Seq2seq for Aspect Sentiment Triplet Extraction
方面情感三元组(Aspect Sentiment Triplet Extraction, ASTE)任务旨在抽取用户评论中表达观点的方面情感三元组。一个三元组包含以下三个部分：
- aspect term: 情感所针对的目标对象，一般是被评价实体的某个方面项，常被称作方面词、属性词
- opinion term: 具体表达情感的词或短语，常被称作情感词
- sentiment polarity: 用户针对aspect term所表达的情感倾向，类别空间为
{POS, NEG, NEU}
生成式方法，(1)首先将ASTE任务转化为一个seq2seq任务，(2) 然后使用一个seq2seq的模型(如t5)来建模。
输入: average to good thai food but incredibly slow and rude delivery
输出: thai food | average to good | positive ; delivery | incredibly slow | negative ; delivery | rude | negative
seed=42
| 14res | 14lap | 15res | 16res | |
|---|---|---|---|---|
| precision | 0.7490 | 0.6727 | 0.6326 | 0.7362 |
| recall | 0.7364 | 0.6188 | 0.6247 | 0.7763 |
| f1-score | 0.7426 | 0.6446 | 0.6286 | 0.7557 |
seed=44
| 14res | 14lap | 15res | 16res | |
|---|---|---|---|---|
| precision | 0.7342 | 0.6681 | 0.6265 | 0.7113 |
| recall | 0.7103 | 0.5691 | 0.6330 | 0.7237 |
| f1-score | 0.7220 | 0.6146 | 0.6297 | 0.7175 |
Requirements
transformers==4.26.1
pytorch==1.10.1
pytorch-lightning==1.9.3
spacy==3.5.0
预处理数据
python t5_convert.py --raw_data_dir data/raw --output_data_dir data/t5 --dataset 14res
训练模型
chmod +x bash/*
bash/train_extractor.sh -d 14res -c 3 -o {YOUR_OUTPUT_DIR}
- 数据来自 https://github.com/xuuuluuu/SemEval-Triplet-data
- Zhang W, Li X, Deng Y, et al. Towards generative aspect-based sentiment analysis[C]//Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers). 2021: 504-510.