Mitigating stereotypes in word embedding through sentiment modulation

Abstract

Word embedding is an influential framework to quantify the meaning of a word, which is widely used in machine learning at a pre-processing level for natural language processing (NLP). However, word embedding trained with a large number of contexts encodes not only general syntactic and semantic meaning of a word, but also the stereotypes and biases that people may have. This thesis proposes a method to indirectly mitigate the stereotypes in the trained word embedding by modulating the dimension of sentimental attributes in a human entity without imposing equal probability on the compatible social groups. To prevent the word embedding from creating problematic predictions such as a stereotype threat, we modulate the strength of the association between a human entity and sentimental attribute and indirectly reduce the gender bias of the embedding model. We show that the proposed method preserves the overall embedding performance. We also confirm that increasing the strength of the association between human entities and sentimental attributes amplifies the model bias through experiment.