Question recommendation method and device and computer equipment

Abstract

The invention relates to a question recommendation method and device and computer equipment. The method comprises the steps of obtaining a question input by a user, and identifying a question intention category of the question input by the user; wherein the user input question is a question input by the user account for the target event; determining a target entity matched with the key entity of the user question according to the knowledge graph entity set matched with the question intention category; processing the target entity and the question intention category through a pre-trained language model to obtain a user recommendation question; and recommending the user recommendation problem. By adopting the method, the question recommendation effect of the intelligent question-answering system can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a problem recommendation method, apparatus, and computer device. Background Technology

[0002] Intelligent question-answering systems are used to answer questions posed by humans in natural language and are now widely used.

[0003] However, the question recommendation function of traditional intelligent question answering systems often relies on rule-based or simple model parsing methods. When a user asks a question, if the user's wording is not precise enough or if industry jargon is used, the question answering recommendation function often has difficulty accurately understanding the overall semantics of the user's input question. This makes it impossible for the question answering recommendation function to accurately match relevant recommended questions, resulting in the user receiving irrelevant or inefficient answers.

[0004] Therefore, traditional intelligent question-answering systems suffer from poor question recommendation performance. Summary of the Invention

[0005] Therefore, it is necessary to provide a question recommendation method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can improve the question recommendation effect of intelligent answering systems in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a problem recommendation method, including:

[0007] Obtain user input questions and identify the question intent category of the user input questions; wherein, the user input questions are questions entered by the user account in response to a target event;

[0008] Based on the knowledge graph entity set that matches the question intent category, target entities that match the key entities of the user question are determined; wherein, the knowledge graph entity set includes at least two entities extracted from a preset target knowledge graph; the target knowledge graph is a knowledge graph associated with the target event; and the key entities of the user question are key entities in the user input question;

[0009] The target entity and the question intent category are processed by a pre-trained language model to obtain user-recommended questions;

[0010] The user recommendation question is recommended.

[0011] In one embodiment, the question intent category is at least one of the preset intent categories associated with the target event, and each preset intent category corresponds to a preset entity category. The method further includes:

[0012] Entity extraction is performed on the target knowledge graph to obtain at least two entities to be classified; each entity to be classified is associated with at least one question category in the target event;

[0013] According to at least one question category associated with each entity to be classified, each entity to be classified is classified to obtain the classified entity corresponding to each preset entity category;

[0014] Based on the categorized entities corresponding to each preset entity category, construct a knowledge graph entity set that matches each preset intent category.

[0015] In one embodiment, the step of constructing a knowledge graph entity set matching each preset intent category based on the categorized entities corresponding to each preset entity category includes:

[0016] The categorized entities corresponding to each preset entity category are obtained and vectorized to obtain the entity vector representation of the categorized entities corresponding to each preset entity category.

[0017] Based on the entity vector representations corresponding to each preset entity category, an entity vector library matching each preset intent category is constructed; the entity vector library is used to represent the corresponding knowledge graph entity set.

[0018] In one embodiment, the preset intent category includes at least one idea graph category and at least one sub-intent category extending from each of the idea graph categories, wherein the question intent category for identifying the user input question includes:

[0019] Among the at least one idea graph category, identify the target idea graph category that matches the user input question;

[0020] Among at least one sub-intent category extending from the target intent category, the target sub-intent category that matches the user input question is identified as the question intent category of the user input question.

[0021] In one embodiment, the step of processing the target entity and the question intent category through a pre-trained language model to obtain a user-recommended question includes:

[0022] The target entity and the question intent category are input into the pre-trained language model to generate a set of recommended questions; the set of recommended questions includes at least one candidate recommended question that contains the target entity and conforms to the question intent category;

[0023] At least one candidate recommendation question is selected from the set of recommendation questions as the user recommendation question.

[0024] In one embodiment, selecting at least one candidate recommendation question from the set of recommendation questions as the user recommendation question includes:

[0025] Obtain the recommendation value of each candidate recommendation question in the recommendation question set; the recommendation value is used to characterize the degree of matching between the corresponding candidate recommendation question and the user input question and the target entity, respectively.

[0026] At least one candidate recommendation question ranked first in the recommendation question set is selected as the user recommendation question.

[0027] In one embodiment, obtaining the recommendation value of each candidate recommendation question in the recommendation question set includes:

[0028] For any candidate recommendation question in the set of recommendation questions, the candidate recommendation question and the user input question are input into a language model. By calling the recommendation fit scoring function of the language model, a recommendation fit score for the candidate recommendation question is generated. The recommendation fit score is used to characterize the appropriateness of the corresponding candidate recommendation question in the current recommendation scenario.

[0029] If the recommendation consistency score of any candidate recommendation question is greater than or equal to a preset score threshold, the recommendation consistency score of any candidate recommendation question is adjusted according to the entity similarity corresponding to the target entity to obtain the recommendation value of any candidate recommendation question; wherein, the entity similarity is the similarity between the target entity and the key entity.

[0030] In one embodiment, the method further includes:

[0031] If the recommendation compliance score of any candidate recommendation question is less than the preset scoring threshold, then any candidate recommendation question is removed from the set of recommendation questions.

[0032] In one embodiment, adjusting the recommendation compliance score of any candidate recommendation question based on the entity similarity corresponding to the target entity to obtain the recommendation value of any candidate recommendation question includes:

[0033] Convert the entity similarity corresponding to the target entity into the corresponding entity similarity score;

[0034] The recommended value for any candidate recommendation question is obtained by taking a weighted average of the entity similarity score and the recommendation conformity score.

[0035] In one embodiment, determining the target entity matching the key entity of the user question based on the knowledge graph entity set that matches the question intent category includes:

[0036] Based on the key entities of the user's question, entity retrieval is performed in the knowledge graph entity set that matches the question intent category to obtain a retrieval entity set; the retrieval entity set includes a preset number of retrieval entities related to the key entities of the user's question.

[0037] The target entity is selected from the set of search entities based on the similarity between each search entity and the key entity of the user question; the similarity between the target entity and the key entity of the user question is greater than a preset similarity threshold.

