NS-ST-GraphRAG: Neuro-Symbolic Spatio-Temporal GraphRAG for Literary Knowledge Processing

Zheng Kui Lin

cs.CLSep 4, 2026

NS-ST-GraphRAG: Neuro-Symbolic Spatio-Temporal GraphRAG for Literary Knowledge Processing

Authors: Zheng Kui Lin

Organizations: Joint Laboratory for Port Big Data and Intelligent Applications, Dalian Ocean University, Dalian 116023, China

Abstract

Long-form literary narratives pose a distinctive information-processing challenge for retrieval-augmented generation: relevant evidence is distributed across chapters, relations evolve over narrative time, and correct answers may depend jointly on temporal, spatial, and relational constraints. We propose NS-ST-GraphRAG, a neuro-symbolic spatio-temporal GraphRAG framework that integrates ontology-guided extraction, deterministic constraint checking, dual temporal coordinates, spatial scene attributes, and dynamic sub-graph retrieval. Instead of retrieving from a single corpus-level graph, the framework selects the graph state valid for the temporal and spatial scope of a query and grounds generated answers in traceable evidence. We further introduce Red-Chamber-QA, to our knowledge the first open multi-hop question-answering benchmark for classical Chinese literature, with time-, space-, and general-question categories, per-part evidence spans, and deterministic shortcut controls. On a 120-question held-out split, NS-ST-GraphRAG achieves mechanical answer reproduction of 0.733 versus 0.675 for the frozen window baseline and 0.083 for a closed-book model (McNemar exact p = 0.092, directionally favorable but not significant); semantic-judge accuracy is 0.866 versus 0.850. The pre-specified constrained-category condition of H2 is not supported by the delivered comparison. These results show how temporal graph representation, constrained extraction, and auditable evaluation integrate into a unified framework for verifiable knowledge processing over long-form narrative.

Figures & tables

Figure 1: Overview of the NS-ST-GraphRAG pipeline, organized into three swimlanes. Module A (OGE + NSCL): a chapter

c_{k}

enters LLM extraction; candidate triples pass through the constraint layer

C(\mathcal{T})

; the dashed feedback arrow marks the bounded repair loop, which re-prompts extraction on violating spans at most

R_{max}

times, and accepted triples leave the module. Module B (STGC): accepted triples undergo entity resolution, receive dual temporal coordinates (chapter index

t_{c}

and narrative time

t_{n}

), and are stored as versioned slices

G

. Module C (DSGR): a query

Q

is parsed into temporal and spatial constraints

Q_{t},Q_{s}

by a lightweight parser; the matching slice

G_{Q}

is retrieved by index lookup; LLM generation produces the grounded answer with provenance. Solid arrows carry data flow; dashed arrows carry feedback.

Figure 2: The neuro-symbolic constraint layer (NSCL). Candidate triples

\mathcal{T}

from extraction are evaluated against the domain ontology

\mathcal{O}

, which branches into hard rules (genealogical, master-servant, and title/gender consistency) and soft rules (atypical patterns). The filtering function

C(\mathcal{T})

applies a binary reject/accept decision for hard rules and a penalty score against a fixed threshold for soft rules. Hard-rule rejections form a violation package carrying the offending rule and the evidence span, which feeds the dashed feedback loop back to extraction, capped at

R_{max}

repair passes; hard violations still present after the cap are written to the uncertainty log and excluded from the graph

G

. Soft-rule penalties above the threshold mark the triple with a low-confidence attribute instead of excluding it (not depicted).

Figure 3: Dual temporal coordinate system. The top axis is the coarse chapter index

t_{c}

with chapter markers (Ch 17, Ch 18, Ch 19); the dashed vertical lines mark chapter boundaries. The bottom axis is the fine narrative time

t_{n}

with story events as intervals: Event B (the Imperial Consort’s Visit) is drawn as a semi-transparent block deliberately overlapping the Ch 17 / Ch 18 boundary, illustrating why chapter indices alone cannot locate events. Below the axes, a knowledge graph edge

e

(Baoyu attends the Visit) carries a validity interval

\tau(e)

rendered as the red band; the dashed connectors anchor its endpoints to the narrative timeline so that

\tau(e)

aligns exactly with Event B’s extent. This alignment is the mechanism by which time-constrained queries select their slice

G_{Q}

Figure 4: The dual-track divergence. Semantic correctness (DPC-judge, light bars) masks mechanical-verification failure: among answers with valid judge verdicts, the text-augmented baselines achieve DPC-judge rates of

0.87

–

0.92

, while reproducing the golden fragments or evidence spans exactly on only

0.13

–

0.23

of questions (DPC-mech, dark bars). The oracle ceiling scores 1.0 on both tracks, sanity-checking the metric logic.

