BatsResearch/LexC-Gen

Generate synthetic labeled data for extremely low-resource languages using bilingual lexicons. # LexC-Gen: Generating Data for Extremely Low-Resource Languages with Large Language Models and Bilingual Lexicons LexC-Gen (which stands for **lexicon-conditioned generation**) generates labeled classification data for extremely low-resource languages using open-access LLMs and bilingual lexicons. The key idea is to prompt LLMs to generate English (or any high-resource-language) task data using words from bilingual lexicons so the generated dataset can be better word-translated into low-resource languages. Here, we explore sentiment analysis and topic classification for 17 extremely low-resource languages. 💻 This repository provides code for LexC-Gen used for generating sentiment analysis and topic classification data using Gatitos bilingual lexicons and reproducing our paper. But our code are written in a modular manner so it can be easily adapted to other tasks or bilingual lexicons. 🧩 We provide our main results for NusaX and SIB-200 in `paper_results/` with average accuracy and standard deviations (over 5 runs) reported, which can serve as a form of sanity check for other researchers reproducing our work. ⭐️ **Our work is accepted to the Findings of EMNLP 2024!** --- ## Table of Contents - Setup - Download Task Datasets and Bilingual Lexicons - Task Datasets - Bilingual Lexicons - LexC-Gen Code - 1. CTG-Training - 2. Lexicon-Conditioned Task Data Generation - 3. Input-Label Filtering and Word-to-Word Translation - Evaluation - 1. Generate Validation Dataset for Task Finetuning - 2. Evaluate on Local LexC-Gen Data - Evaluate on HF LexC-Gen Data - Baselines - Cross-Lingual Zero-Shot - Word Translation - Notes - Bibtex --- ## Setup ```bash git clone github.com cd LexC-Gen pip3 install -r requirements.txt mkdir outputs # create outputs directory in LexC-Gen folder for storing all generated data artifacts ``` ## Download Task Datasets and Bilingual Lexicons ### Task Datasets We experiment with NusaX (Winata et al. …