Search

CN-121980028-A - Novel telecom complaint identification method and system

CN121980028ACN 121980028 ACN121980028 ACN 121980028ACN-121980028-A

Abstract

The invention relates to a novel telecom complaint identification method and system, wherein the method inputs structured complaint data acquired from a telecom complaint form into a large model to generate a complaint text profile set with consistent semantics, inputs DeepSeek the complaint text profile set into a pre-training model to conduct complaint service identification, adopts a layered parameter defrosting adjustment strategy to dynamically defrost a transducer parameter at the bottom layer of the DeepSeek pre-training model, extracts semantic features of the complaint text profile set through a task-oriented dual-mode feature extraction mechanism based on the dynamically defrosted transducer parameter, and inputs the semantic features into a small classifier at the tail end of the DeepSeek pre-training model to output different types of complaint service identification results. Therefore, the telecommunication complaint recognition provided by the invention has the advantages of high recognition precision, low deployment cost and high calculation efficiency.

Inventors

  • CHENG YU
  • LIN JIN

Assignees

  • 福建福诺移动通信技术有限公司

Dates

Publication Date
20260505
Application Date
20251209

Claims (8)

  1. 1. A novel telecommunications complaint recognition method, comprising: Collecting structured complaint data from a telecommunication complaint form, and inputting the structured complaint data into a large model to generate a semantically coherent complaint text profile set; Inputting DeepSeek the complaint text profile set into a pre-training model for complaint service identification, dynamically thawing the transducer parameters at the bottom layer of the DeepSeek pre-training model by adopting a layered parameter thawing adjustment strategy to obtain dynamic thawed transducer parameters, extracting semantic features of the complaint text profile set based on the dynamic thawed transducer parameters through a task-oriented dual-mode feature extraction mechanism, and inputting the semantic features into a small classifier at the tail end of the DeepSeek pre-training model to output different types of complaint service identification results.
  2. 2. The novel telecommunications complaint recognition method of claim 1 wherein inputting the structured complaint data into a large model to generate a semantically coherent set of complaint text profiles includes: Acquiring a complaint treatment timeliness score, a personnel specialty score and a complaint result rationality score from the structured complaint data, inputting the complaint treatment timeliness score, the personnel specialty score and the complaint result rationality score into a first formula for calculation to obtain a comprehensive evaluation score, wherein the first formula is as follows: ; Wherein Z represents the comprehensive evaluation score, Represents a timeliness score for the complaint treatment, A person's expertise score is indicated, A rationality score representing a complaint result; Adding the comprehensive evaluation score to the structured complaint data to obtain added structured complaint data, and converting the added structured complaint data into a natural language instruction set according to a predefined instruction template; The set of natural language instructions is input into a large model to generate a semantically coherent set of complaint text profiles.
  3. 3. The method of claim 1, wherein the bottom layer comprises 12 layers, the dynamically thawing the fransformer parameters of the DeepSeek pre-training model bottom layer using a hierarchical parameter thawing adjustment strategy, the dynamically thawing the fransformer parameters comprising: Calculating the data quantity of the complaint text profile set, simultaneously acquiring the edge equipment resource quantity, judging whether the edge equipment resource quantity exceeds a first resource threshold value, simultaneously judging whether the data quantity exceeds a first data threshold value, and freezing the layer 1 to layer 8 transducer parameters if the edge equipment resource quantity exceeds the first resource threshold value or the data quantity exceeds the first data threshold value, and thawing the layer 9 to layer 12 transducer parameters only to obtain the thawed layer 9 to layer 12 transducer parameters; If the resource amount of the edge equipment does not exceed the first resource threshold value and the data amount does not exceed the first data threshold value, judging whether the data amount exceeds the second data threshold value, if the data amount exceeds the second data threshold value, freezing the layer 1 to layer 9 transducer parameters, and thawing the layer 10 to layer 12 transducer parameters to obtain thawed layer 10 to layer 12 transducer parameters; And if the data quantity does not exceed the second data threshold value, freezing the Transformer parameters of the layers 1 to 10, and thawing the Transformer parameters of the layers 11 to 12 only to obtain the thawed Transformer parameters of the layers 11 to 12.
  4. 4. The method of claim 1, wherein the extracting semantic features of the complaint text profile set by a task-oriented dual-mode feature extraction mechanism based on dynamically defrosted transducer parameters, and inputting the semantic features into a mini-classifier at the end of the DeepSeek pre-training model comprises: Semantic coding is carried out on the complaint text profile set based on the dynamically unfrozen transducer parameters, so that a hidden state is obtained; Acquiring a current task type, when the current task type is complaint service type identification, calculating an average value of the hidden states by adopting a hidden state average method, and inputting the average value serving as a semantic feature of the complaint text profile set into a small classifier positioned at the tail end of the DeepSeek pre-training model; And when the current task type is the service quality type identification, the global semantic association reserved by the hidden state is used as the semantic feature of the complaint text profile set to be input into a small classifier positioned at the tail end of the DeepSeek pre-training model.
  5. 5. The novel telecommunications complaint recognition method of claim 1, wherein the inputting the semantic features into a mini-classifier at the end of the DeepSeek pre-training model to output different types of complaint service recognition results includes: Inputting the semantic features into a small classifier positioned at the tail end of the DeepSeek pre-training model to obtain different types of complaint service identification results; Counting the total type number of the complaint service identification results, simultaneously introducing a temperature regulation function to regulate the temperature of the complaint service identification results based on the total type number to obtain multi-type probability distribution containing confidence, selecting the complaint service identification result with the largest confidence from the multi-type probability distribution, and outputting the complaint service identification result, wherein the temperature regulation function is as follows: ; Wherein, the Representing a multi-type probability distribution of type i, Represents the complaint service recognition result of type i, N represents the total type number of the complaint service recognition result, Representing the temperature coefficient.
  6. 6. The method for identifying telecommunication complaints according to claim 5, wherein the step of obtaining a multi-type probability distribution containing confidence, and selecting a complaint service identification result output with the highest confidence from the multi-type probability distribution comprises: counting the single quantity of each type in the complaint service identification result obtained by the small classifier, and calculating the ratio of the single quantity of each type in the total type quantity to obtain the type ratio of each type; Introducing a dynamic weight distribution strategy into the multi-type probability distribution, dynamically distributing a corresponding weight coefficient for each type according to the type duty ratio of each type, and multiplying the weight coefficient of each type by a corresponding confidence coefficient to obtain an optimized multi-type probability distribution; And selecting the complaint service identification result with the maximum confidence from the optimized multi-type probability distribution and outputting the complaint service identification result.
  7. 7. The novel telecommunications complaint recognition method of claim 1, wherein the mini-classifier is a five-layer fully connected neural network having a structure sequence of 768-dimensional first input layer, 512-dimensional first hidden layer, 256-dimensional second hidden layer, 128-dimensional third hidden layer and 64-dimensional fourth hidden layer.
  8. 8. A novel telecommunications complaint recognition system comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1 to 7 when the computer program is executed.

