ChunQiuTR: Time-Keyed Temporal Retrieval in Classical Chinese Annals

Yihao Wang; Zijian He; Jie Ren; Keze Wang

cs.CLApr 8, 2026

ChunQiuTR: Time-Keyed Temporal Retrieval in Classical Chinese Annals

Authors: Yihao Wang, Zijian He, Jie Ren, Keze Wang

Organizations: Sun Yat-Sen University · Shaanxi Normal University

Abstract

Retrieval shapes how language models access and ground knowledge in retrieval-augmented generation (RAG). In historical research, the target is often not an arbitrary relevant passage, but the exact record for a specific regnal month, where temporal consistency matters as much as topical relevance. This is especially challenging for Classical Chinese annals, where time is expressed through terse, implicit, non-Gregorian reign phrases that must be interpreted from surrounding context, so semantically plausible evidence can still be temporally invalid. We introduce ChunQiuTR, a time-keyed retrieval benchmark built from the Spring and Autumn Annals and its exegetical tradition. ChunQiuTR organizes records by month-level reign keys and includes chrono-near confounders that mirror realistic retrieval failures. We further propose CTD (Calendrical Temporal Dual-encoder), a time-aware dual-encoder that combines Fourier-based absolute calendrical context with relative offset biasing. Experiments show consistent gains over strong semantic dual-encoder baselines under time-keyed evaluation, supporting retrieval-time temporal consistency as a key prerequisite for faithful downstream historical RAG. Our code and datasets are available at github.com/xbdxwyh/ChunQiuTR.

Figures & tables

Figure 1: A query about a specific month can retrieve same-month commentary that repeats the date phrase, or adjacent-month near-miss events with confusable wording, so a retrieval-augmented model answers fluently but at the wrong time.

Figure 2: Overview of the ChunQiuTR construction. ( Left ) Time-key alignment yields event-level records, augmented with chrono-near counterfactual hard negatives from later histories. ( Right ) We define P/G/W temporal queries and leak-free, reign-aligned splits.

Figure 3: Overview of our Calendrical Temporal Dual-encoder (CTD). A shared Transformer dual-encoder encodes queries and records into embeddings. Temporal heads place each text on a unified regnal calendar axis (latent time scalar), supporting (a) absolute context injection and (b) relative biasing to form $s^{\text{CTD}}_{ij}$ . (c) Interval-overlap supervision marks in-batch multi-positives by query–record overlap and trains a multi-positive contrastive loss.

Sparse retrieval
Method	Pub	Type	R@1	R@5	R@10	MRR@10	nDCG@10
BM25	–	Sparse	0.3962	0.5209	0.5620	0.4487	0.3404
BM25+TimeKDE	–	Sparse	0.4943	0.6086	0.6709	0.5456	0.4222
SPLADE-IDF ${}_{(\text{ZS})}$	arXiv’24	Sparse	0.1361	0.2765	0.3569	0.1971	0.1596
SPLADE- $\ell_{0}$ ${}_{(\text{ZS})}$	SIGIR’25	Sparse	0.0006	0.0309	0.0587	0.0132	0.0143
Fusion / late interaction

Table 1: Test-set retrieval performance on our ChunQiuTR benchmark under the official evaluation protocol.

Subset	Records	Queries	FT baseline		CTD (ours)
Subset	Records	Queries	MRR	R@1	MRR	R@1
Qi Ji (part)	268	92	0.2081	0.1848	0.2304	0.2065
Jin Ji (part)	820	119	0.1598	0.1345	0.1751	0.1597

Table 2: Cross-corpus pilot on processed Zizhi Tongjian subsets. No target-corpus training is performed.

Method	Single-month ( $\|Q\|=1$ )		Multi-month ( $\|Q\|>1$ )
	R@1	MRR@10	R@1	MRR@10
BM25	0.397	0.413	0.396 $\downarrow$ -0.001	0.499 $\uparrow$ +0.086
ColBERT-LFM2 ${}_{(\text{ZS})}$	0.298	0.312	0.386 $\uparrow$ +0.088	0.490 $\uparrow$ +0.178
mE5-Large ${}_{(\text{ZS})}$	0.259	0.279	0.337 $\uparrow$ +0.078	0.422 $\uparrow$ +0.143
BERT-base ${}_{(\text{FT})}$	0.497	0.516	0.525 $\uparrow$ +0.027	0.621 $\uparrow$ +0.106
CTD ${}_{\text{BERT-base}}$ (Ours)	0.509	0.530	0.685 $\uparrow$ +0.176	0.744 $\uparrow$ +0.214

Table 3: Impact of query span on retrieval performance on the test set. We compare single-month (

|Q|=1

) and multi-month (

|Q|>1

) queries; for multi-month queries, the numbers in parentheses give absolute changes relative to single-month queries.

