Cross-lingual Biography Enrichment via Claim Extraction and Alignment

Yifei Song; Ziyang Chen; Emil Sayilov; Claire Gardent

cs.CLAug 24, 2026

Cross-lingual Biography Enrichment via Claim Extraction and Alignment

Authors: Yifei Song, Ziyang Chen, Emil Sayilov, Claire Gardent

Organizations: CNRS/LORIA and Université de Lorraine · Université Paris Dauphine - PSL · ICube Laboratory, Université de Strasbourg

Abstract

English Wikipedia is often treated as the default encyclopedic source, yet non-English Wikipedia editions can contain richer locally grounded information for long-tail figures. We study cross-lingual biography enrichment: enriching an existing English biography with facts supported by a non-English biography about the same person. Focusing on women from non-English-speaking contexts, we introduce CLAW-4L, a benchmark consisting of 300 Wikipedia biography pairs linking an English biography with its French, Chinese or Azerbaijani counterpart, along with claim annotations and a fine-grained claim-pair relation corpus. We propose a claim-based enrichment framework that extracts English claims from both biographies, aligns them to identify enrichment evidence from the non-English biography, and rewrites the English biography using the selected claims. Our results show that non-English Wikipedia biographies provide valuable evidence for improving English biography coverage, while lower-resource settings remain challenging.

Figures & tables

Figure 1: Generation trade-off across three representations of non-English evidence. Raw : the generator receives the original English biography and the raw non-English biography. Translation : the non-English biography is first translated into English, and the generator receives the original English biography and the translation. Claims : the generator receives the original English biography and English enrichment claims selected through cross-lingual claim extraction and alignment. Faint points show individual language–generator settings; large points average over three non-English languages and three open-source generators. Translation and Claims increase supported additions over Raw, while the Claims setting substantially reduces hallucination.

Figure 2: Overview of our claim-based framework for cross-lingual biography enrichment. The framework first extracts claims from paired English and non-English biographies into a shared English space, then performs claim-pair alignment to identify non-English-side enrichment candidates not already covered by the English side. The selected enrichment claims are used as structured evidence for controlled English biography enrichment.

Pair	N	Tok (EN/X)	$\mathbf{r_{t}^{norm}}$	Sent (EN/X)	$\mathbf{r_{s}}$	Claim (EN/X)	$\mathbf{r_{c}}$	Max T X	Max C X
EN $\rightarrow$ FR	100	465 / 1867	4.19	16.1 / 48.4	4.64	29.8 / 101.8	3.42	12852	705
EN $\rightarrow$ ZH	100	489 / 1041	2.83	15.8 / 27.2	2.01	29.3 / 72.4	2.47	8772	662
EN $\rightarrow$ AZ	100	544 / 1967	3.20	19.4 / 55.2	3.48	33.6 / 93.8	2.79	28420	1363
EN $\rightarrow$ X	300	500 / 1625	3.41	17.1 / 43.6	3.38	30.9 / 89.3	2.89	28420	1363

Table 1: Summary statistics of CLAW-4L . We report average token (Tok), sentence (Sent), and claim (Claim) counts for English (EN) and non-English (X) biographies, along with normalized token ratios (

r_{t}^{norm}

), sentence ratios (

r_{s}

), and claim ratios (

r_{c}

). Non-English biographies are substantially richer across all levels (e.g., 3.4

\times

tokens and 2.9

\times

claims on average). The maximum non-English biography length reaches 28,420 tokens, and the number of extracted claims reaches 1,363 per instance, highlighting the context length challenges of claim-level alignment and biography generation.

Method	EN		FR		ZH		AZ		Avg.
Method	A-F1	E-F1	A-F1	E-F1	A-F1	E-F1	A-F1	E-F1	A-F1	E-F1
X-FactScore	52.89	27.83	57.89	35.23	52.51	26.30	54.04	39.51	54.33	32.22
X-VeriScore	68.82	46.92	70.01	52.91	68.12	42.94	66.30	55.11	68.31	49.47
X-DnDScore	41.91	11.61	46.16	16.08	44.75	16.02	44.76	22.87	44.40	16.65
X-FactCheck-GPT	59.60	43.70	62.13	49.87	54.70	49.82	52.91	44.38	57.34	46.94
X-Claimify	75.17	72.44	78.05	73.56	71.12	71.91	70.52	66.80	73.72	71.18

Table 2: Claim extraction framework comparison on CLAW-4L-CX using GPT-5.1 as the common backbone. A-F1 denotes Aligned-F1, which counts exact and partial alignment as matches; E-F1 denotes Exact-Aligned-F1, which counts only exact alignment. Avg. is the macro-average over EN, FR, ZH, and AZ.