[0038] Secondly, this application also provides a problem recommendation device, the device comprising:

[0039] The acquisition module is used to acquire user input questions and identify the question intent category of the user input questions; wherein, the user input questions are questions entered by the user account in response to a target event;

[0040] The determination module is used to determine target entities that match key entities in the user's question based on a knowledge graph entity set that matches the question intent category; wherein, the knowledge graph entity set includes at least two entities extracted from a preset target knowledge graph; the target knowledge graph is a knowledge graph associated with the target event; and the key entities in the user's question are key entities in the user's input question.

[0041] The output module is used to process the target entity and the question intent category through a pre-trained language model to obtain user-recommended questions;

[0042] The recommendation module is used to provide recommendations for the user's recommendation question.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-mentioned recommended method for solving the problem.

[0044] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described recommended method for solving the problem.

[0045] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described problem-recommended method.

[0046] The aforementioned question recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product acquire user input questions from a user account for a target event, identify the question intent category of the user input question, and determine target entities matching key entities in the user input question from a knowledge graph entity set that matches the question intent category. The knowledge graph entity set includes at least two entities extracted from a preset target knowledge graph, which is associated with the target event. Then, based on the target entities and the question intent category, a user recommendation question is output. Thus, by effectively identifying the question intent category of the user input question and recalling target entities matching key entities in the user input question from the knowledge graph entity set that matches the question intent category, and by utilizing the target entities and question intent category, a user recommendation question matching the current question scenario is effectively recommended to the user. Simultaneously, since the user recommendation question contains entities extracted from the target knowledge graph, the recommended questions output by the intelligent question answering system can be effectively answered through the target knowledge graph, avoiding irrelevant or inefficient answers for the user and effectively improving the question recommendation effect of the intelligent answering system. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a diagram illustrating the application environment of a problem recommendation method in one embodiment.

[0049] Figure 2 This is a flowchart illustrating a problem recommendation method in one embodiment;

[0050] Figure 3 This is a flowchart illustrating a problem recommendation method in another embodiment;

[0051] Figure 4 A system architecture diagram of a problem recommendation system in one embodiment;

[0052] Figure 5 A structural block diagram of a problem recommendation device in one embodiment;

[0053] Figure 6 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] In this application, the collection and processing of relevant data have strictly complied with the requirements of relevant national laws and regulations, obtained the informed consent or separate consent of the personal information subject, and carried out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0056] The problem recommendation method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with intelligent question-answering system 104 via a network. A data storage system can store the data that intelligent question-answering system 104 needs to process. The data storage system can be integrated into intelligent question-answering system 104 or placed in the cloud or on other network servers. Specifically, a user can input a user input question for a target event through terminal 102, and terminal 102 sends the user input question to intelligent question-answering system 104. Intelligent question-answering system 104 obtains the user input question and identifies the question intent category; the user input question is a question entered by the user account for the target event. Intelligent question-answering system 104 determines the target entity matching the key entity of the user question based on a knowledge graph entity set that matches the question intent category; the knowledge graph entity set includes at least two entities extracted from a preset target knowledge graph; the target knowledge graph is a knowledge graph associated with the target event; the key entity of the user question is the key entity in the user input question. Intelligent question-answering system 104 processes the target entity and question intent category through a pre-trained language model to obtain a user-recommended question; intelligent question-answering system 104 recommends a user-recommended question.

[0057] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle systems, and projection devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, and smart glasses. The intelligent question-answering system 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0058] In one exemplary embodiment, such as Figure 2 As shown, a problem recommendation method is provided, which is then applied to... Figure 1 The following explanation uses the intelligent question-answering system 104 as an example, including the following steps S202 to S206. Wherein:

[0059] Step S202: Obtain the user input question and identify the question intent category of the user input question.

[0060] Among them, user input questions are questions entered by the user account in response to the target event.

[0061] The target event can be any event that triggers a user's account to input a question into the intelligent question-answering system. Specifically, the target event can be a product malfunction event, an event that raises a question about a target object, etc. The target object can refer to a specific object or thing.

[0062] In practice, users can input questions about a target event through the user interaction front-end page of the intelligent question-answering system, allowing the system to retrieve these questions. Additionally, the intelligent question-answering system is equipped with an API interface, which allows users to input questions about a target event (such as a product malfunction event) through third-party systems, enabling the system to retrieve these questions as well.

[0063] Then, the intelligent question-answering system can identify the category of the user's question intent. Specifically, the intelligent question-answering system can call the intent recognition function in the language model to identify the category of the user's question intent.

[0064] In practical applications, the question intent category can be at least one of the preset intent categories associated with the target event. Taking a product malfunction event as an example, the preset intent categories can include at least malfunction-related question intents and parameter-related question intents. For malfunction-related question intents, the intelligent question-answering system can also identify its sub-intents, including root cause-related question intents, troubleshooting-related question intents, or multiple intents that simultaneously include both categories. For parameter-related question intents, the intelligent question-answering system can also identify its sub-intents, which can include model parameter query intents, model-specific parameter value query intents, or multiple intents that simultaneously include both categories.

[0065] Step S202: In the knowledge graph entity set that matches the question intent category, identify the target entity that matches the key entity of the user question.

[0066] The knowledge graph entity set includes at least two entities extracted from a pre-defined target knowledge graph. In practical applications, intelligent question-answering systems can obtain the aforementioned knowledge graph entity set by pre-extracting entities from the target knowledge graph and classifying the extracted entities. In practice, each pre-defined intent category corresponds to one knowledge graph entity set. It should be noted that the construction of the knowledge graph entity set will be further explained below.

[0067] The target knowledge graph is the knowledge graph associated with the target event. In practical applications, when the target event is a product failure event, the target knowledge graph can refer to the product failure knowledge graph. Specifically, the product failure knowledge graph can be stored using a graph database.

[0068] In this context, the key entities in a user's question can refer to the key entities in the user's input question. For example, if the user inputs the question "The wind turbine is malfunctioning, how do I maintain it?", then the key entities in the user's question can include, but are not limited to, "wind turbine," "malfunction," and "maintenance."

[0069] In practice, when the intelligent question answering system identifies the question intent category of the user's input question, it can determine the knowledge graph entity set that matches the question intent category of the user's input question from multiple preset knowledge graph entity sets.

