Customer manager resource analysis method, system and device based on large model and medium

Abstract

The invention relates to the technical field of telecommunication network communication resource management, in particular to a customer manager resource analysis method, system and device based on a large model and a medium. The method comprises the following steps: firstly, inputting a natural language question obtained from a user side into an intention recognition module, and judging a query type; secondly, according to a query type, retrieving to obtain a closest question-answer pair, and obtaining query data; calling a three-dimensional weighted similarity algorithm according to the query data to obtain three pieces of data with the highest similarity; and finally, inputting the three pieces of data with the highest similarity into the large model for integration, generating a query result and returning the query result to the user side. Through the word segmentation capability of a large model, automatic analysis, matching and storage of client addresses are realized, the address operation process is simplified, self-service point supplementation is added, and the acceptance timeliness is improved; the hub effect of the customer manager between the customer and the operation and maintenance personnel is enhanced, manual intervention is reduced, and the customer experience degree and the communication efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of telecommunication network communication resource management, and in particular to a method, system, device and medium for analyzing customer manager resources based on a large model. Background Art

[0002] In the cloud network resource management business process, business personnel have defects such as difficult operation, complexity, and low efficiency. Due to the low intelligence of the system, customers' interest in the product is greatly reduced, resulting in the loss of many potential customer groups.

[0003] There are several types of problems in the cloud network resource management system: 1. Low efficiency. After understanding customer needs, account managers and marketing personnel need to communicate with the construction department and maintenance department on resource conditions many times and evaluate related costs based on product access plans. However, due to the lack of relevant tools and means, it is impossible to complete the evaluation and output of plans at a single point. The plan output takes a long time, making it difficult to give customers an accurate answer and requiring multiple visits.

[0004] 2. Low level of intelligence: Currently, many systems are still traditional query systems and have not introduced artificial intelligence technology, so they lack novelty.

[0005] The vector library retrieval accuracy is low. In the existing systems that introduce RAG vector library technology, almost all of them use simple cosine similarity to calculate the correlation of embedded vectors, which leads to low similarity matching accuracy.

[0006] With the existing network resource analysis methods, after understanding customer needs, account managers and marketing personnel need to communicate with the construction department and maintenance department on resource conditions many times and evaluate related costs based on product access plans. However, due to the lack of relevant tools and means to support them, it is impossible to complete the evaluation and output of plans at a single point. The plan output takes a long time, making it difficult to give customers an accurate answer. Multiple visits are required, and there are problems such as difficulty in obtaining resource information, long investigation and evaluation time, and reliance on back-end support. Summary of the invention

[0007] In view of the problems of existing network resource analysis methods, such as difficulty in obtaining resource information, long investigation and evaluation time, and reliance on back-end support, the present invention proposes a customer manager resource analysis method, system, equipment and medium based on a big model; the method uses the word segmentation capability of the big model to realize automatic parsing of customer addresses, matching and warehousing, simplifying the address operation process, while increasing self-service filling points and improving the timeliness of acceptance; it strengthens the hub role of customer managers between customers and operation and maintenance personnel, reduces manual intervention, and improves customer experience and communication efficiency.

[0008] The specific implementation contents of the present invention are as follows: A method for analyzing customer manager resources based on a large model specifically comprises the following steps: Step S1: inputting the natural language question obtained from the user end into the intention recognition module to determine the query type; Step S2: according to the query type, the most similar question-answer pair is retrieved, and query data is obtained according to the question-answer pair; Step S3: Based on the query data, a three-dimensional weighted similarity algorithm is called to obtain the three pieces of data with the highest similarity; Step S4: Input the three data with the highest similarity into the big model for integration, generate query results and return them to the user end.

[0009] In order to better implement the present invention, further, the step S1 specifically includes the following steps: Step S11: inputting the natural language question obtained from the user end into the intention recognition module; Step S12: supervise and fine-tune the intent recognition module based on the acquired corpus to determine the query type. If the input natural language question is an indicator query type, output 0; if the input natural language question is an analysis query type, output 1.

[0010] In order to better implement the present invention, further, if the query type is an indicator query type, the step S2 specifically includes the following steps: Step S21A: input the natural language question into the vector library, and call the three-dimensional weighted similarity algorithm to retrieve the most similar question-answer pair; Step S22A: Obtain the SQL statement from the most similar question-answer pair and insert it into the prompt word as a few-shot example; Step S23A: inputting the few-shot example prompt words inserted into the SQL statement into the large model to generate the correct SQL statement; Step S24A: Input the correct SQL statement into the SQL database and query to obtain the corresponding indicator data.

