English Writing Intelligent System based on Recurrent Neural Networks

The author proposes a sequence annotation model based on recurrent neural networks to address the impact of noise such as grammar errors on sequence information extraction in English writing corpora, and propose a sequence annotation based method for detecting and correcting grammar errors in English writing. The recurrent neural network (RNN)-based sequence annotation model leverages character, word, and sequence data integration, incorporating coarse-grained learning to enhance the robustness of the annotation process. This method divides the annotation procedure into coarse and fine stages, providing a more resilient approach. In the context of English grammar error detection and correction, the sequence annotation model is employed to effectively label and address grammar errors, and detects and corrects the original words according to the annotation results, avoiding the problem of traditional methods requiring manual extraction of a large number of features. The experimental results show that the sequence annotation model based on recurrent neural networks proposed by the author achieves an accuracy of 95.62% in part of speech annotation of ESL corpus; Meanwhile, in the part of speech tagging of news corpus, the accuracy of this model reaches 96.50%; In the CONLL2003 named entity recognition task, the F1 value reached 90.27%; Utilizing the sequence annotation model for English grammar error detection and correction tasks, the author's approach yields notable improvements over existing methods. Specifically, in the context of overall error correction, the F1 value of the author's method surpasses the UIUC method by 4% and outperforms the Corpus GEC method by an equivalent margin. When focusing on the correction of preposition errors, the author's method excels even further, achieving a 20% higher F1 value compared to the UIUC method and a 12% improvement over the Corpus GEC method.