Data identification method and device based on artificial intelligence, computer equipment and medium

Abstract

The invention belongs to the technical field of artificial intelligence, and relates to a data identification method and device based on artificial intelligence, computer equipment and a storage medium, and the method comprises the steps: collecting the real-time call data of a user in a call process with the user; converting the real-time call data to obtain a call text, and performing feature extraction on the call text to obtain text feature data; performing feature extraction on the real-time call data to obtain feature data; fusing the text feature data and the feature data to obtain a comprehensive feature vector; comparing the comprehensive feature vector with a knowledge base based on an intelligent agent to obtain a case matching result; and reasoning the comprehensive feature vector and the case matching result based on the target large language model to generate a recommendation opportunity recognition result, and outputting the recommendation opportunity recognition result. In addition, the recommendation opportunity recognition result can be stored in the block chain. The method can be applied to a recommendation opportunity recognition scene in the field of financial science and technology, and the accuracy of recommendation opportunity recognition is effectively improved based on the target large language model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and can be applied to the financial technology field, particularly to data recognition methods, devices, computer equipment and storage media based on artificial intelligence. Background Technology

[0002] In traditional telephone-based insurance operations, agents rely primarily on their experience, insurance expertise, and communication skills to determine the appropriate timing for product recommendations during conversations with customers. While this approach allows agents to grasp the timing to some extent, it is heavily influenced by subjective factors. Different agents may judge the timing differently, making it difficult to guarantee the stability and accuracy of recommendations.

[0003] In addition, some organizations also use Natural Language Processing (NLP) technology to analyze customer conversations and combine it with simple rule models to identify the timing of recommendations. However, traditional NLP technology has limited depth of semantic understanding, and simple rule models are difficult to fully cover the complex and ever-changing situations in auto insurance business, resulting in low accuracy in identifying the timing of recommendations. For example, in auto insurance renewal scenarios, customers may mention in a phone call that their vehicle usage frequency has decreased recently. Traditional methods, based solely on this simple statement, may not be able to accurately determine whether the customer still has a need for renewal and when to recommend a suitable renewal product. If the customer has only reduced usage in the short term but will still have a high usage frequency and protection needs in the future, traditional methods may miss the best recommendation time due to incorrect judgment, or recommend unsuitable insurance products, affecting the customer's willingness to renew and the insurance company's business expansion.

[0004] The inaccurate timing of recommendations in this traditional model results in poor matching between recommended products and customers' actual needs, reducing customer willingness to buy and satisfaction, and consequently impacting the competitiveness of insurance companies in the auto insurance market. Therefore, there is an urgent need for a more accurate method for identifying recommendation timing to improve accuracy, meet personalized customer needs, and enhance the overall effectiveness of insurance services. Summary of the Invention

[0005] The purpose of this application is to provide a data recognition method, apparatus, computer device, and storage medium based on artificial intelligence to solve the technical problem of inaccurate identification of recommendation timing in existing technologies.

[0006] Firstly, an artificial intelligence-based data recognition method is provided, including: During a call with a user, real-time call data corresponding to that user is collected. The real-time call data is transformed to obtain the corresponding call text, and the call text is then subjected to feature extraction to obtain the corresponding text feature data. Based on a preset multimodal feature extraction strategy, features are extracted from the real-time call data to obtain the corresponding feature data; The text feature data and the feature data are fused together to obtain the corresponding comprehensive feature vector; Based on the preset intelligent agent, the comprehensive feature vector is compared with the pre-built knowledge base to obtain the corresponding case matching result; Based on the preset target large language model, the comprehensive feature vector and the case matching result are subjected to reasoning processing to generate the corresponding recommendation timing recognition result; The recommendation timing identification results are then processed for output.

[0007] Secondly, an artificial intelligence-based data recognition device is provided, comprising: The data acquisition module is used to collect real-time call data corresponding to the user during a call. The first processing module is used to transform and process the real-time call data to obtain the corresponding call text, and to extract features from the call text to obtain the corresponding text feature data. The first extraction module is used to extract features from the real-time call data based on a preset multimodal feature extraction strategy to obtain corresponding feature data. The fusion module is used to fuse the text feature data and the feature data to obtain the corresponding comprehensive feature vector; The comparison module is used to compare the comprehensive feature vector with a pre-built knowledge base based on a preset intelligent agent to obtain the corresponding case matching result; The reasoning module is used to perform reasoning processing on the comprehensive feature vector and the case matching result based on the preset target large language model, and generate the corresponding recommendation timing recognition result. The output module is used to process the recommendation timing identification results.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described artificial intelligence-based data recognition method.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned artificial intelligence-based data recognition method.

