Problem classification method and device based on large language model, equipment and medium

Abstract

The invention provides a problem classification method and device based on a large language model, equipment and a medium, and the method comprises the steps: converting a problem into a dynamic high-dimensional vector through a pre-trained large language model, and capturing context semantics; a density-centroid dual-drive clustering algorithm is adopted to realize multi-dimensional semantic aggregation; generating a human readable category label in combination with the keyword co-occurrence network and a large language model; a priority order is generated through frequency statistics and key semantic weight enhancement; and finally, outputting a machine analyzable / manual visual decision report. The method has the beneficial effects that based on the dynamic vectorization and unsupervised clustering algorithm of the large language model, the manual intervention requirement is eliminated, and the time consumed for processing ten thousand-level problems is shortened; synonymous / polysemy semantic association is captured through vector space similarity measurement, and the classification accuracy is obviously improved compared with traditional keyword matching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of large language model, in particular to a question classification method, device, equipment and medium based on large language model. BACKGROUND

[0002] With the popularity of the Internet and intelligent terminal devices, the amount of product feedback submitted by users through online platforms presents exponential growth. Enterprises need to identify high-frequency problem types from massive texts to optimize product experience, but traditional methods have significant defects. Existing technologies mainly rely on manual classification or rule-based keyword matching: manual classification requires a large amount of human resources to label problems one by one, which takes several hours or even several days, and is highly subjective and difficult to unify standards; the keyword matching scheme can be automatically processed, but it cannot cope with the complexity of natural language, such as synonymous expressions (such as "login failure" and "unable to log in"), polysemy (such as "stuttering" can refer to interface delay or payment blockage), and context-dependent problems (such as "payment unsuccessful" needs to distinguish between network failure or password error), resulting in a classification accuracy rate of less than 40%. In addition, existing technologies can only output basic category counts, lack of summary description of the nature of the problem (such as unable to automatically generate accurate labels such as "account authentication anomaly"), and cannot quantify the priority of each type of problem. This makes it difficult for enterprises to extract effective decision-making basis from statistical results, often delaying core problem repair due to resource allocation misfocus.

[0003] Current systems attempt to introduce machine learning models (such as SVM, Naive Bayes) for classification, but still require manual annotation of training data, and the model has weak generalization ability. Emerging large language models have semantic understanding capabilities, but when directly used for massive problem classification, they face input length limitations (such as supporting only 512 tokens), high response delays (several seconds per inference), and hallucination risks (error classification rate exceeding 15%), which cannot meet enterprise-level real-time analysis needs. The existing technical system has systematic deficiencies in automation, accuracy, and decision support.

[0004] Therefore, there is an urgent need for a technical solution that combines semantic understanding, automatic clustering, and intelligent decision support to break through the dual bottleneck of efficiency and accuracy. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a question classification method, device, equipment and medium based on large language model, aiming to improve the inference efficiency and accuracy of large language model.

[0006] To solve the above technical problems, the technical solution adopted by the present application is: a question classification method based on large language model, comprising:

[0007] Receiving a multi-source heterogeneous user question text dataset, using the semantic encoding layer of a pre-trained large language model to automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness;

[0008] Based on the vector space similarity measurement results between the dynamic high-dimensional vector representations, a multi-dimensional semantic aggregation of user questions is performed using an unsupervised clustering algorithm to form multiple question categories with semantic consistency;

[0009] For each semantic consistency category, the common semantic features of intra-class user questions are automatically extracted through the text induction capability of the large language model, and a summary description label reflecting the core intent of the category is generated;

[0010] Real-time calculation of the frequency distribution of each semantic consistency category, and automatic priority ranking of the categories according to the preset decision logic;

[0011] Fusion of the summary description label and priority ranking result to generate a structured statistical report for decision analysis.

[0012] Further, before the step of receiving a multi-source heterogeneous user question text dataset, using the semantic encoding layer of a pre-trained large language model to automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness, it includes:

[0013] Noise filtering of user question text is performed through a rule engine and regular expression matching mechanism to remove URLs, special symbols, and non-text characters;

[0014] A dictionary-based text normalization tool is applied to perform at least one semantic unification process such as spelling correction, synonym replacement, and abbreviation expansion.

