Dialogue intention recognition method and device based on large model, equipment and medium

Abstract

The invention relates to the technical field of voice semantics, and discloses a dialogue intention recognition method and device based on a large model, equipment and a medium, and the method comprises the steps: recognizing a target first-level intention label of a target query text through SFT; determining a target knowledge block according to the knowledge base and the target first-level intention label; and determining a dialogue intention according to the RAG and the target knowledge block. Through the above mode, effective preliminary intention recognition is performed on the target query text through the SFT, the target first-level intention label is determined, the target second-level intention label is accurately determined in combination with the preset knowledge base and the RAG model, the intention recognition accuracy and fineness are improved, and the intention recognition efficiency is improved in the second-level intention recognition process. The method can focus on the knowledge content highly related to the target query text, and reduces the burden of processing irrelevant information by the RAG model. The method can be applied to the business fields of financial science and technology, medical health, old-age care and the like, and the user intention recognition accuracy of the intelligent customer service system is improved.

Description

Technical Field

[0001] The present application relates to the field of speech semantics technology, and in particular to a method, apparatus, device and medium for recognizing conversation intentions based on a large model. Background Art

[0002] In today's digital age, both the fintech and healthcare sectors generate massive amounts of data, and the demand for natural language processing (NLP) technology is growing. Within the fintech sector, with the rapid development of internet finance, financial institutions face a large number of tasks, including customer inquiries, transaction requests, and risk assessments. Traditional manual customer service methods are no longer able to meet the high-concurrency, personalized service demands. Natural language processing (NLP) technology is widely used in intelligent customer service systems to automatically answer customer inquiries and process transaction requests. However, existing intent recognition technologies often struggle to accurately understand users' diverse expressions and underlying needs when handling complex tasks such as financial product inquiries and risk assessments. This is particularly true when multi-level intent classification is involved, which can lead to inaccurate classifications.

[0003] In the healthcare, health, and elderly care sectors, emerging service models such as telemedicine and health management are constantly emerging. Patients consult doctors about their conditions and seek health advice through online platforms. Intelligent customer service systems must accurately understand patients' condition descriptions and consultation intentions so that doctors can provide precise medical advice. Furthermore, knowledge in the healthcare, health, and elderly care sectors is highly specialized and complex, requiring intent recognition technology to integrate with a vast medical knowledge base to assist in diagnosis and answering questions.

[0004] Existing intent recognition technologies mostly rely on single, large-parameter models for classification. This approach consumes significant computing resources, making it difficult to meet real-time requirements for applications. Furthermore, it lacks generalizability across domains and scenarios. Therefore, improving the accuracy of user intent recognition in intelligent customer service systems in areas such as fintech and healthcare and elderly care has become a pressing technical challenge. Summary of the Invention

[0005] This application provides a large-model-based conversation intent recognition method, device, equipment, and medium to improve the accuracy of intelligent customer service systems in identifying user intent in business areas such as financial technology, healthcare, and elderly care.

[0006] In a first aspect, the present application provides a method for identifying conversational intent based on a large model, the method comprising:

[0007] Identify the target-level intent label of the target query text through the supervised fine-tuning model SFT;

[0008] Determining a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label;

[0009] According to the preset retrieval enhancement generation model RAG and the target knowledge block, the target secondary intention label of the target query text is determined to determine the dialogue intention according to the target secondary intention label.

[0010] In a second aspect, the present application further provides a large-model-based conversation intention recognition device, comprising:

[0011] The target first-level intent label recognition module is used to identify the target first-level intent label of the target query text through the supervised fine-tuning model SFT;

[0012] A target knowledge block determination module is used to determine the target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label;

[0013] The target secondary intention label determination module is used to determine the target secondary intention label of the target query text based on the preset retrieval enhancement generation model RAG and the target knowledge block, so as to determine the dialogue intention based on the target secondary intention label.

[0014] In a third aspect, the present application also provides a computer device comprising a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the large model-based conversation intention recognition method as described above when executing the computer program.

[0015] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor implements the large model-based conversation intention recognition method as described above.