Non-English-side status	Avg.	Share
Covered by EN	12.4	13.9%
Non-English-additive	23.4	26.2%
Conflict	10.2	11.4%
Unmatched	43.4	48.6%
Selected	77.0	86.2%

Table 3: Average count, ratio and status of non-English-side claims in Claw-4L 400 biography pairs. Covered by EN includes exact matches and English-more-specific alignments. Selected includes non-English-additive, conflict, and unmatched claims. Full language-level statistics are in Appendix F .

Figure 3: Overall generation trade-off across languages, generators, and non-English-evidence formats. Each cell reports Bal.; higher is better. Bolded cells mark the best evidence format for each language–generator pair.

Appendix figures & tables49 assets

Supplementary material from the paper’s appendix.

Appendix

Token Budgets
Task	Prompt	EN	non-English	Out.	Seed
Translation	4,096	4,096	–	8,192	42
Enrichment	2,048	4,096	2,048	8,192	42
Verification	8,192	–	–	8,192	42
Translation / Enrichment
Model	Temp.	Top- $p$	Top- $k$	Pres.	Rep.

Appendix

Table 4: Hyperparameters for generation and LLM-based verification experiments. The upper panel reports token budgets, where EN denotes the English biography, non-English denotes the non-English biography or enrichment context, and Out. denotes the maximum number of generated tokens. The lower panels report decoding settings for generation and verification. Pres. denotes presence penalty and Rep. denotes repetition penalty. Qwen and Gemma generation runs use thinking disabled.

Pair	Inflation Factor $b_{L}$
EN $\rightarrow$ ZH	0.91
EN $\rightarrow$ FR	1.41
EN $\rightarrow$ AZ	1.36

Appendix

Table 5: Language-specific token inflation factors estimated from FLORES+ using the google/mt5-small tokenizer. Values above 1 indicate that the non-English language tends to produce more tokens than English for semantically aligned text; values below 1 indicate the opposite.

Country	Before	After
France	1500	300
China	1500	300
Azerbaijan	434	300
Total	3434	900

Appendix

Table 6: Candidate-pool sizes change before/after tier-aware selection.

Country	N	Artist	Athlete	Politician	Scientist	Other
France	100	32 (32.00%)	26 (26.00%)	26 (26.00%)	10 (10.00%)	6 (6.00%)
China	100	30 (30.00%)	35 (35.00%)	19 (19.00%)	3 (3.00%)	13 (13.00%)
Azerbaijan	100	32 (32.00%)	29 (29.00%)	7 (7.00%)	7 (7.00%)	25 (25.00%)
Total	300	94 (31.33%)	90 (30.00%)	52 (17.33%)	20 (6.67%)	44 (14.67%)

Appendix

Table 7: Occupation-bucket distribution in CLAW-4L, reported by country and in total. We consider four major categories ( artist , athlete , politician , scientist ), with remaining biographies grouped as other .

Split	Pair	N	Tokens (EN / non-EN)	Mean $r_{t}$	Mean $r_{t}^{norm}$	Sent (EN / non-EN)	Mean $r_{s}$
Candidate Pool	EN $\rightarrow$ FR	300	479.22 / 1673.88	5.88	4.16	16.02 / 42.53	4.47
	EN $\rightarrow$ ZH	300	498.84 / 1090.05	2.75	3.00	16.92 / 28.59	2.24
	EN $\rightarrow$ AZ	300	543.36 / 1713.26	4.43	3.26	18.43 / 47.05	3.74
	EN $\rightarrow$ X	900	507.14 / 1492.39	4.35	3.47	17.12 / 39.39	3.48
CLAW-4L	EN $\rightarrow$ FR	100	465.44 / 1867.14	5.92	4.19	16.08 / 48.41	4.64
	EN $\rightarrow$ ZH	100	489.23 / 1041.28	2.60	2.83	15.75 / 27.18	2.01

Appendix

Table 8: Main token- and sentence-level statistics for the 900-instance candidate pool and CLAW-4L. Raw token ratios

r_{t}

denote non-English-to-English token ratios,

r_{t}^{norm}

denotes ratios normalized by language-specific inflation factors estimated from parallel data, and sentence ratios

r_{s}

denote non-English-to-English sentence ratios computed from cleaned biographies using a multilingual neural sentence segmentation model (SaT).

