Logo Lanfrica

andreuni/Language-Recognition

Domaine:

natural language processing
Créateur:
and
Hôte:
This project focuses on language recognition, with a particular emphasis on low-resource languages. It leverages a combination of high-resource datasets (e.g., WiLI-2018) and low-resource datasets (Bible texts from ScriptureEarth) to develop and evaluate models for text classification. # Language Recognition This project focuses on language recognition, with a particular emphasis on low-resource languages. It leverages a combination of **high-resource datasets** (e.g., WiLI-2018) and **low-resource datasets** (Bible texts from ScriptureEarth) to develop and evaluate models for text classification. ## Key Features: ### Datasets: - **High-resource:** 22 languages, preprocessed from the WiLI-2018 dataset. - **Low-resource:** 25 languages, sourced and preprocessed from ScriptureEarth. ### Preprocessing: - Text cleaning (removal of punctuation, numbers, whitespaces, and more). - Sentence segmentation and merging into a unified format. ### Model Architecture: - **Doc2Vec** for feature extraction. - **Bi-LSTM** combined with a linear layer for text classification. ### Results: - High accuracy achieved on both high- and low-resource datasets. - Detailed performance metrics for various scenarios (e.g., excluding certain languages). --- This repository contains the code, preprocessing scripts, and trained models for researchers and developers interested in improving language recognition, especially for underrepresented languages.