RACL

Code and datasets of our paper "Relation-Aware Collaborative Learning for Unified Aspect-Based Sentiment Analysis" accepted by ACL 2020.

1. Requirements

To reproduce the reported results accurately, please install the specific version of each package.

• python 3.6.10
• tensorflow-gpu 1.5.0
• numpy 1.16.4
• scikit-learn 0.22.2

2. Usage

We incorporate the training and evaluation of RACL in the train_racl.py. Just run it as below.

CUDA_VISIBLE_DEVICES=0 python train_racl.py --task res14 --load 0

We use the datasets pre-processed by IMN. The results of 5 runs should be like this.

Res14	AE-F1	OE-F1	SC-F1	ABSA-F1
round1	85.38	85.31	74.72	70.80
round2	85.41	85.03	74.58	70.72
round3	85.20	85.55	74.26	70.43
round4	85.20	85.29	74.19	70.43
round5	85.68	85.42	74.56	70.96
AVG	85.37	85.32	74.46	70.67

We haven't yet integrated the BERT-based version at present. Since we create a separate project for RACL-BERT, it takes time to merge it to this repo. We may update it in late July.

3. Checkpoints

If you have problems in training RACL, you can also use the released pre-trained weights on the following links.

• Google Drive : Click here.
• Baidu Cloud : Click here. (password:78qy)

Unzip the pre-trained weights in the checkpoint folder, and execute the command as below.

CUDA_VISIBLE_DEVICES=0 python train_racl.py --task res14 --load 1

The results after loading weights should be like this.

Dataset	AE-F1	OE-F1	SC-F1	ABSA-F1
Res14	85.33	86.09	76.31	71.61
Lap14	82.81	77.91	73.04	62.59
Res15	71.76	76.74	67.74	61.20

4. Embeddings

We have generated the word-idx mapping file and the global-purpose and domain-specific embeddings in the data folder. If you want to generate them from scratch, follow the steps below.

• Download glove.840B.300d.txt, rename it to glove_emb.txt, and put it in the folder like ./data/res14.
• Download domain-specific vectors proposed by DE-CNN, rename it to domain_emb.txt, and put it in the folder like ./data/res14. Notice that restaurant_emb.vec is for datasets Res14 and Res15, and laptop_emb.vec is for Lap14.
• Specify the task (i.e., Res14) in embedding.py, then run it.
• Word embeddings will be generated in the corresponding folder, e.g., ./data/res14/glove_embedding.npy.

5. Need Better Results?

If you still need a better performance of RACL, you can increase the $hop_num$ argument in train_racl.py. Stacking layers to 5~6 can introduce 1~2% absolute improvements on ABSA-F1.

6. Citation

If you find our code and datasets useful, please cite our paper.

@inproceedings{chen2020racl,
  author    = {Zhuang Chen and Tieyun Qian},
  title     = {Relation-Aware Collaborative Learning for Unified Aspect-Based Sentiment Analysis},
  booktitle = {ACL},
  year      = {2020}
}

We will update the page information in time. 🏁