Figure 4: Visualization of Qualitative Examples.

Variant	$\mathcal{L}_{\text{multi}}$	$b^{\text{time}}_{ij}$	$\mathbf{c}_{x}$	R@1	MRR@10
FT baseline	–	–	–	0.5771 —-	0.6044 —-
+ $\mathcal{L}_{\text{multi}}$	$\checkmark$	–	–	0.5820 $\uparrow$ +0.0049	0.6107 $\uparrow$ +0.0063
+ Bias	$\checkmark$	$\checkmark$	–	0.5898 $\uparrow$ +0.0127	0.6135 $\uparrow$ +0.0091
+ Ctx	$\checkmark$	–	$\checkmark$	0.5850 $\uparrow$ +0.0079	0.6134 $\uparrow$ +0.0090
Full (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	0.5923 $\uparrow$ +0.0152	0.6194 $\uparrow$ +0.0150

Table 4: Ablation study on the test set under the same evaluation protocol as Table 1 . For each metric, the second line reports the change relative to the FT baseline.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Text (excerpt + short gloss)	Source	Label
夏，五月，郑伯克段于鄢。
Annals anchor: Zheng subdued Duan at Yan in month 5.	《春秋》 Chunqiu 经文 / annals	event
夏，五月，郑伯克段于鄢。克之者何？杀之也。杀之则曷为谓之克？大郑伯之恶也。段者何？郑伯之弟也。何以不称弟？当国也。
Gongyang reads ke as killing and amplifies moral blame.	《春秋公羊传》 Gongyang zhuan 传文 / zhuan	event
夏，五月，郑伯克段于鄢。克者何？能也。何能也？能杀也。何以不言杀？见段之有徒众也。段，郑伯弟也……段失子弟之道矣，贱段而甚郑伯也。
Guliang stresses armed revolt, conduct, and intensified censure.	《春秋谷梁传》 Guliang zhuan 传文 / zhuan	event

Appendix

Table 5: Aligned materials under the reign-based time key “鲁隐公元年五月” for “郑伯克段于鄢”. The annals give the anchor, the three zhuan provide aligned expansions, and later sources yield chrono-near non-target paraphrases.

Original chronicle snippet	Cue / update	Normalized time key
元年，春，王正月。	Initialize a new reign-year (隐公元年), month = 正月	鲁隐公元年正月
三月，公及邾仪父盟于蔑。	New month (三月), inherit current reign-year	鲁隐公元年三月
二年，春，公会戎于潜。	New year (二年) under the same ruler, reset month to 正月	鲁隐公二年正月
元年，春，王正月，公即位。	New ruler detected (桓公), reset year to 元年 and month to 正月	鲁桓公元年正月

Appendix

Table 6: Representative mappings from raw chronicle phrases to normalized reign-based time keys.

Figure 5: Example of event-level grouping under the reign-based time key 「鲁隐公元年十二月」. The model is prompted to split mixed passages and group aligned commentary snippets by event.

Figure 6: LLM-assisted reverse matching from paraphrastic event titles in Lü Zuqian’s Chunqiu ZuoZhuan Shuos (《春秋左氏传说》) to Zuo zhuan passages. We show the full text-only prompt given to a classical-Chinese LLM (DeepSeek), together with one concrete example: for ruler Yin of Lu, year 1, and the subtitle “祭仲谏郑庄封叔段” from Chunqiu ZuoZhuan Shuos , the model proposes the most likely Zuo zhuan passage. The model is required to output only the original Zuo text, or the sentinel token NONE when it cannot decide.

Figure 7: Representative temporal query types in ChunQiuTR, including point queries, past/future windows, around windows, and explicit ranges over reign-based month keys.

Template groups (index range)	Type
BASE (1–12)	point, content-oriented
BASE (13–20)	point, existence / no-event
MONTH_PAST (21–26)	window, past
MONTH_FUTURE (27–31)	window, future
MONTH_AROUND (32–36)	window, around
MONTH_RANGE (37–41)	window, range

Appendix

Table 7: Query template groups used in ChunQiuTR.

