Fine-tuned mT5_multilingual_XLSum for abstractive summarization of Bangla-language resumes, section by section. Developed as part of the Onneshon thesis project at Bangladesh University of Professionals (BUP).

Given the raw sentences of a resume section (Objective, Experience, Education, or Skill), the model generates a fluent 2–3 sentence Bangla summary with strong semantic alignment to human-verified references.

Model Details

Property	Value
Base model	`csebuetnlp/mT5_multilingual_XLSum`
Architecture	MT5ForConditionalGeneration (encoder-decoder)
Parameters	12 encoder layers, 12 decoder layers, 768 hidden dim, vocab 250,112
Language	Bengali (bn)
Task	Abstractive summarization (per section)
Language token	`<2bn>` prepended to all inputs (required by mT5_multilingual_XLSum)
Training data	317 resume-summary pairs (80% of Onneshon dataset)
Validation data	79 resume-summary pairs (20% of Onneshon dataset)
Epochs	5 (with early stopping, patience=2)
Learning rate	5e-4
Batch size	4
Max input tokens	512
Max output tokens	128

Dataset

Trained on Onneshon — an original Bangla resume dataset of 100 annotated resumes spanning 20+ professions.
Published on Mendeley Data: DOI: 10.17632/4md7bx6fd7.1

Reference summaries were generated using GPT-OSS-120B (via OpenRouter) and human-verified, producing 396 section-level summary pairs across 4 categories: Objective, Experience, Education, Skill.

Evaluation Results

Evaluated against human-verified abstractive reference summaries using a Bengali-aware ROUGE tokenizer and semantic similarity:

Metric	Score
ROUGE-1	0.6171
ROUGE-2	0.4361
ROUGE-L	0.5384
Semantic Similarity (paraphrase-multilingual-MiniLM-L12-v2)	0.8184

Note on ROUGE scores: ROUGE measures exact word overlap. Since references are abstractive paraphrases and Bangla has rich morphology, ROUGE can still underestimate quality relative to semantic similarity. Semantic similarity of 0.818 — the highest among all 4 methods evaluated — confirms this model produces summaries most aligned in meaning with human-verified references.

Per-section ROUGE-1 and Semantic Similarity:

Section	ROUGE-1	Semantic Sim
Objective	0.6697	0.9120
Experience	0.4892	0.7209
Education	0.7557	0.8585
Skill	0.5487	0.7782

Comparison: BanglaT5 vs mT5

Metric	BanglaT5	mT5	Better for
ROUGE-1	0.620	0.617	BanglaT5 (marginal)
Semantic Sim	0.808	0.818	mT5
Objective Sem	0.848	0.912	mT5
Experience R1	0.504	0.489	BanglaT5

mT5 produces more fluent, semantically natural summaries (higher semantic sim overall). BanglaT5 has higher word-level fidelity (higher ROUGE) for factual sections. Use mT5 when fluency and meaning alignment are priorities.

Usage

⚠️ Important: This model requires the Bengali language token <2bn> prepended to every input. Without it, the model defaults to non-Bengali output.

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

LANG_TOKEN = "<2bn>"

tokenizer = AutoTokenizer.from_pretrained("your-username/bangla-resume-summarizer-mt5", use_fast=False)
model     = AutoModelForSeq2SeqLM.from_pretrained("your-username/bangla-resume-summarizer-mt5")
model.eval()

def summarize(text):
    # Prepend Bengali language token — required for this model
    input_text = LANG_TOKEN + " " + text.strip()
    inputs = tokenizer(input_text, return_tensors="pt", truncation=True, max_length=512)
    outputs = model.generate(
        inputs["input_ids"],
        max_length=128,
        min_length=20,
        num_beams=4,
        length_penalty=1.2,
        repetition_penalty=2.0,
        no_repeat_ngram_size=3,
        early_stopping=True
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Example — Objective section
text = "অভিজ্ঞতা সম্পন্ন সফ্টওয়্যার ইঞ্জিনিয়ার হিসেবে একটি ডায়নামিক টিমে যোগদান করা যেখানে আমার জাভা এবং স্প্রিং ফ্রেমওয়ার্কের দক্ষতা কাজে লাগিয়ে প্রতিষ্ঠানের সাফল্যে অবদান রাখতে পারি"
print(summarize(text))

Input Format

Feed one section at a time. Join multiple sentences with । (Bengali danda). The <2bn> token is added automatically in the usage example above:

<2bn> sentence_1 । sentence_2 । sentence_3

Limitations

Trained on only 317 pairs — a small dataset by deep learning standards. Outputs may be generic for unusual professions.
Optimized for resume text. Performance on other Bangla document types is untested.
Not suitable for very long inputs (>512 tokens); truncation will occur.
The <2bn> language token is mandatory — omitting it will cause incorrect (non-Bengali) output.
Compared to BanglaT5, this model has lower ROUGE on factual sections (Experience, Skill) but higher semantic similarity overall.

Citation

If you use this model, please cite the Onneshon dataset and XL-Sum:

bibtex

@misc{onneshon2026,
  title     = {Onneshon: A Bangla Resume NLP Dataset},
  author    = {Tanvir and Shruti Khisa and Shaira Akther Diba and Fazli Rabbi Noor},
  year      = {2026},
  doi       = {10.17632/4md7bx6fd7.1},
  publisher = {Mendeley Data}
}

@inproceedings{hasan-etal-2021-xl,
  title     = {XL-Sum: Large-Scale Multilingual Abstractive Summarization for 44 Languages},
  author    = {Hasan, Tahmid and Bhattacharjee, Abhik and Islam, Md. Saiful and Mubasshir, Kazi and Li, Yuan-Fang and Kang, Yong-Bin and Rahman, M. Sohel and Shahriyar, Rifat},
  booktitle = {Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021},
  year      = {2021},
  pages     = {4693--4703}
}

Project

Part of the Onneshon thesis project — a Bangla NLP pipeline for resume processing.

Dataset: Mendeley Data DOI: 10.17632/4md7bx6fd7.1
Institution: Bangladesh University of Professionals (BUP), Dhaka, Bangladesh
Supervisor: Rumana Yasmin, Lecturer, CSE Department, BUP

resume-summerization-classification/bangla_resume_summarizerMT5

Bangla Resume Summarizer — mT5