ju-resplande / transformers

Transformers for Assin 1 RTE and TweetSentBR

Transformers

Portuguese BERT base, BERT multilingual base and RoBERTa large evaluation on ASSIN 1 rte and TweetSentBR using Transformers in addition to ASSIN 1 sts and ASSIN2 evaluation.

Original README

TweetSentBR formatted data is not available due to Twitter Policy.

Instructions

1. Install requirements

pip install -r ./examples/requirements.txt

2. Update Transformers package to support these tasks

   pip install --upgrade .

3. Run task

   Replace {TASK_TYPE} for assin and tweesent

   Replace {TASK} for assin-ptbr-rte, assin-ptbr-rte. Leave in blank for tweetsent.

   a) For BERT multilingual

   bash run_{TASK_TYPE}.sh {TASK} PT bert-base-multilingual-cased

   b) For Portuguese BERT

   bash run_{TASK_TYPE}.sh {TASK} PT neuralmind/bert-base-portuguese-cased

   c) For RoBERTa

   bash run_{TASK_TYPE}.sh {TASK} EN

Results

run_{TASK_TYPE}.sh outputs predictions.json in output/{MODEL}/{TASK}.

• Task evaluations and ASSIN XML in output/{MODEL}.

• Task evaluation scripts in {TASK_TYPE}_eval.yaml

• ASSIN xml formatting script in assin_xml.yaml

XML ASSIN Similarity labels were not modified.