Different pretraining tasks for pretrained NLP models.

In recent years, there are a huge number of pretrained NLP models created by researchers. In pretraining the model, a huge amount of data is taken and passed through a model with specific task. These tasks are called as pre-training task. In this article we gonna look at different pre-training tasks:

Depending on whether the pretrained model is encoder only or encoder-decoder framework, there are various pretraining tasks:

  1. Masked Language modeling

  2. Span Correction

Cross-lingual text-to-text pretraining tasks:

  1. Translation Pair Span correction

  2. Translation Span Correction

  3. Machine Translation

5. Machine Translation

Machine Translation is used as a text-to-text pretraining task for Seq-to-Seq learning. It was used in paper ‘Cross-Lingual Natural Language Generation via Pre-training’. Let e and f denote a sentence and its corresponding translation. We directly use e and f as the input and output sequences, respectively. The loss function of MT is
Lmt(e,f) = L(e -> f)

Shubham Garg
Shubham Garg
Machine Learning Leader

My research interests include utilizing AI to solve ambiguous problems - Natural Language Generation, Conversational AI, techable and explainable AI.