WeChat group-based online medical community (WGMC) is increasingly accepted by the public because of its high efficiency, convenience, and shared advantages in seeking medical resources. However, the problem to separate a complete dialogue relationship from the chat records is arising because efficient online community management is based on clear dialogue relationships and clear topics. To solve the problem, we proposed a hybrid three-stages BERT method to recognize the dialogue relationships in "Home of Love" --- a central nervous system tumor online healthy community WeChat group. First, based on the social support theory, a multi-layer BiLSTM model is proposed to classify the conversation scenes into five classes. Then, two domain adaptation methods for transfer learning are designed to optimize the BERT pre-training model for specific tasks using the "Haodaifu" as the training corpus. Finally, a hybrid BERT method based on the text scene information and the pre-training model is proposed to recognize the dialogue relationships, and its feasibility is verified by manual labeling. The results show that adding more prior knowledge to the dialogue recognition model by extracting the social support scene information can effectively improve the classification ability and stability of the model.

This extremely valuable collection of texts in the Lahu language represents the language and culture in the 1960’s, a time when the heritage language and culture were still vibrant and not yet globalized, hence the title Window on a Vanished World. It is also one of the largest collections of texts in any Tibeto-Burman language. The texts are available as a book and online with the audio (originally from 1960’s magnetic tape). This is a massive achievement for all involved in the recording, conversion, and editing.