AI & ML interests

Neural Machine Translation, Sentence Embeddings, Low-Resource Languages

Lazarus NLP is a collective initiative to revive the dying languages of Indonesia through speech and language technology.

logo

Projects

NusaBERT: Teaching IndoBERT to be multilingual and multicultural!

This project aims to extend the multilingual and multicultural capability of IndoBERT. We expanded the IndoBERT tokenizer on 12 new regional languages of Indonesia, and continued pre-training on a large-scale corpus consisting of the Indonesian language and 12 regional languages of Indonesia. Our models are highly competitive and robust on multilingual and multicultural benchmarks, such as IndoNLU, NusaX, and NusaWrites.

IndoT5: T5 Language Models for the Indonesian Language

IndoT5 is a T5-based language model trained specifically for the Indonesian language. With just 8 hours of training on a limited budget, we developed a competitive sequence-to-sequence, encoder-decode model capable of fine-tuning tasks such as summarization, chit-chat, and question-answering. Despite the limited training constraints, our model is competitive when evaluated on the IndoNLG (text generation) benchmark.

Indonesian Sentence Embedding Models

We trained open-source sentence embedding models for Indonesian, enabling applications such as information retrieval (useful for retrieval-augmented generation!) semantic text similarity, and zero-shot text classification. We leverage existing pre-trained Indonesian language models like IndoBERT and state-of-the-art unsupervised techniques and established sentence embedding benchmarks.

Indonesian Natural Language Inference Models

Open-source lightweight NLI models that are competitive with larger models on IndoNLI benchmark, with significantly less parameters. We applied knowledge distillation methods to small existing pre-trained language models like IndoBERT Lite. These models offer efficient solutions for tasks requiring natural language inference capabilities while minimizing computational resources such as cross-encoder-based semantic search.

Many-to-Many Multilingual Translation Models

Adapting mT5 to 45 languages of Indonesia, we developed a robust baseline model for multilingual translation for languages of Indonesia. This facilitates further fine-tuning for niche domains and low-resource languages, contributing to greater linguistic inclusivity. Our models are competitive with existing multilingual translation models on the NusaX benchmark.