A Causal-Regularity and Deep Learning-Based Conjoint Inference

Deep learning is based on the characteristics of the metadata information from input data to train and test data. It deals with high-dimensional and raw data. Through layer-by-layer stacking transformation, deep neural networks are trained, new rules being learnt. Therefore, even if the new data is given, the networks can automatically transform the data to an accurate knowledge representation based the new rules. The text classification process of deep learning, however, is subject to requirements from specific environments. If the premises cannot be satisfied, the robustness and accuracy of classification will be greatly weakened. Meanwhile, deep learning corresponds to a process of black-box training, and it is impossible to describe feature information artificially or explain the training results in many scenarios. If deep learning is to have a highly intelligent learning ability akin to that of human beings, it is necessary to add to it a high-level reasoning and judgment process similar to the human cognitive process. Motivated by causal regularity, we proposed a conjoint reasoning framework, named Causal Regularity Inference Framework (CRIF), which is based on the combination of causal regularity and deep learning to achieve text classification. Causal inference meets the requirements of high-level information reasoning. It can determine the correlations and causalities between the inputs and find out confounding feature words at the same time. Causal inference makes the training of deep learning more reasonable and explicable. Therefore, effective feature extraction can be carried out in the text processing to remove invalid confounding words, and reduce the data redundancy in training process and improve the accuracy of the text classification.