FedParsing: a Semi-Supervised Federated Learning Model on Semantic Parsing

Anonymous

Abstract: Although many semantic parsing models have been proven to work effectively on "NL-to-SQL", the limitation of annotated datasets remains a great challenge. In many semi-supervised models, while they use unlabeled data to greatly improve the model accuracy, they fail to take data privacy of users into account. In this work, we focus on improving the performance of the semantic parsing model and protecting the users’ data privacy without increasing the size of the labeled dataset. Our new model, which is named FedParsing, is a semi-supervised Federated Learning model. In order to solve the difficulty on convergence of traditional semi-supervised Federated Learning model, we incorporate the Mean Teacher algorithm and apply the Exponential Moving Average algorithm to update model parameters. Experiments on WikiSQL show that with extra unlabeled data, our model performs better than supervised training model and traditional semi-supervised Federated Learning model, which proves the effectiveness of FedParsing model.