Figure 5: Citation-faithfulness failure taxonomy for the text-augmented baselines. MIXED (partially verbatim, partially paraphrased evidence) is the dominant failure class, showing that LLMs default to rewriting even under few-shot verbatim enforcement; NON-VERBATIM and ERROR complete the loss.

Figure 6: A time-constrained query traced through Module C using the running example “Who served Wang Xifeng in the Garden before the investigation?” (1) The parser extracts a spatial constraint

Q_{s}

= Garden and a temporal constraint

Q_{t}

= before the investigation. (2) Index lookup maps

Q_{t}

to a narrative-time interval and

Q_{s}

to a location ID. (3) Slice retrieval fetches the temporal slice

G_{Q}

and intersects it with the scene-to-entity spatial index. (4) Context assembly selects triples within the token budget, prioritizing shortest paths between the question’s entities so multi-hop chains survive truncation. (5) Grounded generation answers from the selected triples and returns provenance coordinates

(t_{c},t_{n},\sigma)

. No graph query language statement (e.g., Cypher) is generated at any step (illustrative trace, no numeric results).

Figure 7: Diagnostic protocol for error attribution. The root node is a QA failure. Level 1 stratifies the failed question by category: time-constrained, space-constrained, or general. Level 2 attributes the failure to one of four buckets: a parser error (wrong constraint

Q_{t}

Q_{s}

extracted), a retrieval error (correct parse, wrong slice

G_{Q}

, or budget truncation of an answer path), a generation error (correct slice, LLM hallucinated or failed to reason), or a constraint-filter over-rejection (a correct parse whose slice filter discards the answer path). Each bucket feeds its own metric, so a weak component cannot masquerade as another, and the H2/H3 conclusions remain attributable to the mechanism they test.

Figure 8: The timeline-window diagnostic on TQ-3: the same retriever, with the timeline window on, returns the gold evidence neighborhood (Xiuyan asking Miaoyu to divine, Chapter 95); with the window off, it returns an unrelated earlier scene (the Grandmother Jia and Granny Liu visiting Green Lattice Nunnery). Same retriever, only the timeline window filter differs.

Figure 9: Dual iteration trajectories on the same development set: the window baseline’s five mechanism configurations (Table 5 ) and the system’s five recorded retrieval rounds (Table 7 ). The interventions differ by construction—window composition for the baseline, graph retrieval for the system; only the final configurations are frozen for the campaign, and the system’s frozen R5 context is about seven thousand tokens of graph-anchored neighborhoods against the baseline’s two-thousand-token window.

Figure 10: Conceptual view of the evolution-analytics case study. The horizontal axis is narrative time; the dashed vertical line marks the Grand View Garden investigation, the dividing event. Before the event, the social network of Wang Xifeng (large central node) shows high centrality and a wide spatial range, drawn as dense connections to household members. After the event, her node is smaller and peripheral. A low-status node (marked “node?” in red) is drawn bridging two clusters after the event, representing hypothesis H4b: a hidden information channel maintained by a servant or messenger even as the mistress’s own centrality declines. The question mark is deliberate: the figure states an open, falsifiable hypothesis, not a claimed finding; no numeric values appear anywhere in the figure (illustrative only).

Graph Retrieval-Augmented Generation

Graph Retrieval-Augmented Generation connects language models to knowledge graphs to retrieve relational evidence for grounded, multi-hop answers.

See all 232 papers

Jun 4, 2026

Narrative Knowledge Weaver: Narrative-Centric Retrieval-Augmented Reasoning for Long-Form Text Understanding

Long-form narrative QA requires reasoning over evolving story worlds rather than isolated passages: answers may depend on earlier goals, changing character states, social relations, causal triggers, temporal position, and later consequences. Existing retrieval and graph-augmented generation methods improve evidence access, but their units--chunks, entities, relations, summaries, or tool actions--do not directly encode how evidence functions in a story. We introduce Narrative Knowledge Weaver(NKW), a source-grounded framework that aligns textual evidence, atomic facts, canonical graph structure, entity profiles, interactions, episodes, and storylines. At query time, NKW uses text, graph, and narrative tools with post-retrieval reading skills to assemble evidence and audit actor, scope, polarity, state, and temporal constraints. Across STAGE, FairytaleQA, and QuALITY, NKW is strongest on screenplay-level story-world QA while remaining competitive on more passage-centered benchmarks. Ablations, question-type analyses, graph-asset statistics, and case studies show complementary benefits for character, scene, temporal, causal, and narrative-progression reasoning.