Pool	Source	# Raw units	Total chars (K)	Avg. len.	Median	Min	Max
Positive	All (annals + three zhuan )	6641	293.7	44.22	–	–	–
Positive	Chunqiu annals	1532	19.2	12.52	10	2	71
Positive	Gongyang zhuan	1776	44.5	25.05	10	2	671
Positive	Zuo zhuan	1547	189.1	122.20	48	3	2658
Positive	Guliang zhuan	1786	41.0	22.93	11	2	467
Negative	All exegetical layers	9227	1014.9	109.99	–	–	–

Appendix

Table 8: Raw record-level character-length statistics before sentence splitting. Character counts are reported in thousands (K). Positive records come from the Chunqiu annals and the three traditional zhuan , while negative records come from later exegetical layers.

Figure 8: Month-level coverage, gap counts, and gap ratios for the normalized Chunqiu timeline (overall and per Lu ruler).

Split	# months	# records	# queries	Avg. ground-truth recs/query	# event recs	# no-event recs	# neg. comments
Train	2424	16027	13053	7.3	5360	1209	9458
Validation	295	2049	1520	6.8	626	152	1271
Test	317	2096	1653	7.2	782	149	1165
Total	3036	20172	16226	7.2	6768	1510	11894

Appendix

Table 9: Final benchmark statistics and splits over month-level time keys, record-level retrieval units, and queries. The “Avg. ground-truth recs/query” column reports the average number of labeled relevant records per query in each split, and the last three columns break down records by type ( event , no_event , and neg_comment ).

Work / layer	Role in ChunQiuTR	Digital source (edition)	License note / release plan
Chunqiu (春秋)	base corpus	Chinese Wikisource ( Siku Quanshu edition; page revision oldid recorded)	Work pages: tagged PD-old . Platform text: CC BY-SA 4.0 (Wikimedia Terms of Use). We release derived metadata (time keys, alignments, queries/qrels, indices) and scripts to re-fetch raw texts from recorded oldid revisions.
Zuo zhuan (左氏傳)	base corpus
Gongyang zhuan (公羊傳)	base corpus
Guliang zhuan (穀梁傳)	base corpus
Gu Donggao, Chunqiu Dashibiao (顧栋高《春秋大事表》)	chrono-near paraphrastic negatives
Wei Liaoweng, Chunqiu Zuozhuan Yaoyi (魏了翁《春秋左傳要義》)	chrono-near discursive negatives

Appendix

Table 10: Text sources and licensing. All digitized texts are retrieved from Chinese Wikisource ( Siku Quanshu editions), with page revision IDs ( oldid ) recorded for traceability. See Appendix A for alignment and preprocessing.

Backbone	Variant	#GPU	Time / run	GPU-hours
BERT-base-chinese	FT baseline	1	$\approx$ 15 min	$\approx$ 0.25
BERT-base-chinese	CTD (full)	1	$\approx$ 19 min	$\approx$ 0.32
Qwen3-Embed-0.6B	FT baseline	2	$\approx$ 45 min	$\approx$ 1.50
Qwen3-Embed-0.6B	CTD (full)	2	$\approx$ 45 min	$\approx$ 1.50

Appendix

Table 11: Compute cost for representative fine-tuning runs on ChunQiuTR. Time is wall-clock per run; GPU-hours are computed as (#GPU)

\times

(time in hours).

Variant	$\mathcal{L}_{\text{multi}}$	$b^{\text{time}}_{ij}$	$\mathbf{c}_{x}$	R@1	MRR@10
FT baseline	–	–	–	0.5088 —-	0.5597 —-
+ $\mathcal{L}_{\text{multi}}$	$\checkmark$	–	–	0.5178 $\uparrow$ +0.0090	0.5685 $\uparrow$ +0.0088
+ Bias	$\checkmark$	$\checkmark$	–	0.5384 $\uparrow$ +0.0296	0.5776 $\uparrow$ +0.0179
+ Ctx	$\checkmark$	–	$\checkmark$	0.5620 $\uparrow$ +0.0532	0.5961 $\uparrow$ +0.0364
Full (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	0.5826 $\uparrow$ +0.0738	0.6193 $\uparrow$ +0.0596

Appendix

Table 12: Ablation on the test set with bert-base-chinese .

