CN-122019771-A - Opinion feedback agent execution method and system based on DeepSeek large model

CN122019771ACN 122019771 ACN122019771 ACN 122019771ACN-122019771-A

Abstract

The invention discloses a method and a system for executing opinion feedback agents based on DeepSeek large models, which realize semantic word segmentation, AI classification, service classification and property classification through DeepSeek large models, further realize automatic and accurate labeling classification of feedback content, automatically intercept invalid feedback such as spitting grooves, abuse and the like by combining a RAG knowledge base and a rule engine, realize intelligent filtration, reduce manual processing amount, introduce a machine learning algorithm to evaluate the effectiveness, urgency and influence range of the feedback, intelligently push high-value feedback to corresponding departments, develop value detection reminding and improve processing efficiency.

Inventors

LU FENG
LIU KAIFEI
LI HAONAN

Assignees

东方财富信息股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (9)

1. The opinion feedback agent executing method based on DeepSeek big model is characterized by comprising the following steps: Acquiring user feedback text information, including an APP form, a customer service work order and a social media message; Matching the user feedback text information with the dynamic domain dictionary, if the word which does not appear in the dynamic domain dictionary is used as a new word to update the dynamic domain dictionary, adopting a bidirectional attention mechanism if the matching is successful, carrying out context weighting processing on the user feedback text information according to a matching result to obtain weighted information data, and enhancing the recognition capability of the implicit semantics; Based on DeepSeek large models, emotion basic labels and intention basic labels are obtained according to weighted information data, based on an RAG knowledge base, payment fault service line labels and logistics delay service line labels are matched according to the weighted information data, based on a machine learning model, the effectiveness, the urgency and the influence range of feedback are evaluated according to the weighted information data; According to a preset keyword library and regular expressions, performing preliminary filtering on emotion basic tags, intention basic tags, payment fault service line tags, logistics delay service line tags, effectiveness, urgency and influence range of feedback, and performing secondary filtering on data identification implying negative emotion after preliminary filtering; extracting text features, behavior features and business features of the data after secondary filtering, and simultaneously predicting value scores and evaluating feedback values by priority labels; And pushing emotion basic labels, intention basic labels, payment fault service line labels, logistics delay service line labels, feedback effectiveness, urgency and influence range, secondary filtering data and feedback value to a customer service system, a work order system or a management layer billboard.
2. The opinion feedback agent execution method based on DeepSeek big models of claim 1, wherein the user feedback text information is obtained by adopting multi-source access modes such as API, file uploading and the like.
3. The opinion feedback agent execution method based on DeepSeek big models of claim 1, wherein when the dynamic domain dictionary is updated, the weight corresponding to the new word is calculated by: Wherein alpha, beta and gamma are weight coefficients, and are obtained through historical data training, and TF-IDF represents word frequency-inverse document frequency.
4. The opinion feedback agent execution method of claim 1, wherein classification thresholds are dynamically adjusted according to the accuracy rates corresponding to the emotion base tag and intent base tag, the payment fault line tag and logistics delay line tag, the validity of feedback, urgency and impact range.
5. The opinion feedback agent execution method based on DeepSeek big models of claim 4, wherein the classification threshold calculation formula is as follows: Wherein, the In order to adjust the threshold value of the threshold value, In order to adjust the threshold value before it is adjusted, Accuracy is the current classification Accuracy for learning rate.
6. The opinion feedback agent execution method of claim 1, wherein the text features comprise keyword density, sentence complexity, behavioral features comprise user historical feedback frequency, and business features comprise business line urgency level.
7. An opinion feedback agent execution system based on a DeepSeek large model, employing an opinion feedback agent execution method based on a DeepSeek large model as recited in claim 1, the system comprising: The data access intelligent agent is used for acquiring text information fed back by a user; The semantic analysis agent is used for matching the user feedback text information with the dynamic domain dictionary, if the dynamic domain dictionary does not appear, the dynamic domain dictionary is updated by taking the word as a new word, if the matching is successful, a bidirectional attention mechanism is adopted, the context weighting processing is carried out on the user feedback text information according to the matching result, and the recognition capability of the implicit semantic is enhanced; The classification decision agent is used for obtaining emotion basic labels and intention basic labels according to weighted information data through an AI classification layer based on DeepSeek large models, matching payment fault service line labels and logistics delay service line labels according to weighted information data through a service classification layer based on RAG knowledge base, and evaluating the effectiveness, urgency and influence range of feedback according to weighted information data through a property classification layer based on machine learning models; the filtering execution agent is used for performing preliminary filtering on emotion basic tags, intention basic tags, payment fault service line tags, logistics delay service line tags, effectiveness, urgency and influence range of feedback according to preset keyword libraries and regular expressions through a rule engine layer, performing secondary filtering on data identification hidden negative emotion after preliminary filtering by adopting a semantic enhancement layer of a fine tuning DeepSeek model The value evaluation intelligent agent is used for extracting text features, behavior features and business features of the data after secondary filtering through the multi-feature fusion model, and simultaneously predicting value scores and priority labels to evaluate feedback values by adopting a value scoring model constructed by a multi-task learning framework; and the output scheduling agent is used for pushing emotion basic labels, intention basic labels, payment fault service line labels, logistics delay service line labels, feedback effectiveness, urgency and influence range, secondary filtering data and feedback value to a customer service system, a work order system or a management layer billboard.
8. The opinion feedback agent execution method based on DeepSeek big models of claim 7, wherein the fine tuning DeepSeek model loss function is as follows: Wherein, the As a real tag it is possible to provide a real tag, In order to predict the probability of a probability, For the regularization coefficient(s), Is the L2 norm of the model parameters.
9. The opinion feedback agent execution method based on the DeepSeek large model of claim 7, wherein the value scoring model loss function is as follows: Wherein, the In order to predict the loss of a value score, The loss is predicted for the priority label, Is a balance coefficient.