Qiuyu Tian, Fengyi Chen, Yiding Li +9

Southeast University, Nanjing, China · Nanjing Normal University, Nanjing, China · ZhuiWen Technology Co., Ltd., Beijing, China +1

Retrieval-Augmented Generation Long-Context QA Question Answering Knowledge Representation Multi-Hop Reasoning

Ask PDF

Apr 15, 2026

TCAR-Gen: Temporal Graph Retrieval with Evidence Fusion for Knowledge-Grounded Generation

Retrieval-augmented generation systems struggle with temporal reasoning and evidence fusion when answering complex questions over historical criminal case narratives. Existing approaches either retrieve independently of query semantics or fail to integrate multiple evidence sources coherently. We propose Temporal Context Augmented Retrieval Generation (TCAR-Gen), a framework that combines query-conditioned graph neural networks, temporal evidence fusion, and chain-of-trees reasoning to ground answer generation in retrieved evidence. On the Victorian Crime Diaries benchmark, TCAR-Gen achieves 0.3738 Recall@5, outperforming Vanilla RAG, Temporal RAG, GraphRAG-C, and GraphRAG-T across seven query types including multi-hop reasoning and counterfactual questions. Ablation studies reveal that the context graph, temporal penalty mechanism, and query conditioning are critical components. Cross-model evaluation across five language model (GPT-OSS 20B to TinyLlama 1.1B) demonstrates that TCAR-Gen maintains robust retrieval coverage at smaller model scales, though generation quality degrades substantially with reduced model capacity. Our work shows that explicit temporal modelling and multi-branch evidence fusion are essential for faithful, reasoning-intensive question answering over knowledge-grounded corpora.

Sidra Nasir, Muhammad Noman Zahid, Rizwan Ahmed Khan

Dipartimento di Informatica, Università di Verona, Italy · School of Advanced Studies, University of Camerino, Italy · Department of Computer Science, School of Mathematics and Computer Science, Institute of Business Administration (IBA), Karachi, Pakistan

Graph Retrieval-Augmented Generation Temporal QA Graph-Based Retrieval Multi-Hop QA

Ask PDF

Apr 6, 2026

GROUNDEDKG-RAG: Grounded Knowledge Graph Index for Long-document Question Answering

Retrieval-augmented generation (RAG) systems have been widely adopted in contemporary large language models (LLMs) due to their ability to improve generation quality while reducing the required input context length. In this work, we focus on RAG systems for long-document question answering. Current approaches suffer from a heavy reliance on LLM descriptions resulting in high resource consumption and latency, repetitive content across hierarchical levels, and hallucinations due to no or limited grounding in the source text. To improve both efficiency and factual accuracy through grounding, we propose GroundedKG-RAG, a RAG system in which the knowledge graph is explicitly extracted from and grounded in the source document. Specifically, we define nodes in GroundedKG as entities and actions, and edges as temporal or semantic relations, with each node and edge grounded in the original sentences. We construct GroundedKG from semantic role labeling (SRL) and abstract meaning representation (AMR) parses and then embed it for retrieval. During querying, we apply the same transformation to the query and retrieve the most relevant sentences from the grounded source text for question answering. We evaluate GroundedKG-RAG on examples from the NarrativeQA dataset and find that it performs on par with a state-of-the art proprietary long-context model at smaller cost and outperforms a competitive baseline. Additionally, our GroundedKG is interpretable and readable by humans, facilitating auditing of results and error analysis.

Tianyi Zhang, Andreas Marfurt

getAbstract · Lucerne University of Applied Sciences and Arts Switzerland

Question Answering Document QA Graph Retrieval-Augmented Generation LLM Grounding KG Construction

Ask PDF

Abstract

Graph Retrieval-Augmented Generation

Explore similar work

Narrative Knowledge Weaver: Narrative-Centric Retrieval-Augmented Reasoning for Long-Form Text Understanding

TCAR-Gen: Temporal Graph Retrieval with Evidence Fusion for Knowledge-Grounded Generation

GROUNDEDKG-RAG: Grounded Knowledge Graph Index for Long-document Question Answering