Candidate question recommendation method for intelligent dialogue system and related device

Abstract

The invention belongs to the field of artificial intelligence, and discloses a candidate question recommendation method for an intelligent dialogue system and a related device.Firstly, a deep learning model is adopted for conducting semantic coding and intention classification on an original question of a user, a semantic vector and an intention label are generated, and the semantic limitation of traditional keyword matching is broken through; screening the candidate question database by using the intention label to form a primary screening set, and narrowing the retrieval range; semantic matching of problem levels is achieved through semantic vector similarity calculation; and finally, performing dynamic weighted sorting by integrating multi-dimensional features such as semantic similarity, user portrait matching degree, question popularity and type adjustment factors to form a personalized recommendation list. By adopting the method, the accuracy of question recommendation and the scene generalization ability are effectively improved, so that the recommendation result not only meets the real-time semantic demand of the user, but also gives consideration to personalized preference and business scene characteristics.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, particularly to the field of intelligent dialogue systems, and in particular to a candidate question recommendation method and related devices for an intelligent dialogue system. Background Art

[0002] With the booming rise of enterprise-level AI services, consulting scenarios where enterprise users ask questions and intelligent customer service answers them have become widely used. In daily enterprise operations, intelligent customer service systems (intelligent dialogue systems) play a crucial role in rapidly responding to customer needs and improving service efficiency. Accurately understanding user intent and providing relevant question recommendations are key to improving user experience and operational efficiency, and are also a key direction for the current development of intelligent enterprise services.

[0003] However, existing intelligent dialogue systems have many shortcomings in terms of question recommendation. On the one hand, existing solutions are mostly based on static rules or keyword matching, which leads to a misalignment between recommended questions and users' real needs, and cannot achieve accurate understanding and real-time recommendations at the problem level. In particular, there is a lack of intelligent discrimination and differentiated processing capabilities between structured process-type problems and semi-structured scenarios. On the other hand, traditional methods rely on complex feature engineering and massive labeled data, and have problems such as complex feature engineering, difficult cold start, and high dependence on labeled data. When adapting to internal enterprise application scenarios with medium and low traffic, they face the difficulties of high scene migration costs and high implementation barriers. In addition, the existing system is based on keyword matching and lacks flexible context understanding capabilities. Scene migration relies on manual rule reconstruction, and recommended questions are prone to deviate from the core needs of users.

[0004] This shows that the existing intelligent customer service question recommendation technology cannot achieve accurate understanding of the problem level and is prone to deviate from the core needs of users. Summary of the Invention

[0005] The present invention provides a candidate question recommendation method and related devices for an intelligent dialogue system. Based on a collaborative optimization framework that integrates semantic similarity and user profiles, this method achieves accurate question recommendation in user questioning scenarios, effectively improving enterprise customer service interaction efficiency and user satisfaction.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a candidate question recommendation method for an intelligent dialogue system, comprising: The original question input by the user is semantically encoded and the intent is classified and identified to obtain the semantic vector and intent label of the original question; Based on the semantic vector and intent label of the original question, search the pre-built candidate question database to obtain a preliminary screening set of candidate questions; Based on the semantic vector of the original question, semantic matching is performed with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score; Combined with the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weights of the candidate questions, the candidate questions in the initial screening set are prioritized to obtain a personalized candidate question list, so as to recommend the personalized candidate question list to the user.

[0007] A further improvement of the present invention is that the semantic encoding and intent classification identification of the original question input by the user to obtain the semantic vector and intent label of the original question includes: Use the pre-built language understanding model to semantically encode the original question input by the user to obtain the semantic vector of the original question; A multi-category classification algorithm is used to classify and identify the intent of the original question input by the user to obtain the intent label of the original question.

[0008] A further improvement of the present invention is that, based on the semantic vector and the intention label of the original question, a search is performed in a pre-built candidate question database to obtain a preliminary screening set of candidate questions, including: According to the semantic vector and intent label of the original question, a vector and text hybrid retrieval method is used to retrieve the candidate questions in the pre-built candidate question database to obtain a preliminary screening set of candidate questions; among them, the vector and text hybrid retrieval is implemented based on text retrieval and vector retrieval.

[0009] A further improvement of the present invention is that the process of constructing the candidate question database includes: Collect historical questions, including internal frequently asked questions documents, business frequently asked questions, and operational process questions and answers; A pre-trained large language model is used to vectorize historical question texts, and a semantic index structure is constructed based on a vector similarity search library to obtain a candidate question database. Among them, the vector index in the candidate question database is updated through an offline scheduled task.

[0010] A further improvement of the present invention is that the semantic vector based on the original question is semantically matched with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score, including: The cosine similarity and Euclidean distance correlation algorithms are used to semantically match the semantic vector of the original question with the candidate questions in the initial screening set of candidate questions to obtain the semantic similarity score.

[0011] A further improvement of the present invention is that, before prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, the method further includes: Obtain user portrait matching based on dynamic user tags of user portraits, including: Collect users' static and dynamic data; Using an in-memory database cache layer in conjunction with a distributed stream processing platform message queue, user profiles are updated in real time based on static and dynamic data. Based on user behavior data, combined with time-decay weights and department-tag mapping tables, the user's candidate question preferences are calculated to generate dynamic user tags that are updated in real time. According to dynamic user tags, the user profile matching degree of candidate questions is obtained.

[0012] A further improvement of the present invention is that the candidate questions in the initial screening set are prioritized by combining the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors and corresponding weights of the candidate questions, including: Collect the semantic similarity score, user profile matching degree, question heat factor and question type adjustment factor of the candidate question; wherein the question heat factor is obtained based on the recent click frequency of the candidate question; the question type adjustment factor is a temperature coefficient set based on the question type of the candidate question; Based on the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weight of the candidate questions, the comprehensive score corresponding to each candidate question in the initial screening set of candidate questions is solved. The specific formula is as follows:

[0013] Where Softmax() represents the scoring function; W1, W2, W3, and W4 represent the semantic similarity score, user profile matching, question heat factor, and question type adjustment factor, respectively; α, β, and γ are adjustable hyperparameters representing different weight coefficients; T(W4) represents the temperature adjustment function for the question type adjustment factor; Based on the comprehensive scores corresponding to each candidate question, the candidate questions are prioritized from high to low scores to obtain a personalized candidate question list.

