Code for EMNLP2021 paper: "Learning Implicit Sentiment in Aspect-based Sentiment Analysis with Supervised Contrastive Pre-Training"

In this repository, we provide code for Superived ContrAstive Pre-Training (SCAPT) and aspect-aware fine-tuning, retrieved sentiment corpora from YELP/Amazon reviews, and SemEval2014 Restaurant/Laptop with addtional implicit_sentiment labeling.

SCAPT aims to tackle implicit sentiments expression in aspect-based sentiment analysis(ABSA). In our work, we define implicit sentiment as sentiment expressions that contain no polarity markers but still convey clear human-aware sentiment polarity.

Here are examples for explicit and implicit sentiment in ABSA:

SCAPT gives an aligned representation of sentiment expressions with the same sentiment label, which consists of three objectives:

• Supervised Contrastive Learning (SCL)
• Review Reconstruction (RR)
• Masked Aspect Prediction (MAP)

Sentiment representation and aspect-based representation are taken into account for sentiment prediction in aspect-aware fine-tuning.

• cuda 11.0
• python 3.7.9
  • lxml 4.6.2
  • numpy 1.19.2
  • pytorch 1.8.0
  • pyyaml 5.3.1
  • tqdm 4.55.0
  • transformers 4.2.2

Retrieved sentiment corpora contain millions-level reviews, we provide download links for original corpora and preprocessed data. Download if you want to do pre-training and further use them:

These pickle files can also be generated from json files by the preprocessing method:

bash preprocess.py --pretrain

We have already combined the opinion term labeling to the original SemEval2014 datasets. For example:

    <sentence id="1634">
        <text>The food is uniformly exceptional, with a very capable kitchen which will proudly whip up whatever you feel like eating, whether it's on the menu or not.</text>
        <aspectTerms>
            <aspectTerm term="food" polarity="positive" from="4" to="8" implicit_sentiment="False" opinion_words="exceptional"/>
            <aspectTerm term="kitchen" polarity="positive" from="55" to="62" implicit_sentiment="False" opinion_words="capable"/>
            <aspectTerm term="menu" polarity="neutral" from="141" to="145" implicit_sentiment="True"/>
        </aspectTerms>
        <aspectCategories>
            <aspectCategory category="food" polarity="positive"/>
        </aspectCategories>
    </sentence>

implicit_sentiment indicates whether it is an implicit sentiment expression and yield opinion_words if not implicit. The opinion_words lebaling is credited to TOWE.

Both original and extended fine-tuning data and preprocessed dumps are uploaded to this repository.

Consequently, the structure of your data directory should be:

├── Amazon
│   ├── amazon_laptops.json
│   └── amazon_laptops_preprocess_pretrain.pkl
├── laptops
│   ├── Laptops_Test_Gold_Implicit_Labeled_preprocess_finetune.pkl
│   ├── Laptops_Test_Gold_Implicit_Labeled.xml
│   ├── Laptops_Test_Gold.xml
│   ├── Laptops_Train_v2_Implicit_Labeled_preprocess_finetune.pkl
│   ├── Laptops_Train_v2_Implicit_Labeled.xml
│   └── Laptops_Train_v2.xml
├── MAMS
│   ├── test_preprocess_finetune.pkl
│   ├── test.xml
│   ├── train_preprocess_finetune.pkl
│   ├── train.xml
│   ├── val_preprocess_finetune.pkl
│   └── val.xml
├── restaurants
│   ├── Restaurants_Test_Gold_Implicit_Labeled_preprocess_finetune.pkl
│   ├── Restaurants_Test_Gold_Implicit_Labeled.xml
│   ├── Restaurants_Test_Gold.xml
│   ├── Restaurants_Train_v2_Implicit_Labeled_preprocess_finetune.pkl
│   ├── Restaurants_Train_v2_Implicit_Labeled.xml
│   └── Restaurants_Train_v2.xml
└── YELP
    ├── yelp_restaurants.json
    └── yelp_restaurants_preprocess_pretrain.pkl

The pre-training is conducted on multiple GPUs.

• Pre-training [TransEnc|BERT] on [YELP|Amazon]:

  python -m torch.distributed.launch --nproc_per_node=${THE_CARD_NUM_YOU_HAVE} multi_card_train.py --config config/[yelp|amazon]_[TransEnc|BERT]_pretrain.yml

Model checkpoints are saved in results.

• Directly train [TransEnc|BERT] on [Restaurants|Laptops|MAMS] As [TransEncAsp|BERTAsp]:

  python train.py --config config/[restaurants|laptops|mams]_[TransEnc|BERT]_finetune.yml

• Fine-tune the pre-trained [TransEnc|BERT] on [Restaurants|Laptops|MAMS] As [TransEncAsp+SCAPT|BERTAsp+SCAPT]:

  python train.py --config config/[restaurants|laptops|mams]_[TransEnc|BERT]_finetune.yml --checkpoint PATH/TO/MODEL_CHECKPOINT

Model checkpoints are saved in results.

• Evaluate [TransEnc|BERT]-based model on [Restaurants|Laptops|MAMS] dataset:

  python evaluate.py --config config/[restaurants|laptops|mams]_[TransEnc|BERT]_finetune.yml --checkpoint PATH/TO/MODEL_CHECKPOINT

Our model parameters:

If you found this repository useful, please cite our paper:

@inproceedings{li-etal-2021-learning-implicit,
    title = "Learning Implicit Sentiment in Aspect-based Sentiment Analysis with Supervised Contrastive Pre-Training",
    author = "Li, Zhengyan  and
      Zou, Yicheng  and
      Zhang, Chong  and
      Zhang, Qi  and
      Wei, Zhongyu",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.22",
    pages = "246--256",
    abstract = "Aspect-based sentiment analysis aims to identify the sentiment polarity of a specific aspect in product reviews. We notice that about 30{\%} of reviews do not contain obvious opinion words, but still convey clear human-aware sentiment orientation, which is known as implicit sentiment. However, recent neural network-based approaches paid little attention to implicit sentiment entailed in the reviews. To overcome this issue, we adopt Supervised Contrastive Pre-training on large-scale sentiment-annotated corpora retrieved from in-domain language resources. By aligning the representation of implicit sentiment expressions to those with the same sentiment label, the pre-training process leads to better capture of both implicit and explicit sentiment orientation towards aspects in reviews. Experimental results show that our method achieves state-of-the-art performance on SemEval2014 benchmarks, and comprehensive analysis validates its effectiveness on learning implicit sentiment.",
}