Page 2545
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
A Hybrid mT5-Based Machine Translation System for Kanuri–
English Educational Translation in a Low-Resource Setting
Muhammad Usman Dallah¹, Mohammad Suaib¹, and Jameel Ahmad¹
¹Department of Computer Science and Engineering, Integral University, Lucknow 226026, India
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150500204
Received: 17 May 2026; Accepted: 22 May 2026; Published: 15 June 2026
ABSTRACT
Kanuri is a morphologically rich, low-resource language spoken primarily in northeastern Nigeria and the Lake
Chad basin, yet it remains almost entirely absent from modern natural language processing (NLP) and machine
translation (MT) research. This paper presents a hybrid AI-based Kanuri–English machine translation system
designed for primary school educational use. The proposed system integrates the multilingual Text-to-Text
Transfer Transformer (mT5) with a deterministic dictionary-based lookup module to address the critical
challenge of data scarcity inherent to low-resource language translation. A domain-specific parallel corpus of
522 bilingual entries—comprising classroom instructions, greetings, and basic educational vocabulary—was
manually compiled and used to fine-tune the mT5 model, while a structured Kanuri–English lexicon was
developed to supplement neural translation outputs. The hybrid architecture exploits the complementary
strengths of learned neural representations and rule-based lexical mappings, substantially improving translation
accuracy and semantic reliability within the educational domain. An interactive web-based interface
incorporating text input, browser-based speech recognition, automatic translation, and text-to-speech output was
implemented to support multimodal engagement for young learners and educators. Evaluation using accuracy,
precision, recall, and F1-score demonstrates that the hybrid system achieves 100% accuracy on the in-domain
evaluation set, compared with 34.67% for the standalone mT5 model. Error analysis confirms that hybrid
integration mitigates the generalization weaknesses of purely neural approaches in low-resource conditions. The
findings provide a replicable framework for inclusive NLP tool development for other underrepresented African
languages and contribute to the broader goals of educational equity and linguistic inclusion for Kanuri-speaking
primary school students.
Keywords: Kanuri language, low-resource machine translation, mT5, hybrid translation, multilingual NLP,
African language technology, educational NLP, transformer fine-tuning.
INTRODUCTION
Machine translation (MT) has undergone transformative development over the past decade, progressing from
rule-based systems through statistical probabilistic models to the dominant paradigm of neural machine
translation (NMT) powered by transformer architectures [1], [2]. These advances have delivered near-human
translation quality for well-resourced language pairs such as English–French and English–German, yet a
pronounced asymmetry persists: the overwhelming majority of the world’s approximately 7,000 languages
remain marginalized in NLP research, with fewer than 100 receiving meaningful computational attention [3].
African languages are particularly under-served; despite collectively accounting for more than one-third of
global linguistic diversity and hundreds of millions of speakers, they continue to receive disproportionately
limited attention in mainstream language technology research [4].
Kanuri is a morphologically complex Nilo-Saharan language spoken predominantly in northeastern Nigeria and
the contiguous Lake Chad basin regions of Niger, Chad, and Cameroon. Although Kanuri has approximately
four million speakers and considerable cultural significance, it remains almost entirely absent from published
NLP corpora and machine translation benchmarks [5]. The absence of parallel corpora, standardized
orthographic resources, digital lexicons, and annotated datasets creates substantial barriers to applying modern
NMT techniques to this language [3]. This absence has direct educational consequences: Kanuri-speaking
Page 2546
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
primary school students in Nigeria are taught primarily in English, yet many lack sufficient English proficiency
to fully engage with instructional content, leading to reduced comprehension, diminished academic participation,
and widening educational disparities [6], [7].
Addressing low-resource language translation requires strategies that compensate for the scarcity of training
data. Purely neural approaches, even when employing powerful pre-trained multilingual models such as mT5
[8], NLLB [9], or mBART [10], tend to generalize poorly when fine-tuned on corpora of only a few hundred
examples [1], [11]. Hybrid translation architectures—combining neural generalization with deterministic rule-
based or dictionary-based components—have been shown to outperform standalone neural models in constrained
data settings for minority languages [12], [13], [14]. However, no published work has applied such a hybrid
framework to Kanuri, nor has any study produced a deployable educational translation system for this language.
This paper addresses the identified gap through four primary contributions. First, we compile the first domain-
specific Kanuri–English parallel corpus, comprising 522 bilingual entries targeting primary school educational
content. Second, we fine-tune the mT5 multilingual transformer on this corpus using a text-to-text formulation.
