figoyu/e2e-tbsa

A Unified Model for Opinion Target Extraction and Target Sentiment Prediction (AAAI 2019)

E2E-TBSA

Source code of our AAAI paper on End-to-End Target/Aspect-Based Sentiment Analysis.

Requirements

• Python 3.6
• DyNet 2.0.2 (For building DyNet and enabling the python bindings, please follow the instructions in this link)
• nltk 3.2.2
• numpy 1.13.3

Data

• rest_total consist of the reviews from the SemEval-2014, SemEval-2015, SemEval-2016 restaurant datasets.
• (IMPORTANT) rest14, rest15, rest16: restaurant reviews from SemEval 2014 (task 4), SemEval 2015 (task 12) and SemEval 2016 (task 5) respectively. We have prepared data files with train/dev/test split in our another project, check it out if needed.
• (IMPORTANT) DO NOT use the rest_total dataset built by ourselves again, more details can be found in Updated Results.
• laptop14 is identical to the SemEval-2014 laptop dataset.
• twitter is built by Mitchell et al. (EMNLP 2013).
• We also provide the data in the format of conll03 NER dataset.

Parameter Settings

• To reproduce the results, please refer to the settings in config.py.

Environment

• OS: REHL Server 6.4 (Santiago)
• CPU: Intel Xeon CPU E5-2620 (Yes, we do not use GPU to gurantee the deterministic outputs)

Updated results (IMPORTANT)

• The data files of the rest_total dataset are created by concatenating the train/test counterparts from rest14, rest15 and rest16 and our motivation is to build a larger training/testing dataset to stabilize the training & faithfully reflect the capability of the ABSA model. However, we recently found that the SemEval organizers directly treat the union set of rest15.train and rest15.test as the training set of rest16 (i.e., rest16.train), and thus, there exists overlap between rest_total_train.txt and rest_total_test.txt, which makes this dataset invalid. When you follow our works on this E2E-ABSA task, we hope you DO NOT use this rest_total dataset any more but change to the officially released rest14, rest15 and rest16. We have prepared data files with train/dev/test split in our another project, check it out if needed.

• To facilitate the comparison in the future, we re-run our models following the settings in config.py and report the results on rest14, rest15 and rest16:

  Model	rest14	rest15	rest16
  E2E-ABSA (OURS)	67.10	57.27	64.31
  He et al. (2019)	69.54	59.18	-
  Liu et al. (2020)	68.91	58.37	-
  BERT-Linear (OURS)	72.61	60.29	69.67
  BERT-GRU (OURS)	73.17	59.60	70.21
  BERT-SAN (OURS)	73.68	59.90	70.51
  BERT-TFM (OURS)	73.98	60.24	70.25
  BERT-CRF (OURS)	73.17	60.70	70.37
  Chen and Qian (2019)	75.42	66.05	-
  Liang et al. (2020)	72.60	62.37	-

Citation

If the code is used in your research, please star this repo and cite our paper as follows:

@inproceedings{li2019unified,
  title={A unified model for opinion target extraction and target sentiment prediction},
  author={Li, Xin and Bing, Lidong and Li, Piji and Lam, Wai},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  pages={6714--6721},
  year={2019}
}