[0011] In order to better implement the present invention, further, if the query type is a classification query type, the step S2 specifically includes the following steps: Step S21B: input the natural language question into the vector library, and call the three-dimensional weighted similarity algorithm to retrieve the most similar question-answer pair; Step S22B: insert the first three analysis-class data in the closest question-answer pair into the prompt word as a few-shot example; Step S23B: input the inserted complete prompt word into the large model to generate a corresponding analysis report.

[0012] In order to better implement the present invention, further, the step S3 specifically includes the following steps: Step S31: calling the embedding model to embed the query data into three different dimensions; Do not go to S32: calculate the similarity between the obtained natural language question vector and the data vector obtained from the vector library, and obtain three similarity data; Step S33: weighting similarity data, calculating three-dimensional weighted similarity; Step S34: According to the three-dimensional weighted similarity, three pieces of data with the highest similarity are obtained.

[0013] In order to better implement the present invention, further, the step S4 specifically includes the following steps: Step S41: according to the query type corresponding to the query data, input the query data into the prompt word template corresponding to the query type, and call the large model to format the query data format; Step S42: If no relevant data is found, empty data is input into the prompt word template corresponding to the query type, and an error message indicating that the query failed or the cause of the failure cannot be obtained is output.

[0014] Based on the above-mentioned customer manager resource analysis method based on a big model, in order to better realize the present invention, further, a customer manager resource analysis system based on a big model is proposed, which is used to execute the above-mentioned customer manager resource analysis method based on a big model, including a judgment unit, a query unit, a similarity calculation unit, and an analysis unit; The judgment unit is used to input the natural language question obtained from the user end into the intention recognition module to judge the query type; The query unit is used to retrieve the most similar question-answer pair according to the query type, and obtain query data according to the question-answer pair; The similarity calculation unit is used to call a three-dimensional weighted similarity algorithm according to the query data to obtain three pieces of data with the highest similarity; The analysis unit is used to input the three data with the highest similarity into the large model for integration, generate query results and return them to the user end.

[0015] Based on the above-mentioned big model-based customer manager resource analysis method, in order to better realize the present invention, further, an electronic device is proposed, including a memory and a processor; a computer program is stored on the memory; when the computer program is executed on the processor, the above-mentioned big model-based customer manager resource analysis method is implemented.

[0016] Based on the above-proposed big model-based customer manager resource analysis method, in order to better realize the present invention, further, a computer-readable storage medium is proposed, on which computer instructions are stored; when the computer instructions are executed on the above-mentioned electronic device, the above-proposed big model-based customer manager resource analysis method is implemented.

[0017] The present invention has the following beneficial effects: (1) The present invention uses the SQL generation capability of the large model and combines it with the needs of users to achieve automatic SQL generation and data query, thereby improving the timeliness of acceptance.

[0018] (2) The present invention builds a bridge for real-time interaction with the resource system and uses big model technology to quickly analyze resource data, thereby achieving minute-level prediction of customer service activation capabilities and significantly shortening the response cycle.

[0019] (3) The present invention realizes functions such as instant demand confirmation, address matching, resource confirmation and customer business analysis through human-computer interaction, thereby improving the work efficiency of account managers and shortening the response time to customer needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic flow chart of the customer manager resource analysis method based on a large model provided by the present invention. DETAILED DESCRIPTION

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be understood that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, and therefore should not be regarded as limiting the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technical personnel in this field without making creative work are within the scope of protection of the present invention.

[0022] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "disposed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0023] Embodiment 1: This embodiment proposes a customer manager resource analysis method based on a large model, which specifically includes the following steps: Step S1: Input the natural language question obtained from the user into the intention recognition module to determine the query type.

[0024] Furthermore, the step S1 specifically includes the following steps: Step S11: inputting the natural language question obtained from the user end into the intention recognition module; Step S12: supervise and fine-tune the intent recognition module based on the acquired corpus to determine the query type. If the input natural language question is an indicator query type, output 0; if the input natural language question is an analysis query type, output 1.

[0025] Step S2: According to the query type, the most similar question-answer pair is retrieved, and query data is obtained based on the question-answer pair.

[0026] Furthermore, if the query type is an indicator query type, step S2 specifically includes the following steps: Step S21A: input the natural language question into the vector library, and call the three-dimensional weighted similarity algorithm to retrieve the most similar question-answer pair; Step S22A: Obtain the SQL statement from the most similar question-answer pair and insert it into the prompt word as a few-shot example; Step S23A: inputting the few-shot example prompt words inserted into the SQL statement into the large model to generate the correct SQL statement; Step S24A: Input the correct SQL statement into the SQL database and query to obtain the corresponding indicator data.