[0070] Then, the intelligent question-answering system can identify key entities in the user's input question. Specifically, the intelligent question-answering system can use entity recognition algorithms to identify key entities in the user's input question. Specifically, the intelligent question-answering system can input entity recognition instructions into a large language model, enabling the large language model to identify the key entities in the user's input question. In practical applications, entity recognition instructions can be constructed using open-source large models and context-based learning prompts; the prompts include a task description for entity recognition, few-shot examples, and output format requirements, as shown below:

[0071] (1) Task Description:

[0072] """

[0073] You are an entity recognition expert. Please identify the entities in the user's query.

[0074] """

[0075] (2) Examples of few-shots are as follows:

[0076] """

[0077] Input: How to handle a fan malfunction?

[0078] Output: {"output": "Fan failure"}

[0079] """

[0080] (3) Output format requirements: Please output the results in JSON (a content format).

[0081] When an intelligent question-answering system identifies key entities in a user's input question, it can use these key entities to perform entity retrieval within a knowledge graph entity set that matches the question's intent category, thus retrieving target entities that match the key entities in the user's input question. In practical applications, the similarity between the target entity and the key entity in the user's input question meets a preset similarity condition.

[0082] In practical applications, intelligent question-answering systems can recall several search entities from a knowledge graph entity set that matches the question's intent category. Then, the system refines these search entities, and based on this refined ranking, determines the target entity that matches the key entity in the user's input question. It should be noted that the above process will be further explained below.

[0083] In practical applications, the entities mentioned above can all be represented using corresponding entity vectors (embeddings) to quickly identify the target entity that matches the key entity in the user's input question from the knowledge graph entity set that matches the question intent category.

[0084] Step S206: Process the target entity and question intent category through a pre-trained language model to obtain the user-recommended question.

[0085] In practice, the intelligent question-answering system can call the question generation function of the large language model, and output user-recommended questions by simulating real user questions, based on the target entity and the category of question intent.

[0086] The intelligent question-answering system can input question generation instructions into a large language model, enabling the model to output recommended questions based on the target entity and the question intent category. In practical applications, question generation instructions can be constructed using open-source large models and context-based learning prompts. The prompts include a task description for entity recognition, few-shot examples, and output format requirements, as detailed below:

[0087] (1) Task Description:

[0088] """

[0089] You are an expert who generates recommendation questions based on intent and entities. The questions must be based on intent and contain the entities themselves. You cannot fabricate, create, or modify the entities.

[0090] """

[0091] (2) Examples of few-shots are as follows:

[0092] """

[0093] Intention: Troubleshooting / Repair Issues

[0094] Entity: ['Fan malfunction']

[0095] Output: {"output": "How to solve the fan malfunction?"}

[0096] """

[0097] (3) Output format requirements: Please output the results in JSON (a content format).

[0098] Step S208: Recommend user-recommended question.

[0099] In practice, once the intelligent question-answering system identifies a user-recommended question, it can recommend that question to the user's account. In real-world applications, the system can send the recommended question to the terminal, allowing it to be displayed on the user-interactive front-end page (e.g., a dialog interface) where the system interacts with the user. The user can then instruct the system to provide an answer to the recommended question.

[0100] In the above question recommendation method, user input questions for a target event are obtained from the user's account. The question intent category of the user input question is identified, and target entities matching key entities in the user input question are determined from a knowledge graph entity set that matches the question intent category. This knowledge graph entity set includes at least two entities extracted from a preset target knowledge graph, which is associated with the target event. Then, based on the target entities and the question intent category, a user recommendation question is output. Thus, by effectively identifying the question intent category of the user input question and recalling target entities matching key entities in the user input question from the knowledge graph entity set that matches the question intent category, and by utilizing the target entities and question intent category, a user recommendation question matching the current question scenario is effectively recommended to the user. Furthermore, since the user recommendation question contains entities extracted from the target knowledge graph, the recommended questions output by the intelligent question answering system can be effectively answered through the target knowledge graph, avoiding irrelevant or inefficient answers for the user and effectively improving the question recommendation effect of the intelligent answering system.

[0101] In an exemplary embodiment, the question intent category is at least one of the preset intent categories associated with the target event, and each preset intent category corresponds to a preset entity category. The method further includes: extracting entities from the target knowledge graph to obtain at least two entities to be classified; at least one question category associated with each entity to be classified in the target event; classifying each entity to be classified according to the at least one question category associated with each entity to be classified to obtain the classified entities corresponding to each preset entity category; and constructing a knowledge graph entity set matching each preset intent category based on the classified entities corresponding to each preset entity category.

[0102] As mentioned above, the question intent category is at least one of the preset intent categories associated with the target event. In practical applications, each preset intent category corresponds to a preset entity category. For example, in the case of a product malfunction event, a malfunction-type question intent corresponds to a malfunction-type entity, and a parameter-type question intent corresponds to a parameter-type entity.

[0103] In its implementation, during the construction of knowledge graph entity sets matching each preset intent category, the intelligent question-answering system can extract entities from the target knowledge graph to obtain at least two entities to be classified; each entity to be classified is associated with at least one question category in the target event. For example, entity A to be classified can be associated with both root cause of faults and fault elimination issues.

[0104] Then, the intelligent question-answering system can classify each entity according to at least one question category associated with it, obtaining the classified entities corresponding to each preset entity category. For example, the intelligent question-answering system can extract fault-type entities and parameter-type entities from the product fault graph using Cypher statements (a query language for graph databases). Fault-type entities can be classified into fault root cause entities, fault elimination entities, and entities that simultaneously contain both fault root cause and fault elimination entities (i.e., the intersection of fault root cause entities and fault elimination entities, and mixed entities of fault root cause and fault elimination). Parameter-type entities can be classified into product model entities and product parameter entities.

[0105] Intelligent question-answering systems can construct knowledge graph entity sets matching predefined intent categories based on the categorized entities corresponding to each predefined entity category. For example, an intelligent question-answering system can construct a knowledge graph entity set matching the intent of a fault root cause type question based on the fault root cause entity; an intelligent question-answering system can construct a knowledge graph entity set matching the intent of a fault elimination type question based on the fault elimination entity; and an intelligent question-answering system can construct a knowledge graph entity set matching the intent of a mixed fault root cause and fault elimination type question based on a mixed fault root cause and fault elimination entity.