[0014] A further improvement of the present invention is that after prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, the method further includes: Collect user feedback on the personalized candidate question list and regularly evaluate the recommendation effect based on the feedback; Dynamically adjust the weights of semantic similarity score, user profile matching, question heat factor and question type adjustment factor based on the recommendation effect.

[0015] In a second aspect, the present invention further provides a candidate question recommendation system for an intelligent dialogue system, comprising: The semantic recognition module is used to semantically encode the original question input by the user and classify the intent to obtain the semantic vector and intent label of the original question; The retrieval module is used to search the pre-built candidate question database based on the semantic vector and intent label of the original question to obtain a preliminary screening set of candidate questions; The semantic matching module is used to perform semantic matching with candidate questions in the initial screening set of candidate questions based on the semantic vector of the original question to obtain a semantic similarity score; The question recommendation module is used to prioritize the candidate questions in the initial screening set based on their semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights, and obtain a personalized candidate question list to recommend the personalized candidate question list to users.

[0016] In a third aspect, the present invention further provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned candidate question recommendation method for an intelligent dialogue system.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a candidate question recommendation method for intelligent dialogue systems. First, a deep learning model is used to semantically encode and classify the user's original question, generating a semantic vector and intent label. This overcomes the semantic limitations of traditional keyword matching. The intent label is then used to filter a database of candidate questions to form a preliminary screening set, narrowing the search scope. Semantic matching at the question level is then achieved through semantic vector similarity calculation. Finally, a dynamic weighted ranking is performed based on multi-dimensional features such as semantic similarity, user profile matching, question popularity, and type adjustment factors to form a personalized recommendation list. This method addresses intent understanding bias through a semantic encoding layer. A coarse screening mechanism based on intent labels reduces scenario migration costs. A multi-factor fusion ranking model replaces manual feature engineering, combining static knowledge (question popularity) with dynamic features (user profile) to enhance recommendation adaptability. This method effectively improves the accuracy and scenario generalization of question recommendations, ensuring that recommendation results are tailored to users' real-time semantic needs while also taking into account personalized preferences and business scenario characteristics. Furthermore, a pre-trained model reduces reliance on labeled data, significantly alleviating the semantic bias, cold start difficulties, and high rule maintenance costs associated with traditional methods.

[0018] Preferably, the present invention uses Sentence-BERT or fine-tuned BERT to acquire deep semantic representations and enhance problem representation capabilities; a multi-classification algorithm establishes an intent labeling system. Pre-trained models reduce reliance on labeled data, and dual-path processing preserves common semantics while enhancing domain intent recognition, effectively alleviating the cold start dilemma and improving cross-scenario generalization capabilities.

[0019] Preferably, the present invention uses a hybrid engine combining Elasticsearch (text search) and Faiss (vector search), taking into account the advantages of both literal and semantic matching. This hybrid search strategy overcomes the limitations of a single search mode, significantly improving the relevance of the initial screening set while maintaining recall, laying a high-quality data foundation for subsequent refined sorting.

[0020] Preferably, the present invention constructs a semantic index based on vectorized processing of large models, and an offline scheduled update mechanism ensures the real-time nature of the knowledge base. This method maps sentences into a semantic space, reduces maintenance costs through automated index updates, and addresses the efficiency bottleneck of traditional manual feature engineering methods.

[0021] Preferably, the present invention quantifies semantic relevance using cosine / Euclidean distance to establish an interpretable matching metric. This method transforms abstract semantic relationships into comparable logarithmic values ​​through mathematical modeling, providing standardized input for multi-dimensional fusion and avoiding the subjective bias of traditional manual rules.

[0022] Preferably, the present invention uses a Kafka+Redis architecture to enable real-time updates of user profiles, while time decay and department tag mapping enhance dynamic feature capture. This overcomes the lag inherent in static profiles, enabling the recommendation system to perceive real-time changes in user status and improving the timeliness and accuracy of personalized recommendations.

[0023] Preferably, the present invention uses a weighted formula that integrates semantics, user preferences, question popularity, and business type, combined with a temperature coefficient to adjust the weight of specific questions. This method balances semantic relevance with business operational requirements. Softmax normalization ensures the operability of multi-objective optimization, and the parameter adjustment mechanism enhances adaptability to different scenarios.

[0024] Preferably, in the present invention, a feedback-driven dynamic parameter adjustment mechanism forms a closed-loop optimization. By continuously collecting user behavior data and automatically adjusting feature weights, the system is able to self-evolve, reducing the cost of manual rule maintenance and resolving the pain point of traditional methods that require rule reconstruction for scenario migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A schematic diagram of the process flow of the data preprocessing stage provided by an embodiment of the present invention; Figure 2 A schematic diagram of the process of the online questioning stage provided by an embodiment of the present invention; Figure 3 A flowchart of a method for recommending candidate questions for an intelligent dialogue system provided by the present invention; Figure 4 This is a schematic diagram of the structure of a candidate question recommendation system for an intelligent dialogue system provided by the present invention. DETAILED DESCRIPTION

[0026] In order to further understand the content of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the embodiments are only for explaining the present invention and are not intended to limit it.

[0027] As described in the background technology, in intelligent dialogue systems, current traditional solutions mostly rely on keyword matching, and brute force search has high computational complexity. It focuses more on the current question context and cannot be well applied to enterprise business scenarios. It lacks a deep understanding of the user's historical behavior analysis and business needs, and the recommendation results are likely to deviate from the user's true intentions. It can be seen that in traditional technical solutions, based on the user's original question, feature keywords are extracted and limited question templates are predefined; based on the keyword matching template, candidate questions are generated, and finally the candidate questions are sorted based on a single dimension of static business rules. It can be seen that there is a lack of personalized feature analysis for users, and it is impossible to dynamically respond to business priorities, which may cause problems such as separation from the business logic that users care about.