Description

Novel telecom complaint identification method and system Technical Field The invention relates to the technical field of data processing, in particular to a novel telecom complaint identification method and system. Background With the rapid development of telecommunication services such as 5G and internet of things, the demand of users for telecommunication services is increasingly raised, which directly results in the situation that telecommunication complaint data presents explosive growth and is more complex and diversified. However, the existing telecom complaint recognition system has the following main problems in practical application: The model has the defects of poor adaptability, that is, the traditional machine learning model such as a Support Vector Machine (SVM), a random forest and the like is seriously dependent on manual feature engineering when telecommunication complaint data are processed, a technician is required to manually mine key features in the data, the model is difficult to adapt to a complex semantic mode of emerging business complaints, and the recognition accuracy rate of the novel type complaints is low. And (II) the deployment cost is high, the calculation efficiency is low, and part of the existing schemes adopt a multi-model fusion or cascade architecture for pursuing the recognition precision, and the design improves the recognition effect to a certain extent, but introduces additional model interaction and calculation cost, so that the deployment cost is increased, the system response delay is caused, and the real-time requirement cannot be met. Therefore, a technical solution that combines high recognition accuracy, low deployment cost and high computing efficiency is needed. Disclosure of Invention The invention aims to solve the technical problems that the invention provides a novel telecom complaint identification method and system, which realize high identification precision, low deployment cost and high calculation efficiency. In order to solve the technical problems, the invention adopts the following technical scheme: in a first aspect, the present invention provides a novel telecommunications complaint recognition method comprising: Collecting structured complaint data from a telecommunication complaint form, and inputting the structured complaint data into a large model to generate a semantically coherent complaint text profile set; Inputting DeepSeek the complaint text profile set into a pre-training model for complaint service identification, dynamically thawing the transducer parameters at the bottom layer of the DeepSeek pre-training model by adopting a layered parameter thawing adjustment strategy to obtain dynamic thawed transducer parameters, extracting semantic features of the complaint text profile set based on the dynamic thawed transducer parameters through a task-oriented dual-mode feature extraction mechanism, and inputting the semantic features into a small classifier at the tail end of the DeepSeek pre-training model to output different types of complaint service identification results. The method has the advantages that the structural complaint data is converted into the complaint text profile set with consistent semantics through the large model, the fragmentation defect of the structural complaint data is eliminated, and the follow-up DeepSeek pre-training model can be focused on semantic logic rather than format analysis when the follow-up DeepSeek pre-training model is used for identifying complaint services. The hierarchical parameter defrosting adjustment strategy is adopted to dynamically defrost the Transformer parameters at the bottom layer of the DeepSeek pre-training model, instead of using the full quantity of parameters to participate in the extraction of semantic features, the deployment cost is reduced, the extraction efficiency and the resource utilization rate of the semantic features are improved, and meanwhile, the small classifier at the tail end of the DeepSeek pre-training model is used for outputting different types of complaint service identification results to replace the full quantity reasoning of the DeepSeek pre-training model, so that the identification precision is reserved, the operation and maintenance difficulty is reduced, the effects of high identification precision, low deployment cost and high calculation efficiency are realized on the whole flow of complaint service identification, and the real-time requirement is met. Optionally, the inputting the structured complaint data into a large model to generate a semantically coherent set of complaint text profiles includes: Acquiring a complaint treatment timeliness score, a personnel specialty score and a complaint result rationality score from the structured complaint data, inputting the complaint treatment timeliness score, the personnel specialty score and the complaint result rationality score into a first formula for calculation to obtain a comprehensive eval