A Two-Channel End-to-End Network Based on Dynamic Corpus of Knowledge Graph for Intelligent Recognition of Traditional Chinese Medicine Terminology

Accurate analysis of Traditional Chinese Medicine (TCM) terminology is pivotal in facilitating effective communication between TCM practitioners and patients, thereby enabling precise diagnosis and treatment. TCM terminology includes two forms: speech and text. However, current deep learning methods for TCM terminology recognition are hindered by insufficient corpus and defects of the end-to-end learning framework, which leads to the low accuracy of TCM terminology recognition. To solve the above problems, this paper first combines the information of text and picture of TCM terminology and proposes an extended model of TCM terminology corpus. Joint optimization of text and picture-based knowledge graph TCM terminology corpus expansion model is incorporated, and the traditional corpus is supplemented by incrementally constructing a dynamic TCM terminology corpus. Secondly, the text-speech end-to-end conversion mechanism is used to realize the synchronous incremental expansion of the TCM dynamic speech corpus. After that, the TCM dynamic speech corpus is deeply trained through a unified streaming and non-streaming two-pass end-to-end model, to realize the accurate recognition of the speech of TCM terminology. Additionally, to mitigate the readability challenges posed by redundant words in TCM pronunciation, a directed acyclic graph (DAG) and dynamic programming (DP) based framework is proposed for redundant word detection. The results show that the accuracy of the speech recognition algorithm proposed in this paper for the speech of TCM terminology is increased by 12.92%, 11.18%, and 19.95% compared with three public speech recognition engines of Iflytek, Aliyun, and Baidu, respectively. In addition, the detection accuracy of redundant words reaches 90.67% on average. This method is able to overcome the limitations of TCM terminology recognition. It provides a comprehensive solution with significantly higher recognition accuracy compared to existing methods.