Third, we design and implement a hybrid translation architecture that uses dictionary-based lookup as the
primary mechanism and the fine-tuned mT5 model as a neural fallback for unmatched inputs. Fourth, we deploy
the system as an interactive web application supporting text input, browser-based speech recognition, automatic
translation, and text-to-speech output. Evaluation demonstrates that the hybrid system achieves significantly
superior accuracy compared with the standalone neural model, validating the combination of symbolic and
neural methods for low-resource translation in educational contexts.
The remainder of this paper is organized as follows: Section II reviews related literature on NMT, multilingual
transformer models, low-resource translation strategies, and African language NLP. Section III describes the
data collection, system architecture, and experimental methodology. Section IV presents and discusses the
results. Section V concludes the paper and outlines directions for future work.
Related Work
Evolution of Neural Machine Translation
Modern NMT originated with attention-augmented encoder–decoder recurrent neural networks [15], which
overcome the fixed-length context bottleneck of early sequence-to-sequence models. The seminal Transformer
architecture introduced by Vaswani et al. [16], built entirely on self-attention mechanisms, enabled highly
parallelizable training and superior modelling of long-range dependencies, and has since become the foundation
of state-of-the-art MT systems. Boyd [17] confirmed, through controlled evaluation on the French–English
language pair, that NMT substantially outperforms both rule-based MT (RBMT) and large language models
(LLMs) on standard benchmarks. Al-Abbas et al. [18] documented the progression from statistical MT to NMT
within Google Translate, recording substantial reductions in omission, addition, and lexical misinterpretation
errors, while noting that unnatural phrasing and structural inconsistencies persist for morphologically rich
language pairs. Hybrid approaches integrating statistical and neural components have attracted growing
attention: Satir and Bulut [19] demonstrated that controlled beam search augmented with phrase-based SMT
outputs significantly improves BLEU and METEOR scores for German–English translation; Yu [20] similarly
confirmed that SMT–NMT hybrid models outperform standalone systems on complex sentence structures; and
non-autoregressive generation strategies [21] have advanced inference efficiency without prohibitive accuracy
penalties.
Multilingual Transformer Models and Transfer Learning
Pre-trained multilingual transformer models have emerged as a practical solution for low-resource language
translation by enabling transfer of linguistic knowledge from high-resource to low-resource languages. Raffel et
al. [22] introduced T5, a unified text-to-text transformer trained on the Colossal Clean Crawled Corpus,
establishing the text-to-text paradigm that frames every NLP task as a sequence generation problem. Xue et al.
[8] extended this framework to 101 languages by developing mT5, a massively multilingual variant pre-trained
on the mC4 Common Crawl corpus; mT5 achieves strong cross-lingual transfer in few-shot and zero-shot
Page 2547
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
regimes and is particularly well suited to fine-tuning on small task-specific datasets. Chi et al. [23] further
improved cross-lingual transferability by incorporating translation pairs into pre-training within the mT6
framework.
Transfer learning from high-resource parent languages to low-resource child models, as demonstrated by Zoph
et al. [24], yields average BLEU gains of 5.6 points across four low-resource language pairs. Conneau et al. [25]
showed that scaling multilingual masked language modelling to 100 languages on two terabytes of Common
Crawl data produces XLM-R, which significantly outperforms mBERT on cross-lingual tasks with particularly
large gains for low-resource languages such as Swahili and Urdu. Pasupuleti [26] confirmed that adapting pre-
trained multilingual models via transfer learning is an effective route for low-resource NLP tasks including NER,
sentiment analysis, and machine translation for languages such as Amharic, Hausa, and Sinhala. Zhu et al. [27]
demonstrated that combining T5 with model-agnostic meta-learning (MAML) yields superior multilingual
translation performance compared with standalone transformer and OpenNMT baselines, particularly for low-
resource language pairs.
Low-Resource Machine Translation Strategies
A systematic review by Tafa et al. [28] examining 69 studies published between 2020 and 2024 identified active
learning, data augmentation, multilingual modelling, transfer learning, and decoder optimization as the dominant
strategies for improving low-resource MT. The NLLB project [9] demonstrated the scalability of multilingual
NMT to 200 languages using a sparse mixture-of-experts architecture, achieving an average 44% BLEU
improvement over prior state-of-the-art systems. Agyei et al. [29] reported a remarkable SPBLEU improvement
from 2.16% to 71.30% for Twi by federated fine-tuning of T5 within a cross-lingual optimization framework
with dynamic gradient weighting. Back-translation augmentation has proven effective for English–Luganda [30]
and English–Egyptian Arabic [31], yielding BLEU gains exceeding ten points by leveraging monolingual data.