[0015] Further, the vector space similarity measurement results are obtained through multiple distance calculation strategies, including:

[0016] The cosine similarity between vectors is calculated to measure the semantic direction consistency;

[0017] The Euclidean distance between vectors is calculated to measure the semantic space tightness;

[0018] The direction consistency and space tightness are combined for dynamic weighted fusion to generate a comprehensive similarity score.

[0019] Further, the unsupervised clustering algorithm is a dual-driven clustering model based on density and centroid, and its execution process includes:

[0020] Core semantic dense areas are identified through a density clustering algorithm;

[0021] Based on the density distribution of the core region, the seed points of the centroid clustering algorithm are initialized;

[0022] Iterative optimization of cluster boundaries until semantic convergence conditions are met.

[0023] Further, the operation of automatically extracting the common semantic features of the intra-class user questions specifically includes:

[0024] Extracting the keyword co-occurrence network of all questions within the class;

[0025] Identifying high-frequency strong-association semantic triples based on the co-occurrence network;

[0026] Inputting the triples into a large language model to generate natural language description labels consistent with human cognitive habits.

[0027] Further, the structured statistical report is output in a machine-parsable and human-readable dual format, including:

[0028] Stacked bar chart of the frequency proportion of each question category;

[0029] Priority-sorted category-frequency mapping table;

[0030] Key decision recommendation abstract containing description labels.

[0031] Further, the priority-sorted decision logic is:

[0032] High priority is given to high-frequency categories;

[0033] Weight enhancement strategy is performed on categories involving preset key semantics;

[0034] Output the final ranking list with weight adjustment coefficient.

[0035] The application also provides a question classification device based on a large language model, comprising:

[0036] A data input and vectorization module is used to receive a multi-source heterogeneous user question text dataset, and automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness ability by using the semantic encoding layer of a pre-trained large language model;

[0037] A semantic clustering and classification module is used to perform multi-dimensional semantic aggregation of user questions based on the vector space similarity measurement results between the dynamic high-dimensional vector representations, and form a plurality of question categories with semantic consistency by using an unsupervised clustering algorithm;

[0038] The semantic label intelligent generation module is configured to automatically extract common semantic features of user questions in each semantic consistency category through a text induction capability of a large language model, and generate a summary description label reflecting a core intention of the category.

[0039] The frequency statistics and priority sorting module is configured to calculate a frequency distribution of each semantic consistency category in real time, and automatically sort the categories according to a preset decision logic according to the frequency.

[0040] The result generation and output module is configured to fuse the summary description label and the priority sorting result, and generate a structured statistical report for decision analysis.

[0041] The application further provides a computer device including a memory and a processor, the memory stores a computer program, and the processor implements the question classification method based on a large language model as described above when executing the computer program.

[0042] The application further provides a storage medium storing a computer program, and the computer program can implement the question classification method based on a large language model as described above when executed by a processor.

[0043] The application has the advantages that: the dynamic vectorization and unsupervised clustering algorithm based on a large language model eliminate the need for manual intervention and shorten the time consumption for processing millions of questions; the vector space similarity measurement captures synonymous / multiple semantic associations, and the classification accuracy is significantly improved compared with traditional keyword matching; the summary description label and the weight priority sorting are fused, the statistical report is executable, and the efficiency of enterprise resource allocation is effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0044] The specific structure of the application will be described in detail below with reference to the drawings.

[0045] Figure 1 The question classification method based on a large language model of the embodiment of the application is a flowchart.

[0046] Figure 2 The vector space similarity measurement result acquisition flowchart of the embodiment of the application is a flowchart.

[0047] Figure 3 The unsupervised clustering algorithm execution flowchart of the embodiment of the application is a flowchart.

[0048] Figure 4 The common semantic feature extraction flowchart of the embodiment of the application is a flowchart.

[0049] Figure 5 The priority sorting flowchart of the embodiment of the application is a flowchart.