[0027] Furthermore, if the query type is a classification query, step S2 specifically includes the following steps: Step S21B: input the natural language question into the vector library, and call the three-dimensional weighted similarity algorithm to retrieve the most similar question-answer pair; Step S22B: insert the first three analysis-class data in the closest question-answer pair into the prompt word as a few-shot example; Step S23B: input the inserted complete prompt word into the large model to generate a corresponding analysis report.

[0028] Step S3: Based on the query data, a three-dimensional weighted similarity algorithm is called to obtain the three pieces of data with the highest similarity.

[0029] The step S3 specifically comprises the following steps: Step S31: calling the embedding model to embed the query data into three different dimensions; Do not go to S32: calculate the similarity between the obtained natural language question vector and the data vector obtained from the vector library, and obtain three similarity data; Step S33: weighting similarity data, calculating three-dimensional weighted similarity; Step S34: According to the three-dimensional weighted similarity, three pieces of data with the highest similarity are obtained.

[0030] Step S4: Input the three data with the highest similarity into the big model for integration, generate query results and return them to the user end.

[0031] Furthermore, the step S4 specifically includes the following steps: Step S41: according to the query type corresponding to the query data, input the query data into the prompt word template corresponding to the query type, and call the large model to format the query data format; Step S42: If no relevant data is found, empty data is input into the prompt word template corresponding to the query type, and an error message indicating that the query failed or the cause of the failure cannot be obtained is output.

[0032] Working principle: This embodiment first inputs the natural language question obtained from the user end into the intention recognition module to determine the query type; secondly, according to the query type, the most similar question and answer pair is retrieved, and the query data is obtained based on the question and answer pair; then, according to the query data, the three-dimensional weighted similarity algorithm is called to obtain the three data with the highest similarity; finally, the three data with the highest similarity are input into the big model for integration, the query results are generated and returned to the user end; through the word segmentation capability of the big model, the customer address is automatically parsed and matched into the warehouse, the address operation process is simplified, and at the same time, self-service filling points are added to improve the timeliness of acceptance; the role of the account manager as a hub between customers and operation and maintenance personnel is strengthened, manual intervention is reduced, and customer experience and communication efficiency are improved; by utilizing the big model and RAG technology, the intelligent conversion of natural language to SQL statements is realized, reducing the communication cost between the account manager and the customer.

[0033] Embodiment 2: This embodiment is based on the above embodiment 1. Figure 1 As shown, a specific embodiment is described in detail.

[0034] 1. The components involved in this embodiment are: 1) Large model.

[0035] 2) UIE (Universal Information Extraction) model.

[0036] 3) Vector library.

[0037] 4) Traditional SQL database.

[0038] 2. Data required for this embodiment: Question and answer: natural language-sql statement.

[0039] Corpus pair: Natural language-01 binary classification.

[0040] Domain knowledge analysis documents: Data required for root cause analysis, stored in the vector library.

[0041] 3. The implementation process of this embodiment is as follows: Step S1: Obtain a natural language question input by a user from a user terminal.

[0042] Step S2: Intent recognition: Input the natural language question into the intent recognition module, i.e., the UIE (Universal Information Extraction) model, and determine whether it is an indicator type query or an analysis type query. Use the corpus to supervise and fine-tune the UIE model. If the input natural language question is an indicator query type, the model outputs 0. If the input natural language question is an analysis query type, the output is 1. According to the category output in this step, enter the following two types of queries.

[0043] (1) Indicator query: The natural language question is input into the vector library, and the closest question-answer pair is retrieved using the three-dimensional weighted similarity algorithm.

[0044] Insert the SQL statement retrieved from the most similar question-answer pair into the prompt word as a few-shot example.

[0045] Input the few-shot prompt words for inserting SQL statements into the LLM large model (for example: Qwen2) to generate the correct SQL statements.

[0046] Input the SQL statement generated by the large model into the indicator database to query the corresponding indicator data.

[0047] (2) Analytical queries: The natural language question is input into the vector library, and the closest question-answer pair is retrieved using the three-dimensional weighted similarity algorithm.

[0048] Insert the first three analysis class data in the closest question-answer pair into the prompt word as a few-shot example.

[0049] Input the inserted complete prompt words into the large model to generate the corresponding analysis report.

[0050] Step S3: Three-dimensional weighted similarity algorithm.