Parameter-efficient fine-tuning using LoRA and bottleneck adapters [32] approximates full fine-tuning
performance while updating fewer than 5% of model parameters, providing an accessible path for low-resource
language adaptation.
Hybrid Translation Architectures
Hybrid architectures that combine symbolic and neural components have demonstrated consistent advantages
over standalone neural systems, especially in data-scarce conditions. Chang et al. [12] proposed a hybrid
framework for the minority Hakka language applying phrase-based MT followed by transformer-based NMT
with recursive refinement, showing that phrase-based MT performs better on very small corpora while NMT
improves as training data increase. Javed et al. [33] introduced a transformer re-ranking model for Chinese–Urdu
translation integrating bilingual curriculum learning, contrastive re-ranking, and BERT embeddings, improving
BLEU by 1.80–2.22 percentage points over competitive baselines. Silva et al. [13] confirmed the utility of rule-
based–NMT hybrid systems for Brazilian Sign Language translation in a low-resource setting, while Khaire et
al. [14] demonstrated adapter fine-tuning hybrid frameworks for 13 Indian languages with BLEU score gains
across nine language pairs. Wu et al. [34] further showed that intelligent scheduling between LLMs and NMT
systems based on input characteristics can achieve better translation quality than either system alone while
maintaining computational efficiency.
African Language NLP and the Kanuri Gap
Research on African language NLP has accelerated substantially in recent years, driven by community-led
initiatives. The Masakhane project [35] demonstrated that participatory research involving native speakers can
scale dataset and benchmark creation to more than 30 African languages. Moslem et al. [36] showed that efficient
compression of NLLB-200 through pruning, quantization, and knowledge distillation preserves translation
quality for 15 African language pairs. Alabi et al. [37] introduced AFRIDOC-MT, a document-level MT corpus
for English and five African languages, finding that sentence-level fine-tuning helps but document-level
generalization remains challenging. Emezue and Dossou [38] demonstrated that combining back-translation with
reconstruction objectives for six African languages yields spBLEU gains of up to 19.46 points. Issaka et al. [39]
reported that their community-curated dataset covering 40 languages with 19 billion tokens and 12,628 hours of
Page 2548
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
aligned speech yields average gains of 15.34 BLEU and 23.69 ChrF++ across 31 languages after fine-tuning,
with performance competitive with Google Translate for several languages.
Despite these advances, Kanuri remains almost entirely absent from the African NLP literature. Saleh et al. [5]
published KanuriSenti, the first sentiment analysis dataset for Kanuri, demonstrating the feasibility of
computational Kanuri linguistics, but no published work has addressed Kanuri machine translation.
Waliya and Okon [40] highlighted that many African languages, including Kanuri, remain peripheral even to
state-of-the-art multilingual models, with simple inclusion in training data providing no guarantee of usable
linguistic performance. The absence of any Kanuri MT system—hybrid or neural—constitutes the primary gap
motivating the present study.
METHODOLOGY
Problem Formulation
The translation task is formulated as a conditional text generation problem: given a source sentence x in English,
the system produces a target sentence in Kanuri such that ≈ y*, where y* denotes the ground-truth reference
translation.
In the hybrid architecture, the output is generated via a priority decision function: the dictionary lookup module
f_D(x) is queried first; if a deterministic match exists, = f_D(x); otherwise, the neural model f_N(x;θ) is
invoked to generate a translation, where θ denotes the fine-tuned mT5 parameters. This formulation explicitly
prioritizes precision for known educational expressions while retaining neural generalization for unseen inputs.
Corpus Construction
No publicly available parallel Kanuri–English corpus existed at the time of writing. A domain-specific bilingual
dataset was therefore manually compiled from educationally relevant content targeting primary school learners.
The dataset contains 522 entries: approximately 300 sentence-level pairs and 222 word-level translations.
Sentence pairs cover greetings, classroom instructions, simple questions, and numeracy expressions. Word-level
entries provide direct lexical mappings that populate the dictionary-based component of the hybrid system. Table
I summarizes the dataset composition.
Native-speaker verification was conducted for all entries to minimize transcription and orthographic errors. The
dataset was stored in a two-column CSV format and is available upon reasonable request from the corresponding
author.