[0050] Figure 6 a large language model-based question classification device block diagram of an embodiment of the present application;

[0051] Figure 7 a schematic block diagram of a computer device of an embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] It should be understood that, when used in the specification and the appended claims, the terms “comprise” and “include” indicate the presence of described features, integers, steps, operations, elements, and / or components, but do not exclude one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0054] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, unless otherwise clearly indicated by the context, the singular forms “a”, “an” and “the” are intended to include the plural forms.

[0055] It should be further understood that the term “and / or” used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0056] Embodiments of the present application are: a large language model-based question classification method, comprising the steps of:

[0057] S1, receiving a multi-source heterogeneous user question text dataset, using a semantic encoding layer of a pre-trained large language model to automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness ability;

[0058] In this embodiment, the dynamic high-dimensional vector generation process is implemented based on a pre-trained language model with a Transformer architecture. Specifically, the semantic encoding layer first breaks down the user's question text into sub-word units, which are then processed deeply through a multi-layer neural network. The self-attention mechanism dynamically analyzes the semantic relationships between words. For example, for the question "payment card is unresponsive at checkout," the model can accurately identify "card" as the verb meaning of "stuttering" rather than the noun meaning, based on the context. In the vector generation stage, dynamic quantization technology is used to compress and optimize high-precision floating-point vectors. This technical solution has successfully improved the processing efficiency of payment-related questions in real-world testing on an e-commerce platform, effectively solving the deficiency of traditional bag-of-words models in capturing contextual semantics.

[0059] S2. Based on the vector space similarity measurement results between the dynamic high-dimensional vector representations, an unsupervised clustering algorithm is used to perform multi-dimensional semantic aggregation on user questions to form multiple question categories with semantic consistency.

[0060] Among them, such as Figure 2 As shown, the vector space similarity measurement results are obtained through a multi-distance calculation strategy, including:

[0061] S21. Calculate the cosine similarity between vectors to measure semantic orientation consistency;

[0062] S22. Calculate the Euclidean distance between vectors to measure the density of semantic space;

[0063] S23. Combine the directional consistency and spatial density to perform dynamic weighted fusion and generate a comprehensive similarity score.

[0064] Among them, such as Figure 3 As shown, the unsupervised clustering algorithm is a dual-driven clustering model based on density and centroid, and its execution process includes:

[0065] S24. Identify core semantically dense regions using density clustering algorithms;

[0066] S25. Based on the density distribution of the core region, initialize the seed points for the centroid clustering algorithm;

[0067] S26. Iteratively optimize the cluster boundary until the semantic convergence condition is met.

[0068] In this embodiment, the execution flow of the dual-drive clustering model is operated in three stages. First, the vector space is scanned by the density clustering algorithm to identify core semantic clusters with a fixed neighborhood radius and minimum density threshold, such as finding 32 dense distributions of "refund application" class problems around the vector coordinates (-0.2, 0.7). Then these density core points are used as the initial seed points of the centroid clustering algorithm, and intelligent initialization is achieved by automatically calculating the optimal clustering number. In the iterative optimization stage, the system performs multiple rounds of vector assignment and centroid position adjustment, and terminates the calculation when the contour coefficient of the class cluster structure is detected to reach the preset threshold. In practical applications, double convergence conditions are set: both the maximum number of iterations is limited to 50 times, and the distance of centroid movement is required to be less than 0.01 units for three consecutive rounds.

[0069] S3. For each semantic consistency category, automatically extract the common semantic features of the user questions in the category through the text induction capability of the large language model, and generate a summary description label reflecting the core intent of the category;

[0070] As shown in Figure 4 The operation of automatically extracting the common semantic features of the user questions in the category specifically includes:

[0071] S31. Extract the keyword co-occurrence network of all questions in the category;

[0072] S32. Identify high-frequency strongly associated semantic triples based on the co-occurrence network;

[0073] S33. Input the triples into the large language model to generate a natural language description label that conforms to human cognitive habits.

