User question clustering method and device based on large model cooperation, equipment and storage medium

Abstract

The application discloses a user question clustering method and device based on large model cooperation, equipment and storage medium, relates to the technical field of language processing, and comprises the following steps: obtaining user question information to be clustered and clustering parameter information; determining corresponding dimension reduction semantic vectors based on the user question information to be clustered; clustering the dimension reduction semantic vectors according to different clustering numbers based on the clustering parameter information, and determining clustering cluster information; inputting a predefined semantic evaluation model based on the clustering cluster information and the clustering parameter information to classify corresponding target clustering clusters, and determining target clustering cluster information; and generating corresponding cluster name labels according to a predefined verb and object naming specification based on the target clustering cluster information. The application performs dimension reduction, clustering and semantic evaluation on the user question information to be clustered and the clustering parameter information according to different clustering numbers, thereby standardizing the naming of the screened clustering clusters, significantly improving the processing efficiency, and directly adapting to various customer service applications.

Description

Technical Field

[0001] This application relates to the field of language processing technology, and in particular to a method, apparatus, device and storage medium for user question clustering based on large model collaboration. Background Technology

[0002] In intelligent customer service applications, there is a massive amount of user question data. To efficiently extract core user intent from this data for intelligent question answering, it is necessary to automatically cluster user questions. However, this must go beyond the surface similarity of the text and truly understand and aggregate semantic meaning. Figure 1 A question posed to [the recipient].

[0003] Currently, existing approaches rely on SimHash and TF-IDF combined with hierarchical clustering based on word frequency statistics, or use LLM to generate labels for text and then fuse the label features with traditional text vectors for clustering. Alternatively, user questions can be directly input into the LLM in batches, with the model directly classifying and naming them, and then merging the results across batches. However, these existing approaches heavily depend on the superficial similarity of words or vector distances, exhibiting poor robustness to synonym substitutions and sentence structure variations. They struggle to distinguish between semantically similar but different intentions, leading to inconsistencies in meaning. Figure 1 The problem is that questions with the same intent are separated, or questions with different intentions are incorrectly merged. Furthermore, the over-reliance on prompts leads to unstable label quality, and feature fusion lacks standards, failing to fundamentally escape the limitations of traditional distance metrics. Additionally, it requires batch-based LLM calls on massive datasets, resulting in extremely high costs. Moreover, the clustering results are unstable and prone to duplication or omissions due to the influence of model context length and batch data distribution. Therefore, how to achieve efficient and accurate user question clustering based on large-scale model collaboration has become an urgent problem to be solved.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a user question clustering method, apparatus, device and storage medium based on large model collaboration, aiming to solve the technical problem of how to perform efficient and accurate user question clustering based on large model collaboration.

[0006] To achieve the above objectives, this application proposes a user-interview clustering method based on large-model collaboration, the method comprising:

[0007] Obtain user information and clustering parameter information for the users to be clustered; Determine the corresponding dimensionality-reduced semantic vector based on the user information to be clustered; Based on the clustering parameter information, the dimensionality-reduced semantic vector is clustered according to different numbers of clusters to determine the cluster information; Based on the cluster information and the clustering parameter information, a predefined semantic evaluation model is input to classify the corresponding target clusters and determine the target cluster information; Based on the target cluster information, corresponding cluster name tags are generated according to a predefined verb and object naming convention.

[0008] In one embodiment, the step of determining the corresponding dimensionality-reduced semantic vector based on the user information to be clustered includes: Based on the set of users to be clustered, text preprocessing is performed to determine preprocessing information. The text preprocessing includes special character processing, character writing processing, and space processing. Based on the preprocessed information, a predefined vector generation model is used to generate a corresponding high-dimensional semantic vector. The high-dimensional semantic vector is normalized to determine the normalized vector; Principal component analysis is used to reduce the dimensionality of the normalized vector to obtain a dimensionality-reduced semantic vector.

[0009] In one embodiment, the step of classifying the corresponding target clusters and determining the target cluster information by inputting the cluster information and the clustering parameter information into a predefined semantic evaluation model includes: A clustering algorithm is used to divide the cluster information according to the number of clusters in the clustering parameter information to determine the initial cluster information; Based on the initial cluster information and the clustering parameter information, a predefined semantic evaluation model is input to classify the corresponding target clusters, thereby obtaining the target cluster information.

[0010] In one embodiment, the step of classifying the corresponding target clusters based on the initial cluster information and the clustering parameter information by inputting a predefined semantic evaluation model to obtain target cluster information includes: Based on the initial cluster information, a predefined semantic evaluation model is input to evaluate stable clusters and clusters to be clustered, thereby determining the target stable clusters and target clusters to be clustered. Based on the target stable cluster and the target cluster to be clustered, the number of clusters in the clustering parameter information is adjusted to determine the target clustering parameter information; Based on the target clustering parameter information, the target clusters to be clustered are classified to obtain the target stable iterative clusters; The target cluster information is obtained based on the target stable cluster and the target stable iterative cluster.

[0011] In one embodiment, the step of evaluating stable clusters and clusters to be clustered based on the initial cluster information input into a predefined semantic evaluation model, and determining the target stable cluster and the target cluster to be clustered, includes: Get prompt word template; Convex sampling is used to select user questions that satisfy predefined rules from the initial cluster information to determine the evaluation input; Based on the evaluation input and the predefined semantic evaluation model of the prompt word template input, stable clusters and clusters to be clustered are evaluated to obtain target stable clusters and target clusters to be clustered.

[0012] In one embodiment, the step of classifying the target clusters to be clustered based on the target clustering parameter information to obtain target stable iterative clusters includes: Obtain the unclassified proportion threshold and the target number of iterations; The proportion of unassigned users is determined based on the target stable cluster and the target cluster to be clustered. Based on the proportion of unassigned users, the unclassified proportion threshold, and the target iteration number, the target clusters to be clustered are classified according to the number of clusters in the target clustering parameter information to obtain target cluster information.

[0013] In one embodiment, the step of generating corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention includes: Based on the target cluster information, multiple user questions are sampled from the corresponding clusters to determine a list of user question representatives; Based on the user representative list input, a predefined large language model is used to name the target clusters according to the predefined verb and object naming conventions, resulting in cluster name labels.