[0106] The technical solution of this embodiment extracts entities from the target knowledge graph to obtain at least two entities to be classified; each entity to be classified is associated with at least one question category in the target event; each entity to be classified is classified according to the at least one question category associated with it to obtain the classified entities corresponding to each preset entity category; based on the classified entities corresponding to each preset entity category, a knowledge graph entity set matching each preset intent category is constructed; thus, the intelligent question answering system can quickly determine the knowledge graph entity set matching the question intent category of the user input question after identifying the question intent category in the user input question, so that the final recalled target entities are effectively matched with the question intent of the user input question, thereby improving the question recommendation effect of the intelligent question answering system.

[0107] In an exemplary embodiment, a knowledge graph entity set matching each preset intent category is constructed based on the categorized entities corresponding to each preset entity category, including: vectorizing the categorized entities corresponding to each preset entity category to obtain entity vector representations of the categorized entities corresponding to each preset entity category; and constructing an entity vector library matching each preset intent category based on the entity vector representations corresponding to each preset entity category.

[0108] Among them, the entity vector library is used to represent the corresponding knowledge graph entity set.

[0109] In its implementation, during the process of constructing a knowledge graph entity set matching each preset intent category based on the categorized entities corresponding to each preset entity category, the intelligent question answering system can vectorize the categorized entities corresponding to each preset entity category to obtain entity vector representations of the categorized entities corresponding to each preset entity category. For example, for fault-related questions, the intelligent question answering system can vectorize three types of entities—fault root cause entity, fault elimination entity, and entity containing both fault root cause and elimination—using an embedding model to obtain the embedding of the fault root cause entity, the embedding of the fault elimination entity, and the embedding of the entity containing both fault root cause and elimination. For parameter-related questions, the intelligent question answering system can vectorize two types of entities—product model entity and product parameter entity—to obtain the embedding of the product model entity and the embedding of the product parameter entity.

[0110] Then, the intelligent question-answering system can construct an entity vector library matching each preset intent category based on the entity vector representation corresponding to each preset entity category, in order to represent the corresponding knowledge graph entity set. Continuing the previous example, the entity vector library may include a fault root cause entity vector library, a fault elimination entity vector library, a hybrid entity vector library of fault root causes and elimination, a product model entity vector library, and a product parameter entity vector library.

[0111] The technical solution of this embodiment, in the process of constructing a knowledge graph entity set matching each preset intent category based on the classified entities corresponding to each preset entity category, performs vectorization processing on the classified entities corresponding to each preset entity category to construct an entity vector library matching each preset intent category, which is used to represent the corresponding knowledge graph entity set. While saving computer storage space, it can also effectively improve the retrieval speed of the intelligent question answering system in the knowledge graph entity set, and realize the rapid determination of the target entity matching the key entity in the user input question in the knowledge graph entity set matching the question intent category.

[0112] In an exemplary embodiment, the preset intent category includes at least one idea graph category and at least one sub-intent category extending from each idea graph category. Identifying the question intent category of the user input question includes: identifying a target idea graph category that matches the user input question in the at least one idea graph category; and identifying a target sub-intent category that matches the user input question in the at least one sub-intent category extending from the target idea graph category, as the question intent category of the user input question.

[0113] In its specific implementation, as described above, the preset intent categories include at least one idea graph category and at least one sub-intent category derived from each idea graph category. During the process of identifying the question intent category of the user's input question, the intelligent question answering system identifies a target idea graph category that matches the user's input question within the at least one idea graph category. Then, the intelligent question answering system identifies a target sub-intent category that matches the user's input question within the at least one sub-intent category derived from the target idea graph category, and uses this sub-intent category as the question intent category of the user's input question.

[0114] For example, an intelligent question-answering system can input a purpose graph recognition instruction into a large language model to instruct the model to identify whether the user's input question belongs to a fault-based or parametric question intent. The purpose graph recognition instruction can be constructed using an open-source large model and a context-based learning prompt. The prompt includes a description of the overall intent recognition task, few-shot examples, and output format requirements, as shown below:

[0115] (1) Task Description:

[0116] """

[0117] As a total intent recognition expert, please determine the category of the user's query: fault-based or parameter-based. The total intent category codes and explanations are as follows:

[0118] Intent code: a, Description: Fault-related problem;

[0119] Intent code: b, Description: Parameter-related problem;

[0120] """

[0121] (2) Examples of few-shots are as follows:

[0122] """

[0123] Example 1:

[0124] Input: How to handle a fan malfunction?

[0125] Output: {"overall-intention": "a"}

[0126] Example 2:

[0127] Input: What is the maximum voltage of product A?

[0128] Output: {"overall-intention": "b"}

[0129] """

[0130] (3) Output format requirements: Please output the results in JSON format.

[0131] Then, the intelligent question-answering system can input sub-intent recognition instructions into the large language model to instruct it to identify which sub-intent of the fault-based question intent or the parameter-based question intent the user's input question belongs to. The sub-intent recognition instructions can be constructed using an open-source large model and a context-based learning prompt. The prompt includes a task description for sub-intent recognition, few-shot examples, and output format requirements, as shown below:

[0132] (1) Task Description:

[0133] """

[0134] You are a sub-intent identification expert. Based on the user's query and overall intent, please determine which sub-category of the larger category their intent belongs to. Intent category codes and explanations are as follows:

[0135] Intent code: a1, Description: Root cause problem;

[0136] Intent code: a2, Description: Troubleshooting / remediation issue;

[0137] Intent code: a3, Description: A multi-intent problem that simultaneously includes root cause and remediation classes;

[0138] Intent code: b1, Description: Intent to query model parameters;

[0139] Intent code: b2, Description: Intent to query model-specific parameter values;

[0140] Intent code: b3, Description: A multi-intent question that simultaneously includes model parameter query and parameter value query;

[0141] """

[0142] (2) Examples of few-shots are as follows:

[0143] """

[0144] Example 1:

[0145] Input: How to handle a fan malfunction?

[0146] Output: {"sub-intention": "a2"}

[0147] Example 2:

[0148] Input: What is the maximum voltage of product A?

[0149] Output: {"sub-intention": "b2"}

[0150] """

[0151] (3) Output format requirements: Please output the results in JSON format.

[0152] The technical solution of this embodiment, in the process of identifying the question intent category of a user input question, identifies a target idea graph category that matches the user input question in at least one idea graph category; and identifies a target sub-intention class that matches the user input question in at least one sub-intention category extended from the target idea graph category, as the question intent category of the user input question. This can decompose the complex intent recognition problem into a more granular recognition logic. This hierarchical intent recognition can effectively reduce the complexity of the question intent recognition task for user input questions and significantly improve the efficiency of identifying the question intent category of user input questions.