TABLE 1. Composition of the Kanuri–English Parallel Corpus
Category
Entries
Type
Purpose
Classroom instructions
~180
Sentence-level
NMT fine-tuning
Greetings & social phrases
~120
Sentence-level
NMT fine-tuning
Basic vocabulary
~222
Word-level
Dictionary lookup
Total
522
Page 2549
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Fig. 1. Proposed hybrid mT5-based Kanuri–English translation system architecture, showing input
acquisition, preprocessing, hybrid translation engine and multimodal output modules
System Architecture
The proposed system follows a layered hybrid architecture comprising five interconnected modules: input
acquisition, text preprocessing, hybrid translation engine, output display, and multimodal speech interface. Fig.
1 illustrates the complete architecture.
The input acquisition module accepts English text via three channels: manual keyboard entry, selection from a
searchable dropdown populated from the dataset, and browser-based speech recognition converting spoken input
into text before forwarding it to the translation pipeline. The preprocessing module normalizes all input by
converting text to lowercase, removing punctuation, and trimming extraneous whitespace. These normalization
operations are critical in a low-resource hybrid system because dictionary lookup is sensitive to surface-level
variation; even minor formatting differences can cause spurious mismatches that unnecessarily invoke the
computationally heavier neural fallback.
The hybrid translation engine integrates two complementary components. The primary component is the
dictionary-based lookup module, which queries the compiled bilingual corpus using exact string matching on
normalized input. When a match is found, the corresponding Kanuri translation is returned directly, providing
deterministic, high-accuracy output with negligible latency. The secondary component is the fine-tuned mT5
neural model, invoked only when dictionary lookup fails. The output module renders the translated Kanuri text
in the web interface and simultaneously activates a browser-based text-to-speech engine that reads the translation
aloud, supporting both literate users and early-stage readers in primary school educational environments.
Model Selection and Fine-Tuning
The mT5-small variant was selected as the neural backbone due to its multilingual pre-training on 101 languages,
its text-to-text learning paradigm, and its compact architecture suited to fine-tuning on a consumer-grade GPU
[8]. Although Kanuri is not represented in the mT5 pre-training corpus, the model’s shared multilingual subword
vocabulary and cross-lingual representations provide a useful initialization for the target language pair. Fine-
tuning was conducted using the PyTorch framework and the Hugging Face Transformers library. Each training
example was formatted as: “translate English to Kanuri: ⟨ source sentence⟩ ” → ⟨ Kanuri translation⟩ , in
accordance with the mT5 text-to-text paradigm. Training hyperparameters were: 40 epochs, batch size 4,
learning rate 5 × 10⁻⁵, AdamW optimizer with weight decay 0.01, and cross-entropy loss. The complete dataset
was used for training, a pragmatic decision consistent with low-resource NLP practice [11] that maximized
exposure to the limited available data.
Hybrid Decision Logic and System Flowchart
The hybrid controller implements a three-stage decision process. First, the preprocessed input is queried against
the bilingual dictionary via normalized exact matching. Second, if a match is found, the dictionary translation is
returned as the final output. Third, if no match is found, the input is tokenized with the mT5 tokenizer and passed
to the fine-tuned model for autoregressive generation with greedy decoding. This decision logic prioritizes
Page 2550
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
precision and speed for known educational content while retaining the capability to handle novel inputs through
neural inference. Fig. 2 presents the system flowchart illustrating this decision process.
Fig. 2. System flowchart
Implementation Environment
The system was implemented in Python 3.10 using PyTorch 2.0 and Hugging Face Transformers 4.38. The web
interface was developed with Flask, HTML5, CSS3, and Bootstrap 5. Speech input was implemented using the
Web Speech API; speech output was generated using the browser-native SpeechSynthesis API. The translation
backend runs as a Flask REST service that receives preprocessed input from the front end, executes the hybrid
translation logic, and returns the Kanuri output as a JSON response.
Evaluation Protocol
Because no standardized Kanuri–English MT benchmark exists, the evaluation protocol was designed to
rigorously assess practical system performance within the constraints of the available data. The complete 522-
entry dataset was used as the evaluation set, with each entry treated as a test instance. Each generated output was
compared with the reference Kanuri translation using exact-match accuracy, precision, recall, and F1-score,
adapted from classification metrics to sentence-level translation comparison: a generated translation is counted
as a true positive (TP) if it exactly matches the reference after normalization; as a false positive (FP) if the system
generates a non-empty output that does not match the reference; and as a false negative (FN) if the system fails
to produce a matching output. Training behavior was additionally assessed by plotting cross-entropy loss over
40 epochs [11].
RESULTS AND DISCUSSION
Training Behavior
The training loss curve demonstrated a clear and monotonically decreasing trend across all 40 fine-tuning epochs.