[0016] The present application discloses a method, apparatus, device and medium for recognizing conversation intention based on a large model, wherein the method for recognizing conversation intention based on a large model includes recognizing a target first-level intention label of a target query text through a supervised fine-tuning model SFT; determining a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label; and determining a target second-level intention label of the target query text based on a preset retrieval enhancement generation model RAG and the target knowledge block, so as to determine the conversation intention based on the target second-level intention label. Through the above-mentioned method, the present application performs effective preliminary intent recognition on the target query text through SFT, determines the target first-level intention label, and accurately determines the target second-level intention label by combining the preset knowledge base and the RAG model, thereby improving the accuracy and precision of intent recognition. In the process of second-level intent recognition, it can focus on knowledge content that is highly relevant to the target query text, reduce the burden of the RAG model on processing irrelevant information, and thereby improve the accuracy of intelligent customer service systems in recognizing user intentions in business fields such as financial technology and medical health and elderly care. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0018] Figure 1 This is a schematic flow chart of a method for identifying conversation intentions based on a large model provided in the first embodiment of the present application;

[0019] Figure 2 This is a schematic flowchart of a method for identifying conversation intentions based on a large model provided in the second embodiment of the present application;

[0020] Figure 3 A schematic block diagram of a large model-based conversation intention recognition device provided in an embodiment of the present application;

[0021] Figure 4 A schematic block diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0023] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0024] It should be understood that the terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] The embodiments of the present application provide a method, apparatus, device, and medium for recognizing conversational intent based on a large model. The method can be applied to a server, effectively performing preliminary intent recognition on a target query text through SFT, determining the target's first-level intent label, and accurately determining the target's second-level intent label by combining a preset knowledge base with a RAG model. This improves the accuracy and precision of intent recognition. During the second-level intent recognition process, it can focus on knowledge content that is highly relevant to the target query text, reducing the burden on the RAG model to process irrelevant information. This improves the accuracy of intelligent customer service systems in recognizing user intent in business areas such as financial technology and healthcare and elderly care.

[0027] The following describes some embodiments of the present application in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features therein may be combined with each other.

[0028] See also Figure 1 , Figure 1 This is a schematic flow chart of a large-model-based method for identifying conversational intent, as provided in the first embodiment of this application. This large-model-based method can be applied to intelligent customer service systems to improve their accuracy in identifying user intent in areas such as fintech and healthcare.

[0029] like Figure 1 As shown, the large model-based conversation intention recognition method specifically includes steps S10 to S30.

[0030] Step S10: Identify the target first-level intent label of the target query text through the supervised fine-tuning model SFT;

[0031] Specifically, in the field of financial technology business, this embodiment is explained by taking automobile insurance claims as an example.

[0032] Collect a large amount of text data related to auto insurance claims. This data can come from sources such as auto insurance company customer service records, online consultation platforms, and claims-related forums. Ensure that the data covers a wide range of customer inquiries and questions, such as the claims process, required documents, and compensation standards.

[0033] Based on the needs of auto insurance claims, a complete set of first-level intent tagging systems is defined. For example, "claims process consultation," "claims materials consultation," "compensation amount consultation," or "claims progress inquiry" can be used.

[0034] Supervised Fine-Tuning (SFT) requires identifying a suitable pre-trained language model, such as the qwen2-7B model, as the starting point for supervised fine-tuning. Using a labeled training dataset, the pre-trained model's parameters are fully fine-tuned. During fine-tuning, the model's parameters are continuously adjusted using a backpropagation algorithm, enabling the model to accurately predict the corresponding first-level intent label based on the input target query text. A portion of the untrained data is used as a test set to evaluate the fine-tuned model until it can effectively identify the target first-level intent label for the target query text.

[0035] Step S20: determining a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label;

[0036] Specifically, in the healthcare and elderly care sectors, various healthcare information, such as disease symptoms, diagnostic methods, treatment methods, medication information, and health advice, is collected in advance from authoritative medical textbooks, clinical guidelines, healthcare websites, and medical research papers. This information is then organized according to the healthcare knowledge system and divided into different themes and categories, for example, by disease type, health problem, etc.