[0153] In an exemplary embodiment, outputting a user-recommended question based on the target entity and the question intent category includes: inputting the target entity and the question intent category into a language model, generating a set of recommended questions by calling the question generation function of the language model; and selecting at least one candidate recommended question from the set of recommended questions as the user-recommended question.

[0154] The recommended question set includes at least one candidate recommended question that contains the target entity and conforms to the question intent category.

[0155] In its specific implementation, when the intelligent question answering system outputs recommended questions to users based on the target entity and the question intent category, the intelligent question answering system can input the target entity and the question intent category into a large language model, and by calling the question generation function of the language model, generate at least one candidate recommended question that contains the target entity and conforms to the question intent category, so as to obtain a set of recommended questions.

[0156] Then, the intelligent question-answering system can select at least one candidate recommendation question from the recommended question set as the user's recommendation question according to preset question filtering conditions. Specifically, the intelligent question-answering system can detect whether the candidate recommendation questions generated by the large language model are related to the user's input question and whether they are consistent with the current recommendation scenario, and further determine whether to use the candidate recommendation question as the user's recommendation question to recommend to the user's account.

[0157] The technical solution of this embodiment, by inputting the target entity and the question intent category into the language model, and by calling the question generation function of the language model, generates a set of recommended questions. This can effectively utilize the ability of the language model to generate natural language questions, generating multiple natural language questions containing the target entity and conforming to the question intent category. This makes the questions recommended to the user by the intelligent question answering system understandable to the user, avoiding the situation where irrelevant recommended questions are given to the user, thus preventing the user from receiving irrelevant or inefficient answers.

[0158] In an exemplary embodiment, selecting at least one candidate recommendation question from the set of recommendation questions as the user recommendation question includes: obtaining the recommendation value of each candidate recommendation question in the set of recommendation questions; and selecting at least one candidate recommendation question with the highest recommendation value from the set of recommendation questions as the user recommendation question.

[0159] The recommendation value is used to characterize the degree of matching between the corresponding candidate recommendation question and the user input question and the target entity.

[0160] In its implementation, the intelligent question-answering system selects at least one candidate recommendation question from the set of recommended questions as the user's recommended question. It then obtains the matching degree between each candidate recommendation question and the user's input question and the target entity, quantifies these matching degrees, and maps them to corresponding recommendation values ​​to obtain the recommendation value for each candidate recommendation question. Next, the intelligent question-answering system selects at least one candidate recommendation question from the set with the highest recommendation value as the user's recommended question. Specifically, the intelligent question-answering system can sort the candidate recommendation questions in descending order according to their recommendation values. Then, the system can select the top N candidate recommendation questions as the user's recommended question. In practical applications, N is an integer greater than or equal to 1; optionally, N can be equal to 3.

[0161] The technical solution of this embodiment involves selecting at least one candidate recommendation question from the set of recommendation questions as the user recommendation question, obtaining the recommendation value of each candidate recommendation question in the set of recommendation questions, and selecting at least one candidate recommendation question with the highest recommendation value in the set of recommendation questions as the user recommendation question. This achieves fine ranking of multiple candidate recommendation questions generated by a large model, and recommends user recommendation questions that are as close as possible to the current scenario and meet the user's preferences.

[0162] In an exemplary embodiment, obtaining the recommendation value of each candidate recommendation question in the recommendation question set includes: for any candidate recommendation question in the recommendation question set, inputting the candidate recommendation question and the user input question into a language model, generating a recommendation consistency score for any candidate recommendation question by calling the recommendation consistency scoring function of the language model; if the recommendation consistency score of any candidate recommendation question is greater than or equal to a preset score threshold, adjusting the recommendation consistency score of any candidate recommendation question according to the entity similarity corresponding to the target entity to obtain the recommendation value of any candidate recommendation question.

[0163] Among them, entity similarity is the similarity between the target entity and the key entity.

[0164] Optionally, if the recommendation compliance score of any candidate recommendation question is less than a preset scoring threshold, then any candidate recommendation question is removed from the recommendation question set.

[0165] The recommendation fit score characterizes the suitability of any candidate recommendation question in the current recommendation scenario. In other words, the recommendation fit score can also characterize the user account's preference for any candidate recommendation question.

[0166] In its implementation, the intelligent question-answering system obtains the recommendation values ​​of each candidate question in the recommendation question set through the following steps: The system inputs any candidate question and the user-input question into a language model, and generates a recommendation fit score for each candidate question by calling the language model's recommendation fit scoring function. In practical applications, the intelligent question-answering system can input a question verification and scoring instruction into a large language model, instructing the model to verify and score any candidate question to obtain its recommendation fit score. This question verification and scoring instruction can be constructed using an open-source large model and a context-based learning prompt. The prompt includes a description of the scoring task, few-shot examples, and output format requirements, as shown below:

[0167] (1) Task Description:

[0168] """

[0169] You are a scoring expert. Please validate and score the given user input question and the generated candidate recommendation questions. The score range is 0-100.

[0170] """

[0171] (2) Examples of few-shots are as follows:

[0172] """

[0173] User input question: How to handle the abnormal overheating of a certain product's fan?

[0174] The generated candidate recommendation question is: How to solve the fan malfunction of a certain product?

[0175] Output: {"answer": "100 points"}

[0176] """

[0177] (3) Output format requirements: Please output the results in JSON format.

[0178] Then, the intelligent question answering system can determine whether the recommendation conformity score of the candidate recommendation question is greater than or equal to the preset score threshold m, where the preset score threshold m can be set to 80 points, so as to define whether the candidate recommendation question conforms to the current scenario and whether it meets user preferences, etc. If the recommendation conformity score of the candidate recommendation question is less than the preset score threshold, the intelligent question answering system can determine that the generated candidate recommendation question fails the verification, and the intelligent question answering system can remove the candidate recommendation question from the recommendation question set.

[0179] If the recommendation consistency score of a candidate recommendation question is greater than or equal to a preset score threshold, the intelligent question answering system can determine that the generated candidate recommendation question has passed the verification. At this time, the intelligent question answering system can obtain the similarity between the target entity in the candidate recommendation question and the key entity in the user input question, and obtain the entity similarity corresponding to the target entity. The intelligent question answering system can use the entity similarity corresponding to the target entity to adjust the recommendation consistency score of any candidate recommendation question and obtain the recommendation value of any candidate recommendation question.

