Respecting Temporal-Causal Consistency: Entity-Event Knowledge Graphs for Retrieval-Augmented Generation

📝 Paper Summary

Graph-based RAG pipeline Complex question answering Temporal reasoning

E2RAG improves question answering on narrative texts by constructing separate entity and event graphs linked by a bipartite mapping, preserving the chronological context that standard RAG and KG-RAG methods lose.

Core Problem

Existing RAG methods falter on narrative documents because unstructured retrieval misses temporal structure, while standard Knowledge Graph RAG collapses entity mentions into single nodes, erasing the evolving chronological context needed for reasoning.

Why it matters:

Novels, biographies, and legal histories rely heavily on the timeline of events; flattening this structure makes answering 'what happened after X?' impossible
Current KG-RAG approaches merge all information about an entity (e.g., 'Napoleon') into one node, losing the distinction between 'Napoleon in 1804' and 'Napoleon in 1815'
Standard embedding-based retrieval struggles with causal queries where the answer depends on a specific sequence of prior events rather than just semantic similarity

Concrete Example: In a mystery novel, if a character is innocent in Chapter 1 but commits a crime in Chapter 10, a standard KG merges these facts, confusing the model. E2RAG maintains separate event nodes linked by time, allowing it to correctly identify the character's status at a specific moment.

Key Novelty

Entity-Event Dual-Graph RAG (E2RAG)

Constructs two distinct subgraphs: an Entity Graph for static relationships and an Event Graph for chronological actions
Links these graphs via a bipartite mapping, where entities participate in specific events
Retains temporal order in the Event Graph, allowing the system to traverse the narrative timeline to answer causal and temporal questions

Evaluation Highlights

Outperforms state-of-the-art unstructured and KG-based RAG baselines across the ChronoQA benchmark
Achieves notable gains on causal and character consistency queries specifically
Demonstrates robust performance on long-context narrative understanding where temporal sequence is critical

Breakthrough Assessment

8/10

Significant advance in handling temporal structure in RAG, a known weakness of current systems. The dual-graph approach offers a structural solution to the 'context collapse' problem in standard KGs.

⚙️ Technical Details

Problem Definition

Setting: Question answering over long-form narrative documents with inherent temporal structures (e.g., novels, plays)

Inputs: Natural language question q and a narrative document corpus D

Outputs: Answer a based on the temporal and causal logic within D

Pipeline Flow

Graph Construction (Entity Graph + Event Graph)
Query Processing (Identify entities and temporal constraints)
Dual-Graph Retrieval (Traverse Event Graph for timeline, Entity Graph for details)
Answer Generation (Synthesize retrieved paths into answer)

System Modules

Graph Constructor

Parse narrative text to build Entity Subgraph (static facts) and Event Subgraph (temporal actions), linking them via participation edges

Model or implementation: LLM-based extractor (e.g., GPT-4 or specialized IE model)

Dual-Graph Retriever

Traverse the linked graphs to find relevant event chains and entity states matching the query

Model or implementation: Graph traversal algorithm

Generator

Generate natural language answer from the retrieved graph context

Model or implementation: LLM (e.g., GPT-4, Llama-3)

Novel Architectural Elements

Dual-graph topology separating 'what is' (Entities) from 'what happens' (Events) to prevent context collapse
Explicit bipartite linking mechanism connecting static entities to dynamic event nodes

Modeling

Base Model: Not explicitly reported in the paper summary (generic LLM used for generation)

Compute: Not reported in the paper

Comparison to Prior Work

vs. Unstructured RAG: E2RAG explicitly models time and causality, whereas unstructured RAG relies on semantic similarity which often fails for temporal ordering
vs. KG-RAG: Standard KG-RAG merges all entity info into one node; E2RAG spreads entity states across time-linked event nodes
vs. GraphRAG: GraphRAG focuses on global summary via community detection; E2RAG focuses on precise temporal traversal for narrative QA [not cited in paper]

Limitations

Construction of the dual graph is computationally expensive and requires high-quality extraction
Dependency on the quality of the underlying LLM for event extraction; errors in extraction propagate to retrieval
May struggle with extremely long dependencies if the event chain becomes too sparse

Reproducibility

Code: https://github.com/Tencent/ChronoQA

Code and benchmark data for ChronoQA are publicly available at https://github.com/Tencent/ChronoQA. The paper describes the graph construction method, but specific prompt templates for extraction may need to be retrieved from the repo.

📊 Experiments & Results

Evaluation Setup

QA over narrative documents requiring temporal and causal reasoning

Benchmarks:

ChronoQA (Temporal/Causal Narrative QA) [New]

Metrics:

Accuracy
Temporal Consistency Score
Causal Consistency Score
Statistical methodology: Not explicitly reported in the paper

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
E2RAG outperforms baselines on the newly constructed ChronoQA benchmark, particularly on questions requiring temporal reasoning.
ChronoQA	Accuracy	Not reported in the paper	Not reported in the paper	Positive (Qualitative)

Main Takeaways

Standard KG-RAG fails on narratives because it collapses time; E2RAG's event graph preserves it.
Unstructured RAG lacks the mechanism to reason about 'before' and 'after' relationships in complex stories.
Separating entities and events into dual graphs allows for more precise retrieval of character states at specific time points.

📚 Prerequisite Knowledge

Prerequisites

Knowledge Graphs (nodes, edges, relations)
Retrieval-Augmented Generation (RAG) basics
Graph Neural Networks (GNNs) or graph traversal algorithms

Key Terms

RAG: Retrieval-Augmented Generation—AI systems that answer questions by first searching for relevant documents

KG-RAG: Knowledge Graph RAG—using structured knowledge graphs instead of plain text chunks for retrieval

E2RAG: Entity-Event RAG—the proposed framework using dual graphs for entities and events

bipartite mapping: A connection scheme where nodes in one set (Entities) connect only to nodes in the other set (Events), effectively linking actors to their actions

unstructured RAG: Standard RAG that retrieves raw text chunks based on vector similarity

temporal consistency: The requirement that answers must respect the chronological order of events in the story

causal consistency: The requirement that answers must respect cause-and-effect relationships (e.g., X happened because Y happened first)