Repo for Exploiting Unlabeled Data for Target-Oriented Opinion Words Extraction

Exploiting Unlabeled Data for Target-Oriented Opinion Words Extraction. Yidong Wang*, Hao Wu*, Ao Liu, Wenxin Hou, Zhen Wu, Jindong Wang, Takahiro Shinozaki, Manabu Okumura, Yue Zhang. Accepted by COLING 2022.

---

Quick Start

We provide preprocessed data and files of the 1st stage Pseudo Opinion Targets Labeling for a quick start.

Google Drive: Link

Download Preprocessed Data & Pretrained Model

Please download the pretrained files from the Google Drive and unzip them in the following structure:

|  
|-data  
|   |  
|   |-14res
|       |
|       |-...(many preprocessed data files)
|   |-15res
|   |-16res
|   |-14lap
|
|-towe_model
    |
    |-backup
    |   |
    |   |-senti_14lap_bert.pt
    |   |-senti_14res_bert.pt
    |   |-senti_15res_bert.pt
    |   |-senti_16res_bert.pt
    |
    |-main.py
    |-.....(other .py)

Training Scripts

Please download and unzip the scripts.
Here we take senti_14res.sh for instance, which means the experiments for 14res dataset.

# Enter the towe-ssl folder
bash [the path to senti_14res.sh] 1 # _1 is the order of the GPU you want to use._

This is the bash code of the senti_14res.sh.

gpu=$1
ds=14res

export CUDA_VISIBLE_DEVICES=${gpu}
mkdir logs_100_senti 

for seed in 1 2 3
do
    for senti_thr in 0.9 0.7 0.5
    do
        for conf_thr in 0.9 0.7 0.5
        do
            python main.py --batch_size 16 --u_batch_size 96 --ds $ds --cur_run_times 1 --confidence_mask_tresh $conf_thr --senti_model senti_${ds}_bert.pt --senti_thr $senti_thr --epochs 100 --strategy 2 --seed $seed > ./logs_100_senti/senti_${ds}_${conf_thr}_${senti_thr}_${seed}.log 
        done
    done 
done

---

If You Want to Train from Scratch

This part shows the full workflow of this work, i.e., from original data to the final result.
If you want to run the code fast, we recommend you use the preprocessed data and pretrained model we provided above.

Data Preparation

1. Cleaning non-English data from Amazon and Yelp datasets:

python ./data_preprocessing/cleaning_non_en.py

2. Preprocessing raw data.

python ./pre_amazon_yelp_raw.py --balance False

Pseudo Opinion Word Labeling

3. Prepare data. --ds: 14lap or 14~16res. (The same below)

python ./data_preprocessing/pre_tar_label.py  --ds 14lap

4. Select unlabeled data.

python ./data_preprocessing/choose_uda_tar_size.py --aug_size 3000000 --ds 14lap --raw amazon

5. Do the data augmentation.

python ./data_preprocessing/pre_pseudo_tar_aug.py --aug_mode mix --ds 14lap

6. The final preprocessing for BERT.

python ./pseudo_labeling/pre_bert_aug.py --ds 14lap --cur_run_times 1

7. Pretrain Pseudo Opinion Labeling model with Supervised Learning.

python main.py --seed 1 --ds 14lap --ssl False 

8. Labeling pseudo opinion words with pretrained models.

python main.py --make_pseu_target True --eval_bs 512 --test_model tar_14lap_bert_1.pt --cur_run_times 1 --ds 14lap

TOWE Model Traning

9. Split unlabeled data for towe model.

python ./data_preprocessing/split_unlabel.py --ds 14lap --cur_run_times 1

10. Do the data augmentation for towe model.

python ./data_preprocessing/pre_towe_aug.py --aug_mode mix --ds 14lap --cur_run_times 1
python ./towe_model/pre_bert_aug.py --ds 14lap --cur_time 1

11. Train towe model.

python main.py --batch_size 16 --u_batch_size 96 --ds 14lap --cur_run_times 1 --confidence_mask_tresh 0.9 --senti_model senti_14lap_bert.pt --senti_thr 0.9 --ulb_size 200000 --epochs 100 --strategy 2 --ssl True --seed 0

---

Other Details

Datasets

TOWE

The datasets of TOWE.
14lap~16res: https://github.com/NJUNLP/TOWE

Amazon

For 14lap dataset unlabeled data.
http://deepyeti.ucsd.edu/jianmo/amazon/categoryFilesSmall/Electronics_5.json.gz

(if you can't open this url,please goto https://nijianmo.github.io/amazon/index.html and click Electronics 5-core (6,739,590 reviews) in "Small" subsets for experimentation)

YELP

For 14~16res datasets unlabeled data.
https://www.kaggle.com/yelp-dataset/yelp-dataset

Experiments Logs

We also provide the meta logs of our experiments to show the reliability of this paper.
You can unzip the exp_results.rar to check that.

Pseudo Labeling Model

The pretrained pseudo labeling model is provided as well, though we don't recommend you reproduce the experiment from scratch.

Resource Cost

We train our model on a single 3090GPU. The memory cost is 22K mb and the time cost is about 80mins / 100epochs.

---

Citation

If you used the datasets or code, please cite the following papers:

[1]. Yidong Wang*, Hao Wu*, Ao Liu, Wenxin Hou, Zhen Wu, Jindong Wang, Takahiro Shinozaki, Manabu Okumura, Yue Zhang. Exploiting Unlabeled Data for Target-Oriented Opinion Words Extraction. In COLING, 2022.

[2]. Zhifang Fan, Zhen Wu, Xin-Yu Dai, Shujian Huang, Jiajun Chen. Target-oriented Opinion Words Extraction with Target-fused Neural Sequence Labeling. In Proceedings of NAACL, 2019.

---