[0028] In order to solve the above problems, this embodiment provides a candidate question recommendation method for an intelligent dialogue system. In the scenario where an enterprise builds an intelligent dialogue based on a large model, this method can further generate and output recommended questions on the basis of outputting user question responses by fusing the semantic relevance of the question text with the user portrait, thereby making accurate recommendations to guide questions, making it easier for users to efficiently start the next round of dialogue interaction.

[0029] In order to better understand the present technical solution, the technical terms involved in this invention are explained as follows: Correlation Algorithm: Correlation algorithm is a statistical method used to measure the strength and direction of the relationship between two or more variables.

[0030] Portfolio: A labeling model based on user data that abstracts a user model by collecting and analyzing multi-dimensional data such as user personal information, preferences, and behavioral habits.

[0031] Vector Search: Improves knowledge matching efficiency by converting text into high-dimensional vectors and performing retrieval based on similarity.

[0032] Cosine Similarity: This metric measures the degree of directional similarity between two vectors by calculating the cosine of the angle between them. It focuses on the direction of the vector rather than its magnitude (length) and is widely used in text classification, recommendation systems, information retrieval, and other fields.

[0033] Large Language Models: Large Language Models are deep learning models trained on large amounts of text data, capable of understanding and generating natural language text. Through large-scale pre-training, the Transformer architecture, multi-task learning, and contextual understanding, large language models excel in tasks such as text generation, machine translation, question-answering systems, text summarization, sentiment analysis, and dialogue systems.

[0034] Retrieval Augmented Generation (RAG): This model architecture combines a retrieval module with a generation module. Its core concept is to generate more accurate and richer answers by retrieving relevant information from an external knowledge base and combining it with the power of a generative model. This approach overcomes the shortcomings of large language models in terms of factual accuracy and knowledge coverage.

[0035] Pre-trained language models: Mainstream pre-trained language models such as BERT, RoBERT, and Sentence-BERT (SBERT) possess a deep understanding of context and can encode natural language questions into high-dimensional semantic vectors. They are widely used in tasks such as FAQ (Frequently Asked Questions) matching, question classification, and intent recognition, significantly improving the accuracy of semantic matching. BERT is a pre-trained language model that learns contextual representations of text using a bidirectional Transformer architecture, achieving breakthrough results in natural language processing tasks. Sentence-BERT is a sentence embedding technology improved upon the BERT model.

[0036] Semantic similarity algorithms: After encoding the question text into a vector using a language model, similarity is often scored using methods such as cosine similarity and Euclidean distance. This technology supports high-quality matching in unsupervised or small sample conditions, and is well-suited for questions with frequently updated internal knowledge bases and non-standard structures.

[0037] User portrait modeling: User portrait technology has been maturely applied in ToC e-commerce and recommendation systems and is now gradually being introduced into enterprise intelligent question-answering systems. User portraits include static attributes such as department, role, position, and business permissions; and dynamic behaviors such as question history, clickthrough rate, preferred question types, and typical business trajectory. User portraits can introduce personalized weighting constraints into recommendation algorithms, enabling personalized service experiences for each individual.

[0038] Retrieval-augmented generative (RAG) architecture enhances the capabilities of question-answering systems through a "retrieval + generation" approach. It first uses vector retrieval technology to retrieve relevant documents from the knowledge base. These documents, along with the user's question, are then fed into a generative model (such as GPT) to generate high-quality answers. This mechanism is particularly well-suited for scenarios where enterprise FAQs are fragmented, frequently updated, and have irregular data structures, and has become a key area of ​​focus for upgrading intelligent customer service systems.

[0039] Vector retrieval and hybrid indexing mechanism: Based on multi-dimensional vector semantic representation, it uses vector databases (such as Faiss and Milvus) to achieve efficient retrieval. It also combines cluster indexes, hash indexes, and fine-grained filtering indexes to build a hybrid structure to improve response speed and accuracy. This is one of the important foundational technologies for enterprise-level deployment.

[0040] This embodiment provides a candidate question recommendation method for an intelligent dialogue system, including: The original question input by the user is semantically encoded and the intent is classified and identified to obtain the semantic vector and intent label of the original question; Based on the semantic vector and intent label of the original question, search the pre-built candidate question database to obtain a preliminary screening set of candidate questions; Based on the semantic vector of the original question, semantic matching is performed with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score; Combined with the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weights of the candidate questions, the candidate questions in the initial screening set are prioritized to obtain a personalized candidate question list, so as to recommend the personalized candidate question list to the user.

[0041] The following is a more detailed description of the candidate question recommendation method provided in this embodiment: This embodiment provides a candidate question recommendation method for an intelligent dialogue system, and the specific steps are as follows: S1. The process of building the candidate question database and semantic vector retrieval system, the process of building the user portrait, the process of semantic analysis and feature extraction of the original question, the process of ranking multi-factor recommendation questions, the process of building the recommendation service interface, and the process of feedback and evaluation.

[0042] S11. Construction of question database and semantic vector retrieval system.

[0043] Developers need to collect and standardize data sources such as internal FAQs, business FAQs, and operational process Q&A. They then use pre-trained language models (such as Sentence-BERT and E5) to vectorize the question text and establish a semantic index structure based on Faiss to support fast queries of the top-N similar questions. Furthermore, to support dynamic updates of the embedding model in the future, it is necessary to design an interface for generating and storing the embedding model, along with supporting offline scheduled tasks.

[0044] S12. Build user portraits and collect static and dynamic behavior tags.

[0045] User portraits are the foundation of intelligent recommendation systems. Their purpose is to build an accurate user feature database by collecting and analyzing user behavior data (including static and dynamic data) in real time, thereby providing a decision-making basis for question recommendations.