[0180] In this embodiment, for any candidate recommendation question in the recommendation question set, the candidate recommendation question and the user input question are input into a language model. The recommendation consistency score of the candidate recommendation question is generated by calling the recommendation consistency scoring function of the language model. If the recommendation consistency score of any candidate recommendation question is greater than or equal to a preset score threshold, the recommendation consistency score of the candidate recommendation question is adjusted according to the entity similarity corresponding to the target entity to obtain the recommendation value of the candidate recommendation question. A large model can be used to test and score whether the generated candidate recommendation question meets the user's preferences, and the test and score results of the candidate recommendation question are used as the basis for selecting user recommendation questions from the candidate recommendation questions.

[0181] In an exemplary embodiment, the recommendation compliance score of any candidate recommendation question is adjusted based on the entity similarity corresponding to the target entity to obtain the recommendation value of any candidate recommendation question. This includes: converting the entity similarity corresponding to the target entity into a corresponding entity similarity score; and performing a weighted average of the entity similarity score and the recommendation compliance score to obtain the recommendation value of any candidate recommendation question.

[0182] In its implementation, the intelligent question-answering system adjusts the recommendation accuracy score of any candidate recommendation question based on the entity similarity to the target entity to obtain the recommendation value for that candidate recommendation question. Specifically, this process includes: converting the entity similarity to the target entity into a corresponding entity similarity score. For example, if the entity similarity to the target entity is 0.9, the corresponding entity similarity score is 90. Then, the intelligent question-answering system can perform a weighted average of the entity similarity score and the recommendation accuracy score to obtain a comprehensive score for the candidate recommendation question, which serves as the recommendation value for that candidate recommendation question.

[0183] In practical applications, the overall score for a problem can be expressed as:

[0184]

[0185] in, The score weight corresponding to the recommendation conformity score is set to a default value of 0.5; This is the score weight corresponding to the entity similarity score, with a default value of 0.5; The recommendation fit score is used to evaluate the candidate recommendation question. The entity similarity score is given to the target entity in the candidate recommendation problem.

[0186] The technical solution of this embodiment converts the entity similarity corresponding to the target entity into a corresponding entity similarity score; and performs a weighted average of the entity similarity score and the recommendation conformity score to obtain the recommendation value of any candidate recommendation question. In this way, the matching degree between the candidate recommendation question and the user input question and the target entity can be combined simultaneously to effectively quantify the suitability of the candidate recommendation question in the current recommendation scenario or the user account's preference for the candidate recommendation question.

[0187] In an exemplary embodiment, determining the target entity that matches the key entity in the user input question within the knowledge graph entity set that matches the question intent category includes: performing entity retrieval in the knowledge graph entity set that matches the question intent category based on the key entity to obtain a retrieval entity set; the retrieval entity set includes a preset number of retrieval entities related to the key entity; and selecting the target entity from the retrieval entity set based on the similarity between each retrieval entity and the key entity.

[0188] Among them, the similarity between the target entity and the key entity is greater than the preset similarity threshold.

[0189] In its specific implementation, the question-answering recommendation system determines the target entities matching the key entities within the knowledge graph entity set that matches the question intent category. Specifically, this includes: the system performs entity retrieval or entity recall within the knowledge graph entity set that matches the question intent category, based on the key entities, to obtain a preset number of retrieval entities related to the key entities, i.e., a retrieval entity set. For example, if the question-answering recommendation system identifies the user's input question as belonging to the root cause category or the fault elimination category, the system can retrieve M retrieval entities matching the key entities in the user's input question from the root cause entity vector library and the fault elimination entity vector library; where M can be a positive integer greater than 0; optionally, M can be set to 10.

[0190] Then, the question-answering recommendation system can select target entities from the set of search entities based on the similarity between each search entity and the key entity. For example, it can calculate the cosine similarity between each search entity and the key entity, with the cosine similarity value ranging from 0 to 1. The system can then identify search entities with a cosine similarity greater than a similarity threshold T as target entities. The similarity threshold T is set to 0.8. Specifically, the system can also sort the search entities in descending order based on their cosine similarity, and then identify one or more of the top-ranked search entities as target entities.

[0191] The technical solution of this embodiment uses key entities to perform entity retrieval in a knowledge graph entity set that matches the question intent category, obtaining a preset number of retrieval entities related to the key entities. This can effectively achieve rapid screening of potentially relevant entities in the entity recall stage, transferring the computational burden from the fine-ranking stage, ensuring system response speed, and selecting target entities with similarity greater than a preset similarity threshold from the retrieval entity set based on the similarity between each retrieval entity and the key entity. This effectively reduces the computational cost in the entity fine-ranking stage and improves the efficiency of target entity acquisition.

[0192] In another embodiment, such as Figure 3 As shown, a problem recommendation method is provided, which is then applied to... Figure 1 Taking the intelligent question-answering system 104 in the example, the explanation includes the following steps:

[0193] Step S302: Obtain the user input question; the user input question is the question entered by the user account in response to the target event.

[0194] Step S304: Identify the target idea graph category that matches the user input question in at least one idea graph category of the preset intent categories.

[0195] Step S306: In at least one sub-intent category extending from the target intention graph category, identify the target sub-intent category that matches the user input question as the question intent category of the user input question.

[0196] Step S308: Based on the key entities of the question input by the user, perform entity retrieval in the knowledge graph entity set that matches the question intent category to obtain a retrieval entity set; the retrieval entity set includes a preset number of retrieval entities related to the key entities.

[0197] Step S310: Select the target entity from the search entity set based on the similarity between each search entity and the key entity; the similarity between the target entity and the key entity is greater than a preset similarity threshold.

[0198] Step S312: Input the target entity and question intent category into the language model, and generate a recommended question set by calling the question generation function of the language model; the recommended question set includes at least one candidate recommended question that contains the target entity and conforms to the question intent category.

[0199] Step S314: Obtain the recommendation value of each candidate recommendation question in the recommendation question set; the recommendation value is used to characterize the degree of matching between the corresponding candidate recommendation question and the user input question and the target entity.