[0051] Each input natural language question is embedded into three vectors of different dimensions using the embedding model: (512, 1024, 2048).

[0052] The similarity between the embedded vector and the vector in the vector library is calculated, and the similarity between vectors of the same vector dimension is calculated. Among them, the single-dimensional similarity calculation method is cosine similarity. Since each natural language question data has three vector representations of different dimensions (512, 1024, 2048), the similarity data of the two natural language question data A1 and A2 are calculated to be three (x1, x2, x3).

[0053] Then perform a weighted calculation on (x1, x2, x3) to obtain the final three-dimensional weighted similarity of A1 and A2: X=a1*x1+a2*x2+a3*x3 Among them, a1, a2, and a3 are custom hyperparameters. In this embodiment, the default values ​​are a1=0.4, a2=0.2, and a3=0.4.

[0054] Finally, the three data with the highest similarity X are returned.

[0055] Step S4: After obtaining the data, a generative answer is generated through the large model, and the indicator data is integrated into a table format and output to the customer.

[0056] (1) Indicator query: After finding the data with the highest similarity in the previous step, input the data into the fixed indicator query prompt word template and use the big model to format the data.

[0057] If no relevant data is found in the previous step, the empty data is still entered into the fixed indicator query prompt template. The prompt will indicate that if the large model gets empty data, the output data query fails.

[0058] (2) Analytical queries: The root cause analysis data output in the previous step is input into the fixed analysis query prompt word template, and the root cause analysis data is structurally processed and language optimized using the generation capability of the large model.

[0059] If no relevant domain knowledge analysis document data is found in the previous step, the empty data is still entered into the fixed analysis query prompt word template. The prompt word will indicate that if the large model obtains empty data, the output cannot obtain the cause analysis.

[0060] Working principle: This embodiment improves market share, strengthens the role of account managers as a hub between customers and operation and maintenance personnel, reduces manual intervention, and improves customer experience and communication efficiency based on the coverage of business openings in various provinces and the requirements for the proportion of operators by the Ministry of Telecommunications and Government Enterprises. This embodiment aims to create an efficient resource analysis assistant for account managers by combining large model technology. The assistant implements functions such as instant demand confirmation, address matching, resource confirmation, and customer business analysis through human-computer interaction, thereby improving the work efficiency of account managers, shortening customer demand response time, and improving customer satisfaction. This embodiment not only solves the pain points of account managers in actual work, but also enhances the market competitiveness of enterprises through technological innovation.

[0061] This embodiment is based on the background of the daily workflow of account managers, focusing on the practical problems and pain points faced by account managers in the marketing process, such as the difficulty in obtaining resource information, the long time for investigation and evaluation, and the reliance on back-end support, thus confirming the capability scope of this embodiment.

[0062] 1. Ability to use the word segmentation capability of a large model to complete address matching.

[0063] 2. When the address does not exist, the user's installation address can be structurally analyzed and automatically supplemented.

[0064] 3. It is possible to limit the product type and the access method to return the construction period of the resources that are not available.

[0065] 4. It can display the resources of each access method at the address.

[0066] 5. Ability to conduct comprehensive analysis of the client’s current business.

[0067] 6. Be able to answer questions and answers related to daily marketing regulations.

[0068] Core functions and innovations: (1) Intelligent address management and matching: Through the word segmentation capability of the big model, the customer address can be automatically parsed and matched into the database, simplifying the address operation process. At the same time, self-service filling points are added to improve the timeliness of acceptance.

[0069] (2) Real-time service provisioning capability prediction: Build a bridge for real-time interaction with resource systems, use big model technology to quickly analyze resource distribution, and achieve minute-level prediction of customer service provisioning capabilities, significantly shortening the response cycle and improving customer satisfaction.

[0070] (3) Optimization of labor costs: Through the instant mutual assistance model, the back-end support manpower can be reduced by 50%, thereby reducing operating costs.

[0071] The other parts of this embodiment are the same as those of the above-mentioned embodiment 1, and thus will not be described in detail.

[0072] Embodiment 3: This embodiment, based on any one of the above-mentioned embodiments 1-2, proposes a customer manager resource analysis system based on a big model, which is used to execute the above-mentioned customer manager resource analysis method based on a big model, including a judgment unit, a query unit, a similarity calculation unit, and an analysis unit; The judgment unit is used to input the natural language question obtained from the user end into the intention recognition module to judge the query type; The query unit is used to retrieve the most similar question-answer pair according to the query type, and obtain query data according to the question-answer pair; The similarity calculation unit is used to call a three-dimensional weighted similarity algorithm according to the query data to obtain three pieces of data with the highest similarity; The analysis unit is used to input the three data with the highest similarity into the large model for integration, generate query results and return them to the user end.

