCI index]
  R -->|single anchor, object| O[OCCI index]
  TEG --> C{Facts found?}
  S --> C
  O --> C
  C -->|no| SEM[Semantic retrieval fallback]
  C -->|yes| F[Causal evidence filter]
  SEM --> F
  F --> L[LLM answers]

Evidence

The metric is Hits@1. The quoted gains are overall improvements over the best of more than 10 baselines on each benchmark. The multi-step figure (62.97%) covers that question category only. The text does not name the two benchmarks or the base LLM. It also does not say whether the gains are absolute or relative. Related work names TimeR4 and RTQA as LLM-based TKGQA methods built on semantic retrieval. The text also mentions ablation and case studies, but the fetched portion contains no tables.

Limits

The fetched text breaks off early in the methodology section. Results tables, ablation numbers, the efficiency setup, and any limitations section are missing. No code, model, or licence is mentioned.

The design depends on questions that map cleanly onto a single relation, a role, and cues from a fixed set (before, after, same, first, last). The text does not show how it handles queries outside that grammar, beyond falling back to semantic retrieval. The method also assumes knowledge already stored as timestamped quadruples, not free text.