[0200] Step S316: Select at least one candidate recommendation question with the highest recommendation value from the set of recommendation questions as the user recommendation question.

[0201] Step S316: Recommend user-recommended question.

[0202] It should be noted that the specific limitations of the above steps can be found in the specific limitations of a problem recommendation method described above.

[0203] For the convenience of those skilled in the art, Figure 4 An exemplary system architecture diagram for a question recommendation system is also provided; wherein, the question recommendation system includes a user question input module, an overall intent recognition module, a sub-intent recognition module, an entity recognition and vectorization module, a vector library retrieval module, a similarity calculation module, a recommended question generation module, a recommended question verification module, a recommended question score calculation and ranking module (recommended question filtering module), a recommended question output module, a knowledge graph storage module, a knowledge graph retrieval module, and an entity vector library module.

[0204] Specifically, the knowledge graph storage module is used to store the knowledge graph associated with the target event using a graph database; the knowledge graph retrieval module is used to extract entities from the knowledge graph associated with the target event using the Cypher statement to obtain at least two entities to be classified, classify each entity to be classified, and obtain the classified entities corresponding to each preset entity category; the entity vector library module is used to vectorize the classified entities corresponding to each preset entity category and construct an entity vector library that matches each preset intent category.

[0205] The user question input module is used to obtain user input questions; the overall intent recognition module is used to identify the target idea graph category that matches the user input question in at least one idea graph category of the preset intent categories; the sub-intent recognition module is used to identify the target sub-intent class that matches the user input question in at least one sub-intent category extended from the target idea graph category, as the question intent category of the user input question.

[0206] The entity recognition and vectorization module is used to identify key entities in the user input question; the vector library retrieval module is used to perform entity retrieval in the knowledge graph entity set that matches the question intent category based on the key entities in the user input question, and obtain the retrieval entity set; the retrieval entity set includes a preset number of retrieval entities related to the key entities; the similarity calculation module is used to select the target entity in the retrieval entity set based on the similarity between each retrieval entity and the key entity; the similarity between the target entity and the key entity is greater than a preset similarity threshold;

[0207] The recommended question generation module is used to input the target entity and question intent category into the language model, and generate a set of recommended questions by calling the question generation function of the language model.

[0208] The recommendation question verification module is used to input any candidate recommendation question and the user input question into the language model, and generate a recommendation consistency score for any candidate recommendation question by calling the recommendation consistency scoring function of the language model; the recommendation question score calculation and ranking module is used to adjust the recommendation consistency score of any candidate recommendation question according to the entity similarity corresponding to the target entity, and obtain the recommendation value of any candidate recommendation question; at least one candidate recommendation question with the highest recommendation value is selected from the recommendation question set as the user recommendation question; the recommendation question output module is used to output and store the user recommendation question.

[0209] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0210] Based on the same inventive concept, this application also provides a problem recommendation apparatus for implementing the problem recommendation method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, the specific limitations in one or more problem recommendation apparatus embodiments provided below can be found in the limitations of the problem recommendation method described above, and will not be repeated here.

[0211] In one exemplary embodiment, such as Figure 5 As shown, a problem recommendation device is provided, comprising:

[0212] The acquisition module 510 is used to acquire user input questions and identify the question intent category of the user input questions; wherein, the user input questions are questions entered by the user account in response to a target event;

[0213] The determining module 520 is used to determine a target entity that matches a key entity in the user's question based on a knowledge graph entity set that matches the question intent category; wherein, the knowledge graph entity set includes at least two entities extracted from a preset target knowledge graph; the target knowledge graph is a knowledge graph associated with the target event; and the key entity in the user's question is a key entity in the user's input question.

[0214] Output module 530 is used to process the target entity and the question intent category through a pre-trained language model to obtain user-recommended questions;

[0215] The recommendation module 540 is used to recommend the user recommendation question.

[0216] In one embodiment, the question intent category is at least one of the preset intent categories associated with the target event, and each preset intent category corresponds to a preset entity category. The device is further configured to extract entities from the target knowledge graph to obtain at least two entities to be classified; at least one question category associated with each entity to be classified in the target event; classify each entity to be classified according to the at least one question category associated with each entity to be classified to obtain the classified entities corresponding to each preset entity category; and construct a knowledge graph entity set matching each preset intent category based on the classified entities corresponding to each preset entity category.

[0217] In one embodiment, the device is further configured to perform vectorization processing on the categorized entities corresponding to each preset entity category to obtain entity vector representations of the categorized entities corresponding to each preset entity category; and to construct an entity vector library matching each preset intent category based on the entity vector representations corresponding to each preset entity category; the entity vector library is used to represent the corresponding knowledge graph entity set.

[0218] In one embodiment, the preset intent category includes at least one idea map category and at least one sub-intent category extended from each idea map category. The acquisition module 510 is specifically used to identify a target idea map category that matches the user input question in the at least one idea map category; and to identify a target sub-intent category that matches the user input question in the at least one sub-intent category extended from the target idea map category, as the question intent category of the user input question.

[0219] In one embodiment, the output module 530 is specifically used to input the target entity and the question intent category into the pre-trained language model to generate a recommended question set; the recommended question set includes at least one candidate recommended question that contains the target entity and conforms to the question intent category; and at least one candidate recommended question is selected from the recommended question set as the user recommended question.

[0220] In one embodiment, the output module 530 is specifically used to obtain the recommendation value of each candidate recommendation question in the recommendation question set; the recommendation value is used to characterize the degree of matching between the corresponding candidate recommendation question and the user input question and the target entity; and at least one candidate recommendation question with the highest recommendation value is selected from the recommendation question set as the user recommendation question.

[0221] In one embodiment, the output module 530 is specifically configured to, for any candidate recommendation question in the recommendation question set, input the candidate recommendation question and the user input question into a language model, and generate a recommendation consistency score for the candidate recommendation question by calling the recommendation consistency scoring function of the language model; the recommendation consistency score is used to characterize the appropriateness of the corresponding candidate recommendation question in the current recommendation scenario; if the recommendation consistency score of the candidate recommendation question is greater than or equal to a preset score threshold, the recommendation consistency score of the candidate recommendation question is adjusted according to the entity similarity corresponding to the target entity to obtain the recommendation value of the candidate recommendation question; wherein, the entity similarity is the similarity between the target entity and the key entity.

