cahya-wirawan / indonesian-language-models

Indonesian Language Models and its Usage

Home Page:https://cahya-wirawan.github.io/indonesian-language-models

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Indonesian Language Models

The language model is a probability distribution over word sequences used to predict the next word based on previous sentences. This ability makes the language model the core component of modern natural language processing. We use it for many different tasks, such as speech recognition, conversational AI, information retrieval, sentiment analysis, or text summarization.

For this reason, many big companies are competing to build large and larger language models, such as Google BERT, Facebook RoBERTa, or OpenAI GPT3, with its massive number of parameters. Most of the time, they built only language models in English and some other European languages. Other countries with low resource languages have big challenges to catch up on this technology race.

Therefore the author tries to build some language models for Indonesian, started with ULMFiT in 2018. The first language model has been only trained with Indonesian Wikipedia, which is very small compared to other datasets used to train the English language model.

Universal Language Model Fine-tuning (ULMFiT)

Jeremy Howard and Sebastian Ruder proposed ULMFiT in early 2018 as a novel method for fine-tuning language models for inductive transfer learning. The language model ULMFiT for Indonesian has been trained as part of the author's project while learning FastAI. It achieved a perplexity of 27.67 on Indonesian Wikipedia.

Transformers

Ashish Vaswani et al. proposed Transfomer in the paper Attention Is All You Need. It is a novel architecture that aims to solve sequence-to-sequence tasks while handling long-range dependencies with ease.

At the time of writing (March 2021), there are already more than 50 different types of transformer-based language models (according to the model list at huggingface), such as BERT, GPT2, Longformer, or MT5, built by companies and individual contributors. The author built also several Indonesian transformer-based language models using Huggingface Transformers Library and hosted them in the Huggingfaces model hub.

About

Indonesian Language Models and its Usage

https://cahya-wirawan.github.io/indonesian-language-models

License:MIT License


Languages

Language:Jupyter Notebook 97.8%Language:Python 2.2%