Method	Including pure no-event queries (fix neg =0, ne =1)						Hard-negative robustness (fix ne=1, dq=1 )
Method	dq=1		dq=0		$\Delta$		neg=0		neg=1		$\Delta$
	R@1	MRR@10	R@1	MRR@10	R@1	MRR@10	R@1	MRR@10	R@1	MRR@10	R@1	MRR@10
BM25	0.240	0.296	0.418	0.470	$\uparrow$ +0.178	$\uparrow$ +0.173	0.240	0.296	0.228	0.282	$\downarrow$ -0.012	$\downarrow$ -0.014
ColBERT-LFM2 ${}_{(\text{ZS})}$	0.220	0.277	0.333	0.386	$\uparrow$ +0.113	$\uparrow$ +0.109	0.220	0.277	0.222	0.278	$\uparrow$ +0.002	$\uparrow$ +0.001
BERT-base ${}_{(\text{FT})}$	0.378	0.439	0.560	0.606	$\uparrow$ +0.182	$\uparrow$ +0.167	0.378	0.439	0.306	0.374	$\downarrow$ -0.072	$\downarrow$ -0.065
CTD ${}_{\text{BERT-base}}$ (Ours)	0.407	0.456	0.583	0.619	$\uparrow$ +0.175	$\uparrow$ +0.163	0.407	0.456	0.407	0.456	0.000	0.000

Appendix

Table 13: No-event and hard-negative behavior under protocol variations on the test set. Left: dq=1 vs. dq=0 (fix ne =1, neg =0), where dq drops or keeps pure no_event queries. Right: neg=0 vs. neg=1 (fix ne =1, dq =1), where neg injects neg_comment passages into the gallery.

\Delta

denotes the within-method change.

Mode	Family	Validation					Test
Mode	Family	R@1	R@5	R@10	MRR@10	nDCG@10	R@1	R@5	R@10	MRR@10	nDCG@10
neg0_ne0_dq1	all	0.0407	0.1200	0.2232	0.0798	0.0635	0.0654	0.1466	0.2042	0.1014	0.0677
neg0_ne0_dq1	point	0.0375	0.1148	0.2459	0.0775	0.0756	0.0710	0.1696	0.2387	0.1156	0.0934
neg0_ne0_dq1	window	0.0430	0.1239	0.2065	0.0815	0.0546	0.0610	0.1283	0.1768	0.0902	0.0472
neg0_ne1_dq0	all	0.6329	0.6697	0.7191	0.6522	0.4921	0.5935	0.6497	0.6945	0.6206	0.4588
neg0_ne1_dq0	point	0.5200	0.5440	0.6057	0.5347	0.5380	0.4922	0.5307	0.5869	0.5138	0.5103

Appendix

Table 14: Full protocol grid results for CTD - Qwen3-Embed-0.6B . Mode names follow neg / ne / dq as defined in Section B.2.3 .

Mode	Family	Validation					Test
Mode	Family	R@1	R@5	R@10	MRR@10	nDCG@10	R@1	R@5	R@10	MRR@10	nDCG@10
neg0_ne0_dq1	all	0.0159	0.0456	0.0952	0.0331	0.0205	0.0096	0.0366	0.0672	0.0234	0.0192
neg0_ne0_dq1	point	0.0258	0.0468	0.0937	0.0382	0.0265	0.0079	0.0394	0.0690	0.0226	0.0235
neg0_ne0_dq1	window	0.0086	0.0448	0.0964	0.0294	0.0161	0.0110	0.0344	0.0657	0.0240	0.0158
neg0_ne1_dq0	all	0.4474	0.5678	0.5980	0.4989	0.3794	0.4180	0.5408	0.5735	0.4696	0.3543
neg0_ne1_dq0	point	0.4434	0.4709	0.4789	0.4556	0.4612	0.4152	0.4475	0.4506	0.4284	0.4339

Appendix

Table 15: Full protocol grid results for BM25, reported in the same format as Table 14 .

Figure 9: Top-1 near-miss cases on ChunQiuTR (test set). We show the top–5 results from a baseline (left) and ours (middle), with the ground-truth set (right). $\star$ marks ground-truth passages; ✓indicates a ground-truth hit in top–5.

Figure 10: Online LLM without evidence grounding on a month-level point query from ChunQiuTR. For the query “鲁隐公元年十二月”, the gold month contains two entries ( 祭伯来 and 公子益师卒 ). Without evidence, the model either predicts an empty month or returns an incomplete answer.

Figure 11: Evidence-bounded RAG for the same query as Fig. 10 . With a small evidence pack containing the gold month records and confusable materials, the model recovers both gold entries and grounds the answer in cited evidence.

Subset	Approx. coverage	Representative reign titles	Records	Queries
Qi Ji (part)	479–489 CE	建元, 永明	268	92
Jin Ji (part)	265–279 CE	泰始, 咸宁	820	119

Appendix

Table 16: Basic statistics of the processed Zizhi Tongjian subsets used in the cross-corpus pilot.