[0074] In this embodiment, the implementation process of the semantic label generation includes three key technical links. First, build a keyword co-occurrence network, extract high-frequency words and their association relationships in the category through statistical analysis methods, such as identifying the strong co-occurrence pattern of core words such as "refund", "time limit", "handling fee" in the order cancellation class questions. Then, perform semantic triple mining, use natural language processing technology to analyze the question syntax structure, and extract triple expressions such as "(refund, status, not arrived)" that reflect the essential relationship. Finally, through carefully designed text generation instructions, guide the large language model to convert triples into refined natural language descriptions. The specific instructions require the model to summarize the essence of the question based on semantic relationship frequency data using no more than 8 characters of daily language. In the SaaS product analysis scene, the labels generated by this method such as "account authentication exception" have a consistency of 94.7% with expert annotations, significantly reducing the workload of manual induction.

[0075] S4. Real-time calculation of the frequency distribution of each semantic consistency category, and automatic priority sorting of the categories according to the preset decision logic;

[0076] wherein, as shown in Figure 5 the priority ranking decision logic is:

[0077] S41, assigning high priority to high-frequency categories;

[0078] S42, performing a weight enhancement strategy on categories involving preset key semantics;

[0079] S43, outputting a final ranking list with weight adjustment coefficients.

[0080] In this embodiment, the priority ranking system operates with a hierarchical decision logic. The basic layer calculates the original frequency of each type of problem, for example, payment failure accounts for 18.3% of the total. The enhancement layer loads a preset key semantic dictionary, including important keywords such as security risk and financial loss. When the preset key semantics are detected within the category, the weight promotion mechanism is automatically triggered. The specific promotion amplitude is dynamically calculated according to the proportion of key semantic hits, for example, the weight coefficient of the account theft problem containing 3 security keywords is increased to 1.36. The final ranking score is obtained by multiplying the original frequency and the weight coefficient, forming a priority list with risk markers. After applying this scheme to a cross-border e-commerce platform, the response speed of high-risk problems can be effectively shortened, and customer satisfaction indicators have significantly improved, effectively solving the defect that traditional methods cannot distinguish the severity of problems.

[0081] S5, fusion of the summary description label and the priority ranking result, generating a structured statistical report for decision analysis.

[0082] The structured statistical report is output in a machine-parsable and human-readable dual format, including:

[0083] Stacked chart of frequency proportion of each problem category;

[0084] Category-frequency mapping table ranked by priority;

[0085] Key decision suggestion abstract containing description labels.

[0086] In this embodiment, a dual-track output scheme is adopted for generating the structured statistical report. The machine-readable part adopts a standardized data structure, including structured fields such as problem categories, frequency values, priority indexes, etc., and supports direct parsing and invocation by an automated system. The human-readable part is presented through visualization techniques, including a dynamically updated stacked bar chart of frequency percentages, a priority ranking table with color markings, and automatically generated key decision-making suggestions. Among them, the decision-making suggestion module comprehensively analyzes the top three priority problems and gives specific optimization direction guidance, such as preferentially fixing the abnormal payment channel. Enterprise application verification shows that this scheme can shorten the time for decision-makers to study the report, improve the efficiency of formulating optimization schemes, and completely change the current situation of insufficient operability of traditional statistical reports.

[0087] In a specific embodiment, before the step of receiving the multi-source heterogeneous user problem text data set and automatically converting each user problem text into a dynamic high-dimensional vector representation with context awareness using the semantic encoding layer of the pre-trained large language model, it includes:

[0088] Filter the noise of the user problem text through a rule engine and a regular expression matching mechanism, and remove URLs, special symbols, and non-text characters;

[0089] Apply a text normalization tool based on a dictionary to perform at least one semantic unification process such as spelling correction, synonym replacement, and abbreviation expansion.

[0090] In this embodiment, the data preprocessing system realizes text purification through the dual-level linkage of a rule engine and a semantic gateway. In the noise filtering stage, the rule engine loads a regular expression group to achieve precise cleaning: the URL cleaner uses a specific pattern to identify and strip network links, the special symbol stripper intelligently retains functional punctuation based on the non-text character blacklist, such as the question mark in "Password error?", and the whitespace compressor solves the common typesetting chaos problem in mobile input, compressing consecutive line breaks into a single space.实测在电商日志处理中,降低无效字符残留率。文本规范化阶段部署动态语义网关:拼写纠错模块融合音形码混合算法,如"支fu失败"校正为"支付失败":先计算拼音相似度,再检测形近字混淆,最终结合上下文选择最优解,同义词熔合器加载领域定制词库,如金融场景将"登陸"统一为"登录",缩写扩展器通过预置术语表解决歧义,如"APP支付ID错误"扩展为"应用程序支付身份凭证错误"。