[0037] The organized knowledge is further subdivided into specific knowledge blocks. Each knowledge block contains a relatively independent knowledge point or theme, such as "Symptoms of Hypertension", "Dietary Control of Diabetes", etc. Detailed information is annotated for each knowledge block, including its first-level intent label (such as "Disease Diagnosis Consultation", "Treatment Plan Consultation", etc.), keywords, etc. In the index of the knowledge base, a preliminary screening is performed based on the target first-level intent label to find the range of knowledge blocks marked with the same or related first-level intent labels. For example, under the label of "Disease Diagnosis Consultation", knowledge blocks related to the diagnosis of various diseases are screened out.

[0038] The keywords and phrases are extracted from the target query text, and the keywords matching and semantic similarity calculation methods are combined to further find the knowledge blocks that are semantically consistent with the target query text within the range of the initially screened knowledge blocks. For example, the target query text is about the possible causes of symptoms such as chest pain and difficulty breathing. Through keyword matching and semantic analysis, disease diagnosis knowledge blocks related to "chest pain and difficulty breathing" are retrieved from the knowledge base, such as "symptom manifestations of heart disease" and "common symptoms of lung disease", which are used as target knowledge blocks.

[0039] In step S30, the target secondary intent label of the target query text is determined according to the pre-set retrieval augmented generation model RAG and the target knowledge block, so as to determine the dialogue intent according to the target secondary intent label.

[0040] Specifically, the content of the target knowledge block is integrated with the target query text. For example, the target query text is a patient's consultation about the possible causes of chest pain and difficulty breathing, and the target knowledge block is related knowledge about "common symptom manifestations of heart disease". The description of the symptoms of heart disease in the knowledge block is combined with the patient's question to form a coherent prompt: "Patient asks about the possible causes of chest pain and difficulty breathing."

[0041] The problem to be solved is clearly described in the prompt, and the RAG (Retri eval-Augmented Generation) model is guided to generate the target secondary intent label. For example, in the above prompt, it is explicitly stated that "please judge the patient's specific disease diagnosis consultation intent according to the relevant knowledge".

[0042] The retrieval module analyzes the prompt and extracts the key information. According to the extracted key information, the retrieval module further retrieves related knowledge information in the pre-set medical and health knowledge base. These knowledge information may include more information about the symptoms of heart disease, diagnosis methods, and other possible causes, which supplement and perfect the information of the target knowledge block.

[0043] The generation module of the RAG model generates the target secondary intent label corresponding to the target query text according to the content in the prompt and the additional knowledge information retrieved. For example, the generated target secondary intent label can be "heart disease diagnosis consultation" or "coronary heart disease diagnosis consultation". In the actual medical and health consultation scenario, if it is in a multi-round dialogue, the context information of the previous dialogue will also be combined for generation to ensure that the generated secondary intent label can maintain coherence and consistency with the previous dialogue content, and accurately grasp the real intent changes of the patient in the whole dialogue process.

[0044] The application discloses a dialogue intention recognition method based on a large model, which comprises the following steps: identifying a target primary intention label of a target query text by a supervised fine-tuning model SFT; determining a target knowledge block corresponding to the target primary intention label according to a preset knowledge base and the target primary intention label; determining a target secondary intention label of the target query text according to a preset retrieval and enhancement generation model RAG and the target knowledge block, so as to determine a dialogue intention according to the target secondary intention label. In the foregoing manner, the application enables the RAG model to more accurately understand the semantics and context of the target query text by retrieving information with a high degree of association with the target knowledge block, so as to more reliably determine the target secondary intention label. In combination with the highly relevant associated knowledge information and the target query text, the RAG model can perform more in-depth semantic understanding and logical reasoning, thereby improving the accuracy of an intelligent customer service system in identifying user intentions in the fields of financial technology, medical and health care, and the like.

[0045] Based on Figure 1 In the embodiment shown in the figure, before step S20, the following steps are included:

[0046] According to the target query text, determine relevant text information;

[0047] According to the preset business logic and the preset intention recognition requirement, classify all the relevant text information, determine the text category and text level of each relevant text information, generate a preset primary intention label and a preset secondary intention label;

[0048] According to each preset primary intention label and each preset secondary intention label, generate a preset knowledge block, and establish the preset knowledge base according to each preset knowledge block.