[0014] Furthermore, to achieve the above objectives, this application also proposes a user-interview clustering device based on large-model collaboration, the user-interview clustering device based on large-model collaboration comprising: The acquisition module is used to acquire user information to be clustered and clustering parameter information; The processing module is used to determine the corresponding dimensionality-reduced semantic vector based on the user information to be clustered; The processing module is also used to cluster the dimensionality-reduced semantic vector according to different numbers of clusters based on the clustering parameter information, and determine the cluster information; The execution module is used to classify the corresponding target clusters based on the cluster information and the clustering parameter information input into a predefined semantic evaluation model, and determine the target cluster information; The execution module is also used to generate corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention.

[0015] Furthermore, to achieve the above objectives, this application also proposes a user-interview clustering device based on large model collaboration, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the user-interview clustering method based on large model collaboration as described above.

[0016] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the user-defined clustering method based on large model collaboration as described above.

[0017] One or more technical solutions proposed in this application have at least the following technical effects: This embodiment proposes a user question clustering method based on large-scale model collaboration. The method involves acquiring user question information to be clustered and clustering parameters; determining corresponding dimensionality-reduced semantic vectors based on the user question information; clustering the dimensionality-reduced semantic vectors according to different numbers of clusters based on the clustering parameters to determine cluster information; inputting the cluster information and clustering parameters into a predefined semantic evaluation model to classify the corresponding target clusters to determine target cluster information; and generating corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention. This application improves processing efficiency by reducing the user question information to low-dimensional semantic vectors, performing multiple rounds of clustering on the low-dimensional semantic vectors based on clustering parameters, and classifying the corresponding target clusters using a semantic evaluation model. It retains semantically consistent clusters and automatically generates corresponding name tags using a predefined verb and object naming convention, significantly reducing computational costs, improving the efficiency of large-scale data processing, ensuring high consistency of clustering results at the intent level, and directly adapting to downstream applications such as customer service intent classification model training and intelligent question answering. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the user clustering method based on large model collaboration in this application; Figure 2This is a flowchart illustrating an embodiment of the user clustering method based on large model collaboration in this application. Figure 3 This is a flowchart illustrating Embodiment 2 of the user clustering method based on large model collaboration in this application. Figure 4 This is a schematic diagram of the module structure of the user clustering device based on large model collaboration in an embodiment of this application; Figure 5 This is a schematic diagram of the device structure of the hardware operating environment involved in the user clustering method based on large model collaboration in the embodiments of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] The main solution of this application embodiment is as follows: obtaining user query information to be clustered and clustering parameter information; determining the corresponding dimensionality-reduced semantic vector based on the user query information to be clustered; clustering the dimensionality-reduced semantic vector according to different numbers of clusters based on the clustering parameter information to determine cluster information; inputting the cluster information and the clustering parameter information into a predefined semantic evaluation model to classify the corresponding target clusters to determine target cluster information; and generating corresponding cluster name tags according to a predefined verb and object naming convention based on the target cluster information.

[0025] In this embodiment, for ease of description, the following description will focus on identifying user clustering devices based on large model collaboration.

[0026] Because existing technologies heavily rely on superficial similarity based on lexical or vector distance, they are poorly robust to synonym substitution and sentence structure changes, making it difficult to distinguish between semantically similar but differently intentional issues, leading to... Figure 1 The problem is that questions with the same intent are separated, or questions with different intents are incorrectly merged. Furthermore, the problem relies too much on prompt words, resulting in unstable label quality. The feature fusion lacks standards and fails to fundamentally break free from the limitations of traditional distance metric logic. At the same time, it requires batch calls to LLM on massive amounts of data, which is extremely costly. Moreover, the clustering results are unstable and prone to duplication or omissions due to the influence of model context length and batch data distribution.

[0027] This application provides a solution, such as Figure 1 As shown, Figure 1 This is a flowchart illustrating the user question clustering method based on large-model collaboration in this application. Based on the user question information and clustering parameters, the user question is transformed into a 1024-dimensional high-dimensional vector using the BGE Embedding model. This vector is then reduced to a 128-dimensional low-dimensional semantic vector using methods such as PCA and cached. Based on the number of clusters k set in the clustering parameters, the reduced-dimensional semantic vector is input into an improved K-Means algorithm for initial clustering, yielding initial cluster partitioning results. The initial cluster information and clustering parameters are then input into a semantic consistency evaluator built on LLM, which performs a binary classification of each cluster as good or bad. Semantically consistent target clusters are retained, while bad clusters or samples that do not meet the threshold are placed in a pool for re-clustering. The value of k is dynamically adjusted to enter the next iteration. Iteration continues until the proportion of unassigned samples falls below a set threshold or the maximum number of iterations is reached. According to a predefined verb-object naming convention, an interpretable cluster name label is automatically generated for each confirmed good cluster using LLM.

[0028] As can be seen from the above embodiments, this application improves processing efficiency by reducing the user information to be clustered into low-dimensional semantic vectors, performing multiple rounds of clustering on the low-dimensional semantic vectors based on clustering parameter information, and classifying the corresponding target clusters using a semantic evaluation model, retaining semantically consistent clusters, and automatically generating corresponding name tags according to preset naming conventions for verbs and objects, which significantly reduces computational costs, improves the efficiency of large-scale data processing, ensures high consistency of clustering results at the intent level, and directly adapts to downstream applications such as customer service intent classification model training and intelligent question answering.

[0029] Based on this, embodiments of this application provide a user-interview clustering method based on large-model collaboration, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the user clustering method based on large model collaboration in this application.

[0030] In this embodiment, the user clustering method based on large model collaboration includes steps S10~S50: Step S10: Obtain user information to be clustered and clustering parameter information; It should be noted that the user question information to be clustered is the original set of user question texts that need to be processed for intent clustering. That is, short text corpus accumulated in interactive scenarios such as intelligent customer service, online consultation, and user feedback. For example, in the scenario of bank customer service, users ask a variety of natural language questions such as how to change the reserved mobile phone number, how to check the transfer limit, and how much the annual fee of a credit card is. These questions differ in form, expression, and sentence structure, but belong to the same type of user intent, such as changing the mobile phone number or checking the limit. The clustering parameter information is the configuration parameter that controls the clustering process. It may include the initial number of clusters, the threshold for the proportion of unassigned samples, the maximum number of iterations, the proportion of variance retained by vector dimensionality reduction, and the cache capacity threshold. These parameters can be preset or dynamically adjusted according to the actual scenario and data scale to balance clustering accuracy and computational efficiency.