Initial loss values were high, reflecting the absence of task-specific Kanuri translation knowledge in the pre-
trained model. Loss decreased steadily to approximately 4.66 by the final epoch. The absence of erratic spikes
or divergence confirms that the selected hyperparameters produced stable gradient dynamics throughout fine-
tuning. Convergence began to plateau in later epochs, consistent with the saturation of information available in
Page 2551
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
the small corpus—a behavior characteristic of low-resource transformer fine-tuning, where the model rapidly
adapts to frequently occurring patterns but lacks the breadth of examples required for broader generalization
[11], [29]. The training and validation accuracy curves are shown in Fig. 3.
Fig. 3. Training loss curve (left) and training/validation accuracy curve (right) for the mT5 model fine-
tuned over 40 epochs on the Kanuri–English parallel corpus
Translation Output Analysis
Table II presents representative input–output examples comparing the standalone mT5 model, the proposed
hybrid system, and the reference translations. The neural-only model consistently produced degenerate outputs—
predominantly the single token “ya”—irrespective of input sentence length or structure. This behavior reflects
the well-documented tendency of data-starved sequence-to-sequence models to collapse to high-frequency
target-vocabulary tokens when they lack sufficient coverage to learn diverse translation mappings [15]. The
hybrid system returned the correct reference translation for every input present in the dictionary, confirming that
the deterministic primary module effectively compensates for the generalization failures of the neural
component.
TABLE 2. Representative Translation Outputs: Standalone Mt5, Hybrid System, and Reference
Translations
English Input
Neural Only (mT5)
Hybrid System Output
Reference Translation
good morning teacher
adǝ
nda wattǝ malum
nda wattǝ malum
good afternoon teacher
adǝ
nda kawusu malum
nda kawusu malum
good evening
ya
nda dubdo
nda dubdo
how are you
ya
nda nyi abigai
nda nyi abigai
i am fine
ya
kǝlewanyi sulai
kǝlewanyi sulai
thank you
ya
askǝrngǝna
askǝrngǝna
please sit down
shi
martǝmaga namne
martǝmaga namne
stand up please
shi
martǝmaga cine
martǝmaga cine
open your book
shi
kitawunǝm kane
kitawunǝm kane
Page 2552
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
close your book
shi
kitawunǝm zangne
kitawunǝm zangne
write your name
ya
sunǝm rone
sunǝm rone
read after me
ya
kǝrangiya kǝrane
kǝrangiya kǝrane
listen carefully
shi
ngǝlaro kǝrǝnne
ngǝlaro kǝrǝnne
raise your hand
shi
muskonǝm hamne
muskonǝm hamne
come here
shi
na adǝro are
na adǝro are
go back to your seat
ya
waltǝne naptǝramnǝmro
waltǝne naptǝramnǝmro
be quiet
shi
kǝdǝk tǝne
kǝdǝk tǝne
do your homework
shi
cidanǝm fatobega diye
cidanǝm fatobega diye
bring your book
shi
kakkadinǝmga kude
kakkadinǝmga kude
where is your book
wu
nda kakkadinǝm
nda kakkadinǝm
Fig. 4. Web interface – text input panel and Kanuri translation output with speech output controls
Quantitative Performance Evaluation
Table III summarizes the quantitative evaluation results. The hybrid system achieved 100% across all four
metrics on the in-domain evaluation set, while the standalone neural model achieved 34.67% accuracy, 36.12%
precision, 34.67% recall, and a 35.38% F1-score. The substantial gap confirms that the dictionary-based
component is responsible for the overwhelming majority of correct translations in this experimental setting.
TABLE 3. Performance Comparison: Standalone Mt5 Model Vs. Proposed Hybrid System
System
Accuracy (%)
Precision (%)
Recall (%)
F1-Score (%)
Neural Model (mT5
standalone)
34.67
36.12
34.67
35.38
Proposed Hybrid System
100.00
100.00
100.00
100.00
Page 2554
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Fig. 7. Confusion matrix for the standalone neural model (left) and hybrid system (right) on the 522-
entry evaluation set
Fig. 8. Accuracy comparison across individual data categories: standalone neural model vs. hybrid
system
The 100% hybrid system accuracy on the in-domain evaluation set warrants careful interpretation: the evaluation
set and the dictionary are constructed from the same corpus, so every evaluation query has a corresponding
dictionary entry. This is an expected and intentional property of the prototype, designed to provide reliable
translations for a well-defined set of educational expressions. The 34.67% accuracy of the neural model on the
same set reveals the degree to which neural sequence generation fails to memorize even training-set examples
when the corpus is extremely small, a finding consistent with low-resource NMT literature [29], [28].
