Towards English - Mizo Legal Machine Translation: Corpus Creation, Analysis, and Benchmarks

Telem Joyson Singh; Sanabam Boyshana; Sanasam Ranbir Singh; Priyankoo Sarmah

cs.CLOct 21, 2026

Towards English - Mizo Legal Machine Translation: Corpus Creation, Analysis, and Benchmarks

Authors: Telem Joyson Singh, Sanabam Boyshana, Sanasam Ranbir Singh, Priyankoo Sarmah

Organizations: Indian Institute of Technology Guwahati, Assam, India

Abstract

Indian court proceedings are held in English, so speakers of languages with limited translation technology have limited access to them. We present the first English Mizo parallel dataset of Indian court judgments. It contains 13,168 sentence pairs from the Gauhati High Court and the Supreme Court of India. Professional native Mizo translators translated complete documents; sentence pairs were manually aligned, and two independent reviewers checked the results. Our study finds that English words are borrowed in a very selective way: official and legal terms are almost always kept in English, but words for everyday things like actions, situations, and results are usually expressed in native Mizo. Mizo diacritics are also used inconsistently across the corpus. When we fine tune the NLLB 200 1.3B model on our training data, it scores 27.47 BLEU on a balanced legal test set of 1,000 sentence pairs, an improvement of 7.16 BLEU over the base model. However, it still uses English words about three times more often than a human translator does.

Figures & tables

Statistic	English	Mizo
Sentence pairs	$13,168$	$13,168$
Running tokens	$286,590$	$291,751$
Vocabulary (types)	$10,317$	$11,996$
Type–token ratio	0.036	0.041
Hapax legomena	$2,476$	$4,253$
as % of types	24.00	35.45

Table 1: Descriptive statistics for the English–Mizo corpus.

Top- $k$ types	English	Mizo
$100$	52.73	51.55
$500$	73.98	74.96
$1,000$	82.09	83.11
$5,000$	96.69	96.34

Table 2: Cumulative token coverage by vocabulary rank (%).

Figure 1: Zipf distribution of token frequencies.

Figure 2: Vocabulary growth (Heaps’ law).

Measure	Value
Shared word types	$6,612$
Mizo tokens also in English vocab.	$55.30\text{\,}\mathrm{\%}$
Mean token-level Jaccard (per-pair)	$0.4315$

Table 3: Corpus-level lexical overlap (English–Mizo).

Term	En	Mz	Ret. (%)
High retention: institutions and instruments
CrPC	17	17	100.0
FIR	46	46	100.0
court	$1,616$	$1,616$	100.0
police	230	230	100.0
IPC	47	47	100.0

Table 4: Retention of English legal terms in the Mizo text. Ret. is the proportion of Mizo segments to English segments that include the term.

Character	Name	Count
â	a with circumflex	294
ṭ	t with dot below	282
ê	e with circumflex	171
û	u with circumflex	94
î	i with circumflex	41

Table 5: Inventory of Mizo-specific orthographic characters on the Mizo side.

Base form	Attested variants	Freq.
chu	chu / chû	$5,744$
te	te / tê	$2,973$
tur	tur / tûr	$1,858$
dan	dan / dân	$1,416$
he	he / hê	$1,257$
awm	awm / âwm	$1,238$

Table 6: Most frequent Mizo types attested with inconsistent circumflex marking. Frequency is the combined count of both variants.

Plain	Freq.	Dotted	Freq.
tan	277	ṭan	12
tha	111	ṭha	27
tul	109	ṭul	16
tawng	13	ṭawng	4

Table 7: Realisation of the t / ṭ contrast for selected lexemes.

Split	Lang.	Sent.	Tokens	Types	TTR
Train	English	$11,522$	$245,838$	$11,929$	0.049
Train	Mizo	$11,522$	$248,039$	$13,205$	0.053
Test	English	$1,000$	$21,990$	$3,968$	0.180
Test	Mizo	$1,000$	$23,150$	$3,669$	0.159

Table 8: Statistics of the training and test splits.

Figure 3: Sentence types in the $1,000$ -pair benchmark test set.

Figure 4: Test set coverage by training set vocabulary.

