Hierarchical Multitask Learning Model for Chinese Grammatical Error Detection

Grammatical error detection is important in natural language processing and has developed rapidly in recent years. Most existing studies based on neural networks have achieved good performances. However, similar to other studies pertaining to natural language processing, such as part-of-speech tagging and named entity recognition, severe data imbalance is encountered: the number of negative samples is significantly greater than the number of positive samples. This results in the tendency of learning negative samples during model training. Hence, we define the Chinese grammatical error detection problem as a hierarchical task and propose a hierarchical multitask learning model. The first and second levels comprise grammatical error detection and error type classification tasks, respectively. In the former task, the model only detects grammatical errors and does not distinguish error types, and it is known as a coarse-grained network. In the latter task, the model classifies the detected grammatical errors into their corresponding categories, and it is known as a fine-grained network. By doing this, the positive class information in the first-level task can be increased, and the accuracy of the second-level task further improved. Therefore, we cast each level as a subtask and developed a multitask learning framework based on the two subtasks. Experimental results show that the performance of the proposed method is better than that of the basic model, particularly in term of position level.