ACM multimedia 2022 | counterfactual measurement and elimination of social prejudice in visual language pre training model

Paper title ：

Counterfactually Measuring and Eliminating Social Bias in Vision-Language Pre-training Models

Thesis link ：

https://arxiv.org/submit/4382207

Code and dataset ：

https://github.com/VL-Bias/VL-Bias

In this work , We aim at VLP Multimodal characteristics of the model , Based on the counterfactual sample generation, the social prejudice of each mode and the social prejudice of multimodal integration are measured . Measuring social prejudice in the pre training model is essentially the concept of measuring prejudice （ such as , Gender and other social attributes ） And the concept of goals .

We propose a bias measurement method based on counterfactual CounterBias, For each sample, counterfactual samples on bias attributes are generated , Then use the modeling changes of the target attributes before and after the counterfactual model to measure the relationship between the concept of bias and the concept of target .

Because there is no ready-made data set for analysis VLP Social prejudice in the model , We proposed VL-Bias Data sets , It is used in this work and in the community VLP Understanding of social prejudice in the model .VL-Bias contain 24K Picture and text right , The target attribute includes 52 Two actions and 13 A career . stay VL-Bias On dataset , Use CounterBias, We tested two typical VLP structure , Single stream VLP And Shuangliu VLP. Key observations include ：（1） The test of VLP There are social biases in the model , Both visual mode and text mode exist ;（2） Included in VLP The social prejudice in the model is basically consistent with the influence of human stereotypes ;（3） Tested single stream architecture VLP And dual flow architecture VLP The bias in the two modes shows different consistency ;（4） The test of VLP The model shows better performance than BERT Stronger gender bias .

Inherited CounterBias Counterfactual thinking in , We propose a simple but effective method FiarVLP eliminate VLP Social prejudice in the model . say concretely , We first generate counterfactual samples of biased concepts in visual and linguistic modalities . then MLM The task minimizes the probability difference between the prediction of facts and counterfactual , So as to prevent the correlation between model learning deviation and target concept . The result of the experiment proves that FairVLP The effectiveness of the .