tabularis-ai/Synthetic-Data-Generation-Pipeline-for-Low-Resource-Swahili-Sentiment-Analysis

# Synthetic Data for Low-Resource Swahili Language Sentiment Analysis This repository contains the code and dataset for our paper on **"Synthetic Data for Low-Resource Swahili Language Sentiment Analysis: A Controllable Pipeline with LLM Judging"**. ## Abstract Swahili, a vital African lingua franca, is under-resourced in Natural Language Processing (NLP), hindering technological progress for its over 100 million speakers. To help close this gap, we introduce a controllable synthetic data pipeline that generates culturally grounded Swahili text and scores it with LLM-as-a-judge across linguistic quality, cultural relevance, sentiment alignment, and instruction adherence. The pipeline defines fine-grained generation criteria, produces diverse candidates with multiple LLMs, and filters aggressively to retain only high-quality samples. Using the resulting corpus, we continue fine-tuning multilingual sentiment classifiers and observe consistent macro–F1 improvements on AfriSenti–Swahili over zero-shot baselines. This demonstrates that quality-controlled synthetic supervision can reliably transfer sentiment capability to a low-resource language. ## Authors **Samuel Gyamfi, Alfred Kondoro, Yankı Öztürk, Richard H. Schreiber, Vadim Borisov** *tabularis.ai* ## Resources - **Dataset**: tabularisai/swahili-sentime… - **Model**: huggingface.co ## Repository Contents - `src/proposed_pipeline/pipeline.py` - The synthetic data generation script with LLM-as-a-judge pipeline - `requirements.txt` - Python dependencies ## Installation ### Prerequisites - Python 3.8 or higher - OpenRouter API key (get one at openrouter.ai) ### Setup 1. Clone this repository: ```bash git clone github.com cd Synthetic-Data-Generation-Pipeline-for-Low-Resource-Swahili-Sentiment-Analysis ` …