Sohag Gain
// education
BSc in Computer Science & Engineering
Jamalpur Science & Technology University · February 2025
CGPA: 3.55 / 4.00
BSc Computer Science & Engineering graduate researching how classical machine learning and transformer architectures perform across Bangla and English text-classification tasks.
Jamalpur Science & Technology University · February 2025
CGPA: 3.55 / 4.00
A large-scale unified evaluation of classical models and transformer architectures, with a morphological-complexity analysis for Bangla.
The thesis examines which model families best balance accuracy and practical deployment needs across two news-classification datasets: 434,563 Bangla articles in 15 categories and 127,600 English articles in four categories.
Logistic Regression, Naive Bayes, Support Vector Machine, Random Forest, and XGBoost were compared using Bag-of-Words, one-hot, and TF-IDF features.
BiLSTM, monolingual BERT, and a hybrid LSTM+BERT model were tested under the same train, validation, and test protocol.
The thesis measured Bangla inflectional density and found that 98.8% of articles fell within the medium-to-high complexity ranges, providing context for model-performance differences.
The thesis states limits that matter when applying its results beyond the evaluated datasets.
The Bangla corpus comes from one news source and has substantial category imbalance. The English AG News set has four broad categories, so cross-language scores should not be read as a direct comparison of language difficulty.
The inflectional-density metric uses a curated set of 17 suffixes. It does not capture all derivation, sandhi, or compounding; a fuller morphological analyzer could provide a richer linguistic analysis.
Results use a single seeded run without variance estimates or significance tests. Category-level performance and comparisons with multilingual or instruction-tuned models remain useful directions for follow-up study.
The reported results below are from the thesis evaluation, not a claim about every future dataset or deployment.
Best Bangla test accuracy, reached by the monolingual BERT model.
Best English test accuracy, reached by BERT in the reported evaluation.
Strong Bangla / English results from the best classical TF-IDF + SVM baseline, supporting lower-resource deployment decisions.
Supporting records are organized for application-specific sharing. Personal identity documents are kept private.
University, HSC, and SSC transcripts and testimonials are available to support academic history.
A Medium of Instruction certificate and three recommendation letters are available for scholarship applications.
Supporting certificates document programming-contest participation, volunteer activity, and a Robotics & IoT workshop.
Use the academic CV with the thesis, official transcripts, program-specific statement of purpose, and referee letters.