Research on customer service conversation classification technology based on multi-source information

Inhalt:
At present, Meituan uses the BERT model to perform binary classification judgments on customer service conversations. The input form of the conversation is often to convert the text in the form of a dialogue into a sentence/short text as the input of the model, and the downstream is often a single-sentence classification task. Inputs in the form of paragraphs have some flaws, such as the model’s inability to recognize the role attribution of sentences. And the sentence in the form of dialogue is different from the complete text, there is no continuity before and after. In response to the above problems, we propose a pre-trained model text classification algorithm based on fusion of multi-source information. By integrating role information and keyword information, the model indicators are improved. Finally, the feasibility of the improvement is verified by comparing the experimental results and analyzing the improvement effect.