An adversarial aspect-based sentiment analysis (ABSA) benchmark, dubbed advABSA for both aspect-based sentiment classification (SC) and opinion extraction (OE).
In response to the concerning robustness issue of ABSA, Arts is proposed, which contains datasets crafted only for adversarial evaluaiton on SC but not for OE. So we additionally craft datasets for adversarial evaluaiton on OE following their track. These gathered datasets form advABSA. That is, advABSA can be decomposed to two parts, where the first part is Arts-[domain]-SC reused from Arts and the second part is Arts-[domain]-OE newly produced by us.
In addition, we also provide stdABSA containing datasets from SemEval14 for standard evaluation, namely Sem14-[domain]-SC and Sem14-[domain]-OE. So corresponding performance drops can be measured properly.
advABSA is also available at HuggingFace Datasets, please visit here.
If you find advABSA useful, please kindly star this repositary and cite our paper as follows:
@inproceedings{ma-etal-2022-aspect,
title = "Aspect-specific Context Modeling for Aspect-based Sentiment Analysis",
author = "Ma, Fang and Zhang, Chen and Zhang, Bo and Song, Dawei",
booktitle = "NLPCC",
month = "sep", year = "2022",
address = "Guilin, China",
url = "https://arxiv.org/pdf/2207.08099.pdf",
}The benchmark is mainly processed by Fang Ma.