Lang	Country	N	Total Claims	Mean Claims	Median Claims	Mean Sentence Coverage
EN	Overall	300	9261	30.87	20.00	0.966
	Azerbaijan	100	3359	33.59	24.00	0.964
	China	100	2926	29.26	19.50	0.967
	France	100	2976	29.76	17.00	0.968
non-EN	Overall	300	26798	89.33	43.00	0.960
	Azerbaijan	100	9379	93.79	45.50	0.947

Appendix

Table 9: Claim extraction statistics for CLAW-4L. Mean Sentence Coverage averages, over biography instances, the fraction of source sentences that yield at least one extracted claim.

Figure 4: Relationship between source sentence count (x) and extracted claim count (y) per biography in CLAW-4L (EN on top, non-English language on bottom). Each point denotes one biography; dashed lines show least-squares fits (EN: $y=1.70x+1.88$ , $R^{2}=0.961$ ; non-English: $y=1.82x+9.16$ , $R^{2}=0.941$ ). Claim extraction generally yields more claims than source sentences, with mean claims-per-sentence ratios of 1.89 (EN) and 2.14 (non-English).

Figure 5: Distribution of extracted claim counts per biography in CLAW-4L, comparing English (EN) and non-English biographies in a single overlaid histogram (shared bins; y-axis shows instance share). EN is more concentrated at lower claim counts, while non-English biographies exhibit a heavier right tail, indicating more high-claim instances.

Figure 6: Log-scale comparison of extracted claim counts per biography between English and non-English texts in CLAW-4L. Each point is one biography; the diagonal dashed line indicates parity $(y=x)$ . Most points lie above the diagonal, showing that the non-English biographies generally yield more extracted claims than their English counterparts, while the log scale improves visibility across the long-tailed count range.

Figure 7: Human annotation interface for CLAW-4L-CX (1).

Figure 8: Human annotation interface for CLAW-4L-CX (2).

Lang	Aligned			Exact-Aligned
Lang	P	R	F1	P	R	F1
AZ	91.4	91.0	91.1	74.5	73.9	73.9
EN	98.9	100.0	99.3	81.2	80.7	80.3
FR	98.1	98.5	98.0	87.6	85.8	85.7
ZH	98.6	98.7	98.6	81.7	82.9	81.9

Appendix

Table 10: Post-hoc inter-annotator agreement for CLAW-4L-CX reference claim annotations. We compare claim sets produced by different annotators for the same sentence using the selected claim-pair classifier. A claim is counted as matched if it has at least one aligned or exact-aligned counterpart in the other annotator’s set. Scores are reported as percentages.

Lang	Country	N	Mean $n_{\text{claims}}$	B1	B2	B3	B4
EN	Azerbaijan	100	2.67	25	25	25	25
	China	100	2.61	25	25	25	25
	France	100	2.60	25	25	25	25
	Global	300	2.63	75	75	75	75
non-EN	Azerbaijan	100	2.68	25	25	25	25
	China	100	2.84	25	25	25	25

Appendix

Table 11: Claim-count distribution of the final selected sentences in CLAW-4L-CX. Within each language side, rows report country-level statistics followed by a global summary. Here, B1, B2, B3, and B4 denote sentences yielding 1, 2, 3, and 4 or more extracted claims, respectively.