[0049] Specifically, the target query text is analyzed by natural language processing to identify keywords and phrases. For example, if the target query text is "how to apply for a credit card", the keywords are "apply" and "credit card". In the database in the financial field, the text information matching the keywords is retrieved, the most relevant text information to the target query text is selected according to the relevance ranking, such as "credit card application process", "credit card application conditions", etc., and the text information of the same category is integrated into a knowledge block, including a primary and secondary intention label, such as a "credit card application" block containing credit card application conditions, process, etc.

[0050] Based on Figure 1 In the embodiment shown in the figure, step S20 comprises the following steps:

[0051] According to the target primary intention label, determine a primary intention label index;

[0052] The target knowledge block is determined according to the first-level intention label index and the preset intention label corresponding to each of the preset knowledge blocks.

[0053] Specifically, in the healthcare and elderly care sectors, inverted indexing technology is used, using keywords and phrases within knowledge blocks as indexing criteria to establish a mapping relationship between knowledge blocks and index terms. For example, keywords such as "hypertension," "symptoms," and "manifestations" are indexed and associated with the knowledge block "Symptoms of hypertension," allowing queries for these keywords to quickly locate the corresponding block.

[0054] See also Figure 2 , Figure 2 This is a schematic flow chart of a large-model-based method for identifying conversational intent, provided in the second embodiment of the present application. This large-model-based method can be applied to intelligent customer service systems. By retrieving information highly associated with the target knowledge block, the RAG model can more accurately understand the semantics and context of the target query text, thereby more reliably determining the target secondary intent label. By combining highly relevant associated knowledge information and the target query text, the RAG model can perform deeper semantic understanding and logical reasoning, thereby improving the accuracy of intelligent customer service systems in identifying user intent in business areas such as financial technology and healthcare and elderly care.

[0055] based on Figure 1 The embodiment shown, this embodiment Figure 2 As shown, step S30 includes steps S301 to S302.

[0056] Step S301: searching the preset knowledge base for associated knowledge information corresponding to the target knowledge block through the built-in retrieval module of the RAG, wherein the degree of association between the associated knowledge information and the target knowledge block is greater than a preset association threshold;

[0057] Specifically, the built-in retrieval module of RAG receives the target knowledge block. For example, in the financial field, the target knowledge block may be knowledge related to "credit card application conditions and procedures". The retrieval module uses the keywords and phrases in the target knowledge block as clues to search in the preset financial knowledge base to find other knowledge information related to the target knowledge block.

[0058] For each piece of knowledge retrieved, its relevance to the target knowledge block is calculated, and the relevance is compared with a preset relevance threshold. Knowledge pieces with a relevance greater than the threshold are selected as relevant knowledge pieces. For example, if the preset relevance threshold is 0.7, only knowledge pieces with a relevance greater than 0.7 are retained.

[0059] Step S302: Determine the target secondary intent label through the RAG according to the associated knowledge information and the target query text.

[0060] Specifically, the obtained related knowledge information is integrated with the target query text to construct a prompt. For example, if the target query text is "how to apply for a credit card," the related knowledge information includes "credit card application materials" and "credit card application review time." The constructed prompt might be "User asks how to apply for a credit card; Related knowledge: The following materials are required to apply for a credit card: proof of identity, proof of income, etc.; the review time is generally 5-10 business days."

[0061] The generation module uses its internal neural network structure and pre-trained language generation capabilities to generate target secondary intent labels corresponding to the target query text. For example, the generated target secondary intent labels may be "credit card application process consultation", "credit card application conditions consultation", etc., which specifically describe the user's detailed consultation intention regarding credit card application.