Model / Setting	Qi Ji (part)				Jin Ji (part)
Model / Setting	MRR	R@1	R@5	R@10	MRR	R@1	R@5	R@10
Qwen3-Embed-0.6B (ZS)	0.0692	0.0217	0.0870	0.1413	0.0691	0.0420	0.0756	0.1345
Qwen3-Embed-0.6B (FT baseline)	0.2081	0.1848	0.2174	0.2391	0.1598	0.1345	0.1849	0.1849
CTD (Ours)	0.2304	0.2065	0.2391	0.2717	0.1751	0.1597	0.1849	0.1849

Appendix

Table 17: Cross-corpus pilot results on two processed Zizhi Tongjian subsets. No target-corpus training is performed.

Item	Value
Total non-empty months	1,533
Months containing multiple events	558
Fraction multi-event	36.41%
Extra human corrections	63
Correction rate among multi-event months	11.29%
Corrections among all non-empty months	4.11%

Appendix

Table 18: Audit statistics for multi-event splitting (LLM proposals + human verification).

Source	#Candidates	Accepted	Rejected	Acceptance
Gu Donggao	899	899	0	100.00%
Kong Yingda	5,286	5,179	107	97.98%
Du Yu	5,373	5,266	107	98.01%
Lü Zuqian	360	336	24	93.33%

Appendix

Table 19: Acceptance rates for later-commentary alignments (LLM candidate proposal + human verification).

Temporal IR

Temporal information retrieval finds information relevant to when events occurred or facts were valid, not just what they describe.

See all 65 papers

Bridging Static and Agentic RAG for Taiwanese Historical Question Answering
Sep 19, 2026Kai-Hsin Chen, Wei-Yu Chen, Xuanjun Chen +1Retrieval-Augmented Generation Agentic RAG
Graduate Institute of Networking and Multimedia, National Taiwan University · Dept. of Computer Science & Information Engineering, National Taiwan University · Graduate Institute of Communication Engineering, National Taiwan University
TEMPS: Temporal Sentence Embeddings for Temporal Information Retrieval
Sep 23, 2026Mourad Hassani, Julien Romero, Amel Bouzeghoub +1Sentence Embeddings Temporal IR
SAMOVAR, Télécom SudParis, Institut Polytechnique de Paris, France · Aldebaran Care, France
IA-RAG: Interval-Algebra-Driven Temporal Reasoning for Dynamic Knowledge Retrieval
Jun 4, 2026Xiaoman Wang, Yaoze Zhang, Wenzhuo Fan +7Retrieval-Augmented Generation Temporal IR
East China Normal University · Shanghai Artificial Intelligence Laboratory · University of Shanghai for Science and Technology +1

Sep 19, 2026J/K move · Enter open · S save

Bridging Static and Agentic RAG for Taiwanese Historical Question Answering

Kai-Hsin Chen, Wei-Yu Chen, Xuanjun Chen, Jyh-Shing Roger Jang

Graduate Institute of Networking and Multimedia, National Taiwan University · Dept. of Computer Science & Information Engineering, National Taiwan University · Graduate Institute of Communication Engineering, National Taiwan University

Retrieval-Augmented Generation Agentic RAG Question Answering RAG Evaluation

Open paper AskPDF arXiv

Agentic retrieval-augmented generation (RAG) enables language models to adapt retrieval based on previously retrieved evidence, but it remains unclear whether such adaptive orchestration consistently outperforms well-designed static pipelines. We conduct a controlled comparison of agentic and static RAG for Taiwanese historical question answering, sharing the same generator and hybrid retrieval backend. Despite similar aggregate performance, the two pipelines differ on 70.83% of questions, with their advantages largely canceling out when averaged. An oracle that selects the better response per question improves the composite score by 0.2417 over the better individual pipeline, revealing substantial headroom for question-level selection. We therefore introduce a post-hoc selector that compares the two responses and their cited evidence, significantly outperforming either individual pipeline and recovering 60.34% of the oracle headroom. These results show that aggregate comparisons can obscure meaningful question-level differences between retrieval strategies, suggesting that exploiting their complementarity may be more fruitful than seeking a universally superior pipeline.

Abstract

Temporal IR

Explore similar work

Bridging Static and Agentic RAG for Taiwanese Historical Question Answering

TEMPS: Temporal Sentence Embeddings for Temporal Information Retrieval

IA-RAG: Interval-Algebra-Driven Temporal Reasoning for Dynamic Knowledge Retrieval