[0046] The specific functional design is as follows: S21. Data Collection and Standardization: First, the system collects and standardizes static user data (such as department and role) and dynamic data (such as question keywords and clicked recommended questions) in real time. During this process, Apache Airflow is used for regular data scheduling to ensure data accuracy and timeliness.

[0047] S22. Dynamic Profile Update: Utilizing a Redis cache layer and Kafka message queues, the system processes user behavior data in real time and updates user profiles. To ensure real-time profiles, the system has designed a low-latency behavior data processing architecture to ensure rapid response to user requests even in high-concurrency scenarios.

[0048] S23. User Tag Generation: Based on user behavior data, combined with time-decay weights and a department-tag mapping table, we calculate user question preferences and generate dynamic user tags that are updated in real time. The key challenge in this step is ensuring real-time updates while minimizing database query pressure. Therefore, we use Redis caching to store the latest user profiles and adjust them based on behavioral changes.

[0049] S3. Semantic parsing and feature extraction process of the original question.

[0050] After a user asks a question, developers need to build a pre-analysis module. Using the language model API or a locally deployed model, the module performs word segmentation, intent identification, keyword extraction, and vectorized encoding on the question text. This module must be packaged as a standalone service, support high-concurrency calls, and output a structured query request body (including a semantic vector, parsed fields, and a profile ID). After this step, the semantic vector and intent label of the original question are obtained.

[0051] S4. The recommendation module integrates multi-factor weights to rank question candidates.

[0052] After receiving the semantic vector and intent label, the recommendation module first obtains the top-N candidate question vector clusters in the Faiss index through the vector retrieval module to obtain a preliminary screening set of candidate questions.

[0053] Then, weight calculation and priority ranking are performed based on the following multiple factors: W1 represents the semantic similarity score (the vector distance between the original question and the candidate question); W2 represents the user profile matching degree (the matching score between the question topic and the user profile preference); W3 represents the question popularity factor (calculated based on click / usage frequency in the past 30 days); W4 represents the problem type adjustment factor (dynamically setting the temperature coefficient T based on industry-wide, business process, indicator-based, and other problem types to control the conservatism and diversity of recommendations). The specific formula is as follows:

[0054] Among them, α, β, and γ are adjustable hyperparameters, and T(W4) is the temperature adjustment function of the problem type adjustment factor.

[0055] Developers need to design configurable weight parameters (such as α, β, γ, etc.), use a unified scoring function (such as weighted Softmax) to perform recommendation sorting, and output the Top-K question results. At this point, the Top-K question results are recommended to users as a personalized candidate question list.

[0056] Exemplarily, the recommendation method for this problem further includes: S5. Build the recommendation service interface and the front-end docking logic process.

[0057] Developers encapsulate the above recommendation module as a RESTful API, providing a callable method that accepts question text and user ID as input and outputs a list of recommended questions (a personalized list of candidate questions). Front-end development requires implementing components for displaying results, logic for tracking clicks on recommended questions, and a fallback mechanism (e.g., returning a FAQ link if no recommendation is made) to ensure a smooth recommendation process.

[0058] S6. Build a behavior feedback collection mechanism and popularity update task (feedback and evaluation process).

[0059] The intelligent dialogue system needs to process user behavior logs regularly (such as daily), update the popularity data of candidate questions, and use key indicators such as recommendation hit rate and click-through rate for the next round of recommendation optimization.

[0060] Developers need to configure scheduling scripts, update the Redis hot cache or persistent database, and update the matching history of "question-user profiles" to form a closed-loop optimization system. To ensure the effectiveness of recommendations, an evaluation module must be developed to support the calculation of metrics such as recommendation accuracy, click-through rate, and average dwell time. A / B experimentation should be supported. By adjusting weight combinations and temperature coefficient strategies, the quality of recommendations under different configurations can be compared, providing a basis for continuous optimization of recommendation strategies.

[0061] Exemplarily, in order to effectively execute the steps of the above-mentioned candidate question recommendation method for an intelligent dialogue system, this embodiment also provides a candidate question recommendation system for an intelligent dialogue system, including: a user portrait module, a semantic understanding module, a recommendation algorithm module, and an effect evaluation module; this system also introduces a real-time feedback mechanism and a performance optimization module to ensure the real-time and accuracy of the recommendation system.

[0062] User profile module. This module is responsible for collecting, storing, and processing user data, and generating and updating user profiles to provide personalized decision-making basis for question recommendations. User profiles include static attributes such as the user's department and role, which are regularly scheduled and updated using Apache Airflow, as well as dynamic attributes such as the user's historical questions and clicked recommended questions, which are updated in real time using a Redis cache layer and Kafka. The user profile module provides personalized decision support for the recommendation algorithm. Its output (user tags) is passed to the recommendation algorithm module to influence the screening and ranking of candidate questions.

[0063] Semantic understanding module. This module is responsible for converting the user's natural language questions into semantic representations that can be processed by the machine. Sentence-BERT or Fine-tuned BERT is used to generate a vector representation of the question (768-dimensional vector), and intent is identified through a multi-category classification algorithm (such as XGBoost), outputting the intent label of the question. This module is not only responsible for semantic conversion, but also passes the identified intent label to the subsequent retrieval module for coarse screening, thereby improving the accuracy of subsequent recommendations. The semantic understanding module provides key semantic information to the recommendation algorithm module through intent identification and semantic vector generation, ensuring that the recommendation results meet the user's actual needs. It should be noted that in this embodiment, the processing logic for question type discrimination is to identify the intent of the user's question based on the Fine-tuned BERT model or the XGBoost multi-classification model, and output the corresponding question type label (such as industry-wide questions, business process questions, indicator query questions), which is passed as a structured intermediate result to the recommendation ranking module to drive the selection of the subsequent multi-factor weight adjustment strategy.