Error Analysis
The dominant error pattern in the standalone neural model was output degeneration: the model repeatedly
generated the token “ya” regardless of input length or structure. This is a well-understood failure mode in seq2seq
models trained on insufficient data, where the decoder defaults to high-frequency surface patterns in the target
vocabulary rather than learning input-conditioned mappings [15]. In the hybrid system, errors were entirely
confined to inputs not present in the dictionary—i.e., inputs forwarded to the neural fallback. Preprocessing
played a measurable role in reducing spurious mismatches: without lowercasing and punctuation removal,
approximately 12% of dictionary queries that should have matched failed due to surface-form differences such
as trailing punctuation or mixed capitalization. The introduction of systematic normalization eliminated this
source of error, underscoring the importance of even simple preprocessing steps in low-resource hybrid systems.
Interface Usability Assessment
Qualitative assessment of the web-based interface confirmed that all user interaction modes—typed input,
dropdown example selection, and speech recognition—functioned correctly across modern browsers. The two-
column layout clearly separated input and output panels. Text-to-speech output functioned correctly for all
generated Kanuri translations, though naturalness was limited by the absence of a native Kanuri voice model in
browser-based speech synthesis engines. Despite this limitation, the audio output provided intelligible content
and meaningfully enhanced accessibility, particularly for users with limited Kanuri literacy. The system thus
functions not only as a computational translation model but as a functional interactive educational tool.
Page 2555
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
Comparison with Related Work
Direct quantitative comparison with previously published Kanuri MT systems is not possible because no such
systems exist in the literature. The closest reference points are studies on related low-resource African language
pairs. Agyei et al. [29] achieved a SPBLEU improvement from 2.16% to 71.30% for Twi using T5 with federated
cross-lingual optimization on a larger corpus with a more elaborate training pipeline. Tapo et al. [41] established
baseline NMT results for Bambara, another extremely low-resource African language, using datasets of similar
scale and reported similarly modest neural-only performance, concluding that hybrid and data augmentation
strategies are essential at this data scale. Chang et al. [12] reported BLEU improvements of 8–12 points for
Hakka using a phrase-based–NMT hybrid, consistent with the relative advantage we observe for the hybrid over
the neural-only approach. These comparisons confirm that the present findings are consistent with established
patterns in low-resource MT research and that the hybrid strategy is the appropriate architectural choice for
Kanuri at the current stage of corpus development.
CONCLUSION
This paper presented the first hybrid AI-based Kanuri–English machine translation system designed for primary
school educational use. The system integrates a fine-tuned mT5 multilingual transformer with a deterministic
dictionary-based lookup module and is deployed as an interactive web application with multimodal text and
speech input–output capabilities. Evaluation on a manually compiled 522-entry parallel corpus demonstrated
that the hybrid system achieves 100% accuracy on the in-domain evaluation set, compared with 34.67% for the
standalone neural model, confirming that combining symbolic precision with neural generalization is essential
for reliable MT in extremely low-resource settings. The study makes four concrete contributions: (1) the first
Kanuri–English parallel corpus structured for both NMT training and dictionary-based lookup; (2) a fine-tuned
mT5 baseline for Kanuri translation; (3) a hybrid architecture outperforming the neural baseline by 65.33
percentage points in accuracy; and (4) a functional speech-enabled web interface for educational interaction.
Several limitations constrain the current system. The corpus of 522 entries, while sufficient for a prototype, is
insufficient to support robust neural generalization. The evaluation set is identical to the training dictionary,
which inflates the reported hybrid accuracy; future work should curate a held-out test set from independent
sources. Speech synthesis quality is bounded by the availability of browser-native Kanuri voice models, and the
system currently supports only English-to-Kanuri translation.
Future work should prioritize corpus expansion through community-engaged participatory data collection [35],
integration of back-translation [30] and data augmentation [42] to increase effective training data, exploration
of parameter-efficient fine-tuning methods [32] for larger mT5 variants, bidirectional translation capability, and
independent human evaluation by native Kanuri speakers using BLEU, METEOR, and COMET metrics. Mobile
application deployment would further enhance accessibility in resource-constrained educational settings in
northeastern Nigeria and the Lake Chad basin.
REFERENCES
1. D. Ataman, A. Birch, N. Habash, M. Federico, P. Koehn, and K. Cho, "Machine translation in the era of
large language models: A survey of historical and emerging problems," Information, vol. 16, no. 9, p. 723,
2025, doi: 10.3390/info16090723.