Lang	Country	$N_{\text{before}}\rightarrow N_{\text{after}}$	$\Delta R_{\mathrm{cov}}\uparrow$	$\Delta R_{\mathrm{dens}}\uparrow$	$\Delta R_{\mathrm{rew}}\uparrow$	$\Delta\mathrm{Score}\uparrow$
EN	Azerbaijan	1837 $\rightarrow$ 100	0.4004	0.1347	0.2813	0.2721
	China	1502 $\rightarrow$ 100	0.4127	0.2447	0.1714	0.2763
	France	1523 $\rightarrow$ 100	0.4190	0.2490	0.1367	0.2682
	Global	4862 $\rightarrow$ 300	0.4107	0.2095	0.1965	0.2722
non-EN	Azerbaijan	5141 $\rightarrow$ 100	0.3177	0.2201	0.4190	0.3189
	China	2635 $\rightarrow$ 100	0.2522	0.2316	0.3657	0.2832

Appendix

Table 12: Shifts in normalized rank-based sentence selection statistics from CLAW-4L to the final selected sentences in CLAW-4L-CX. Here,

N_{\text{before}}

and

N_{\text{after}}

denote the numbers of candidate and selected sentences, respectively. We report

\Delta=\text{after}-\text{before}

for normalized coverage rank, density rank, rewrite-rank, and the final aggregated score, where larger positive values indicate a better trade-off between informativeness and editable difficulty of the selected set.

Category	Rel.	Claim A	Claim B
Alig.	$A{=}B$	Libby Lee Ha-yun attended Ying Wa Girls’ School during her high school years .	Libby Lee attended Ying Wa Girls’ School during her senior secondary school years .
	$A{>}B$	Libby Lee Ha-yun attended Ying Wa Girls’ School in Mid-Levels during her high school years.	Libby Lee attended Ying Wa Girls’ School during her senior secondary school years.
	$B{>}A$	Libby Lee attended Ying Wa Girls’ School during her senior secondary school years.	Libby Lee Ha-yun attended Ying Wa Girls’ School in Mid-Levels during her high school years.
	$A{\leftrightarrow}B$	Libby Lee Ha-yun attended Ying Wa Girls’ School in Mid-Levels during her high school years .	Libby Lee attended Ying Wa Girls’ School in 1991 .
Contr.	$A{\perp}B$	Libby Lee Ha-yun attended Ying Wa Girls’ School during her high school years.	Libby Lee attended Diocesan Girls’ School during her senior secondary school years.
Not Rel.	$A{\dashv}B$	Libby Lee Ha-yun attended Ying Wa Girls’ School during her high school years.	In 1957, Huang Yifan died in London, England, at the age of 61.

Appendix

Table 13: Examples of claim relations. Orange text highlights the additional details that appear only on one side of the claim. Red text highlights the contradictory key words between two claims.

Label	#	BGE	MPNet
Aligned – Exact ( $A=B$ )	100	96.87	96.04
Aligned – Partial ( $A>B$ )	100	91.15	89.79
Aligned – Partial ( $B>A$ )	100	91.15	89.79
Aligned – Partial ( $A\leftrightarrow B$ )	100	85.47	84.27
Aligned	400	91.16	89.97
Contradicted ( $A\perp B$ )	100	88.47	88.47

Appendix

Table 14: Label distribution and semantic similarity in CLAW-4L-RC . BGE and MPNet denote cosine similarity scores computed with BGE-Large-En-v1.5 and All-MPNet-Base-v2 sentence embeddings, respectively; scores are reported on a 0-100 scale.

Figure 9: Comparison of the best-performing classifier from each model family on CLAW-4L-RC across three aggregate metrics (ARC, Align-FG, Overall). GPT-5.1 achieves the strongest performance, while Qwen3.5-9B is the closest open-source model across all metrics.

NLI
Model	Info.	Aligned				Contr.	Not Rel.	Aggregate
Model	Info.	$A{=}B$	$A{>}B$	$B{>}A$	$A{\leftrightarrow}B$	$A{\perp}B$	$A{\dashv}B$	ARC	Align-FG	Overall
MiniCheck-7B	w/o	85.0	82.0	82.0	✗	✗	✗	✗	✗	✗
DeBERTa-v3-large	w/o	90.0	86.0	86.0	✗	95.0	✗	✗	✗	✗
Qwen3
Qwen3-4B	w/o	52.0	83.0	61.0	22.0	63.0	91.0	69.5	54.5	62.0

Appendix

Table 15: Claim-pair relation classification results on CLAW-4L-RC across various classifiers including NLI models and LLMs. w/o denotes using only claim pairs as input, while w/ additionally provides GPT-5.1-parsed infoboxes. All metrics are accuracy (%). ARC is the macro-average over the three top-level relation types, where the Aligned score is first averaged over its four sub-labels. Align-FG measures fine-grained accuracy within the Aligned category (average over its four sub-labels). Overall is the macro-average over all six fine-grained labels. Best and second-best results in each column are marked in bold and underlined, respectively. Entries marked with ✗ are not directly computable because the corresponding classifier does not provide enough label resolution to distinguish that relation type.