[0031] In a specific embodiment, as an optional implementation, the user questions to be clustered can originate from real-time collected online dialogue streams, or from a static corpus extracted from historical logs and deduplicated. In actual systems, these questions typically undergo preliminary cleaning, such as removing irrelevant symbols and standardizing the encoding format, and are input into the clustering process in the form of a list or database table. The clustering parameters can be set to system default values, for example, initial cluster size k=200, unassigned proportion threshold ε=5%, maximum number of iterations 10, dimensionality reduction and variance retention ratio ≥95%, and cache capacity threshold of 1 million records. Users or system administrators can also adjust these parameters according to the data scale and business needs. For example, when processing more than 100,000 user questions, the initial k value can be appropriately increased to enhance the cluster subdivision capability. If real-time requirements are high, the maximum number of iterations can be appropriately reduced to accelerate cluster convergence. At the same time, the clustering parameter information can also be adaptively generated based on the data scale. For example, the initial k value can be dynamically set according to a certain proportion of the total number of questions to be clustered, such as setting one initial cluster for every 500 questions. The unallocated proportion threshold ε can also be adjusted according to the business requirements for clustering integrity, such as setting it to 1% in high recall scenarios.

[0032] In a specific embodiment, as another optional implementation, the system can also support a parameter tuning interface, which allows for the automatic search of the optimal parameter combination through a small-scale validation set or historical clustering effect feedback. For example, the k value in the next iteration can be dynamically adjusted based on the proportion of good clusters in the previous round of clustering results, thereby achieving adaptive optimization of parameters.

[0033] In a specific embodiment, if the user information to be clustered is empty or too few, such as less than the preset minimum number of clustering samples, the system can output a corresponding prompt and skip the clustering process, directly return an empty result or execute other processing strategies, such as directly calling LLM for intent classification.

[0034] Step S20: Determine the corresponding dimensionality-reduced semantic vector based on the user information to be clustered; It should be noted that the dimensionality-reduced semantic vector is a low-dimensional numerical vector representation obtained by performing a series of semantic encoding and compression processes on the original user question text. This vector retains the core semantic information of the original text while significantly reducing the dimension, for example, from 1024 dimensions to 128 dimensions, so as to greatly improve the efficiency of clustering operations and reduce storage overhead.

[0035] In a specific embodiment, as an optional implementation, an Embedding vector compression buffer module can be used to achieve efficient conversion and reuse from the original user question text to low-dimensional semantic vectors. For example, based on the user question set to be clustered, text preprocessing is performed to determine preprocessing information. The text preprocessing includes special character processing, character spelling processing, and space processing. That is, this module can clean and standardize the original user questions, including removing irrelevant special characters (such as punctuation and emoticons), unifying English letters to lowercase, merging consecutive redundant spaces, etc., to ensure that the input text format is consistent. Based on the preprocessing information, a predefined vector generation model is input to generate the corresponding high-dimensional semantic vector. That is, the preprocessed text is input into a pre-trained text embedding model to generate high-dimensional semantic vectors. For example, a model specifically optimized for short Chinese texts (such as BGE-Embedding-zh) can be used to generate high-dimensional semantic vectors for each question. A 1024-dimensional vector is used to capture the semantic and contextual information of a sentence. This high-dimensional semantic vector is normalized to determine the normalized vector. This normalization process involves dividing the generated high-dimensional vector by its magnitude to convert it into a unit vector. All vectors are then mapped onto a unit hypersphere to eliminate the influence of length differences and ensure the consistency of distance metrics. Principal component analysis (PCA) is used to reduce the dimensionality of the normalized vector, resulting in a dimensionality-reduced semantic vector. This allows for the compression of the normalized high-dimensional vector using dimensionality reduction algorithms, significantly reducing the dimensionality while retaining most of the semantic information. The preferred dimensionality reduction algorithm is Principal Component Analysis (PCA). By calculating the principal components of the data, the high-dimensional vector (e.g., 1024-dimensional) is projected onto principal components that retain a preset variance percentage (e.g., ≥95%), resulting in a low-dimensional semantic vector (e.g., 128-dimensional). The dimensionality-reduced vector retains the original semantic discriminative power while greatly improving storage and computational efficiency.

[0036] In a specific embodiment, as another optional implementation method, in order to improve the efficiency of processing massive amounts of data, the system can design a caching mechanism, that is, the preprocessed text is hashed (such as MD5) to generate a unique key, and stored as a key-value pair with the corresponding dimensionality-reduced semantic vector. It supports memory caching (such as Redis, for high-frequency access) and disk persistent storage. At the same time, strategies such as LRU can be used to manage the cache capacity, avoid repeated calculations, and realize the rapid reuse of vectors.

[0037] In specific implementations, the technology selection can be adjusted according to actual resource and performance requirements. For example, for the generation of high-dimensional semantic vectors, other pre-trained models such as Sentence-BERT and SimCSE can be used. For dimensionality reduction algorithms, in addition to principal component analysis (PCA), nonlinear dimensionality reduction methods such as t-SNE and UMAP can also be used, which are suitable for scenarios with particularly complex semantic structures, but the computational cost needs to be weighed. The caching mechanism can be flexibly selected according to the deployment environment. When memory resources are scarce or the amount of data is extremely large, a hierarchical caching strategy can be adopted.

[0038] In a specific embodiment, the system may also have a built-in dimensionality reduction quality verification module. For example, by calculating the average cosine similarity between vectors before and after dimensionality reduction of a batch of samples, the system ensures that the value is not lower than a preset threshold (such as 0.92), thereby ensuring from a quantitative perspective that the core semantic information is not seriously lost during the compression process. If the question to be processed already has a corresponding dimensionality reduction semantic vector in the cache, the system can directly read and reuse it without recalculation.

[0039] In one feasible implementation, step S20 may include steps A11 to A14: Step A11: Perform text preprocessing based on the user question set to be clustered, and determine the preprocessing information. The text preprocessing includes special character processing, character writing processing, and space processing. It should be noted that the preprocessed information is the standardized text data obtained after cleaning and standardizing the original user questions. The purpose is to eliminate non-semantic noise and ensure vector consistency and accuracy.

[0040] Understandably, special character processing removes irrelevant symbols such as punctuation and emoticons, character writing processing unifies English letters to lowercase or uppercase to avoid semantic discrepancies caused by differences in case, and space processing merges multiple consecutive spaces into a single space and removes extra spaces at the beginning and end to ensure clean and consistent text formatting.

