You can have multilingual premise and hypothesis pairs to train.They can be classifiaed into three relations; entailment,contradiction ,and neutral.The last category means that a pair is neither entailment nor contradiction.
A Task is detecting these relations from paired text using TPUs.

A Turtorial notebook is given here(https://www.kaggle.com/anasofiauzsoy/tutorial-notebook). I made some changes to it and finally it scored 0.70933. (Therefore, only additional code will be written below.)

## Changing Turtorial Code

First of all, I run the tutorial code and the score was 0.64542.
It used a BERT model to pretrain and finetuned to answer.

I changed two points for this model.
The first change is model type. I used XLM RoBERTa (jplu/tf-xlm-roberta-base) model and its pretrained parameters like this.

The Secondaly revision is translating all languages in English. I used googetrans libraly for this revision.

The final Code is here (https://github.com/gojiteji/kaggle/blob/master/watson.ipynb).

I’ve trained in a single language because English accounts for about 60% in both train.csv and test.csv dataset this time. I think that the score will increase if I split it up into each language and let it learn the data because XLM RoBERTa is already scaled cross lingual sentence encoder.