Method (Year)	Uses Context	Decomp.	Decontext.	Verifiable Sel.	Disamb.	Execution
FactScore (2023)	✗	✓	✗	✗	✗	Joint
DnDScore (2025)	✓	✓	✓	✗	✓	In-prompt Seq.
VeriScore (2024)	✓	✓	✓	✓	✗	Joint
FactCheck-GPT (2024)	✓	✓	✓	✓	✗	Multi-turn
Claimify (2025)	✓	✓	✓	✓	✓	Multi-turn

Appendix

Table 16: Functional comparison of the English claim extraction frameworks adapted in our experiments. A check mark (✓) indicates that the corresponding operation is explicitly specified in the method prompt or pipeline. Uses Context indicates whether the method explicitly incorporates the surrounding sentence or paragraph context beyond the focal sentence. Execution describes how the operations are organized: Joint uses a single prompt without staged intermediate outputs, In-prompt Seq. performs the steps sequentially within one prompt, and Multi-turn splits them across multiple prompting rounds.

Lang	Method	Aligned			Exact-Aligned
Lang	Method	P	R	F1	P	R	F1
AZ	X-FactScore	71.38	47.19	54.04	37.07	47.73	39.51
	X-VeriScore	68.79	68.00	66.30	57.27	56.31	55.11
	X-DnDScore	64.72	36.87	44.76	19.09	32.95	22.87
	X-FactCheck-GPT	56.63	51.88	52.91	43.69	45.93	44.38
	X-Claimify	69.92	72.93	70.52	67.77	66.83	66.80

Appendix

Table 17: Comparison of different claim extraction methods using GPT-5.1 against human reference claims on CLAW-4L-CX . Scores are percentages.

Lang	Method	Aligned			Exact-Aligned			Avg-F1
Lang	Method	P	R	F1	P	R	F1	Avg-F1
AZ	GPT-5.1	69.92	72.93	70.52	67.77	66.83	66.80	68.66
	Mistral-Small-3.2-24B-IT	64.00	70.29	64.90	59.59	52.42	54.35	59.62
	Gemma-4-31B-IT	66.25	78.45	69.42	58.16	53.27	54.28	61.85
	Qwen3.6-27B	67.30	75.42	69.23	63.90	58.73	59.88	64.56
EN	GPT-5.1	77.72	75.76	75.17	72.82	73.18	72.44	73.81

Appendix

Table 18: Claim extraction quality of X-Claimify using different claim generation models against human reference claims on CLAW-4L-CX . Scores are percentages. Avg-F1 is the average of Aligned F1 and Exact-Aligned F1.

Lang	Model	Aligned			Exact-Aligned
Lang	Model	P	R	F1	P	R	F1
AZ $\rightarrow$ EN	Qwen3.6-27B	92.27	90.53	91.39	34.49	32.92	33.69
	Gemma-4-31B-it	92.66	91.51	92.08	36.64	33.70	35.11
	Mistral-3.2-24B-it	91.08	90.94	91.01	34.57	32.23	33.36
	M2M100-1.2B	78.64	64.86	71.09	27.28	10.37	15.03
	NLLB-200-3.3B	88.08	86.02	87.04	32.33	29.89	31.06

Appendix

Table 19: Claim-based evaluation of candidate non-English-to-English translation models for biography translation. We report precision (P), recall (R), and F1 under both coarse-grained Aligned matching and strict Exact-Aligned matching. Best values within each language pair are shown in bold.

Language Pair	Selected MT Model
AZ $\rightarrow$ EN	Gemma-4-31B-it
FR $\rightarrow$ EN	Mistral-3.2-24B-it
ZH $\rightarrow$ EN	Qwen3.6-27B

Appendix

Table 20: Final translation model selected for each non-English language.