[0064] The recommendation algorithm module is responsible for quickly identifying relevant questions from the candidate question pool and ranking them based on multi-dimensional weighting. First, candidate questions are screened using a hybrid search based on Elasticsearch and Faiss. Initially, the module uses intent tags for coarse screening, followed by refined screening based on multiple factors such as semantic similarity, user profile fit, and question popularity. Question ranking is then performed, combining BERT semantic analysis and the BM25 algorithm to rank the candidate questions after the coarse screening. Different weighting strategies are used for different types of questions (general industry questions, business process questions, and indicator query questions), specifically adjusting the temperature coefficient to influence diversity and accuracy. Finally, by dynamically adjusting weights and incorporating real-time changes in user profiles, the module recommends the most relevant and appropriate questions to users. The recommendation algorithm module uses multi-factor weighting, combining factors such as user profile, semantic analysis, and question popularity, to generate personalized recommended questions and return the recommended results to the user. In this recommendation algorithm module, the question type label output by the semantic understanding module is used to control the temperature coefficient T in the ranking weight, improving the diversity (general category) or accuracy (indicator category) of recommended questions, thus implementing an adaptive ranking mechanism for question types.

[0065] Effectiveness Evaluation Module. This module optimizes the recommendation algorithm through continuous user behavior feedback, continuously improving recommendation quality. This includes: user behavior logging, which regularly evaluates recommendation effectiveness by recording user clicks, selections, and feedback on recommended questions (such as question resolution rate); dynamic weight adjustment, which automatically fine-tunes the weight configuration of various dimensions based on core indicators such as CTR (click-through rate) and question resolution rate; in particular, question popularity is dynamically incorporated into weight adjustment as a key factor. The system tracks and calculates the frequency of questions asked within a certain period of time, and more popular questions are given higher recommendation priority.

[0066] Real-time feedback mechanism module. This module collects and analyzes real-time user feedback to drive continuous optimization of the recommendation system. By integrating Elasticsearch, the system can capture user behavior data in real time and update recommendation results, ensuring that the recommendation system can respond promptly to changes in user needs.

[0067] Softmax function. The Softmax function is a mathematical function widely used in machine learning and deep learning. It is primarily used to convert a vector containing arbitrary real numbers into a probability distribution, meaning that each element of the output is between 0 and 1, and the sum of all elements is 1. This property makes the Softmax function very useful in multi-classification problems because it can convert the model's raw output into class probabilities, facilitating probabilistic interpretation and decision-making.

[0068] For example, the candidate question recommendation method for an intelligent dialogue system provided in this embodiment is applied and implemented, including a data preprocessing stage and an online questioning stage. The specific implementation process is as follows: like Figure 1 As shown in Figure 2, the data preprocessing stage includes: Step 1: Generate candidate question vectors. Use historical questions and knowledge base documents to train the Embedding model and generate vector representations of candidate questions. Use Faiss to build a database of candidate question vectors, then apply clustering techniques to stratify them. This hybrid index structure, combining LSH hashing and refined indexing, is constructed. This provides rapid location and precise screening capabilities for subsequent searches.

[0069] Step 2: User profile construction. Static attributes are used to generate basic user profiles, while dynamic attributes are collected and updated in real time via Kafka and Redis. Time-decaying weights are used to calculate user behavioral preferences and generate dynamic user tags. Ultimately, user profile data is passed to the recommendation algorithm module to support personalized recommendations. The Kafka+Redis architecture combines Kafka (a distributed stream processing platform) and Redis (an in-memory database), leveraging the strengths of both to achieve efficient data processing and fast data access.

[0070] like Figure 2 As shown, the online questioning stage includes: Step 1: Semantic understanding and intent recognition.

[0071] After receiving a user question, we use Fine-tuned BERT to perform semantic analysis on the question and generate a question vector (preferably 768 dimensions). Simultaneously, we use an XGBoost classification model to identify intent and output a label for the question. This label, along with the question vector, is passed to the recommendation algorithm module to support the subsequent recommendation process.

[0072] Step 2: Candidate question screening and user profile integration.

[0073] After receiving the user's question intent label and semantic vector, we first search the candidate question pool using Elasticsearch and Faiss. We then perform a rough screening based on user profiles (e.g., role, department) and historical behavior data (e.g., click preferences) to ensure that the recommended questions semantically match the user's needs.

[0074] Step 3: Question sorting and personalized recommendations.

[0075] After screening, candidate questions are sorted based on semantic similarity, user profile fit, question popularity, and other dimensions. In this step, multi-factor weight configuration for different question types is introduced: Common industry questions (e.g., "What is Porter's Five Forces?"): These questions typically require more diversity to guide users through the functionality of the intelligent customer service. To achieve this, we increase the temperature coefficient, increasing the probability gap between tokens and recommending more diverse questions.

[0076] Business process questions (e.g., "How do I use this financial reporting system?"): These questions require high accuracy, so the temperature coefficient is lowered to ensure the accuracy and consistency of the recommended questions. At the same time, the weight of the question vectors in the standard knowledge base is increased to ensure that the recommended questions are highly relevant to the business process.

[0077] Metrics-based questions (e.g., "What is the company's online sales revenue?"): These questions require user permissions, so we factor in the user's business role (e.g., finance, sales, etc.) during the recommendation process and filter based on permission level. When adjusting weights, we'll adjust them based on the match between the user role and the question type to ensure that the recommended questions are relevant to the user's query.

[0078] Step 4: User Feedback and Optimization. Users' clicks, feedback, and problem resolution on recommended questions are recorded as a basis for evaluating the quality of recommended questions. By calculating metrics such as click-through rate (CTR) and problem resolution rate, the weighting of the recommendation algorithm is regularly adjusted. After each update, the system will provide real-time feedback on user behavior data via Kafka and Elasticsearch, updating the popularity of candidate questions and user profiles to optimize recommendation accuracy.