[0041] Step A12: Based on the preprocessed information, input a predefined vector generation model to generate the corresponding high-dimensional semantic vector; It should be noted that the high-dimensional semantic vector is a high-dimensional numerical vector that maps text into text through a pre-trained text embedding model. This vector can capture the semantic, contextual, and intent information of the sentence.

[0042] Understandably, vector generation models are pre-trained models based on the Transformer architecture, such as BGE and BERT. They gain the ability to understand language through training on a large amount of corpus and can convert text into vector representations with semantic discriminative power.

[0043] Step A13: Normalize the high-dimensional semantic vector to determine the normalized vector; It should be noted that the normalized vector is a unit vector obtained by dividing the high-dimensional semantic vector by its magnitude, so that it falls into the unit hypersphere space, thereby eliminating the influence of vector length on distance metric.

[0044] Understandably, normalization is a common vector standardization operation that ensures that the similarity calculation between different vectors depends only on direction rather than length, thereby improving the fairness and stability of distance comparison during clustering.

[0045] Step A14: Use principal component analysis to reduce the dimensionality of the normalized vector to obtain a dimensionality-reduced semantic vector.

[0046] It is understandable that principal component analysis is a linear dimensionality reduction method. By calculating the eigenvectors (principal components) of the data covariance matrix, the high-dimensional vector is projected into a low-dimensional subspace that retains most of the variance information. The dimensionality reduction semantic vector is to compress the high-dimensional vector into a low-dimensional representation while maintaining the original semantic information, thereby significantly improving storage efficiency and clustering calculation speed, while reducing noise interference caused by dimensional redundancy.

[0047] Step S30: Based on the clustering parameter information, the dimensionality-reduced semantic vector is clustered according to different numbers of clusters to determine the cluster information; It should be noted that the cluster information is a structured data set that represents the classification of user question intent categories, that is, detailed information on each cluster into which all user questions to be processed are divided under specific clustering parameters.

[0048] It is understood that each cluster in the clustering cluster information corresponds to a potential, semantically similar user intent category, which is used to carry and transmit the intermediate clustering results from the original text to the semantic tags. It can exist in the form of a data structure, such as a list, dictionary or database record, containing elements such as cluster identifier, member samples, vector features and state attributes.

[0049] In a specific embodiment, the KMeans clustering module can be used to implement iterative intent clustering based on low-dimensional embedding vectors. The low-dimensional embedding vectors output by the embedding vector compression buffer module are divided into several clusters according to the k value of the current iteration round. That is, variables are initialized according to clustering parameter information, including the initial number of clusters k (e.g., set to 200), the unclassified proportion threshold ε (e.g., 0.05), and the maximum number of iterations (e.g., 10). In each iteration, the improved K-Means algorithm is executed to divide the input dimensionality-reduced semantic vectors according to the current k value to obtain cluster information. After that, semantic consistency evaluation can be performed. Clusters with good evaluation results will be retained and output, and their corresponding vectors will not participate in subsequent calculations. Bad clusters are disassembled, and the vectors corresponding to all user questions contained therein are reassigned to the unassigned set. Subsequently, the system dynamically adjusts the k value of the next round according to the total number of unassigned vectors remaining in the current round based on the semantically driven adaptive search mechanism for the number of clusters. This is expressed as: k = min(total number of unassigned user questions / 5, the current k value), thereby gradually reducing the search range and improving clustering efficiency in the iteration. If the condition is not met, the loop will continue until the preset iteration termination condition is met, so that clustering is performed according to different cluster numbers to obtain cluster information. The iteration termination condition can be that the proportion of unassigned vectors to the total number is lower than the threshold ε or the maximum number of iterations has been reached.

[0050] Step S40: Based on the cluster information and the clustering parameter information, input a predefined semantic evaluation model to classify the corresponding target clusters and determine the target cluster information; It should be noted that the target cluster information is a set of high-quality user intent clusters after screening. Based on the cluster information, the deep semantic understanding capability of the semantic evaluation model is used to perform binary classification evaluation of each cluster as good clusters and bad clusters. After multiple rounds of evaluation, the complete information of the clusters that meet the semantic consistency standard is retained.

[0051] It is understood that the target cluster information can be updated with each iteration round. The initial round may be empty. As each round of evaluation retains good clusters and removes bad clusters, its content gradually becomes richer and more stable until the iteration termination condition is met, and complete target cluster information is obtained.

[0052] In a specific embodiment, as an optional implementation, the LLM semantic evaluator module can be used to quantitatively evaluate the clustering quality. For example, a clustering algorithm is used to divide the cluster information according to the number of clusters in the clustering parameter information to determine the initial cluster information. Based on the initial cluster information and the clustering parameter information, a predefined semantic evaluation model is input to classify the corresponding target clusters to obtain the target cluster information. That is, according to the initial number of clusters set in the clustering parameter information (e.g., k=200), an improved K-Means algorithm is used to divide the low-dimensional semantic vectors processed by module one, assigning each vector to the nearest cluster center to generate the initial cluster information. Subsequently, the LLM semantic evaluation model, that is, a semantic consistency binary classification model built based on open-source LLM, is used to accurately distinguish semantic clusters. Figure 1 To differentiate between statement groups and statement groups with mixed intents, a cluster sampling evaluation can be performed. For each cluster to be evaluated, the semantic vectors of all statements within it are first reduced to a 2D space using PCA. Then, a convex sampling algorithm is used to select 10 representative user questions that cover different expressions within the cluster, forming the evaluation input text. Semantic consistency binary classification is then performed. Through carefully designed prompt word engineering, the sampled question list is encapsulated into a binary classification task. For example, using a fixed prompt template: "You are a sentence classification assistant," determine whether the following user questions belong to the same intent, outputting only "yes" or "no," and submitting this to the semantic evaluation model for processing. The semantic evaluation model then evaluates the semantic consistency of the statements within the cluster. Figure 1 Consistently output the binary classification results of good or bad clusters. At this time, the clusters that are judged as good clusters will be retained as the target clusters, while the bad clusters will be disassembled and re-enter the iterative clustering process to obtain the target cluster information.

[0053] The semantic evaluation model Prompt example is as follows: You are a sentence classification assistant. Determine whether the following user questions belong to the same intent, and answer "yes" or "no". Question list: {A representative user from the cluster asked} Requirements: Output only "Yes" or "No", do not output any other irrelevant content.