Direction	EN	X	Exact	EN-More	X-Add.	Conflict	Unmatched	Selected
Direction	EN	X	$A{=}B$	$A{>}B$	$B{>}A$ / $A{\leftrightarrow}B$	$A{\perp}B$	$A{\dashv}B$	Selected
FR $\rightarrow$ EN	29.8	101.8	5.2	7.2	23.5	9.7	56.2	89.4
AZ $\rightarrow$ EN	33.6	93.8	6.9	8.6	26.0	12.4	39.9	78.3
ZH $\rightarrow$ EN	29.3	72.4	3.7	5.5	20.6	8.4	34.1	63.2
Avg.	30.9	89.3	5.3	7.1	23.4	10.2	43.4	77.0

Appendix

Table 21: Average counts of English claims, non-English claims, and non-English claim status categories per biography pair. Here

A

denotes an English claim and

B

denotes a non-English claim. Exact and EN-More are already covered by English and are not selected. X-Add. contains non-English claims with additional non-conflicting information. Conflict contains classifier-labeled contradictions between non-English claims and retrieved English candidates, while Unmatched contains non-English claims with no retrieved or classifier-relevant English-side counterpart. Selected is the set of Non-English-side enrichment claims used as structured evidence for generation.

Evidence	#Valid $\uparrow$	Halluc.% $\downarrow$
Raw	15.89	29.49
Translation	19.05	29.59
Claims	20.47	23.54

Appendix

Table 22: Average supported additions and hallucination rate across all non-English languages and generators for each evidence format.

Lang.	Case	Entity	Supported/Output
FR	Low	Tanina Mammeri	3/24
FR	High	Juliette Gréco	109/261
ZH	Low	Li Jiaqi	1/6
ZH	High	Xie Bingying	98/207
AZ	Low	Aygün Kazimova	1/536
AZ	High	Franghiz Ali-Zadeh	113/239

Appendix

Table 23: Cases selected for qualitative analysis. Scores are measured under Claims for FR and ZH and Translation for AZ; all three enrichment methods are inspected for every entity.

Pattern and case	Evidence–output comparison
Missing event metadata Juliette Gréco; Raw	Evidence: “made her debut in the play Victor ou les Enfants au pouvoir in November 1946.” Output: “began her career in the theater.”
Temporal error Juliette Gréco; Claims	Evidence: “was born on 7 February 1927.” Output: “was born on 6 February 1927.”
Education / location error Franghiz Ali-Zadeh; Raw	Evidence: “studied piano at the Azerbaijan State Conservatory under Ulfan Khalilov.” Output: “studied at the Moscow Conservatory.”
Over-specific enumeration Xie Bingying; Translation	Evidence: a 2000 Huaxia Publishing House edition of Collected Works of Xie Bingying . Output: three separate volume-specific claims (Vols. 1–3), each attributed to Anhui Literary Publishing House in August 1999.
Repetitive generation Aygün Kazimova; Translation	The 6,384-word output repeatedly cycles through country variants of “She is a winner of the ‘Golden Hit Parade’ award in Russia [Ukraine, Belarus, …].”

Appendix

Table 24: Representative omissions and hallucination patterns. Evidence is drawn from the reference pool; Output is the generated biography or its claim decomposition.

Method	Read.	Coh.	Wiki.
Claims	3.933	3.867	3.822
Raw	3.844	3.911	3.778
Translation	3.756	3.733	3.689

Appendix

Table 25: Human ratings of writing quality (1–5; higher is better), averaged over three annotators and 15 outputs per method. Read., Coh., and Wiki. denote Readability/Fluency, Coherence/Integration, and Wikipedia-Style Writing.

Dimension	Method	B1	B2	B3	B4	B5
Readability	Claims	4.889	4.333	4.000	3.222	3.222
	Raw	4.889	4.333	3.444	3.556	3.000
	Translation	4.889	4.111	3.667	3.667	2.444
Coherence	Claims	4.778	4.222	4.000	3.444	2.889
	Raw	4.889	4.444	3.889	3.667	2.667
	Translation	4.667	4.222	3.889	3.667	2.222

Appendix

Table 26: Writing-quality ratings by method and length-based sampling stratum. Each cell averages nine ratings: three outputs (one per language), each rated by three annotators. Bold marks the highest score within each dimension and bucket, including ties.

