BM25-Augmented Many-Shot Translation for Low-Resource North-Eastern Indian Languages

Aashish Dhawan; Christopher Driggers-Ellis; Dzmitry Kasinets; Christan Grant; Daisy Zhe Wang

cs.CLAug 13, 2026

BM25-Augmented Many-Shot Translation for Low-Resource North-Eastern Indian Languages

Authors: Aashish Dhawan, Christopher Driggers-Ellis, Dzmitry Kasinets, Christan Grant, Daisy Zhe Wang

Organizations: University of Florida

Abstract

This paper describes the University of Florida Gators submission to the WMT26 Low-Resource Indic Language Translation shared task. We adapt the retrieval-augmented many-shot translation pipeline from our AmericasNLP 2026 system to translate between English and eleven North-Eastern Indian languages in both directions. At inference time, BM25 retrieves the most similar parallel examples from a language-specific training bank, and Gemini 2.5 Flash translates the input conditioned on these examples. No model fine-tuning is involved. Training banks combine official WMT26 data with publicly available corpora such as Samanantar and prior WMT shared task releases. A grid search over retrieval count r and development exemplar count d across all 22 language-direction pairs selects the best configuration for each submission.

Figures & tables

Figure 1: Overview of the gators pipeline.

Cat.	Language	ISO	Train	Aug.	Pairs
1	Assamese	As	52,440	3.3 $\times$	171,794
	Mizo	Lus	49,293	1.0 $\times$	50,620
	Khasi	Kha	25,495	0.9 $\times$	22,847
	Manipuri	Mni	22,539	1.0 $\times$	22,544
	Meitei	Mni_Mtei	16,312	1.0 $\times$	16,312 $\dagger$
	Nyishi	Njz	49,613	1.0 $\times$	49,793

Table 1: Sizes of the in-domain training data before and after augmentation. The primary system uses the Train data, while the contrastive system uses the Pairs data. The augmentation ratio is computed as the number of augmented sentence pairs divided by the number of base pairs.

\dagger

Pairs recalculated from Train . They are equal because we use no external data for Mni_Mtei.

Cat.	Language	ISO	$\|\mathcal{D}\|$
1	Assamese	As	500
	Mizo	Lus	500
	Khasi	Kha	500
	Manipuri	Mni	500
	Meitei	Mni_Mtei	500
	Nyishi	Njz	500

Table 2:

|\mathcal{D}|

for each language in the shared task. Note that

|\mathcal{D}|=500

except for some Cat. 2 languagues.

Invariant					dev En $\to$ X				dev X $\to$ En
Cat.	Language	ISO	Aug.	Pairs	$r$	$d$	ChrF	ChrF++	$r$	$d$	ChrF	ChrF++
1	Assamese	As	3.3 $\times$	171,794	80	10	52.95	49.81	40	49	68.80	67.98
	Mizo	Lus	1.0 $\times$	50,620	80	20	53.40	51.90	40	20	63.53	62.53
	Khasi	Kha	0.9 $\times$	22,847	80	49	56.44	55.71	80	49	75.64	75.09
	Manipuri	Mni	1.0 $\times$	22,544	80	49	66.72	62.34	20	49	69.97	67.86
	Meitei	Mni_Mtei	1.0 $\times$	16,312	80	10	41.22	36.86	40	10	55.20	52.85

Table 3: Development contrastive submission results for each combination of language and translation direction. Two optimal

r,d

pairs are selected for each language from our search grid, one per translation direction. Augmentation (Aug.) gives a rough multiplier on the original size of the development set that our contrastive data augmentation provides. Pairs is the total number of contrastive retrieval exemplars

|\mathcal{R}|

after augmentation.

