DialCL-Bias: A Semantic and Contextual Framework to Identify Social Bias in Dialogue
Ziyan Cai, Dingjun Wu, Ping Li · 2023
The content generated by dialogue systems has been found to contain social bias due to the social bias in the corpus and the algorithm design. We propose a Dialogue-based Contrastive Learning for identifying social Bias (DialCL-Bias) framework in dialogue based on two dimensions-relevance and fine-grain. For the identification of social bias relevance, we propose a Dialogue-based Semantic Contrastive Learning (DialSCL) method. DialSCL converts simple dichotomy into triple input and decomposes the classification target based on data annotation rules for achieving semantic-level classification. We use a Key Words and Cosine Similarity (KWCS) method to construct difficult samples and achieve improvement in DialCCL. For the identification of fine-grained social bias, we propose a Dialogue-based Context Contrastive Learning (DialCCL) method. DialCCL simultaneously learns the context-level features of samples and the parameters of classifiers in the same space to naturally generate a classifier. Finally, we measure the degree of bias in the open-domain dialogue corpus and the responses of ChatGPT by using the social bias identification classifiers.