Dimension	$\alpha$	ICC(2,1)	ICC(2,3)
Readability	0.512	0.572	0.801
Coherence	0.507	0.598	0.817
Wikipedia style	0.537	0.592	0.813

Appendix

Table 27: Inter-annotator agreement and reliability of single-rater and three-rater mean scores. Krippendorff’s

\alpha

uses the ordinal distance.

Cross-Lingual Generation
Lang	Model	Claim Growth		Supported Additions		Hallucination			Overall
Lang	Model	#Gain $\uparrow$	Coverage% $\uparrow$	Valid% $\uparrow$	#Valid $\uparrow$	Unsup.% $\downarrow$	Contr.% $\downarrow$	Halluc.% $\downarrow$	Bal. $\uparrow$
AZ	Gemma-4-31B-IT	+17.66	22.55	69.99	12.36	29.36	0.65	30.01	10.79
	Qwen3.6-27B	+23.41	29.90	66.12	15.48	32.81	1.07	33.88	11.26
	Mistral-Small-3.2-24B-IT	+19.41	24.79	71.52	13.88	27.58	0.90	28.48	20.54
FR	Gemma-4-31B-IT	+21.94	24.54	72.43	15.89	27.08	0.49	27.57	30.33

Appendix

Table 28: Full generation results on CLAW-4L . Counts are averages per biography and rates are percentages. #Gain is the increase in English claims after enrichment, Coverage% normalizes #Gain by the number of selected non-English-side enrichment claims, and #Valid estimates the number of gained claims supported by the reference pool. Unsup.% aggregates unsupported, not-relevant, and unknown claims; Halluc.% is the sum of Unsup.% and Contr.%. Bal. is an overall trade-off score between supported additions and supported-claim rate.

Panel of Figure 10 — Figure 10: Prompt for X-Claimify - Selection (1)

Panel of Figure 11 — Figure 11: Prompt for X-Claimify - Selection (2)

Panel of Figure 12 — Figure 12: Prompt for X-Claimify - Selection (3)

Panel of Figure 13 — Figure 13: Prompt for X-Claimify - Disambiguation (1)

Panel of Figure 14 — Figure 14: Prompt for X-Claimify - Disambiguation (2)

Panel of Figure 15 — Figure 15: Prompt for X-Claimify - Disambiguation (3)

Panel of Figure 16 — Figure 16: Prompt for X-Claimify - Decomposition (1)

Panel of Figure 17 — Figure 17: Prompt for X-Claimify - Decomposition (2)

Panel of Figure 18 — Figure 18: Prompt for X-Claimify - Decomposition (3)

Panel of Figure 19 — Figure 19: Prompt for X-Claimify - Decomposition (4)

Panel of Figure 20 — Figure 20: Prompt for Extraction Infobox (1)

Panel of Figure 21 — Figure 21: Prompt for Extraction Infobox (2)

Panel of Figure 22 — Figure 22: Prompt for Alignment (1)

Panel of Figure 23 — Figure 23: Prompt for Alignment (2)

Panel of Figure 24 — Figure 24: Prompt for Alignment (3)

Panel of Figure 25 — Figure 25: System Prompt for Cross-Lingual Generation

Panel of Figure 26 — Figure 26: User Prompt for Cross-Lingual Generation

Panel of Figure 27 — Figure 27: Prompt for LLM-based Biography Translation

Evidence-Grounded Generation

Evidence-grounded generation produces answers and documents whose claims are supported by traceable sources.

See all 149 papers

Oct 11, 2026

KhyontekAI at CheckThat! 2026: Multilingual Dense Retrieval and Cross-Encoder Reranking for Scientific Web Claims