[0079] As can be seen, this embodiment solves the technical problems of enterprise question recommendation systems in terms of personalized accuracy and adaptability to multiple enterprise scenarios from five aspects: semantic understanding, user portrait modeling, multi-factor dynamic fusion calculation, question type adjustment mechanism, and feedback optimization mechanism. The specific principles are: First, multi-level semantic understanding and question vectorization. We use an embedding model to vectorize historical questions and knowledge base questions. We then combine correlation algorithms such as cosine similarity and Euclidean distance for deep semantic matching, replacing traditional keyword matching logic and improving our ability to understand the intent of user input questions. On this basis, we build a vector clustering + hybrid index structure (such as Faiss clustering index + Elasticsearch text inverted index) to accelerate the efficient retrieval of candidate questions.

[0080] Second, user profile modeling and personalized feature injection. User profiles are broken down into static attributes (such as department, position, and business line) and dynamic behavioral characteristics (such as historical clickthroughs, common question types, and recent search keywords). User profiles are constructed as structured vectors and used as auxiliary features in the question recommendation process to dynamically adjust recommendation weights and achieve personalized recommendations for different user roles (e.g., "HR personnel are more interested in questions related to vacation policies, while finance personnel are more interested in questions related to cost indicators").

[0081] Third, multi-factor fusion recommendation mechanism and temperature coefficient adjustment. The recommendation algorithm module introduces a multi-factor fusion weight mechanism, which comprehensively considers the following factors: The semantic similarity between the current input question and the candidate question; The fit between the user persona and the candidate questions; The historical popularity of the candidate question (such as visit frequency / click feedback); The user's response to the recommendation results in the closed-loop feedback (such as click-through rate, satisfaction score, etc.).

[0082] In addition, the system configures differentiated temperature coefficients (T) for different question types (such as general industry questions, standard business process questions, and indicator query questions) to adjust the "divergence" or "concentration" of recommendation results: For standard process issues (such as "reimbursement process"), accuracy should be prioritized, the temperature coefficient should be set low, and high confidence and high coincidence recommendations should be output; For industry-wide or open-ended questions (such as “project management tool recommendations”), the temperature coefficient should be moderately increased to enhance the diversity and exploratory nature of the recommendations.

[0083] The adoption of this mechanism can enhance the system's adaptability to complex question-answer structures, while taking into account both accuracy and coverage.

[0084] Fourth, a closed-loop feedback loop and dynamic optimization mechanism for question recommendations. The system records user behavior in real time, including clicks, jumps, and new question inputs within recommended questions. This behavior data is quickly written and updated through a Redis cache layer. This behavior data is synchronized back to the recommendation module, used to adjust question popularity in real time, optimize the ranking impact of various weighting factors, update the dynamic dimensions of user profiles, and synchronize new questions into the embedding index, forming a cold-start question replenishment mechanism. This creates a closed-loop mechanism for the adaptive evolution of recommendation strategies.

[0085] For example, Figure 3 As shown, this embodiment also provides a candidate question recommendation method for an intelligent dialogue system, including: The original question input by the user is semantically encoded and the intent is classified and identified to obtain the semantic vector and intent label of the original question; Based on the semantic vector and intent label of the original question, search the pre-built candidate question database to obtain a preliminary screening set of candidate questions; Based on the semantic vector of the original question, semantic matching is performed with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score; Combined with the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weights of the candidate questions, the candidate questions in the initial screening set are prioritized to obtain a personalized candidate question list, so as to recommend the personalized candidate question list to the user.

[0086] In this embodiment, the semantic encoding and intent classification identification of the original question input by the user to obtain the semantic vector and intent label of the original question includes: Using a pre-built language understanding model to semantically encode the original question input by the user to obtain a semantic vector of the original question; wherein the language understanding model uses Sentence-BERT or Fine-tuned BERT; A multi-category classification algorithm is used to classify and identify the intent of the original question input by the user to obtain the intent label of the original question.

[0087] In this embodiment, the semantic vector and intent label of the original question are used to search a pre-built candidate question database to obtain a preliminary screening set of candidate questions, including: Based on the semantic vector and intent label of the original question, a vector and text hybrid retrieval method is used to retrieve candidate questions in the pre-built candidate question database to obtain a preliminary screening set of candidate questions; among them, the vector and text hybrid retrieval is implemented based on Elasticsearch and Faiss.

[0088] In this embodiment, the process of constructing the candidate question database includes: Collect historical questions, including internal enterprise FAQs, business FAQs, and operational process Q&A; A pre-trained large language model is used to vectorize historical question texts. A vector index library with a semantic index structure is built based on a vector similarity search library to support obtaining answers based on candidate question queries. Among them, the vector index in the candidate question database is updated through an offline scheduled task.

[0089] In this embodiment, the semantic vector based on the original question is semantically matched with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score, including: The cosine similarity and Euclidean distance correlation algorithms are used to semantically match the semantic vector of the original question with the candidate questions in the initial screening set of candidate questions to obtain the semantic similarity score.

[0090] In this embodiment, before prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, the following steps are also included: Obtain user portrait matching based on dynamic user tags of user portraits, including: Collect users' static and dynamic data; Use the Redis cache layer in conjunction with the Kafka message queue to update user profiles in real time based on static and dynamic data; Based on user behavior data, combined with time-decay weights and department-tag mapping tables, the user's candidate question preferences are calculated to generate dynamic user tags that are updated in real time. According to dynamic user tags, the user profile matching degree of candidate questions is obtained.

[0091] In this embodiment, the candidate questions in the initial screening set are prioritized by combining the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, including: Collect the semantic similarity score, user profile matching degree, question heat factor and question type adjustment factor of the candidate question; wherein the question heat factor is obtained based on the recent click frequency of the candidate question; the question type adjustment factor is a temperature coefficient set based on the question type of the candidate question; Based on the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weight of the candidate questions, the comprehensive score corresponding to each candidate question in the initial screening set of candidate questions is solved. The specific formula is as follows:

[0092] Where Softmax() represents the scoring function; W1, W2, W3, and W4 represent the semantic similarity score, user profile matching, question heat factor, and question type adjustment factor, respectively; α, β, and γ are adjustable hyperparameters representing different weight coefficients; T(W4) represents the temperature adjustment function for the question type adjustment factor; Based on the comprehensive scores corresponding to each candidate question, the candidate questions are prioritized from high to low scores to obtain a personalized candidate question list.