[0062] The present application discloses a method for identifying conversation intentions based on a large model, which includes identifying a target first-level intention label of a target query text through a supervised fine-tuning model SFT; determining a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label; retrieving associated knowledge information corresponding to the target knowledge block in the preset knowledge base through a retrieval module built into the RAG, wherein the degree of association between the associated knowledge information and the target knowledge block is greater than a preset association threshold; and determining the target second-level intention label through the RAG based on the associated knowledge information and the target query text. In the above manner, the present application enables the RAG model to more accurately understand the semantics and context of the target query text by retrieving information with a high degree of association with the target knowledge block, thereby more reliably determining the target second-level intention label. Combined with highly relevant associated knowledge information and the target query text, the RAG model can perform deeper semantic understanding and logical reasoning, thereby improving the accuracy of intelligent customer service systems in identifying user intentions in business fields such as financial technology and medical health and elderly care.

[0063] based on Figure 2 In the embodiment shown, in this embodiment, step S302 includes:

[0064] Integrating the associated knowledge information and the target query text to generate a coherent text sequence;

[0065] Adding question guide characters to the coherent text sequence to generate guide characters;

[0066] The target secondary intent label is determined through the RAG built-in generation module and the guide.

[0067] Specifically, the target query text and associated knowledge information are integrated into a coherent text sequence. Question guide characters (such as "Question:" or "Please ask:") are added to the integrated text sequence to clarify the problem to be solved. The text sequence with added question guide characters is used as a guide to guide the RAG model to generate the target secondary intent label.

[0068] The generated leader is input into the generation module of the RAG model. The generation module of the RAG model processes the information in the leader and uses its internal neural network structure and pre-trained language generation capabilities to generate the target secondary intent label corresponding to the target query text.

[0069] based on Figure 1 In the embodiment shown, in this embodiment, step S10 includes:

[0070] Extracting the intent keywords of the target query text and generating an initial intent tag based on the intent keywords;

[0071] Performing semantic analysis on the initial intent tag through the SFT to generate a semantic analysis result;

[0072] The target context information of the target query text is predicted according to the semantic analysis result, and at least one target first-level intent label is determined from the initial intent labels according to the target context information.

[0073] Specifically, the keyword extraction algorithm is used to extract the intent keywords in the target query text. Based on the extracted intent keywords, combined with predefined label mapping rules or through a simple matching strategy, an initial intent label is generated. The initial intent label is input into the SFT model. The initial intent label is semantically analyzed by the SFT model to extract its semantic features and generate a semantic analysis result. The semantic analysis result can be in vector form or a structured semantic representation.

[0074] Based on the semantic analysis results and the target query text, the target context information is predicted. For example, if the semantic analysis results show that the initial intent label "Credit card application consultation" is related to context information such as "Credit card application conditions" and "Credit card application process," at least one target first-level intent label is determined from the initial intent labels based on the predicted target context information. For example, based on the context information "Credit card application conditions" and "Credit card application process," the target first-level intent label "Product consultation" is determined from the initial intent label "Credit card application consultation."

[0075] In a specific embodiment, predicting target context information of the target query text according to the semantic analysis result, and determining at least one target first-level intent label from the initial intent label according to the target context information includes:

[0076] Performing lexical analysis and syntactic analysis on the target query text by using natural language processing technology to generate lexical analysis results and syntactic analysis results;

[0077] generating the target context information according to the lexical analysis result and the syntactic analysis result;

[0078] Calculating probability distribution information of matching the target context information with each of the initial intent labels;

[0079] The initial intent label whose probability value matching the target context information is greater than or equal to a preset probability threshold is determined as the target first-level intent label.

[0080] Specifically, natural language processing tools are used to segment the target query text, dividing it into individual words or phrases. Part-of-speech tags are then assigned to each word to determine its part of speech within the sentence. Based on the lexical analysis results, a grammatical structure is constructed to show the syntactic relationships between words. Based on the lexical and syntactic analysis results, key information from the target query text is extracted. Combined with this key information, target context is constructed to describe the semantic background and relevant details of the target query text.

[0081] Calculate the similarity between the target context and each initial intent tag. Based on the similarity calculation results, generate probability distribution information. Set an appropriate preset probability threshold based on business needs and application scenarios. For example, the threshold can be set to 0.7, indicating that only initial intent tags with a matching probability greater than or equal to 0.7 will be determined as the target first-level intent tag.

