notesum.ai

Published at November 12

IAE: Irony-based Adversarial Examples for Sentiment Analysis Systems

cs.CL

cs.AI

Released Date: November 12, 2024

Authors: Xiaoyin Yi¹, Jiacheng Huang²

Aff.: ¹Chongqing Key Laboratory of Public Big Data Security Technology, Chongqing, 400000, China; ²School of Computer Science and Technology, Chongqing University of Posts and Telecommunications, Chongqing, 400065, China

Arxiv: http://arxiv.org/abs/2411.07850v1

Dataset	Method	Local Model	Victim model			WMD
Dataset	Method	Local Model	TextCNN	BidLSTM	Bert	WMD
Meituan	Origin	N/A	0.885	0.894	0.934	N/A
	Visual-based	TextCNN	0.312	0.702	0.880	1.497
		BidLSTM	0.260	0.670	0.892	1.787
		Bert	0.610	0.818	0.860	0.384
	Homonym-based	TextCNN	0.326	0.694	0.854	1.551
		BidLSTM	0.304	0.698	0.856	1.784
		Bert	0.606	0.832	0.846	0.411
	Ours	TextCNN	0.464	0.204	0.456	1.197
		BidLSTM	0.324	0.416	0.542	1.041
		Bert	0.700	0.876	0.370	0.808
Amazon	Origin	N/A	0.900	0.936	0.944	N/A
	Visual-based	TextCNN	0.564	0.664	0.870	2.570
		BidLSTM	0.346	0.400	0.914	2.643
		Bert	0.860	0.830	0.836	2.480
	Homonym-based	TextCNN	0.648	0.702	0.844	2.559
		BidLSTM	0.544	0.550	0.870	2.605
		Bert	0.842	0.826	0.834	2.472
	Ours	TextCNN	0.338	0.322	0.602	2.344
		BidLSTM	0.720	0.728	0.538	2.341
		Bert	0.880	0.886	0.670	2.379