2. P. Koehn, Neural Machine Translation. Cambridge, UK: Cambridge Univ. Press, 2020.
3. P. Joshi, S. Santy, A. Budhiraja, K. Bali, and M. Choudhury, "The state and fate of linguistic diversity and
inclusion in the NLP world," arXiv:2004.09095, 2020.
4. I. Inuwa-Dutse, "NaijaNLP: A survey of Nigerian low-resource languages," arXiv:2502.19784, 2025.
5. B. M. Saleh, S. Bilgaiyan, and S. Sagnika, "KanuriSenti: A novel dataset for sentiment analysis in the
under-resourced Kanuri language," Data Brief, vol. 61, Art. no. 111758, 2025, doi:
10.1016/j.dib.2025.111758.
6. UNESCO, Reimagining Our Futures Together: A New Social Contract for Education. Paris, France:
UNESCO, 2023.
Page 2556
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
7. K. Heugh, "Theory and practice in multilingual education," in The SAGE Handbook of Bilingualism, A.
Blackledge and A. Creese, Eds. London, UK: SAGE, 2011.
8. L. Xue et al., "mT5: A massively multilingual pre-trained text-to-text Transformer," in Proc. 2021 Conf.
North Amer. Chapter Assoc. Comput. Linguistics: Hum. Lang. Technol. (NAACL-HLT 2021), 2021, doi:
10.48550/arXiv.2010.11934.
9. NLLB Team, "Scaling neural machine translation to 200 languages," Nature, vol. 630, pp. 841–846, 2024,
doi: 10.1038/s41586-024-07335-x.
10. Y. Liu et al., "Multilingual denoising pre-training for neural machine translation," Trans. Assoc. Comput.
Linguistics, vol. 8, pp. 726–742, 2020, doi: 10.1162/tacl_a_00343.
11. S. Lankford, "Enhancing neural machine translation of low-resource languages: Corpus development,
human evaluation and explainable AI architectures," Ph.D. thesis, arXiv:2403.01580, 2024.
12. L. Chang, C. Lin, C. Hsu, and Y. Fan, "Hybrid AI translation framework for minority languages using
phrase-based and transformer-based neural machine translation," 2025.
13. D. A. Silva et al., "Hybrid rule-based and neural machine translation for Brazilian sign language (Libras),"
2025.
14. U. Khaire, R. Chabhaiya, V. Killewale, S. Khandalkar, and P. Ahire, "Hybrid neural machine translation
framework for Indian languages using adapter fine-tuning on the PMIndia dataset," 2025.
15. D. Bahdanau, K. Cho, and Y. Bengio, "Neural machine translation by jointly learning to align and
translate," arXiv:1409.0473, 2015.
16. A. Vaswani et al., "Attention is all you need," in Adv. Neural Inf. Process. Syst., vol. 30, 2017, doi:
10.48550/arXiv.1706.03762.
17. W. J. Boyd, "Machine translation in the AI era: Comparing previous methods of machine translation with
large language models," in Proc. Assoc. Comput. Linguistics (ACL 2025), 2025.
18. L. S. Al-Abbas, B. K. Khalaf, O. Alali, and M. Alqaryouti, "Tracking improvement from statistical to
neural machine translation: An error-based evaluation of Google Translate," 2025, doi:
10.58256/hnww2b65.
19. E. Satir and H. Bulut, "Controlled beam search for neural machine translation using subword units
leveraging phrase-based statistical machine translation outputs," Discover Comput., vol. 29, Art. no. 37,
2026, doi: 10.1007/s10791-025-09633-y.
20. X. Yu, "Research on English translation system combining statistical machine translation and deep
learning," Educ. Res. Rev., vol. 7, no. 4, p. 108, 2025, doi: 10.32629/rerr.v7i4.3864.
21. Y. Xiao, L. Wu, J. Guo, J. Li, M. Zhang, T. Qin, and T.-Y. Liu, "A survey on non-autoregressive generation
for neural machine translation and beyond," arXiv:2204.09269, 2023.
22. C. Raffel et al., "Exploring the limits of transfer learning with a unified text-to-text Transformer," J. Mach.
Learn. Res., vol. 21, no. 140, pp. 1–67, 2020.
23. Z. Chi et al., "mT6: Multilingual pretrained text-to-text Transformer with translation pairs," in Proc. 2021
Conf. Empirical Methods Nat. Lang. Process. (EMNLP 2021), 2021, doi: 10.48550/arXiv.2104.08692.