[0091] As Figure 6 shown, the embodiment of the present invention also provides a problem classification device based on a large language model, including:

[0092] It should be noted that there seems to be an incomplete sentence in the original text of item (实测在电商日志处理中,降低无效字符残留率.). This may need to be further confirmed and corrected in the original text for a more accurate translation.The data input and vectorization module 10 is configured to receive a multi-source heterogeneous user question text dataset, and automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness by using a semantic encoding layer of a pre-trained large language model;

[0093] The semantic clustering and classification module 20 is configured to perform multi-dimensional semantic aggregation of user questions based on vector space similarity measurement results between the dynamic high-dimensional vector representations, and form a plurality of question categories with semantic consistency by using an unsupervised clustering algorithm;

[0094] The semantic label intelligent generation module 30 is configured to automatically extract common semantic features of user questions in each semantic consistency category by using a text induction capability of a large language model, and generate a summary description label reflecting a core intent of the category;

[0095] The frequency statistics and priority sorting module 40 is configured to calculate a frequency distribution of each semantic consistency category in real time, and automatically sort the categories according to the frequency according to a preset decision logic;

[0096] The result generation and output module 50 is configured to fuse the summary description label and the priority sorting result, and generate a structured statistical report for decision analysis.

[0097] It should be noted that the specific implementation process of the above-described large language model-based question classification device can be clearly understood by those skilled in the art, and can refer to the corresponding description in the foregoing method embodiments. For the convenience and brevity of description, it will not be repeated here.

[0098] The large language model-based question classification device described above can be implemented in the form of a computer program, which can run on a computer device as shown in Figure 7 .

[0099] Please refer to Figure 7 , Figure 7 is a schematic block diagram of a computer device provided by an embodiment of the present application. The computer device 500 can be a terminal or a server, wherein the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, a personal digital assistant, a wearable device, and an electronic device with a communication function. The server can be a stand-alone server or a server cluster composed of multiple servers.

[0100] Please refer to Figure 7 , the computer device 500 includes a processor 502, a memory, and a network interface 505 connected through a system bus 501, wherein the memory can include a non-volatile storage medium 503 and an internal memory 504.

[0101] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, can cause the processor 502 to perform a large language model-based question classification method.

[0102] The processor 502 is configured to provide computing and control capabilities to support the operation of the entire computer device 500.

[0103] The non-volatile storage medium 503 provides an environment for the computer program 5032 stored therein to run, and the computer program 5032, when executed by the processor 502, can cause the processor 502 to perform a large language model-based question classification method.

[0104] The network interface 505 is configured to perform network communication with other devices. Those skilled in the art can understand that the network interface 505 can be configured to perform wired or wireless communication. Figure 7 The structure shown in FIG. 5 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device 500 to which the scheme of the present application is applied. Specifically, the computer device 500 can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0105] The processor 502 is configured to run the computer program 5032 stored in the memory to implement the large language model-based question classification method as described above.

[0106] It should be understood that, in the embodiments of the present application, the processor 502 can be a central processing unit (CPU), and the processor 502 can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0107] Those of ordinary skill in the art can understand that all or part of the processes in the above-described embodiments can be completed by a computer program instructing related hardware. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the above-described embodiments.

[0108] Therefore, the application further provides a storage medium. The storage medium can be a computer-readable storage medium. The storage medium stores a computer program, where the computer program includes program instructions. The program instructions are executed by a processor to enable the processor to perform the large language model-based question classification method described above.

[0109] The storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk, and various computer-readable storage media that can store program codes.

[0110] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, computer software, or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in the above description in a general manner. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0111] In several embodiments provided by the application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of each unit is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed.

[0112] The steps in the method embodiments of the application can be adjusted, combined, and reduced in sequence according to actual needs. The units in the device embodiments of the application can be combined, divided, and reduced according to actual needs. In addition, each functional unit in each embodiment of the application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0113] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a storage medium. Based on such understanding, the technical solutions of the application essentially or say the parts that make contributions to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for enabling a computer device (which can be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the application.