[0082] The probability distribution information is traversed to identify initial intent labels whose probability of matching the target context information is greater than or equal to the preset probability threshold. These initial intent labels are then designated as the target first-level intent labels and used as the final intent recognition results. For example, if the initial intent label "credit card application consultation" has a matching probability of 0.8, which is greater than the preset threshold of 0.7, it is designated as the target first-level intent label.

[0083] See also Figure 3 , Figure 3 This embodiment of the present application provides a schematic block diagram of a large-model-based conversation intention recognition device, which is used to perform the aforementioned large-model-based conversation intention recognition method. The large-model-based conversation intention recognition device can be configured on a server.

[0084] like Figure 3 As shown, the large model-based conversation intention recognition device 400 includes:

[0085] A target first-level intent label identification module 410 is configured to identify a target first-level intent label of a target query text using a supervised fine-tuning model SFT;

[0086] A target knowledge block determination module 420 is configured to determine a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label;

[0087] The target secondary intention label determination module 430 is used to determine the target secondary intention label of the target query text based on the preset retrieval enhancement generation model RAG and the target knowledge block, so as to determine the dialogue intention based on the target secondary intention label.

[0088] Furthermore, the large model-based conversation intention recognition device 400 includes:

[0089] A relevant text information determination module, configured to determine relevant text information based on the target query text;

[0090] A preset label generation module is used to classify all the relevant text information based on preset business logic and preset intent recognition requirements, determine the text category and text level of each relevant text information, and generate a preset first-level intent label and a preset second-level intent label;

[0091] The preset knowledge base establishment module is used to generate preset knowledge blocks according to each of the preset first-level intention tags and each of the preset second-level intention tags, and to establish the preset knowledge base according to each of the preset knowledge blocks.

[0092] Furthermore, the target knowledge block determination module 420 includes:

[0093] a first-level intent label index determining unit, configured to determine a first-level intent label index according to the target first-level intent label;

[0094] The target knowledge block determination unit is configured to determine the target knowledge block according to the first-level intention label index and the preset intention label corresponding to each of the preset knowledge blocks.

[0095] Furthermore, the target secondary intention label determination module 430 includes:

[0096] an associated knowledge information retrieval unit, configured to retrieve associated knowledge information corresponding to the target knowledge block from the preset knowledge base through a retrieval module built into the RAG, wherein the degree of association between the associated knowledge information and the target knowledge block is greater than a preset association threshold;

[0097] A target secondary intention label determination unit is used to determine the target secondary intention label through the RAG according to the associated knowledge information and the target query text.

[0098] Furthermore, the target secondary intention label determination unit includes:

[0099] A coherent text sequence generating subunit, configured to integrate the associated knowledge information and the target query text to generate a coherent text sequence;

[0100] A leader generating subunit, configured to add a question leader character to the coherent text sequence to generate a leader;

[0101] The target secondary intention label determination subunit is used to determine the target secondary intention label through the RAG built-in generation module and the guide.

[0102] Furthermore, the target first-level intention label recognition module 410 includes:

[0103] An initial intent tag generating unit, configured to extract intent keywords from the target query text and generate an initial intent tag based on the intent keywords;

[0104] a semantic analysis result generating unit, configured to perform semantic analysis on the initial intent tag through the SFT to generate a semantic analysis result;

[0105] A target first-level intent label determination unit is used to predict the target context information of the target query text according to the semantic analysis result, and determine at least one target first-level intent label from the initial intent labels according to the target context information.

[0106] Furthermore, the target first-level intention label determination unit includes:

[0107] An analysis subunit, configured to perform lexical analysis and syntactic analysis on the target query text by using natural language processing technology to generate lexical analysis results and syntactic analysis results;

[0108] a target context information generating subunit, configured to generate the target context information according to the lexical analysis result and the syntactic analysis result;

[0109] A probability distribution information calculation subunit, configured to calculate probability distribution information of a match between the target context information and each of the initial intent labels;

[0110] The target first-level intention label determination subunit is used to determine the initial intention label whose probability value matching the target context information is greater than or equal to a preset probability threshold as the target first-level intention label.