[0054] result: In a specific embodiment, as another optional implementation, the semantic evaluation model is not limited to outputting only a binary classification of yes / no. For example, for scenarios where the intent boundary is ambiguous or confidence assessment is required, the semantic evaluation model can be designed to output a semantic consistency score between 0 and 1. In this case, the determination of the target cluster information can be based on a score threshold, such as a good cluster with a score greater than 0.8, allowing for flexible quality control. After all iterations are completed, all determined good clusters are compared pairwise. If two different clusters are found to have highly similar names or whose intents overlap after a second evaluation, they can be merged, and the target cluster information can be updated for further optimization.

[0055] Step S50: Generate corresponding cluster name tags based on the target cluster information according to the predefined verb and object naming convention.

[0056] It should be noted that the cluster name label is a concise and highly generalized text identifier automatically generated from each good cluster in the target cluster information that has passed semantic consistency verification. It is used to identify different user intent categories. More importantly, its standardized format allows it to be directly and seamlessly applied to downstream tasks, such as customer service intent classification model training, knowledge base construction, and intelligent question answering system routing, which greatly improves the operability and practical value of the clustering results.

[0057] Understandably, the verb and object naming convention is a standardized format rule for generating cluster names. It requires that the name generated for each user intent cluster consists of a core verb and a core object connected by a hyphen, in the form of <verb>-<object>. The verb summarizes the core operation or intent expressed by the user's question in that cluster, such as query, modify, process, consult, activate, cancel, etc. The object summarizes the core object or business entity operated on by the user's question in that cluster, such as mobile phone number, transfer limit, bank card balance, credit card annual fee, login password, etc.

[0058] In a specific embodiment, as an optional implementation, naming can be achieved using an LLM cluster namer module. For example, based on the target cluster information, multiple user questions are sampled from the corresponding clusters to determine a representative list of user questions. Based on the user question representative list, a predefined large language model is input to name the target clusters according to the predefined verb and object naming conventions, resulting in cluster name tags. That is, this module automatically generates interpretable cluster name tags for each confirmed good cluster in the target cluster information. Using the large language model, the tag synthesis strictly follows the verb-object naming convention, such as modify-phone number, query-transfer limit. Specifically, cluster sampling can be performed, systematically selecting 20 user questions from each target cluster to ensure coverage of the main expression variations under the intent, forming a representative list of user questions, and naming the clusters. The large language model is called through a pre-encapsulated naming prompt template. The large language model extracts unified actions and objects based on the list content and outputs concise tags that conform to the conventions.

[0059] The large language model is represented as follows: You are a sentence clustering and naming assistant. Based on the relevance and commonalities of the following list of user questions, generate a concise and short "verb-object" label.

[0060] Common tag examples: -Adjustment-Limit -Check-balance -Modify-Mobile Number -Consultation- regarding xx issue Question list: {Cluster sampling represents user questions} Requirements: Output only one line, do not output any other irrelevant content.

[0061] result: In a specific embodiment, as another optional implementation, the naming convention can be fine-tuned or expanded according to different business domains. For example, in some scenarios that require more granular intent, a three-part structure of verb-object-modifier can be adopted, such as query-transfer record-last month. Or, for consultation intents, consultation can be used as the verb prefix. To ensure naming quality, simple verification or post-processing steps can be introduced after generating tags. For example, check whether the generated tags conform to basic grammatical rules, or provide prompts for different clusters with extremely similar names, and have a more complex model perform secondary review and merging to further improve the accuracy and uniqueness of the output tags.

[0062] In one feasible implementation, step S50 may include steps B11-B12: Step B11: Based on the target cluster information, sample multiple user questions from the corresponding clusters to determine the user question representative list; It should be noted that the user question representative list is a finite set of questions that characterize the core intent of the cluster and the diversity of its expressions.

[0063] It is understandable that the user question representative list may contain multiple user questions selected from the cluster system. When sampling, it is necessary to cover different forms of expression under the intent, such as direct inquiry, simplified expression, paraphrasing, etc., and take into account the differences in sentence length, keyword distribution and semantic emphasis, so as to ensure that the list can not only represent the commonality of statements within the cluster, but also reflect the range of changes in their semantic expression.

[0064] Step B12: Based on the user representative list, input a predefined large language model and name the target clusters according to the predefined verb and object naming conventions to obtain cluster name labels.

[0065] Understandably, the large language model is an artificial intelligence model based on massive text pre-training and possessing deep semantic understanding and generation capabilities. By analyzing the correlation and commonalities of all questions in the list, it automatically extracts unified verbs and objects and generates cluster name labels in strict accordance with the standard format.

[0066] This embodiment proposes a user query clustering method based on large model collaboration, which involves: acquiring user query information to be clustered and clustering parameter information; determining the corresponding dimensionality-reduced semantic vector based on the user query information; clustering the dimensionality-reduced semantic vector according to different numbers of clusters based on the clustering parameter information to determine cluster information; inputting the cluster information and the clustering parameter information into a predefined semantic evaluation model to classify the corresponding target clusters to determine target cluster information; and generating corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention. This application addresses the technical challenge of efficient and accurate user question clustering based on large-scale model collaboration. Compared to existing technologies, it improves processing efficiency by reducing the user question information to be clustered into low-dimensional semantic vectors. Multiple rounds of clustering are performed on the low-dimensional semantic vectors based on clustering parameters, and a semantic evaluation model is used to classify the corresponding target clusters. Semantically consistent clusters are retained, and corresponding name tags are automatically generated using a predefined naming convention for verbs and objects. This significantly reduces computational costs, improves the efficiency of large-scale data processing, ensures high consistency of clustering results at the intent level, and directly adapts to downstream applications such as customer service intent classification model training and intelligent question answering.

[0067] Based on the first embodiment of this application, in the second embodiment of this application, the same or similar content as the first embodiment can be referred to the above description, and will not be repeated hereafter.

[0068] In this embodiment, refer to Figure 3 , Figure 3 This is a flowchart illustrating Embodiment 2 of the user clustering method based on large model collaboration in this application. Step S40 specifically includes steps S41 to S42: Step S41: Use a clustering algorithm to divide the cluster information according to the number of clusters in the clustering parameter information to determine the initial cluster information; It should be noted that the initial cluster information is the original cluster partitioning result obtained in the first iteration of the clustering process, which is obtained by directly performing basic clustering on the dimensionality-reduced semantic vectors of the current batch based on the clustering parameter information, without semantic consistency screening.

[0069] It is understandable that each cluster in the initial cluster information only represents a set of samples that are close in vector space distance, and has not yet undergone deep semantic processing. Figure 1 The verification of consistency means that it may contain both truly semantically consistent good clusters and bad clusters that are close in vector distance but have different intentions.

