TCM Translator: A Sequence Generation Approach for Prescribing Herbal Medicines

In traditional Chinese medicine (TCM), herbal prescriptions are accumulated from doctors' clinical experience and play an essential role in treatment process for more than thousands of years. Mining their containing many-to-many relationship between symptoms and herbs are important for both clinical practice and novel prescriptions development. Previously, the research attention in this field is attracted by topic models and multi-label classification approaches, but they don't fully capture the correlations of labels and achieve very promising performance. Alternatively, we problematic it as a machine translation problem, that the source sentence is formed by symptoms and the target sentence consists of herbs. Two sentences are both in weakly order. In terms of this assumption, we propose a novel sequence-to-sequence (seq2seq) architecture, namely TCM translator, to translate symptoms to herbs. Seq2seq, consisting of an encoder and a decoder, is a well known framework for resolving the machine translation problem in natural language processing (NLP) community. It maps a bag of symptoms to the latent space, which is decoded to the target sentence, i.e., herbs. In this paper, transformer is used as our encoder to explore the latent correlations of symptoms and LSTM is implemented as the decoder to generate the herbs sequentially. The experimental results demonstrate that our proposed approach is effective and outperforms other baseline models by a substantial margin.