yongqiangzheng/absa_plm

Fintune BERT and RoBERTa with LoRA in ABSA dataset

NLP

Run

1. download pre-trained models and process dataset

python preprocess.py

2. fine-tune models

You can fine-tune single model with:

python finetune.py --model_name xxx --dataset xxx --lr xxx

or fine-tune all models:

bash run.sh

3. evaluate results and web demo

copy your finetune model to checkpoint folder fist

#python demo.py

Results

top: 3-runs average

bottom: 3-runs best

Model	Lap14		Rest14		MAMS
	Acc(%)	F1(%)	Acc(%)	F1(%)	Acc(%)	F1(%)
BERT	78.63 78.84	74.72 75.01	85.33 85.62	78.44 79.26	83.26 83.61	82.75 83.20
RoBERTa	81.77 81.97	78.31 78.50	87.71 88.30	81.64 82.53	83.16 84.06	82.46 83.44
BERT_LoRA	77.74 78.84	72.73 74.98	84.70 85.00	76.84 77.24	82.44 83.23	81.67 82.58
RoBERTa_LoRA	82.24 82.76	79.27 79.97	87.05 87.59	80.10 81.05	84.03 84.43	83.47 84.03