[0070] In a specific embodiment, the number of clusters to be used in the current round (k) can be read from the clustering parameter information (the initial value is used in the first round, for example, k=200). The set of samples to be clustered in the cluster information is taken as input, k initial points are selected as cluster centers, and then the distance from each sample to all cluster centers is calculated iteratively. The sample is assigned to the nearest cluster center. Based on all the samples assigned to each cluster, the center of the cluster is recalculated, thereby outputting k clusters. Each cluster contains a set of samples and its center vector, i.e., the initial cluster information.

[0071] Step S42: Based on the initial cluster information and the clustering parameter information, input a predefined semantic evaluation model to classify the corresponding target clusters and obtain target cluster information.

[0072] It should be noted that the semantic evaluation model is a binary classifier used to evaluate semantic consistency, determining whether questions belong to the same core user intent at the semantic level.

[0073] Understandably, the semantic evaluation model is not a complex neural network that needs to be trained from scratch. Instead, it directly encapsulates and invokes the capabilities of the existing model through prompt word engineering. This design avoids additional training costs and data labeling requirements while maintaining the powerful learning capabilities of LLM.

[0074] In a specific embodiment, a prompt word template is obtained; convex sampling is used to select user questions that satisfy predefined rules from the initial cluster information to determine the evaluation input; based on the evaluation input and the prompt word template, a predefined semantic evaluation model is used to evaluate stable clusters and clusters to be clustered, resulting in target stable clusters and target clusters to be clustered. This can be achieved by obtaining fixed prompt word templates from a preset configuration library or template file, which typically contain explicit instructions and output format constraints. For each cluster to be evaluated, the semantic vectors corresponding to all statements within it are reduced to a 2-dimensional space using PCA. The convex hull of the cluster's point set is calculated in this low-dimensional space, and approximately 10 user questions are selected from the convex hull boundary and interior according to uniform distribution. These questions need to cover the main distribution range of the cluster in the semantic space, thus forming the evaluation input text. The predefined semantic evaluation model is then used to evaluate the input text. Figure 1 Consistency judgment is performed, and the classification results of good or bad clusters are output. All clusters judged as good clusters are identified as target stable clusters and directly retained in the final result, while clusters judged as bad clusters are classified as target clusters to be clustered, thus obtaining the corresponding target stable clusters and target clusters to be clustered.

[0075] Based on the target stable cluster and the target cluster to be clustered, the number of clusters in the clustering parameter information is adjusted to determine the target clustering parameter information. That is, based on the total number of user questions in the target cluster to be clustered, this value is subtracted from the initial total number of questions in this round to accurately calculate the corresponding number of unassigned user questions. The number of clusters in the clustering parameter information is updated according to a preset dynamic adjustment strategy to obtain the target clustering parameter information, ensuring that the number of clusters will adaptively decrease as the size of the questions to be processed decreases.

[0076] Obtain the unclassified proportion threshold and the target number of iterations; determine the corresponding proportion of unassigned users based on the target stable cluster and the target cluster to be clustered; classify the target cluster to be clustered according to the number of clusters in the target clustering parameter information based on the proportion of unassigned users, the unclassified proportion threshold, and the target number of iterations to obtain the target cluster information, thereby obtaining the preset unclassified proportion threshold (e.g., ε=0.05) and the target maximum number of iterations (e.g., 10 times), based on the target stable cluster and the target cluster to be clustered determined in the previous round of evaluation results. The system iterates through the target clusters, calculating the percentage of unassigned user questions in each cluster. If the calculated percentage of unassigned user questions is lower than the preset unclassified proportion threshold ε, or if the current cumulative iteration count has reached the target maximum iteration count, the iteration terminates. Based on the target stable clusters and the target stable iterative clusters, the system obtains the target cluster information. Otherwise, the system uses the updated target clustering parameter information to re-execute the clustering and subsequent evaluation process for all user questions in the target clusters, entering the next iteration until any of the above termination conditions are met.

[0077] In one feasible implementation, step S42 may include steps C11-C14: Step C11: Based on the initial cluster information, input a predefined semantic evaluation model to evaluate stable clusters and clusters to be clustered, and determine the target stable clusters and target clusters to be clustered; It should be noted that the target stable cluster is a cluster that is judged as a good cluster in the evaluation, that is, all user questions within it are highly consistent in semantics and clearly point to the same core intent. Therefore, it is regarded as a high-quality result and is directly retained to the final output. The target cluster to be clustered is a cluster that is judged as a bad cluster in the evaluation, that is, the semantics of its statements are mixed and the intents are inconsistent. Therefore, it needs to be decomposed and the user questions it contains are reassigned to the unassigned set so that it can be clustered again in the iteration.

[0078] In one feasible implementation, step C11 may include steps D11 to D13: Step D11: Obtain the prompt word template; It should be noted that the prompt word template is a predefined text instruction framework used to ensure the clarity, unambiguity and consistency of semantic evaluation task instructions. The template may include clear role settings, specific task descriptions, strict output format requirements and optional examples.

[0079] Step D12: Using convex sampling, select user questions that satisfy predefined rules from the initial cluster information to determine the evaluation input; It should be noted that the evaluation input is a set of representative user questions generated for semantic consistency evaluation. For each initial cluster to be evaluated, the semantic vectors of all statements in the cluster are reduced to a 2-dimensional space and then a convex sampling algorithm is used. For example, the convex hull of the cluster in the low-dimensional space is calculated, and about 10 questions are uniformly selected from the boundary and interior of the convex hull, so that the selected questions can cover the distribution range of the cluster in the semantic space to the maximum extent.

[0080] Step D13: Based on the evaluation input and the predefined semantic evaluation model of the prompt word template input, evaluate the stable clusters and the clusters to be clustered to obtain the target stable clusters and the target clusters to be clustered.

[0081] It should be noted that stable clusters are objects to be evaluated that are initially generated by the clustering algorithm in the current iteration, while clusters to be clustered are clusters that need to be reprocessed after semantic evaluation.

[0082] Step C12: Based on the target stable cluster and the target cluster to be clustered, adjust the number of clusters in the clustering parameter information to determine the target clustering parameter information; It should be noted that the target clustering parameter information is an updated set of parameters used for the next round of iterative clustering, so that the next round of clustering can process the remaining difficult statements more precisely and efficiently.