Invariant					test En $\to$ X				test X $\to$ En
Cat.	Language	ISO	Aug.	Pairs	$r$	$d$	BLEU	ChrF++	$r$	$d$	BLEU	ChrF++
1	Assamese	As	3.3 $\times$	171,794	80	10	20.77	52.81	40	49	27.27	57.35
	Mizo	Lus	1.0 $\times$	50,620	80	20	21.60	50.69	40	20	33.28	63.36
	Khasi	Kha	0.9 $\times$	22,847	80	49	29.14	53.16	80	49	25.76	52.92
	Manipuri	Mni	1.0 $\times$	22,544	80	49	8.15	39.89	20	49	45.93	72.37
	Meitei	Mni_Mtei	1.0 $\times$	16,312	80	10	16.62	52.85	40	10	38.13	66.16

Table 4: Test contrastive submission results for each combination language and translation direction Pakray et al. (28–29) . Two optimal

r,d

pairs are selected for each language from our search grid, one per translation direction. Augmentation Ratio (Aug.) gives a rough multiplier on the original size of the development set that our contrastive data augmentation provides. Pairs is the total number of contrastive retrieval exemplars

|\mathcal{R}|

after augmentation. Entries in bold indicate that our submission performs best in the shared task for that metric.

Figure 2: Dev-to-test performance changes in the contrastive system. Left: En

\to

X translation performance for dev ( blue ) and test ( orange ) data. Right: X

\to

En translation performance for dev ( blue ) and test ( orange ) data.

Invariant					En $\to$ X			X $\to$ En
Cat.	Language	ISO	Aug.	Pairs	$r$	$d$	ChrF++	$r$	$d$	ChrF++
1	Assamese	As	3.3 $\times$	171,794	80	20	49.78	80	49	79.00
	Mizo	Lus	1.0 $\times$	50,620	80	49	52.54	40	49	77.90
	Khasi	Kha	0.9 $\times$	22,847	80	49	55.23	40	49	82.81
	Manipuri	Mni	1.0 $\times$	22,544	80	20	62.96	0	49	88.88
	Meitei	Mni_Mtei	1.0 $\times$	16,312	80	49	36.39	20	49	55.08

Table 5: Results of repeating our

r,d

grid search on near-duplicate-filtered WMT26 data. Augmentation (Aug.) gives a rough multiplier on the original size of the development set that our contrastive data augmentation provides. Pairs is the total number of contrastive retrieval exemplars

|\mathcal{R}|

after augmentation. Unchanged optimal

r,d

configurations in italics .

Language		En $\to$ X		X $\to$ En
Lang.	ISO	URL	Considerations	URL	Considerations
Assamese	As	URL	Agglutinative Postpositions Bengali-Assamese Script Honorifics Match Example Register Match Example Style No Addtl. En, Expl., Parens Output Only Single Line of As SOV Word Order Unicode NFC Encoding Unique Assamese Characters	URL	Agglutinative Bengali-Assamese Script Match Register of Examples No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En SOV Word Order
Mizo	Lus	URL	Latin Script Match Example Register Match Example Style No Addtl. Explanations Output Only Single Line of Lus Output Std. Lus Orthography SOV Word Order Tonal w/o Changes in Script Tibeto Burman Family Verb-Final Agglutination	URL	No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En
Khasi	Khasi	URL	Austroasiatic Family Different Typology Gendered Prefix Agreement Latin Diacritics No Addtl. Explanations Output Only Single Line of Kha Polysynthetic Morph. Preserve Diacritics VOS Word Order	URL	Gendered Prefix Agreement No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En Polysynthetic Morph. VOS Word Order
Manipuri	Mni	URL	Bengali Script Complex Verb Agglutination Match Example Style NFC Unicode Encoding No Addtl. Explanations No Bengali Output Only Single Line of Mni SOV Word Order Tibeto-Burman Family Unicode NFC Norm.	URL	No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En
Meitei	Mni_Mtei	URL	Agglutinative Match Example Script Match Example Style Meitei Mayek Script No Addtl. Explanations No Bengali Script No Devanagari Script Output Only Single Line of Mtei SOV Word Order Tibeto-Burman Family	URL	No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En

Table 6: We describe the prompting strategies we utilize for each Category 1 language in this table. Links in this table direct the reader to each prompt where it lives in our public Github repository.

