MGAN

Data for our AAAI'19, oral paper "Exploiting Coarse-to-Fine Task Transfer for Aspect-level Sentiment Classification".

2019/02/24 Update: Data resource has been released now!

Descriptions

This dataset acts as highly beneficial source domains to improve the learning of more fine-grained aspect-term level (AT) sentiment analysis. The dataset has three characteristics:

Large-scale: 100k for each domain

Multi-domain: Restaurant, Hotel, Beautyspa

Aspect-category (AC): Coarse-grained asepct

Even with a simple attention-based model for the AT task, our method can achieve the STOA performances by leveraging the knowledge distilled from the AC task.

Data format

Each instance behaves as the format below:

inst1: ID1/sentence1/aspect1/label1

inst2: ID1/sentence1/aspect2/label2

inst3: ID1/sentence1/aspect3/label3

....

The sentence containing multiple aspects are arranged together.

Example

0H0FwmPY78v_5u51r2TQrw i did n't dislike the food , but the menu is n't exactly cohesive ... pizza and asian cuisine . FOOD_SELECTION -1

0H0FwmPY78v_5u51r2TQrw i did n't dislike the food , but the menu is n't exactly cohesive ... pizza and asian cuisine . FOOD_FOOD_DISH -1

0H0FwmPY78v_5u51r2TQrw i did n't dislike the food , but the menu is n't exactly cohesive ... pizza and asian cuisine . RESTAURANT_CUSINE -1

0H0FwmPY78v_5u51r2TQrw i did n't dislike the food , but the menu is n't exactly cohesive ... pizza and asian cuisine . FOOD_FOOD 1

Label Mapping

positive: 1 neutral: 0 negative: -1

Citation

If the data is useful for your research, please be kindly to give us stars and cite our paper as follows:

@article{li2018exploiting,
  title={Exploiting Coarse-to-Fine Task Transfer for Aspect-level Sentiment Classification},
  author={Li, Zheng and Wei, Ying and Zhang, Yu and Zhang, Xiang and Li, Xin and Yang, Qiang},
  conference = {AAAI Conference on Artificial Intelligence},
  year={2019}
}