[0093] In this embodiment, after prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, the following steps are further included: Collect user feedback on the personalized candidate question list and regularly evaluate the recommendation effect based on the feedback; Dynamically adjust the weights of semantic similarity score, user profile matching, question heat factor and question type adjustment factor based on the recommendation effect.

[0094] like Figure 4 As shown, this embodiment also provides a candidate question recommendation system for an intelligent dialogue system, including: a semantic recognition module for semantically encoding and intent classification of the original question input by the user to obtain a semantic vector and intent label of the original question; a retrieval module for searching in a pre-built candidate question database based on the semantic vector and intent label of the original question to obtain a preliminary screening set of candidate questions; a semantic matching module for semantically matching the candidate questions in the preliminary screening set of candidate questions based on the semantic vector of the original question to obtain a semantic similarity score; a question recommendation module for prioritizing the candidate questions in the preliminary screening set of candidate questions based on the semantic similarity score of the candidate question, the user portrait matching degree, the question heat factor and the question type adjustment factor and the corresponding weights to obtain a personalized candidate question list to recommend the personalized candidate question list to the user.

[0095] The present invention also provides an electronic device, comprising: a memory for storing a computer program; and a processor for implementing the steps of the candidate question recommendation method for an intelligent dialogue system when executing the computer program.

[0096] The present invention also provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of the candidate question recommendation method for an intelligent dialogue system.

[0097] When the processor executes the computer program, it implements the above-mentioned steps for recommending candidate questions for the intelligent dialogue system, for example: semantically encoding and classifying the original question input by the user to obtain the semantic vector and intent label of the original question; searching in a pre-built candidate question database based on the semantic vector and intent label of the original question to obtain a preliminary screening set of candidate questions; semantically matching the candidate questions in the preliminary screening set of candidate questions based on the semantic vector of the original question to obtain a semantic similarity score; prioritizing the candidate questions in the preliminary screening set of candidate questions based on the semantic similarity scores, user portrait matching, question heat factors, question type adjustment factors and corresponding weights of the candidate questions to obtain a personalized candidate question list, so as to recommend the personalized candidate question list to the user.

[0098] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to implement the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing preset functions, and the instruction segments are used to describe the execution process of the computer program in the candidate question recommendation device for an intelligent dialogue system. For example, the computer program can be divided into a semantic recognition module, a retrieval module, a semantic matching module and a question recommendation module; the specific functions of each module are as follows: a semantic recognition module, which is used to semantically encode and classify the original question input by the user to obtain the semantic vector and intent label of the original question; a retrieval module, which is used to search in a pre-built candidate question database based on the semantic vector and intent label of the original question to obtain a preliminary screening set of candidate questions; a semantic matching module, which is used to semantically match the candidate questions in the preliminary screening set of candidate questions based on the semantic vector of the original question to obtain a semantic similarity score; a question recommendation module, which is used to prioritize the candidate questions in the preliminary screening set of candidate questions based on the semantic similarity score, user portrait matching, question heat factor and question type adjustment factor of the candidate questions and the corresponding weights to obtain a personalized candidate question list to recommend the personalized candidate question list to the user.

[0099] The candidate question recommendation device for an intelligent dialogue system can be a computing device such as a desktop computer, laptop, PDA, or cloud server. The candidate question recommendation device for an intelligent dialogue system may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the above examples of candidate question recommendation devices for an intelligent dialogue system are not intended to limit the entire candidate question recommendation device for an intelligent dialogue system. The candidate question recommendation device for an intelligent dialogue system may include more components than those described above, or a combination of certain components, or different components. For example, the candidate question recommendation device for an intelligent dialogue system may also include input / output devices, network access devices, buses, and the like.

[0100] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The processor serves as the control center for candidate question recommendation for the intelligent dialogue system, connecting various components of the candidate question recommendation device for the intelligent dialogue system using various interfaces and circuits.

[0101] The memory can be used to store the computer program and / or module, and the processor implements various functions of the candidate question recommendation device for the intelligent dialogue system by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory.

[0102] The memory may primarily include a program storage area and a data storage area. The program storage area may store an operating system and at least one application required for a function (such as sound playback or image playback); the data storage area may store data generated based on the use of the mobile phone (such as audio data and a phone book). Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0103] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the candidate question recommendation method for an intelligent dialogue system.

[0104] If the module / unit integrated into the candidate question recommendation system for an intelligent dialogue system is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0105] Based on this understanding, the present invention implements all or part of the process of the candidate question recommendation method for an intelligent dialogue system by using a computer program to instruct related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the candidate question recommendation method for an intelligent dialogue system. The computer program includes computer program code, which can be in source code form, object code form, executable file, or a pre-set intermediate form.

[0106] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0107] It should be noted that the content contained in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable storage media do not include electrical carrier signals and telecommunication signals.

[0108] In summary, the present invention provides a candidate question recommendation method and related device for an intelligent dialogue system, which has the following advantages: First, a multi-factor weighted fusion recommendation model was built. This invention provides a recommendation model based on multi-factor weight configuration. It comprehensively considers multiple factors such as semantic similarity, user profile fit, and question popularity. It uses a dynamically adjusted temperature coefficient to balance the accuracy and diversity of recommended questions, solving the problem that traditional single-dimensional models cannot simultaneously take into account general recommendations and personalized needs.

[0109] Second, a dynamic user profile and real-time feedback mechanism is introduced. This invention introduces dynamic user profile calculation and real-time feedback mechanisms. By collecting and updating users' static and dynamic behavior data (such as question logs and click behavior) in real time, and using technologies such as Redis and Kafka, low-latency data updates and real-time feature calculation are achieved.