Language		En $\to$ X		X $\to$ En
Lang.	ISO	URL	Considerations	URL	Considerations
Bodo	Bodo	URL	Agglutinative Morph. Devanagari Script Examples Define Bodo Follow Examples Carefully Grammar Different From Hindi No Addtl. Explanations No Hindi, Sanskrit, or Marathi Not Indo-Aryan Family Output Only Single Line of Bodo SOV Word Order Tibeto-Burman Family Unicode NFC Normalization	URL	Devanagari Script No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En Tibeto-Burman Family
Kokborok	Trp	URL	Agglutinative Morph. Follow Examples Closely Latin Script Match Example Orthography Match Example Script No Addtl. Explanations Output Only Single Line of Trp SOV Word Order Tibeto-Burman Family Very LoRes	URL	No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En
Karbi	Mjw	URL	Agglutinative Extremely LoRes Latin Script Match Example Orthography No Addtl. Explanation Output Only Single Line of Mjw SOV Word Order Tibeto-Burman Family Tonal	URL	No Addtl. Explanations Output Fluent En Output Natural En Output Only Single Line of En
Nagamese	Nag	URL	En and As/Bengali Vocabulary En-Based Creole Fewer Inflections than En Srcs Latin Script Match Example Orthography No Addtl. Explanations No En No SOV Word Order Output Only Nag Creole Output Single Line of Nag Simplified Morphology Simplified En-Based Phonology SVO Word Order	URL	No Addtl. Explanations Normalize Nag Form to Std. En Output Fluent En Output Natural En Output Single Line of Std. En Output Std. En Only
Tagin	Tgj	URL	Agglutinative Extremely LoRes Latin Script Match Example Orthography Njz Relation No Addtl. Explanations Output Only Single Line of Tgj Shared Njz Vocab. SOV Word Order Tibeto-Burman Family	URL	No Addtl. Explanations Output Fluent En Output Natural En Output Only En Output Single Line of En

Table 7: We describe the prompting strategies we utilize for each Category 2 language in this table. Links in this table direct the reader to each prompt where it lives in our public Github repository.

Machine Translation

Machine translation develops computational methods for converting text, speech, and visual content between languages while preserving meaning and purpose.

See all 267 papers

An Empirical Study of Many-Shot In-Context Learning for Machine Translation of Low-Resource Languages
Apr 3, 2026Yinhan Lu, Gaganpreet Jhajj, Chen Zhang +2In-Context Learning Low-Resource MT
Mila – Quebec AI Institute · McGill University · Athabasca University +2
North Small Translate: Advanced Cost-Effective Translation (Cohere CAT+)
Sep 12, 2026Tom Kocmi, Alexandre Bérard, Phil Blunsom +20Machine Translation Language Model Post-Training
Cohere
Neural Machine Translation for Low-Resource Tangkhul--English
Jun 24, 2026Chormi Zimik Vashai, Agniva MaitiLow-Resource MT Low-Resource Language Processing
Independent Researcher · KIIT University

Apr 3, 2026J/K move · Enter open · S save

An Empirical Study of Many-Shot In-Context Learning for Machine Translation of Low-Resource Languages

Yinhan Lu, Gaganpreet Jhajj, Chen Zhang, Anietie Andy+1

Mila – Quebec AI Institute · McGill University · Athabasca University+2

In-Context Learning Low-Resource MT

Open paper AskPDF arXiv

In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks from a few examples, making it promising for languages underrepresented in pre-training. Recent work on many-shot ICL suggests that modern LLMs can further benefit from larger ICL examples enabled by their long context windows. However, such gains depend on careful example selection, and the inference cost can be prohibitive for low-resource language communities. In this paper, we present an empirical study of many-shot ICL for machine translation from English into ten truly low-resource languages recently added to FLORES+. We analyze the effects of retrieving more informative examples, using out-of-domain data, and ordering examples by length. Our findings show that many-shot ICL becomes more effective as the number of examples increases. More importantly, we show that BM25-based retrieval substantially improves data efficiency: 50 retrieved examples roughly match 250 many-shot examples, while 250 retrieved examples perform similarly to 1,000 many-shot examples. We further show that ICL provides additional gains on top of fine-tuning.

Abstract

Machine Translation

Explore similar work

An Empirical Study of Many-Shot In-Context Learning for Machine Translation of Low-Resource Languages

North Small Translate: Advanced Cost-Effective Translation (Cohere CAT+)

Neural Machine Translation for Low-Resource Tangkhul--English