ChronoSynth: Constraint-Aware Question Synthesis over Temporal Knowledge Graphs

Feng Liang; Weixin Zeng; Runhao Zhao; Mingjun Zhou; Shaoshi Yang; Xiang Zhao

cs.CLOct 21, 2026

ChronoSynth: Constraint-Aware Question Synthesis over Temporal Knowledge Graphs

Authors: Feng Liang, Weixin Zeng, Runhao Zhao, Mingjun Zhou, Shaoshi Yang, Xiang Zhao

Organizations: National Key Laboratory of Big Data and Decision, National University of Defense Technology, China · Beijing University of Posts and Telecommunications, China

Abstract

Knowledge graph question generation (KGQG) generates natural-language questions from structured graph evidence and a target answer. When the input graph is temporal, a valid question must also preserve the answer field, time condition, reference event, and temporal order that determine the answer set. Existing KGQG methods and prompting baselines can produce fluent questions that alter these conditions, causing temporal semantic drift. This paper studies temporal KGQG as a constraint-preserving graph-to-question problem and proposes ChronoSynth, a memory-augmented and constraint-aware generation framework. ChronoSynth first constructs a Temporal Intent State to make the intended question semantics explicit. It then retrieves compatible typed question patterns from an indexed pattern memory, adapts them to the input graph, and validates the generated question. Validation feedback guides revisions when a supported inconsistency is detected. Experiments on CRONQUESTIONS and MultiTQ demonstrate state-of-the-art generation performance, with ablations supporting the contributions of its components.

Figures & tables

Figure 1: Motivating examples of temporal semantic drift. The prompting baselines are fluent but change the time condition, reference event, answer field, or order cue.

Figure 2: Overview of ChronoSynth. ChronoSynth constructs a Temporal Intent State, retrieves compatible typed question patterns from an indexed pattern memory, adapts them to the input graph, and validates and refines the generated question.

CRONQUESTIONS
Setting	Method	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE-L	CIDEr
Trained	G2S-AE	0.5011	0.3873	0.3102	0.2577	0.4799	0.2480
	G2S-AE-RL	0.3852	0.2628	0.1792	0.1261	0.3850	0.0625
	RoleAgentQG-G2S	0.3743	0.2771	0.2123	0.1703	0.3686	0.1037
	RoleAgentQG-G2S-AE-RL	0.3842	0.2861	0.2205	0.1778	0.3772	0.1110
Training-free	Qwen3.5-9B	0.4237	0.3122	0.2417	0.1952	0.4029	1.8729

Table 1: Question-generation results on CRONQUESTIONS and MultiTQ. Bold and underlined values denote the best and second best. Relative improvements compare the best ChronoSynth score with the strongest baseline for each metric.

Figure 3: Component ablations on CRONQUESTIONS and MultiTQ. F: complete ChronoSynth; W: w/o Refinement; R: Relation-only Retrieval; M: w/o Pattern Memory.

CRONQUESTIONS
Method	Fluency	Temporal faithfulness	Answerability
GPT-4o-mini	2.313	2.287	2.340
KQG-CoT	2.413	2.360	2.433
Retrieval prompting	2.380	2.360	2.373
ChronoSynth	2.987	2.973	2.967
MultiTQ

Table 2: Human evaluation on 50 matched inputs per dataset. Scores average three raters on a 1–5 scale; bold marks the highest mean.

Training data	CRON- QUESTIONS	MultiTQ
Gold Reference	$98.33\pm 0.42$	$92.02\pm 0.50$
GPT-4o-mini	$92.36\pm 2.88$	$73.36\pm 2.31$
KQG-CoT	$95.15\pm 1.22$	$76.90\pm 1.54$
Vanilla RAG	$88.30\pm 0.64$	$77.71\pm 2.07$
ChronoSynth	$\mathbf{95.49\pm 2.05}$	$\mathbf{86.37\pm 0.10}$

Table 3: Downstream QA EM (%) with Qwen3.5-0.8B. Values are mean

\pm

sample standard deviation across three seeds. Bold marks the best synthetic-data mean.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Temporal rule	Trigger condition	Semantics
plain	entity answer, no explicit temporal comparison	ask a non-extremum factual question
when	time answer, no explicit reference	ask for the event timestamp
first	focus edge has the earliest time in its same-relation family	ask for the earliest event or participant
last	focus edge has the latest time in its same-relation family	ask for the latest event or participant
before	later reference exists	ask for an answer before a reference event
before_last	later reference and latest valid answer before it	ask for the last answer before a reference event

Appendix

Table A1: Temporal-rule inventory.