[0083] Step C13: Based on the target clustering parameter information, classify the target clusters to be clustered to obtain the target stable iterative clusters; It should be noted that the target stable iterative cluster is the set of clustering results obtained by re-dividing the target cluster to be clustered and undergoing semantic evaluation again in the next round of iterative clustering.

[0084] In one feasible implementation, step C13 may include steps E11 to E13: Step E11: Obtain the unclassified proportion threshold and the target number of iterations; It should be noted that the unclassified proportion threshold is an acceptable upper limit for the proportion of unclustered questions. When it is below this threshold, clustering is considered sufficient. The target number of iterations is an absolute upper limit for the number of rounds to prevent infinite loops.

[0085] Step E12: Determine the corresponding proportion of unassigned users based on the target stable cluster and the target cluster to be clustered; It should be noted that the percentage of unassigned user questions is used to quantify the proportion of user questions that are still pending processing in the total.

[0086] Understandably, the percentage of unassigned users can be compared with the unclassified proportion threshold to determine whether the iteration should end prematurely.

[0087] Step E13: Based on the proportion of unassigned users, the unclassified proportion threshold, and the target iteration number, classify the target clusters to be clustered according to the number of clusters in the target clustering parameter information to obtain target cluster information.

[0088] Understandably, the system will compare the real-time calculated proportion of unassigned users with the preset unclassified proportion threshold, and at the same time check whether the number of iterations already performed has reached the target maximum number of iterations. If the proportion of unassigned users is lower than the threshold, or the number of iterations has reached the upper limit, the iteration will terminate, and the system will no longer perform new classification on the target clusters to be clustered. At this time, the target cluster information is equivalent to all the target stable clusters that have been accumulated. Otherwise, the system will start a new round of clustering cycle.

[0089] Step C14: Obtain target cluster information based on the target stable cluster and the target stable iterative cluster.

[0090] It is understandable that the target cluster information is the result of the algorithm's multiple rounds of division, evaluation and optimization of user questions. Each cluster has good intent discrimination, so it can be directly used for standardized naming, or for application scenarios such as training customer service intent classification models and configuring intelligent question answering systems.

[0091] This embodiment proposes a user question clustering method based on large-model collaboration. It employs a clustering algorithm to divide the cluster information according to the number of clusters in the clustering parameters, determining initial cluster information. Based on the initial cluster information and the clustering parameters, a predefined semantic evaluation model is input to classify the corresponding target clusters, obtaining target cluster information. This solves the technical problem of how to perform efficient and accurate user question clustering based on large-model collaboration. Compared to existing technologies, this application significantly improves the semantic accuracy and consistency of clusters by dividing the initial clusters according to the number of clusters in the clustering parameters and using a semantic evaluation model to filter out target clusters. It effectively separates clusters with mixed intentions, significantly improving the reliability and efficiency of large-scale data processing while ensuring accurate clustering of core intentions.

[0092] This application also provides a user clustering device based on large model collaboration; please refer to [reference needed]. Figure 4 The user clustering device based on large model collaboration includes: Module 10 is used to acquire user information to be clustered and clustering parameter information; Processing module 20 is used to determine the corresponding dimensionality-reduced semantic vector based on the user information to be clustered; The processing module 20 is further configured to cluster the dimensionality-reduced semantic vector according to different numbers of clusters based on the clustering parameter information, and determine the cluster information; Execution module 30 is used to classify the corresponding target clusters based on the cluster information and the clustering parameter information by inputting a predefined semantic evaluation model, and to determine the target cluster information; The execution module 30 is also used to generate corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention.

[0093] The processing module 20 is further configured to perform text preprocessing based on the user question set to be clustered, and determine preprocessing information. The text preprocessing includes special character processing, character writing processing, and space processing. Based on the preprocessed information, a predefined vector generation model is used to generate a corresponding high-dimensional semantic vector. The high-dimensional semantic vector is normalized to determine the normalized vector; Principal component analysis is used to reduce the dimensionality of the normalized vector to obtain a dimensionality-reduced semantic vector.

[0094] The execution module 30 is further configured to use a clustering algorithm to divide the cluster information according to the number of clusters in the clustering parameter information to determine the initial cluster information; Based on the initial cluster information and the clustering parameter information, a predefined semantic evaluation model is input to classify the corresponding target clusters, thereby obtaining the target cluster information.

[0095] The execution module 30 is further configured to evaluate stable clusters and clusters to be clustered based on the input of the predefined semantic evaluation model of the initial cluster information, and determine the target stable clusters and target clusters to be clustered. Based on the target stable cluster and the target cluster to be clustered, the number of clusters in the clustering parameter information is adjusted to determine the target clustering parameter information; Based on the target clustering parameter information, the target clusters to be clustered are classified to obtain the target stable iterative clusters; The target cluster information is obtained based on the target stable cluster and the target stable iterative cluster.

[0096] The execution module 30 is also used to obtain a prompt word template; Convex sampling is used to select user questions that satisfy predefined rules from the initial cluster information to determine the evaluation input; Based on the evaluation input and the predefined semantic evaluation model of the prompt word template input, stable clusters and clusters to be clustered are evaluated to obtain target stable clusters and target clusters to be clustered.

[0097] The execution module 30 is also used to obtain the unclassified proportion threshold and the target number of iterations; The proportion of unassigned users is determined based on the target stable cluster and the target cluster to be clustered. Based on the proportion of unassigned users, the unclassified proportion threshold, and the target iteration number, the target clusters to be clustered are classified according to the number of clusters in the target clustering parameter information to obtain target cluster information.

[0098] The execution module 30 is further configured to sample multiple user questions from the corresponding cluster based on the target cluster information to determine a user question representative list; Based on the user representative list input, a predefined large language model is used to name the target clusters according to the predefined verb and object naming conventions, resulting in cluster name labels.

[0099] The user question clustering device based on large model collaboration provided in this application employs the user question clustering method based on large model collaboration in the above embodiments, and can solve the technical problem of how to perform efficient and accurate user question clustering based on large model collaboration. Compared with the prior art, the beneficial effects of the user question clustering device based on large model collaboration provided in this application are the same as the beneficial effects of the user question clustering method based on large model collaboration provided in the above embodiments, and other technical features in the user question clustering device based on large model collaboration are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0100] This application provides a user clustering device based on large model collaboration. The user clustering device based on large model collaboration includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the user clustering method based on large model collaboration in the above embodiment 1.