System	BLEU $\uparrow$	chrF++ $\uparrow$	TER $\downarrow$
NLLB-200 $1.3$ B	20.31	42.56	77.81
NLLB-FT	27.47	50.39	65.68

Table 9: Machine translation results on the

1,000

-pair benchmark.

Stratum	$n$	BLEU $\uparrow$	chrF++ $\uparrow$	TER $\downarrow$
Legal terms	230	19.71	46.43	69.70
Abbreviations	220	34.05	53.58	63.73
Named entities	200	26.33	53.08	62.42
Act names	150	32.18	52.60	63.34
Others	200	21.14	46.27	68.38
Overall	$1,000$	27.47	50.39	65.68

Table 10: NLLB-FT performance by test-set stratum.

Error type	$n$ (%)	Typical case
Over-retention of	50 (21.7)	learned , evidence
English form		kept verbatim
Spurious nativisation	19 (8.3)	affidavit $\rightarrow$ native
Variant equivalent	7 (3.0)	order : thupek /
(not an error)		thuchhuah
Meaning-shifting	5 (2.2)	accused $\rightarrow$

Table 11: Legal-terminology errors in the NLLB-FT output on the

230

-segment legal-term stratum. Percentages are of segments in the stratum;

n

counts affected segments.

Behaviour	$n$ (%)	Example
Transferred correctly	814 (96.7)	Jogen Kalita
Corrupted form	15 (1.8)	Practices $\rightarrow$
		Practises
Omitted	13 (1.5)	entity dropped

Table 12: Named-entity behaviour in the NLLB-FT output, over the

842

entity mentions that the reference transfers verbatim.

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

COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages
Sep 23, 2026Kshetrimayum Boynao Singh, Nitin Kumar Mishra, Palash Pratim Dutta +22Low-Resource MT Machine Translation
IIT Patna · IIT Delhi · IIT Guwahati +3
bharatSetu: a 1M x 22 bridge for indic translation
Oct 11, 2026Haq Nawaz Malik, Nahfid Nissar, Faizan IqbalLow-Resource MT Machine Translation
NepTam: A Nepali-Tamang Parallel Corpus and Baseline Machine Translation Experiments
Mar 14, 2026Rupak Raj Ghimire, Bipesh Subedi, Balaram Prasain +7Low-Resource MT Machine Translation
ILPRL, Kathmandu University, Nepal · Tribhuvan University, Nepal · Dublin City University, Ireland

Sep 23, 2026J/K move · Enter open · S save

COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages

Kshetrimayum Boynao Singh, Nitin Kumar Mishra, Palash Pratim Dutta, Atai Waris Khan+21

IIT Patna · IIT Delhi · IIT Guwahati+3

Low-Resource MT Machine Translation MT Evaluation

Open paper AskPDF arXiv

Machine translation (MT) for Indian languages remains constrained by the limited availability of high-quality, Indic-centric parallel corpora and evaluation benchmarks. Existing multilingual resources are largely constructed from English-pivot content and often fail to capture the linguistic diversity, cultural complexity, and domain-specific characteristics of Indian languages. We present COILD, an Indic-centric parallel corpus comprising over 1.16 million human-translated and human-verified sentence pairs, covering 20 Indian language pairs across the Indo-Aryan, Dravidian, Tibeto-Burman, and Austro-Asiatic language families. The corpus is built entirely from original Indian language sources collected from licensed repositories spanning eight domains with direct real-world applicability. Furthermore, we introduce a domain-centric benchmark comprising 2,000 expert-verified sentences to enable consistent multilingual and cross-lingual evaluation across Indian language pairs. To validate the effectiveness of COILD, we fine-tune two representative multilingual neural machine translation models, IndicTrans2-Distilled and NLLB-200. Experimental results demonstrate consistent improvements across language pairs, domains, automatic evaluation metrics, and human evaluation, highlighting the effectiveness of high-quality Indic-centric supervision. COILD provides a valuable training and evaluation resource for advancing multilingual machine translation and future multilingual language models for Indian languages.

Abstract

Machine Translation

Explore similar work

COILD: An Indic-Centric Parallel Corpus and Benchmark for Machine Translation Across Indian Languages

bharatSetu: a 1M x 22 bridge for indic translation

NepTam: A Nepali-Tamang Parallel Corpus and Baseline Machine Translation Experiments