Field	Meaning	Main use
$a_{x}$	answer field	decides whether the question asks for the source, target, or time field
$y_{x}$	answer type	constrains the wh-word and surface form
$\omega_{x}$	temporal rule	controls temporal semantics such as before , after , and last
$g_{x}$	time granularity	distinguishes year-, month-, and day-level realization
$e_{x}$	focus edge	identifies the fact to verbalize
$h_{x}$	reference entity/event	provides the reference point for retrieval and drafting

Appendix

Table A2: Fields in the Temporal Intent State.

Dataset	Train	Dev	Test
CRONQUESTIONS	212,259	19,014	19,122
MultiTQ	257,687	39,128	36,318

Appendix

Table A3: Numbers of graph-converted records.

Resource	Use	Source
CRONQUESTIONS	temporal benchmark	CronKGQA Saxena et al. (2021)
MultiTQ	temporal benchmark	MultiTQ Chen et al. (2023)
Qwen3.5-9B	local generation	Qwen model card
Qwen3.5-0.8B	downstream reader	Qwen model card
GPT-4o-mini	generation and interpretation	OpenAI
GPT-4.1-mini	supplementary judge	OpenAI

Appendix

Table A4: External resources and their experimental roles.

Temporal rule
Group	Count	$\Delta$ BLEU-4	$\Delta$ ROUGE-L
after_first	1,494	+0.2253	+0.3688
before	27	+0.2256	+0.2837
before_last	1,907	+0.1861	+0.1670
first	1,030	+0.0928	+0.1244
last	879	+0.1095	+0.1355

Appendix

Table A5: Grouped gains on CRONQUESTIONS. Differences are complete ChronoSynth minus w/o Pattern Memory.

Temporal rule
Group	Count	$\Delta$ BLEU-4	$\Delta$ ROUGE-L
after	675	+0.1627	+0.1454
after_first	2,784	+0.1168	+0.1160
after_time	1,331	+0.1152	+0.0948
before_last	8,620	+0.0898	+0.0699
before_time	1,441	+0.1162	+0.0906

Appendix

Table A6: Grouped gains on MultiTQ. Differences are complete ChronoSynth minus w/o Pattern Memory.

Score	Fluency	Temporal faithfulness	Answerability
5	Natural, clear, and grammatical	All required temporal conditions are preserved	Clear question with the correct complete target answer set
4	Minor language issues	Minor issue that does not materially change the temporal meaning	Minor ambiguity
3	Understandable but noticeably awkward	Temporal intent is recognizable but incomplete or ambiguous	Meaningful ambiguity
2	Major language problems	An important temporal component is wrong or missing	Poorly answerable or likely to yield another answer
1	Incoherent or invalid question	Core temporal meaning is reversed, lost, or replaced	Unanswerable, wrong target, or invalid question

Appendix

Table A7: The annotation rubric used in human evaluation.

Figure A1: Generation quality, throughput, and token use with local Qwen3.5-9B.

Figure A2: BLEU-4 versus total tokens per question. Arrows connect KQG-CoT to ChronoSynth.

Source subset	Train	Test
Cron-S	4,172	1,041
Cron-M	1,290	309
Event-S	4,955	1,236
Event-M	1,742	441
Total	12,159	3,027

Appendix

Table A8: ChronoQG data split. S/M denote single/multiple temporal constraints.