Scientific claims shared through social media platforms frequently reference research findings without explicitly linking to the corresponding scientific publications. Automatically retrieving the correct scientific source for such claims is a challenging multilingual retrieval problem involving semantic matching, paraphrasing, noisy social media text, and cross-domain language variation. In this work, we present a multilingual two-stage retrieval framework for CLEF 2026 CheckThat! Task 1 on Source Retrieval for Scientific Web Claims. Our system combines fine-tuned dense retrieval models with cross-encoder reranking for improving multilingual scientific document retrieval. We evaluate multiple transformer-based dense retrieval architectures including multilingual-E5, BGE, and Contriever under different metadata configurations. We further investigate multilingual joint training, retrieval depth analysis, metadata enrichment, and language balancing strategies for improving retrieval effectiveness. Experimental results demonstrate that metadata enrichment consistently improves retrieval effectiveness, while fine-tuned cross-encoder reranking provides additional gains in retrieval precision. Our final multilingual system achieves MRR@5 scores of 0.6222 for English, 0.5112 for German, and 0.6254 for French, with an overall average MRR@5 score of 0.5862.

Bhargab Jyoti Bhuyan, Pritam Deka

Computer Science and Engineering Department, School of Engineering, Tezpur University, Assam, India · KhyontekAI, Guwahati, Assam, India

Scientific Document Retrieval Multilingual IR Dense Retrieval Cross-Encoder Reranking

Ask PDF

Jun 13, 2026

Beyond Monolingual Deep Research: Evaluating Agents and Retrievers with Cross-Lingual BrowseComp-Plus

Deep research agents are increasingly evaluated on their ability to search for evidence, reason over retrieved sources, and produce grounded answers. Existing browsing benchmarks, however, largely assume that the user's query and the supporting evidence are written in the same language, leaving open whether agentic search systems can operate when relevant evidence appears in another language. We introduce XBCP (Cross-lingual BrowseComp-Plus), a controlled benchmark that preserves the English question-and-answer space of BrowseComp-Plus but varies the languages of the supporting documents. XBCP instantiates two complementary settings: in the cross-lingual setting, each query is paired with evidence in a single assigned language. In the multilingual setting, the full evidence corpus is distributed equally and randomly across 12 languages spanning high-resource and low-resource regimes. We evaluate four deep research agents using sparse and dense multilingual retrievers, measuring answer accuracy, evidence recall, search behavior, calibration, citation fidelity, and oracle retrieval. Results reveal substantial degradation when evidence is translated. Even strong, dense retrievers lose evidence recall, and agents become less calibrated and cite evidence less reliably. Notably, accuracy remains lower even when all gold evidence is supplied directly. These findings suggest that cross-lingual deep research exposes both retrieval failures and an independent, agent-side difficulty in integrating language-mismatched evidence.

Yuheng Lu, Qingcheng Zeng, Heli Qi +6

Waseda University · Northwestern University · RIKEN AIP +4

Multilingual Language Model Evaluation Deep Research Agents Multilingual IR Evidence Retrieval

Ask PDF

May 26, 2026

SEEK: Semantic Evidence Extraction via Adaptive ChunKing for Multilingual Fact-Checking

Multilingual fact verification requires evidence that is both relevant and sufficiently complete for reliable factuality prediction. However, existing systems often rely on search snippets, sentence-level evidence, or locally segmented passages, which can miss decisive context and produce fragmented evidence. To overcome these limitations, we propose SEEK, a Semantic Evidence Extraction with an adaptive chunKing framework that constructs coherent evidence chunks from full fact-checking articles by identifying semantic topic transitions and preserving local verification context. The constructed chunks are encoded using a multilingual encoder and then multilingual LLMs are finetuned using LoRA adapter for veracity prediction. Experiments on X-FACT and RU22Fact show that SEEK improves macro-f1 by up to 10% over semantic chunking, 19% over sentence chunking, and 20% over search-snippet baselines. Evidence completeness and significance analyses further show that SEEK preserves richer verification context and enables more reliable multilingual fact-checking.

Gaurav Kumar, Babu Kumar, Ayush Garg +2

Department of Data Science and Engineering Indian Institute of Science Education and Research, Bhopal, India

Document Chunking Automated Fact-Checking Evidence Retrieval

Ask PDF

Abstract

Evidence-Grounded Generation

Explore similar work

KhyontekAI at CheckThat! 2026: Multilingual Dense Retrieval and Cross-Encoder Reranking for Scientific Web Claims

Beyond Monolingual Deep Research: Evaluating Agents and Retrievers with Cross-Lingual BrowseComp-Plus

SEEK: Semantic Evidence Extraction via Adaptive ChunKing for Multilingual Fact-Checking