[0111] It should be noted that those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0112] The above-mentioned device can be realized in the form of a computer program. The computer program can be used in Figure 4 Runs on the computer equipment shown.

[0113] See also Figure 4 , Figure 4 1 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device may be a server.

[0114] See Figure 4 The computer device includes a processor, a memory, and a network interface connected through a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0115] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable the processor to execute any one of the large model-based conversation intention recognition methods.

[0116] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0117] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any large model-based dialogue intention recognition method.

[0118] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0119] It should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or 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, etc.

[0120] In one embodiment, the processor is configured to execute a computer program stored in the memory to implement the following steps:

[0121] Identify the target-level intent label of the target query text through the supervised fine-tuning model SFT;

[0122] Determining a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label;

[0123] According to the preset retrieval enhancement generation model RAG and the target knowledge block, the target secondary intention label of the target query text is determined to determine the dialogue intention according to the target secondary intention label.

[0124] In one embodiment, before determining the target knowledge block corresponding to the target first-level intention label based on the preset knowledge base and the target first-level intention label, it is used to implement:

[0125] Determining relevant text information according to the target query text;

[0126] Classify all the relevant text information based on the preset business logic and preset intent recognition requirements, determine the text category and text level of each of the relevant text information, and generate a preset first-level intent label and a preset second-level intent label;

[0127] A preset knowledge block is generated according to each of the preset first-level intention tags and each of the preset second-level intention tags, and the preset knowledge base is established according to each of the preset knowledge blocks.

[0128] In one embodiment, based on a preset knowledge base and the target first-level intention label, a target knowledge block corresponding to the target first-level intention label is determined to achieve:

[0129] Determine a first-level intent label index based on the target first-level intent label;

[0130] The target knowledge block is determined according to the first-level intention label index and the preset intention label corresponding to each of the preset knowledge blocks.

[0131] In one embodiment, a target secondary intent tag of the target query text is determined based on a preset retrieval enhancement generation model RAG and the target knowledge block, and a conversation intent is determined based on the target secondary intent tag, to achieve:

[0132] Retrieving associated knowledge information corresponding to the target knowledge block from the preset knowledge base through a retrieval module built into the RAG, wherein the degree of association between the associated knowledge information and the target knowledge block is greater than a preset association threshold;

[0133] The target secondary intent label is determined through the RAG according to the associated knowledge information and the target query text.

[0134] In one embodiment, the target secondary intent tag is determined by the RAG based on the associated knowledge information and the target query text, so as to achieve:

[0135] Integrating the associated knowledge information and the target query text to generate a coherent text sequence;

[0136] Adding question guide characters to the coherent text sequence to generate guide characters;

[0137] The target secondary intent label is determined through the RAG built-in generation module and the guide.

[0138] In one embodiment, the supervised fine-tuning model (SFT) is used to identify the target first-level intent label of the target query text to achieve:

[0139] Extracting the intent keywords of the target query text and generating an initial intent tag based on the intent keywords;

[0140] Performing semantic analysis on the initial intent tag through the SFT to generate a semantic analysis result;

[0141] The target context information of the target query text is predicted according to the semantic analysis result, and at least one target first-level intent label is determined from the initial intent labels according to the target context information.

[0142] In one embodiment, target context information of the target query text is predicted based on the semantic analysis result, and at least one target first-level intent tag is determined from the initial intent tags based on the target context information, so as to achieve:

[0143] performing morphological analysis and syntactic analysis on the target query text by using a natural language processing technology to generate morphological analysis results and syntactic analysis results;

[0144] generating the target context information according to the morphological analysis results and the syntactic analysis results;

[0145] calculating probability distribution information of the target context information matching each initial intent label;

[0146] determining the initial intent label with a probability value greater than or equal to a preset probability threshold as the target primary intent label.

[0147] An embodiment of the present application also provides a computer readable storage medium storing a computer program, the computer program including program instructions, and the processor executes the program instructions to implement any one of the dialog intent recognition methods based on a large model provided by the embodiments of the present application.

[0148] The computer readable storage medium can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device. The computer readable storage medium can also be an external storage device of the computer device, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc.

