共 1 条
Generating Realistic Natural Language Counterfactuals
被引:0
|作者:
Robeer, Marcel
[1
,2
]
Bex, Floris
[1
,2
,3
]
Feelders, Ad
[2
]
机构:
[1] Netherlands Police Lab AI, Utrecht, Netherlands
[2] Univ Utrecht, Utrecht, Netherlands
[3] Tilburg Univ, Tilburg, Netherlands
来源:
FINDINGS OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, EMNLP 2021
|
2021年
关键词:
ADVERSARIAL NETWORKS;
D O I:
暂无
中图分类号:
TP18 [人工智能理论];
学科分类号:
081104 ;
0812 ;
0835 ;
1405 ;
摘要:
Counterfactuals are a valuable means for understanding decisions made by ML systems. However, the counterfactuals generated by the methods currently available for natural language text are either unrealistic or introduce imperceptible changes. We propose CounterfactualGAN: a method that combines a conditional GAN and the embeddings of a pretrained BERT encoder to model-agnostically generate realistic natural language text counterfactuals for explaining regression and classification tasks. Experimental results show that our method produces perceptibly distinguishable counterfactuals, while outperforming four baseline methods on fidelity and human judgments of naturalness, across multiple datasets and multiple predictive models.
引用
收藏
页码:3611 / 3625
页数:15
相关论文