[0073] This embodiment also proposes an electronic device, including a memory and a processor; a computer program is stored on the memory; when the computer program is executed on the processor, the above-mentioned customer manager resource analysis method based on a large model is implemented.

[0074] This embodiment also proposes a computer-readable storage medium, on which computer instructions are stored; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned customer manager resource analysis method based on a large model is implemented.

[0075] The other parts of this embodiment are the same as any one of the above-mentioned embodiments 1-2, so they will not be repeated here.

[0076] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for analyzing customer manager resources based on a large model, characterized in that: The specific steps include: Step S1: inputting the natural language question obtained from the user end into the intention recognition module to determine the query type; Step S2: according to the query type, the most similar question-answer pair is retrieved, and query data is obtained according to the question-answer pair; Step S3: Based on the query data, a three-dimensional weighted similarity algorithm is called to obtain the data with the highest similarity; Step S4: Input the data with the highest similarity into the big model for integration, generate generative answers, and integrate the indicator data into a table format and return it to the user end.

2. According to the big model-based customer manager resource analysis method of claim 1, it is characterized in that: The step S1 specifically includes the following steps: Step S11: inputting the natural language question obtained from the user end into the intention recognition module; Step S12: supervise and fine-tune the intent recognition module based on the acquired corpus to determine the query type. If the input natural language question is an indicator query type, output 0; if the input natural language question is an analysis query type, output 1.

3. The method for analyzing customer manager resources based on a large model according to claim 2 is characterized in that: If the query type is an indicator query type, step S2 specifically includes the following steps: Step S21A: input the natural language question into the vector library, and call the three-dimensional weighted similarity algorithm to retrieve the most similar question-answer pair; Step S22A: Obtain the SQL statement from the most similar question-answer pair and insert it into the prompt word as a few-shot example; Step S23A: input the few-shot prompt words inserted into the SQL statement into the large model to generate the correct SQL statement; Step S24A: Input the correct SQL statement into the SQL database and query to obtain the corresponding indicator data.

4. The method for analyzing customer manager resources based on a large model according to claim 2 is characterized in that: If the query type is a classification query, step S2 specifically includes the following steps: Step S21B: input the natural language question into the vector library, and call the three-dimensional weighted similarity algorithm to retrieve the most similar question-answer pair; Step S22B: insert the first three analysis-class data in the closest question-answer pair into the prompt word as a few-shot example; Step S23B: input the inserted complete prompt word into the large model to generate a corresponding analysis report.

5. The method for analyzing customer manager resources based on a large model according to claim 1, characterized in that: The step S3 specifically comprises the following steps: Step S31: calling the embedding model to embed the query data into vectors of three different dimensions; Do not go to S32: calculate the similarity between the embedded vector and the data vector obtained from the vector library to obtain similarity data; Step S33: weighting similarity data, calculating three-dimensional weighted similarity; Step S34: According to the three-dimensional weighted similarity, three pieces of data with the highest similarity are obtained.

6. The method for analyzing customer manager resources based on a large model according to claim 5, characterized in that: The step S4 specifically comprises the following steps: Step S41: according to the query type corresponding to the query data, input the query data into the prompt word template corresponding to the query type, and call the large model to format the query data format; Step S42: If no relevant data is found, empty data is input into the prompt word template corresponding to the query type, and an error message indicating that the query failed or the cause of the failure cannot be obtained is output.

7. A customer manager resource analysis system based on a big model, used to execute the customer manager resource analysis method based on a big model as claimed in claim 1, characterized in that: It includes a judgment unit, a query unit, a similarity calculation unit, and an analysis unit; The judgment unit is used to input the natural language question obtained from the user end into the intention recognition module to judge the query type; The query unit is used to retrieve the most similar question-answer pair according to the query type, and obtain query data according to the question-answer pair; The similarity calculation unit is used to call a three-dimensional weighted similarity algorithm according to the query data to obtain three pieces of data with the highest similarity; The analysis unit is used to input the three data with the highest similarity into the large model for integration, generate query results and return them to the user end.

8. An electronic device, characterized in that: It comprises a memory and a processor; a computer program is stored in the memory; when the computer program is executed on the processor, the customer manager resource analysis method based on a large model as described in any one of claims 1-6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 8, the customer manager resource analysis method based on a large model as described in any one of claims 1-6 is implemented.

Images