[0114] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A problem classification method based on a large language model, characterized in that, include: It receives a multi-source heterogeneous user question text dataset and uses the semantic encoding layer of a pre-trained large language model to automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness. Based on the vector space similarity measurement results between the dynamic high-dimensional vector representations, an unsupervised clustering algorithm is used to perform multi-dimensional semantic aggregation of user questions, forming multiple question categories with semantic consistency. For each semantic consistency category, the text induction capability of the large language model is used to automatically extract the common semantic features of user questions within the category and generate a general description label that reflects the core intent of the category. The frequency distribution of each semantic consistency category is calculated in real time, and the categories are automatically prioritized according to frequency based on preset decision logic. By combining the general description labels with the priority ranking results, a structured statistical report for decision analysis is generated.

2. The problem classification method based on a large language model according to claim 1, characterized in that: Prior to the step of receiving a multi-source heterogeneous user question text dataset and automatically converting each user question text into a context-aware, dynamic, high-dimensional vector representation using the semantic encoding layer of a pre-trained large language model, the following steps are included: By using a rule engine and regular expression matching mechanism, noise is filtered out from user question text, removing URLs, special symbols, and non-text characters; Apply a dictionary-based text normalization tool to perform at least one semantic unification process, including spell correction, synonym replacement, and abbreviation expansion.

3. The problem classification method based on a large language model according to claim 1, characterized in that: The vector space similarity measurement results are obtained through a multi-distance calculation strategy, including: Calculate the cosine similarity between vectors to measure semantic orientation consistency; Calculate the Euclidean distance between vectors to measure the density of semantic space; By combining the directional consistency and spatial density, a dynamic weighted fusion is performed to generate a comprehensive similarity score.

4. The problem classification method based on a large language model according to claim 1, characterized in that: The unsupervised clustering algorithm is a dual-driven clustering model based on density and centroid, and its execution process includes: Identify core semantically dense regions using density clustering algorithms; Based on the density distribution of the core region, initialize the seed points for the centroid clustering algorithm; Iteratively optimize the cluster boundary until the semantic convergence condition is met.

5. The problem classification method based on a large language model according to claim 1, characterized in that: The operation of automatically extracting common semantic features of user questions within a category specifically includes: Extract the keyword co-occurrence network for all questions within the class; Based on the co-occurrence network, high-frequency strongly correlated semantic triples are identified; Input the triples into a large language model to generate natural language description labels that conform to human cognitive habits.

6. The problem classification method based on a large language model according to claim 1, characterized in that: The structured statistical report is output in both machine-parseable and human-readable formats, including: A stacked chart showing the frequency distribution of each problem category; Category-frequency mapping table sorted by priority; A summary of key decision recommendations, including descriptive tags.

7. The problem classification method based on a large language model according to claim 1, characterized in that: The decision logic for prioritization is as follows: High-frequency categories are assigned higher priority; Apply weight enhancement strategies to categories involving preset key semantics; Output the final sorted list with weight adjustment factors.

8. A problem classification device based on a large language model, characterized in that, include: The data input and vectorization module is used to receive multi-source heterogeneous user question text datasets and automatically convert each user question text into a dynamic high-dimensional vector representation with context awareness by utilizing the semantic encoding layer of a pre-trained large language model. The semantic clustering and classification module is used to perform multi-dimensional semantic aggregation of user questions based on the vector space similarity measurement results between the dynamic high-dimensional vector representations, and to form multiple question categories with semantic consistency. The semantic tag intelligent generation module is used to automatically extract the common semantic features of user questions within each semantic consistency category by leveraging the text induction capabilities of a large language model, and generate a general descriptive tag that reflects the core intent of the category. The frequency statistics and priority sorting module is used to calculate the frequency distribution of each semantic consistency category in real time, and automatically prioritize the categories by frequency according to the preset decision logic. The results generation and output module is used to integrate the general description labels and priority ranking results to generate a structured statistical report for decision analysis.

9. A computer device, characterized in that: The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the problem classification method based on a large language model as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, can implement the problem classification method based on a large language model as described in any one of claims 1 to 7.

Images