avinashsai / BERT-Aspect

BERT Fine-tuning for Aspect Based Sentiment Analysis

BERT-Aspect

Code for the paper "Utilizing BERT Intermediate Layers for Aspect Based Sentiment Analysis and Natural Language Inference" https://arxiv.org/pdf/2002.04815.pdf.

Requirements

python>=3.6
transformers==2.9.0
pytorch==1.5.0

Running the code

git clone https://github.com/avinashsai/BERT-Aspect.git
cd PyTorch

python main.py --dataset (laptop/ restaurant)
               --maxlen (Maximum Sentence length (default: 80))
               --numclasses (3 if "conflict" class is not included else 4 (default:3))
               --data-path (path to datasets (default: '../Data/))
               --batch-size (Batch Size (default: 8)
               --numepochs (Number of training epochs (default: 10))
               --runs (Number of average runs to report results (default: 10))
               --model_name (lstm /attention /base)

Note

This code is un-official implementation of the paper. Hence, training details may not be exactly similar. Also, I have made couple of changes due to which results are superior than the reported paper results.

Results

For Laptop dataset:

Model	This Implementation Result (Acc)	Paper Result (Acc)
BERT Base Uncased + Linear	75.44	74.66
BERT Base Uncased + LSTM	76	75.31
BERT Base Uncased + Attention	75.91	75.16

Model	This Implementation Result (F1)	Paper Result (F1)
BERT Base Uncased + Linear	70	68.64
BERT Base Uncased + LSTM	70.6	69.37
BERT Base Uncased + Attention	70.6	68.76

For Restaurant dataset:

Model	This Implementation Result (Acc)	Paper Result (Acc)
BERT Base Uncased + Linear	82.91	81.92
BERT Base Uncased + LSTM	83.04	82.21
BERT Base Uncased + Attention	83.29	82.38

Model	This Implementation Result (F1)	Paper Result (F1)
BERT Base Uncased + Linear	73.2	71.97
BERT Base Uncased + LSTM	73.4	72.52
BERT Base Uncased + Attention	73.6	73.22

For Twitter dataset:

Model	This Implementation Result (Acc)	Paper Result (Acc)
BERT Base Uncased + Linear	70.32	72.46
BERT Base Uncased + LSTM	70.66	73.06
BERT Base Uncased + Attention	69.06	73.35

Model	This Implementation Result (F1)	Paper Result (F1)
BERT Base Uncased + Linear	68.5	71.04
BERT Base Uncased + LSTM	67.1	71.61
BERT Base Uncased + Attention	69	71.88