[0101] The following is for reference. Figure 5This document illustrates a structural diagram of a user clustering device based on large-model collaboration suitable for implementing embodiments of this application. The user clustering device based on large-model collaboration in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The user clustering device based on large model collaboration shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0102] like Figure 5 As shown, a user clustering device based on large model collaboration may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the user clustering device based on large model collaboration. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the user-interactive clustering device based on large model collaboration to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a user-interactive clustering device based on large model collaboration with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented alternatively.

[0103] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0104] The user question clustering device based on large model collaboration provided in this application, employing the user question clustering method based on large model collaboration in the above embodiments, can solve the technical problem of how to perform efficient and accurate user question clustering based on large model collaboration. Compared with the prior art, the beneficial effects of the user question clustering device based on large model collaboration provided in this application are the same as those of the user question clustering method based on large model collaboration provided in the above embodiments, and other technical features in this user question clustering device based on large model collaboration are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0105] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0107] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the user-interview clustering method based on large model collaboration in the above embodiments.

[0108] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0109] The aforementioned computer-readable storage medium may be included in a user-interview clustering device based on large model collaboration; or it may exist independently and not be assembled into a user-interview clustering device based on large model collaboration.

[0110] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a user query clustering device based on large-model collaboration, the user query clustering device based on large-model collaboration performs the following: acquires user query information to be clustered and clustering parameter information; determines the corresponding dimensionality-reduced semantic vector based on the user query information to be clustered; clusters the dimensionality-reduced semantic vector according to different numbers of clusters based on the clustering parameter information to determine cluster information; inputs the cluster information and the clustering parameter information into a predefined semantic evaluation model to classify the corresponding target clusters to determine target cluster information; and generates corresponding cluster name labels based on the target cluster information according to a predefined verb and object naming convention.

[0111] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0112] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0113] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0114] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described user-interview clustering method based on large model collaboration. This solves the technical problem of how to perform efficient and accurate user-interview clustering based on large model collaboration. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the user-interview clustering method based on large model collaboration provided in the above embodiments, and will not be repeated here.

[0115] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A user-interview clustering method based on large-model collaboration, characterized in that, The method includes: Obtain user information and clustering parameter information for the users to be clustered; Determine the corresponding dimensionality-reduced semantic vector based on the user information to be clustered; Based on the clustering parameter information, the dimensionality-reduced semantic vector is clustered according to different numbers of clusters to determine the cluster information; Based on the cluster information and the clustering parameter information, a predefined semantic evaluation model is input to classify the corresponding target clusters and determine the target cluster information; Based on the target cluster information, corresponding cluster name tags are generated according to a predefined verb and object naming convention.

2. The method as described in claim 1, characterized in that, The step of determining the corresponding dimensionality-reduced semantic vector based on the user information to be clustered includes: Based on the set of users to be clustered, text preprocessing is performed to determine preprocessing information. The text preprocessing includes special character processing, character writing processing, and space processing. Based on the preprocessed information, a predefined vector generation model is used to generate a corresponding high-dimensional semantic vector. The high-dimensional semantic vector is normalized to determine the normalized vector; Principal component analysis is used to reduce the dimensionality of the normalized vector to obtain a dimensionality-reduced semantic vector.

3. The method as described in claim 1, characterized in that, The step of classifying the corresponding target clusters and determining the target cluster information by inputting the cluster information and the clustering parameter information into a predefined semantic evaluation model includes: A clustering algorithm is used to divide the cluster information according to the number of clusters in the clustering parameter information to determine the initial cluster information; Based on the initial cluster information and the clustering parameter information, a predefined semantic evaluation model is input to classify the corresponding target clusters, thereby obtaining the target cluster information.

4. The method as described in claim 3, characterized in that, The step of classifying the corresponding target clusters and obtaining target cluster information by inputting a predefined semantic evaluation model based on the initial cluster information and the clustering parameter information includes: Based on the initial cluster information, a predefined semantic evaluation model is input to evaluate stable clusters and clusters to be clustered, thereby determining the target stable clusters and target clusters to be clustered. Based on the target stable cluster and the target cluster to be clustered, the number of clusters in the clustering parameter information is adjusted to determine the target clustering parameter information; Based on the target clustering parameter information, the target clusters to be clustered are classified to obtain the target stable iterative clusters; The target cluster information is obtained based on the target stable cluster and the target stable iterative cluster.

5. The method as described in claim 4, characterized in that, The step of evaluating stable clusters and clusters to be clustered based on the initial cluster information input into a predefined semantic evaluation model, and determining the target stable clusters and target clusters to be clustered, includes: Get the prompt word template; Convex sampling is used to select user questions that satisfy predefined rules from the initial cluster information to determine the evaluation input; Based on the evaluation input and the predefined semantic evaluation model of the prompt word template input, stable clusters and clusters to be clustered are evaluated to obtain target stable clusters and target clusters to be clustered.

6. The method as described in claim 4, characterized in that, The step of classifying the target clusters to be clustered based on the target clustering parameter information to obtain target stable iterative clusters includes: Obtain the unclassified proportion threshold and the target number of iterations; The proportion of unassigned users is determined based on the target stable cluster and the target cluster to be clustered. Based on the proportion of unassigned users, the unclassified proportion threshold, and the target iteration number, the target clusters to be clustered are classified according to the number of clusters in the target clustering parameter information to obtain target cluster information.

7. The method as described in claim 1, characterized in that, The step of generating corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention includes: Based on the target cluster information, multiple user questions are sampled from the corresponding clusters to determine a list of user question representatives; Based on the user representative list input, a predefined large language model is used to name the target clusters according to the predefined verb and object naming conventions, resulting in cluster name labels.

8. A user clustering device based on large model collaboration, characterized in that, The device includes: The acquisition module is used to acquire user information to be clustered and clustering parameter information; The processing module is used to determine the corresponding dimensionality-reduced semantic vector based on the user information to be clustered; The processing module is also used to cluster the dimensionality-reduced semantic vector according to different numbers of clusters based on the clustering parameter information, and determine the cluster information; The execution module is used to classify the corresponding target clusters based on the cluster information and the clustering parameter information input into a predefined semantic evaluation model, and determine the target cluster information; The execution module is also used to generate corresponding cluster name tags based on the target cluster information according to a predefined verb and object naming convention.

9. A user clustering device based on large model collaboration, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the user-interview clustering method based on large model collaboration as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the user-interview clustering method based on large model collaboration as described in any one of claims 1 to 7.

Images