[0110] Third, we optimize the diversity of recommended questions based on dynamic temperature coefficient adjustment. This invention provides a dynamic temperature coefficient adjustment mechanism that flexibly adjusts the diversity and accuracy of recommended questions based on the different question types (such as general questions, business questions, and indicator queries) of different enterprise users.

[0111] Fourth, it achieves high versatility and low cost. This invention does not rely on a single large-scale model or specific technology stack. It is highly versatile and can be flexibly applied in different business scenarios. Its system architecture design is highly scalable, significantly reducing resource consumption and computing costs of existing processes.

[0112] Fifth, a closed-loop optimization and question popularity update mechanism. By establishing a comprehensive performance evaluation mechanism, this invention continuously optimizes the quality of recommended questions based on user feedback and behavior (such as click-through rate and question resolution rate), and regularly updates question popularity and recommendation rankings. This enables the system to self-evolve, reduces manual rule maintenance costs, and addresses the pain point of traditional methods that require rule reconstruction for scenario migration.

[0113] The above embodiment is only one of the implementation methods that can realize the technical solution of the present invention. The scope of protection claimed by the present invention is not limited only to this embodiment, but also includes changes, replacements and other implementation methods that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A candidate question recommendation method for an intelligent dialogue system, characterized in that: include: The original question input by the user is semantically encoded and the intent is classified and identified to obtain the semantic vector and intent label of the original question; Based on the semantic vector and intent label of the original question, search the pre-built candidate question database to obtain a preliminary screening set of candidate questions; Based on the semantic vector of the original question, semantic matching is performed with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score; Combined with the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weights of the candidate questions, the candidate questions in the initial screening set are prioritized to obtain a personalized candidate question list, so as to recommend the personalized candidate question list to the user.

2. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: The semantic encoding and intent classification of the original question input by the user to obtain the semantic vector and intent label of the original question includes: Use the pre-built language understanding model to semantically encode the original question input by the user to obtain the semantic vector of the original question; A multi-category classification algorithm is used to classify and identify the intent of the original question input by the user to obtain the intent label of the original question.

3. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: According to the semantic vector and intent label of the original question, a search is performed in the pre-built candidate question database to obtain a preliminary screening set of candidate questions, including: According to the semantic vector and intent label of the original question, a vector and text hybrid retrieval method is used to retrieve the candidate questions in the pre-built candidate question database to obtain a preliminary screening set of candidate questions; among them, the vector and text hybrid retrieval is implemented based on text retrieval and vector retrieval.

4. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: The process of constructing the candidate question database includes: Collect historical questions, including internal frequently asked questions documents, business frequently asked questions, and operational process questions and answers; A pre-trained large language model is used to vectorize historical question texts, and a semantic index structure is constructed based on a vector similarity search library to obtain a candidate question database. Among them, the vector index in the candidate question database is updated through an offline scheduled task.

5. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: The semantic vector based on the original question is semantically matched with the candidate questions in the initial screening set of candidate questions to obtain a semantic similarity score, including: The cosine similarity and Euclidean distance correlation algorithms are used to semantically match the semantic vector of the original question with the candidate questions in the initial screening set of candidate questions to obtain the semantic similarity score.

6. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: Before prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, the following steps are also included: Obtain user portrait matching based on dynamic user tags of user portraits, including: Collect users' static and dynamic data; Using an in-memory database cache layer in conjunction with a distributed stream processing platform message queue, user profiles are updated in real time based on static and dynamic data. Based on user behavior data, combined with time-decay weights and department-tag mapping tables, the user's candidate question preferences are calculated to generate dynamic user tags that are updated in real time. According to dynamic user tags, the user profile matching degree of candidate questions is obtained.

7. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: Prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions includes: Collect the semantic similarity score, user profile matching degree, question heat factor and question type adjustment factor of the candidate question; wherein the question heat factor is obtained based on the recent click frequency of the candidate question; the question type adjustment factor is a temperature coefficient set based on the question type of the candidate question; Based on the semantic similarity score, user profile matching degree, question popularity factor, question type adjustment factor and corresponding weight of the candidate questions, the comprehensive score corresponding to each candidate question in the initial screening set of candidate questions is solved. The specific formula is as follows: Where Softmax() represents the scoring function; W1, W2, W3, and W4 represent the semantic similarity score, user profile matching, question heat factor, and question type adjustment factor, respectively; α, β, and γ are adjustable hyperparameters representing different weight coefficients; T(W4) represents the temperature adjustment function for the question type adjustment factor; Based on the comprehensive scores corresponding to each candidate question, the candidate questions are prioritized from high to low scores to obtain a personalized candidate question list.

8. The candidate question recommendation method for an intelligent dialogue system according to claim 1, characterized in that: After prioritizing the candidate questions in the initial screening set based on the semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights of the candidate questions, the following further comprises: Collect user feedback on the personalized candidate question list and regularly evaluate the recommendation effect based on the feedback; Dynamically adjust the weights of semantic similarity score, user profile matching, question heat factor and question type adjustment factor based on the recommendation effect.

9. A candidate question recommendation system for an intelligent dialogue system, characterized in that: include: The semantic recognition module is used to semantically encode and classify the original question input by the user to obtain the semantic vector and intent label of the original question; The retrieval module is used to search the pre-built candidate question database based on the semantic vector and intent label of the original question to obtain a preliminary screening set of candidate questions; The semantic matching module is used to perform semantic matching with candidate questions in the initial screening set of candidate questions based on the semantic vector of the original question to obtain a semantic similarity score; The question recommendation module is used to prioritize the candidate questions in the initial screening set based on their semantic similarity scores, user profile matching, question popularity factors, question type adjustment factors, and corresponding weights, and obtain a personalized candidate question list to recommend the personalized candidate question list to users.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the candidate question recommendation method for an intelligent dialogue system according to any one of claims 1 to 8 are implemented.

Images