Description

Opinion feedback agent execution method and system based on DeepSeek large model Technical Field The invention belongs to the technical field of intelligent agents, and particularly relates to an opinion feedback intelligent agent executing method and system based on DeepSeek big models. Background In the present digital age, the business scale of enterprises is continuously enlarged, the interaction between users and enterprises is increasingly frequent, the number of user opinion feedback (covering various aspects such as product experience, service problems, function suggestions and the like) presents exponentially increased situations, and the user feedback is an important way for enterprises to know own products and insufficient services and improve the user satisfaction. However, the conventional manner of manually processing user feedback is difficult to adapt to the large-scale feedback processing requirement, and many problems are exposed, so that great challenges are brought to customer service work of enterprises. Currently, there are some user feedback classification tools based on rules or simple models, which attempt to solve the problem of low manual processing efficiency, and these tools mainly classify and process user feedback through preset keywords or simple algorithms. For example, some tools may classify feedback including keywords into corresponding product function categories based on preset keywords related to product functions, and some tools may only classify user feedback in a single emotion dimension, such as determining whether the feedback is positive, neutral, or negative emotion. The prior art has the following problems: 1. The rule engine is limited in that feedback content of a semantically blurred or emerging scene cannot be processed depending on preset keywords. For example, "APP flash back" may be misjudged as "malfunction" or "performance problem" resulting in inaccurate classification. 2. And the classification dimension is single, namely, only emotion or business single dimension classification is supported, value evaluation (such as urgency and influence range) is not combined, feedback priority cannot be distinguished, and efficient feedback processing is not facilitated. 3. The filtering capability is insufficient, the recognition capability of hidden negative emotion (such as irony) or complex abuse (such as harmonic stems) is weak, the problem of false interception or missed interception is prominent, and the customer service work efficiency is affected. Disclosure of Invention The technical scheme of the invention aims to provide a method and a system for executing opinion feedback agent based on DeepSeek big models. The technical scheme of the invention provides an opinion feedback agent execution method based on DeepSeek big models, which comprises the following steps: Acquiring user feedback text information, including an APP form, a customer service work order and a social media message; Matching the user feedback text information with the dynamic domain dictionary, if the word which does not appear in the dynamic domain dictionary is used as a new word to update the dynamic domain dictionary, adopting a bidirectional attention mechanism if the matching is successful, carrying out context weighting processing on the user feedback text information according to a matching result to obtain weighted information data, and enhancing the recognition capability of the implicit semantics; Based on DeepSeek large models, emotion basic labels and intention basic labels are obtained according to weighted information data, based on an RAG knowledge base, payment fault service line labels and logistics delay service line labels are matched according to the weighted information data, based on a machine learning model, the effectiveness, the urgency and the influence range of feedback are evaluated according to the weighted information data; According to a preset keyword library and regular expressions, performing preliminary filtering on emotion basic tags, intention basic tags, payment fault service line tags, logistics delay service line tags, effectiveness, urgency and influence range of feedback, and performing secondary filtering on data identification implying negative emotion after preliminary filtering; extracting text features, behavior features and business features of the data after secondary filtering, and simultaneously predicting value scores and evaluating feedback values by priority labels; And pushing emotion basic labels, intention basic labels, payment fault service line labels, logistics delay service line labels, feedback effectiveness, urgency and influence range, secondary filtering data and feedback value to a customer service system, a work order system or a management layer billboard. Preferably, the user feedback text information is obtained by adopting multi-source access modes such as API, file uploading and the like. Preferably, when the dyn