Method	BLEU-1	BLEU-2	BLEU-3	BLEU-4	ROUGE-L	CIDEr
Direct Prompting	0.4710	0.3836	0.3220	0.2761	0.4625	2.3437
KQG-CoT	0.3987	0.3339	0.2892	0.2557	0.4339	2.6055
Vanilla RAG	0.5625	0.4832	0.4271	0.3834	0.5711	3.4138
ChronoSynth	0.6201	0.5495	0.4979	0.4562	0.6472	4.3373

Appendix

Table A9: Question generation on ChronoQG with local Qwen3.5-9B. Bold marks the highest score.

Method	Time (min)	Tokens /question	Questions /second
Direct Prompting	2.53	1,333	19.91
KQG-CoT	37.82	14,106	1.33
Vanilla RAG	6.49	3,682	7.77
ChronoSynth	51.65	5,012	0.98

Appendix

Table A10: Measured inference cost on ChronoQG.

Knowledge Graphs

Knowledge graphs represent entities and relationships as structured, queryable networks for integrating data and supporting reasoning.

See all 232 papers

ChronoQG: Towards a Temporally Expressive and Hop-Bounded Benchmark for Temporal Knowledge Graph Question Generation
Jul 16, 2026Xuemeng Liu, Zhengpin Li, Wanpeng Tang +2Synthetic Benchmark Generation Temporal KGs
Peking University Beijing, China · Shanghai University of Finance and Economics Shanghai, China
SABET-QA: Temporal Knowledge Graph Question Answering
Aug 20, 2026Brahim Touayouch, Mirette Moawad, Dmitry AkulovKnowledge Graph Question Answering Temporal QA
QuickSort Research, Paris, France · ENS Paris-Saclay, École Polytechnique, France
Temporal-Aware Heterogeneous Graph Reasoning with Multi-View Fusion for Temporal Question Answering
Feb 23, 2026Wuzhenghong Wen, Bowen Zhou, Jinwen Huang +5Knowledge Graph Question Answering Temporal QA
School of Internet of Things, Nanjing University of Posts and Telecommunications · Faculty of Medicine & Health, University of New South Wales, Sydney, Australia · Beijing Information Science and Technology University +2

Jul 16, 2026J/K move · Enter open · S save

ChronoQG: Towards a Temporally Expressive and Hop-Bounded Benchmark for Temporal Knowledge Graph Question Generation

Xuemeng Liu, Zhengpin Li, Wanpeng Tang, Haotong Xie+1

Peking University Beijing, China · Shanghai University of Finance and Economics Shanghai, China

Synthetic Benchmark Generation Temporal KGs

Open paper AskPDF arXiv

Knowledge graph question generation (KGQG) aims to generate natural-language questions from structured graph evidence. Existing KGQG benchmarks, however, are mostly built on static knowledge graphs and do not encode the temporal scopes of graph facts. As a result, they cannot evaluate whether generated questions faithfully preserve temporal validity, event ordering, and answer-determining temporal constraints. In this paper, we study temporal knowledge graph question generation (TKGQG), where a generated question must be faithful to both the support subgraph and the temporal constraints required to identify the target answer. We propose ChronoQG, the first temporally expressive and hop-bounded benchmark construction framework for TKGQG. ChronoQG integrates a comprehensive temporal-constraint taxonomy, topology-temporal subgraph sampling, and trace-grounded question generation to construct temporally faithful questions. The framework produces four benchmark datasets from heterogeneous temporal knowledge graphs, totaling 16,011 verified questions. We evaluate representative LLM-based KGQG methods and prompting baselines across diverse TKGQG settings, including temporal-constraint counts, topological templates, and temporal-constraint types. The results show that existing methods struggle to preserve temporal constraints, especially under multi-constraint settings and harder temporal-constraint types. These findings reveal a clear gap between static KGQG and TKGQG, and establish ChronoQG as a challenging testbed for temporally faithful question generation.

Abstract

Knowledge Graphs

Explore similar work

ChronoQG: Towards a Temporally Expressive and Hop-Bounded Benchmark for Temporal Knowledge Graph Question Generation

SABET-QA: Temporal Knowledge Graph Question Answering

Temporal-Aware Heterogeneous Graph Reasoning with Multi-View Fusion for Temporal Question Answering