POS4ASC

POS4ASC - Exploiting Position Bias for Robust Aspect Sentiment Classification

Our code and data for ACL 2021 paper titled "Exploiting Position Bias for Robust Aspect Sentiment Classification"
Fang Ma, Chen Zhang, and Dawei Song.

Intuition

Robustness Issue

State-of-the-art ASC models are recently shown to suffer from the issue of robustness. Particularly in two common scenarios: 1) out-of-domain (O.O.D.) scenario, and 2) adversarial (Adv.) scenario. A case is given below:

Position Bias Helps

We hypothesize that position bias (i.e., the words closer to a concerning aspect would carry a higher degree of importance) is crucial for building more robust ASC models by reducing the probability of mis-attending. The hypothesis is statistically evidenced by existing benchmarks as below, aspect proximity depicts how close opinions are positioned around their corresponding aspects:
From the observation, we propose two mechanisms for capturing position bias, namely position-biased weight and position-biased dropout, which can be flexibly injected into existing models to enhance word representations, thereby robustness of models.

Requirements

python 3.8
pytorch 1.7.0
nltk
transformer

Usage

Download pretrained GloVe embeddings with this link and extract glove.840B.300d.txt into the root directory.
Train a model with a command similar to those in run.sh or run_bert.sh, optional arguments could be found in run.py.

An example of training with LSTM w/ position-biased weight

python run.py 
    --model_name lstm 
    --train_data_name laptop 
    --mode train 
    --weight

An example of training with Roberta w/ position-biased weight

python run.py 
    --model_name roberta 
    --train_data_name laptop 
    --mode train 
    --learning_rate 1e-5 
    --weight_decay 0.0

Evaluate the model with a command similar to those in run.sh or run_bert.sh, optional arguments could be found in run.py.

An example of evaluation with LSTM w/ position-biased weight on the adversarial test set

python run.py 
    --model_name lstm 
    --train_data_name laptop 
    --test_data_name arts_laptop 
    --mode evluate
    --weight

An example of evaluation with Roberta w/ position-biased weight on the adversarial test set

python run.py 
    --model_name roberta 
    --train_data_name laptop 
    --test_data_name arts_laptop 
    --mode evaluate 
    --weight

Citation

If you use the code in your paper, please kindly star this repo and cite our paper

@inproceedings{ma-etal-2021-exploiting, 
    title = "Exploiting Position Bias for Robust Aspect Sentiment Classification", 
    author = "Ma, Fang and Zhang, Chen and Song, Dawei", 
    booktitle = "Findings of ACL",
    year = "2021", 
    address = "Online", 
    publisher = "Association for Computational Linguistics",
}

Issues

For any issues or suggestions about this work, don't hesitate to create an issue or directly contact me via gene_zhangchen@163.com !

BD-MF / POS4ASC