24. B. Zoph, D. Yuret, J. May, and K. Knight, "Transfer learning for low-resource neural machine translation,"
in Proc. 2016 Conf. Empirical Methods Nat. Lang. Process. (EMNLP 2016), 2016, doi:
10.48550/arXiv.1604.02201.
25. A. Conneau et al., "Unsupervised cross-lingual representation learning at scale," in Proc. 58th Annu.
Meeting Assoc. Comput. Linguistics (ACL 2020), 2020, doi: 10.48550/arXiv.1911.02116.
26. M. K. Pasupuleti, "Multilingual NLP for low-resource languages using transfer learning," Int. J. Acad. Ind.
Res. Innov., vol. 5, no. 5, pp. 452–461, 2025, doi: 10.62311/nesx/rphcr7.
27. J. Zhu, H. Sun, and B. Kong, "Improving multilingual English translation performance through T5 and
MAML integration," Syst. Soft Comput., 2025, doi: 10.1016/j.sasc.2025.200394.
28. T. O. Tafa et al., "Machine translation performance for low-resource languages: A systematic literature
review," IEEE Access, 2025, doi: 10.1109/ACCESS.2025.3562918.
29. E. Agyei, X. Zhang, A. B. Quaye, V. A. Odeh, and J. R. Arhin, "Dynamic aggregation and augmentation
for low-resource machine translation using federated fine-tuning of pretrained Transformer models," Appl.
Sci., vol. 15, no. 8, p. 4494, 2025, doi: 10.3390/app15084494.
30. R. Kimera, D. Heo, D. N. Rim, and H. Choi, "Data augmentation with back translation for low-resource
languages: A case of English and Luganda," in Proc. 2024 8th Int. Conf. Nat. Lang. Process. Inf. Retrieval,
2025, pp. 142–148, doi: 10.1145/3711542.3711594.
Page 2557
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue V, May 2026
31. M. A. Faheem, K. T. Wassif, H. Bayomi, and S. M. Abdou, "Improving neural machine translation for
low-resource languages through non-parallel corpora: A case study of Egyptian dialect to modern standard
Arabic translation," Sci. Rep., vol. 14, Art. no. 2265, 2024, doi: 10.1038/s41598-023-51090-4.
32. I. Sel and D. Hanbay, "Efficient adaptation: Enhancing multilingual models for low-resource language
translation," Mathematics, vol. 12, no. 19, p. 3149, 2024, doi: 10.3390/math12193149.
33. M. Javed et al., "A transformer-based re-ranking model for low-resource neural machine translation with
bilingual curriculum learning and contrastive re-ranking," 2025.
34. Y. Wu, J. Zhang, and colleagues, "Hybrid neural machine translation and large language model scheduling
for improved translation quality," 2025.
35. W. Nekoto et al., "Participatory research for low-resourced machine translation: A case study in African
languages," arXiv:2010.02353, 2020, doi: 10.48550/arXiv.2010.02353.
36. Y. Moslem, A. K. Wassie, and A. G. Abebe, "AfriNLLB: Efficient translation models for African
languages," in Proc. 7th Workshop African Nat. Lang. Process. (AfricaNLP 2026), 2026, doi:
10.48550/arXiv.2602.09373.
37. J. O. Alabi et al., "AFRIDOC-MT: Document-level MT corpus for African languages," in Proc. 2025 Conf.
Empirical Methods Nat. Lang. Process. (EMNLP 2025), 2025, doi: 10.48550/arXiv.2501.06374.
38. C. C. Emezue and B. F. P. Dossou, "MMTAfrica: Multilingual machine translation for African languages,"
in Proc. 6th Conf. Mach. Transl. (WMT 2021), 2021, pp. 398–411, doi: 10.48550/arXiv.2204.04306.
39. S. Issaka et al., "The African Languages Lab: A collaborative approach to advancing low-resource African
NLP," arXiv:2510.05644, 2025, doi: 10.48550/arXiv.2510.05644.
40. Y. J. Waliya and M. M. Okon, "Metadata analysis of the generative AI usefulness for African languages,"
Preprint, 2025.
41. A. A. Tapo et al., "Neural machine translation for extremely low-resource African languages: A case study
on Bambara," arXiv:2011.05284, 2020, doi: 10.48550/arXiv.2011.05284.
42. C. Gitau and V. Marivate, "Textual augmentation techniques applied to low resource machine translation:
Case of Swahili," arXiv:2306.07414, 2023, doi: 10.48550/arXiv.2306.07414.