[0149] The above merely describes the specific embodiments of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for identifying conversation intentions based on a large model, characterized in that: include: Identify the target-level intent label of the target query text through the supervised fine-tuning model SFT; Determining a target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label; According to the preset retrieval enhancement generation model RAG and the target knowledge block, the target secondary intention label of the target query text is determined to determine the dialogue intention according to the target secondary intention label.

2. The method for identifying conversation intention based on a large model according to claim 1, characterized in that: Before determining the target knowledge block corresponding to the target first-level intention label based on the preset knowledge base and the target first-level intention label, the method includes: Determining relevant text information according to the target query text; Classify all the relevant text information based on the preset business logic and preset intent recognition requirements, determine the text category and text level of each of the relevant text information, and generate a preset first-level intent label and a preset second-level intent label; A preset knowledge block is generated according to each of the preset first-level intention tags and each of the preset second-level intention tags, and the preset knowledge base is established according to each of the preset knowledge blocks.

3. The method for identifying conversation intention based on a large model according to claim 2, characterized in that: The determining, based on a preset knowledge base and the target first-level intention label, a target knowledge block corresponding to the target first-level intention label includes: Determine a first-level intent label index based on the target first-level intent label; The target knowledge block is determined according to the first-level intention label index and the preset intention label corresponding to each of the preset knowledge blocks.

4. The method for identifying conversation intention based on a large model according to claim 2, characterized in that: The step of determining a target secondary intent tag of the target query text based on the preset retrieval enhancement generation model RAG and the target knowledge block, and determining a dialogue intent based on the target secondary intent tag, includes: Retrieving associated knowledge information corresponding to the target knowledge block from the preset knowledge base through a retrieval module built into the RAG, wherein the degree of association between the associated knowledge information and the target knowledge block is greater than a preset association threshold; The target secondary intent label is determined through the RAG according to the associated knowledge information and the target query text.

5. The method for recognizing conversation intention based on a large model according to claim 4 is characterized in that: The determining the target secondary intent label through the RAG according to the associated knowledge information and the target query text includes: Integrating the associated knowledge information and the target query text to generate a coherent text sequence; Adding question guide characters to the coherent text sequence to generate guide characters; The target secondary intent label is determined through the RAG built-in generation module and the guide.

6. The method for identifying conversation intention based on a large model according to claim 1, characterized in that: The method of identifying the target first-level intent label of the target query text by using the supervised fine-tuning model SFT includes: Extracting the intent keywords of the target query text and generating an initial intent tag based on the intent keywords; Performing semantic analysis on the initial intent tag through the SFT to generate a semantic analysis result; The target context information of the target query text is predicted according to the semantic analysis result, and at least one target first-level intent label is determined from the initial intent labels according to the target context information.

7. The method for identifying conversation intentions based on a large model according to claim 6, characterized in that: The predicting target context information of the target query text according to the semantic analysis result, and determining at least one target first-level intent label from the initial intent labels according to the target context information, includes: Performing lexical analysis and syntactic analysis on the target query text by using natural language processing technology to generate lexical analysis results and syntactic analysis results; generating the target context information according to the lexical analysis result and the syntactic analysis result; Calculating probability distribution information of matching the target context information with each of the initial intent labels; The initial intent label whose probability value matching the target context information is greater than or equal to a preset probability threshold is determined as the target first-level intent label.

8. A conversation intention recognition device based on a large model, characterized in that: include: The target first-level intent label recognition module is used to identify the target first-level intent label of the target query text through the supervised fine-tuning model SFT; A target knowledge block determination module is used to determine the target knowledge block corresponding to the target first-level intention label based on a preset knowledge base and the target first-level intention label; The target secondary intention label determination module is used to determine the target secondary intention label of the target query text based on the preset retrieval enhancement generation model RAG and the target knowledge block, so as to determine the dialogue intention based on the target secondary intention label.

9. A computer device, characterized in that: The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is configured to execute the computer program and implement the large model-based conversation intention recognition method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to implement the large model-based conversation intention recognition method according to any one of claims 1 to 7.

Images