Comparative study of text categorisation models
Controlled comparison of SVM, BiLSTM, BERT and zero-shot text classifiers.
- Sole author (MSc dissertation)
- Individual project, University of Leicester
- Jun – Sep 2026
Overview
A controlled four-way comparison of SVM, BiLSTM, BERT and zero-shot classifiers on accuracy, training cost and inference latency, across 11 model variants on a single shared pipeline.
What I built
- A data pipeline merging 6 source datasets into a unified 318,000-row, 8-category corpus, with encoding repair, label mapping and cross-source deduplication.
- A containerised web application serving all 11 models from one API for live side-by-side prediction.
- Per-model inference latency benchmarking, ranging from 0.0006 ms (SVM) to 306 ms (zero-shot).