Method and device for outputting insurance outbound conversation information based on generative large language model

A method for outputting insurance outbound conversation information based on a generative large language model, comprising:

• acquiring conversation information in outbound consultation on an insurance product;
• classifying the conversation information with a trained generative large language model to obtain a classification result of the conversation information, wherein the generative large language model is trained with a training sample set obtained by performing deduplication on crossing area in formulated conversation sample data;
• and
• determining the insurance outbound conversation information for the outbound consultation based on the classification result and outputting the insurance outbound conversation information, wherein before the classifying the conversation information with a trained generative large language model to obtain a classification result of the conversation information, the method further comprises:
• acquiring historical conversation information, and extracting pieces of conversation sample data from the historical conversation information according to the length of a text sliding window;
• and
• determining the crossing area in the conversation sample data and performing deduplication on the crossing area to obtain the training sample set, to train the generative large language model with the training sample set, wherein the crossing area includes time crossing area and text content crossing area, and wherein the determining the crossing area in the conversation sample data and performing deduplication on the crossing area to obtain the training sample set comprises:
• performing text analysis on the conversation sample data, comparing text of previous conversation sample data with text of the conversation sample data, and determining the text content crossing area;
• deleting target text content corresponding to the text content crossing area, if the target text content is marked as modularized words;
• and
• performing deduplication on target text content corresponding to the text content crossing area, if the target text content is marked as non-modularized words,
• wherein the modularized words are pre-defined text components that do not contribute to the meaning of the conversation information, and
• wherein the non-modularized words are text components that convey technical content related to conversation context of the conversation information;
• formulating the generative large language model to be trained and acquiring the training sample set that has performed deduplication;
• and
• training the generative large language model with the training sample set and completing the training of the generative large language model when a loss value of the generative large language model matches a preset loss threshold, wherein the loss value is calculated when the generative large language model is trained with the conversation sample data which has performed deduplication on target text content.