[0222] In one embodiment, the apparatus further includes a rejection module, configured to reject any candidate recommendation question from the set of recommendation questions if the recommendation compliance score of any candidate recommendation question is less than the preset scoring threshold.

[0223] In one embodiment, the output module 530 is specifically used to convert the entity similarity corresponding to the target entity into a corresponding entity similarity score; and to perform a weighted average of the entity similarity score and the recommendation conformity score to obtain a recommendation value for any candidate recommendation question.

[0224] In one embodiment, the determining module 520 is specifically configured to perform entity retrieval in a knowledge graph entity set that matches the question intent category based on the key entity of the user question, to obtain a retrieval entity set; the retrieval entity set includes a preset number of retrieval entities related to the key entity of the user question; and select the target entity from the retrieval entity set based on the similarity between each retrieval entity and the key entity of the user question; the similarity between the target entity and the key entity of the user question is greater than a preset similarity threshold.

[0225] The modules in the aforementioned recommended device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0226] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 6As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data for implementing the proposed method for solving the aforementioned problem. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for recommending a solution to a problem.

[0227] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0228] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the problem recommendation method described above. The steps of the problem recommendation method here can be steps from one of the problem recommendation methods in the various embodiments described above.

[0229] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the steps of the problem recommendation method described above. The steps of the problem recommendation method here may be steps from one of the problem recommendation methods in the various embodiments described above.

[0230] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, causes the processor to perform the steps of the problem recommendation method described above. The steps of the problem recommendation method here may be steps from one of the problem recommendation methods in the various embodiments described above.

[0231] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0232] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic resistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0233] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0234] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A question recommendation method characterized by comprising: The method comprises: obtaining a user input question and identifying a question intent category of the user input question; wherein the user input question is a question input by a user account for a target event; determining a target entity matching a user question key entity according to a knowledge graph entity set matching the question intent category; wherein the knowledge graph entity set comprises at least two entities extracted from a preset target knowledge graph; the target knowledge graph is a knowledge graph associated with the target event; the user question key entity is a key entity in the user input question; processing the target entity and the question intent category through a pre-trained language model to obtain a user recommended question; recommending the user recommended question.

2. The method of claim 1, wherein, The question intent category is at least one of preset intent categories associated with the target event, and each preset intent category corresponds to a preset entity category; the method further comprises: performing entity extraction on the target knowledge graph to obtain at least two to-be-classified entities; each to-be-classified entity is associated with at least one question category in the target event; classifying each to-be-classified entity according to at least one question category associated with each to-be-classified entity to obtain classified entities corresponding to each preset entity category; constructing a knowledge graph entity set matching each preset intent category according to the classified entities corresponding to each preset entity category.

3. The method of claim 2, wherein, The construction of the knowledge graph entity set matching each preset intent category according to the classified entities corresponding to each preset entity category comprises: performing vectorization processing on the classified entities corresponding to each preset entity category to obtain entity vector representations of the classified entities corresponding to each preset entity category; constructing an entity vector library matching each preset intent category according to the entity vector representations corresponding to each preset entity category; the entity vector library is used to represent the corresponding knowledge graph entity set.

4. The method of claim 2, wherein, The preset intent category comprises at least one main intent category and at least one sub intent category extended from each main intent category; the identification of the question intent category of the user input question comprises: identifying a target main intent category matching the user input question in the at least one main intent category; identifying a target sub intent category matching the user input question in the at least one sub intent category extended from the target main intent category as the question intent category of the user input question.

5. The method of claim 1, wherein, The processing of the target entity and the question intent category through the pre-trained language model to obtain the user recommended question comprises: inputting the target entity and the question intent category into the pre-trained language model to generate a recommended question set; the recommended question set comprises at least one candidate recommended question containing the target entity and conforming to the question intent category; selecting at least one candidate recommended question from the recommended question set as the user recommended question.

6. The method of claim 5, wherein, The selection of at least one candidate recommended question from the recommended question set as the user recommended question comprises: obtain a recommendation value of each candidate recommended question in the recommended question set; the recommendation value is used to represent a matching degree between the corresponding candidate recommended question and the user input question and the target entity respectively; select at least one candidate recommended question with a high ranking recommendation value from the recommended question set as the user recommended question.

7. The method of claim 6, wherein, The method further comprises: For any candidate recommended question in the recommended question set, input the any candidate recommended question and the user input question into a language model to generate a recommendation fitness score of the any candidate recommended question by calling a recommendation fitness score function of the language model; the recommendation fitness score is used to represent a fitness degree of the any candidate recommended question in the current recommendation scenario; In a case where the recommendation fitness score of the any candidate recommended question is greater than or equal to a preset score threshold, adjust the recommendation fitness score of the any candidate recommended question according to an entity similarity corresponding to the target entity to obtain a recommendation value of the any candidate recommended question; the entity similarity is a similarity between the target entity and the key entity.

8. The method of claim 7, wherein, The method further comprises: In a case where the recommendation fitness score of the any candidate recommended question is less than the preset score threshold, eliminate the any candidate recommended question from the recommended question set.

9. The method of claim 7, wherein, The method further comprises: convert the entity similarity corresponding to the target entity into an entity similarity score; perform weighted average on the entity similarity score and the recommendation fitness score to obtain the recommendation value of the any candidate recommended question.

10. The method of claim 1, wherein, The method further comprises: perform entity retrieval in the knowledge graph entity set matched with the question intent category according to the user question key entity to obtain a retrieval entity set; the retrieval entity set includes a preset number of retrieval entities related to the user question key entity; select the target entity from the retrieval entity set according to a similarity between each retrieval entity and the user question key entity; the similarity between the target entity and the user question key entity is greater than a preset similarity threshold.

11. A question recommendation apparatus characterized by comprising: The apparatus comprises: an obtaining module, configured to obtain a user input question, and identify a question intent category of the user input question; the user input question is a question input by a user account for a target event; determining, according to a knowledge graph entity set matched with the question intent category, a target entity matched with a user question key entity; wherein the knowledge graph entity set comprises at least two entities extracted from a preset target knowledge graph; the target knowledge graph is a knowledge graph associated with the target event; the user question key entity is a key entity in the user input question; outputting, by a pre-trained language model, the target entity and the question intent category to obtain a user recommended question; recommending the user recommended question.

12. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method in any one of claims 1 to 10.

Images