MultiFiT is a method for efficient multilingual language model fine-tuning, outperforming existing models with fewer labeled examples.
MultiFiT is a novel method for multilingual text classification, based on Universal Language Model Fine-Tuning (ULMFiT). It addresses the challenge of applying deep learning NLP advances beyond English by enabling efficient fine-tuning of monolingual language models. The approach is designed to be less resource-intensive than models like multilingual BERT and does not require parallel text corpora, making it suitable for low-resource languages and applications where collecting a few hundred training examples in the target language is more feasible than relying on high-resource language data. The method incorporates several key improvements over ULMFiT, including the use of SentencePiece subword tokenization for better handling of morphologically rich languages and open-vocabulary problems, and replacing the AWD-LSTM with a Quasi-Recurrent Neural Network (QRNN) for 2-3x faster training. It also utilizes a cosine variant of the one-cycle policy and label smoothing. MultiFiT demonstrates superior performance in text classification tasks, even with only 100 labeled documents in the target language, compared to multilingual BERT and LASER.
Platforms: Web
View full MultiFiT profile on Tools-Radar | Browse Coding & Developer Tools tools | Alternatives to MultiFiT
Tools-Radar is a free directory of 10,000+ AI tools — discover, compare, and choose the right AI software for your needs. Visit tools-radar.com