A Survey of Language Priors for Visual Question Answering

Abstract

In recent years, with the development of deep learning technology, visual question answering tasks have gradually attracted the attention of scientific researchers. Due to the continuous improvement of relevant large-scale standard data sets, a large number of visual questions answering research results have been released one after another, and the accuracy rate of the visual question answering model based on deep learning on the data set has been continuously improved. Recent studies have found that the previously proposed visual question answering model has different degrees of data set language prior problems, that is, the model is overly dependent on the strong phase between the question and the answer in the training process. Many articles briefly describe various research methods, and look forward to the future development direction of alleviating the prior problem of visual question answering based on the existing research.