[0010] In the aforementioned scheme implemented by the AI-based data recognition method, device, computer equipment, and storage medium, during a call with a user, real-time call data corresponding to the user is first collected; then, the real-time call data is transformed to obtain corresponding call text, and features are extracted from the call text to obtain corresponding text feature data; features are then extracted from the real-time call data based on a preset multimodal feature extraction strategy to obtain corresponding feature data; subsequently, the text feature data and the feature data are fused to obtain a corresponding comprehensive feature vector; subsequently, a preset intelligent agent compares the comprehensive feature vector with a pre-built knowledge base to obtain a corresponding case matching result; further, a preset target large language model is used to reason about the comprehensive feature vector and the case matching result to generate a corresponding recommendation timing recognition result; finally, the recommendation timing recognition result is output. Unlike existing recommendation timing recognition methods, this application, by using a target large language model, performs reasoning processing on the comprehensive feature vector that integrates text feature data and feature data, and compares the case matching result with the knowledge base, which can effectively improve the recognition efficiency and accuracy of recommendation timing recognition, and ensure the accuracy of the obtained recommendation timing recognition result. Attached Figure Description

[0011] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 This is a flowchart of an embodiment of the artificial intelligence-based data recognition method according to this application; Figure 3 This is a schematic diagram of a structure of an embodiment of the artificial intelligence-based data recognition device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0016] like Figure 1 As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0017] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0018] Terminal device 101 can be various electronic devices with a display screen and support web browsing. In addition to laptops 1011, tablets 1012, or mobile phones 1013, terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer, and a desktop computer, etc.

[0019] Server 103 can be a server that provides various services, such as a backend server that provides support for the pages displayed on terminal device 101.

[0020] It should be noted that the AI-based data recognition method provided in this application is generally executed by a server / terminal device, and correspondingly, the AI-based data recognition device is generally installed in the server / terminal device.

[0021] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0022] Continue to refer to Figure 2 The flowchart illustrates an embodiment of the AI-based data recognition method according to this application. The order of steps in the flowchart can be changed, and some steps can be omitted, depending on different needs. The AI-based data recognition method provided in this application can be applied to any scenario requiring data recognition, and therefore can be applied to products in these scenarios, such as data recognition products in the financial insurance field. The AI-based data recognition method includes the following steps: Step S201: During a call with a user, collect real-time call data corresponding to the user.

[0023] In this embodiment, the artificial intelligence-based data recognition method operates on an electronic device (e.g., Figure 1The server / terminal device shown can acquire real-time call data corresponding to the user via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future-developed wireless connection methods. The executing entity of this application is specifically a data identification system, or timing identification system, which can be simply referred to as the system. This application can be applied to telephone-based operational scenarios in the insurance business field to provide agents with accurate product recommendation timing identification results. For example, the aforementioned operational scenario can be a business scenario involving conducting car insurance-related business through telephone channels. This may include scenarios such as answering user inquiries, recommending car insurance products, guiding and processing insurance applications, reminding and following up on renewals, and providing claims consultation and assistance.

[0024] In addition, during calls between agents and users, such as telemarketing and online customer service communication, voice capture devices, such as microphones in telephones or conference systems, can be used to continuously collect voice signals from both the user and the agent, providing real-time call data. This real-time call data includes all content exchanged between the two parties, including questions raised by the user and responses from the agent.

[0025] Step S202: The real-time call data is transformed to obtain the corresponding call text, and the call text is feature extracted to obtain the corresponding text feature data.

[0026] In this embodiment, the collected real-time call data can be processed in real time using a general speech recognition algorithm. It segments the continuous speech stream into small speech segments, and then extracts features from each segment, such as spectral and pitch features. Using these features, the speech recognition algorithm can identify the corresponding text content and use it as the corresponding call text, thus completing the conversion and processing of the aforementioned real-time call data.

[0027] In addition, the process of extracting features from the aforementioned call text to obtain corresponding text feature data includes: Data preprocessing: Preprocessing the call text, including text cleaning (removing noise, typos, etc.), word segmentation, and part-of-speech tagging, to facilitate subsequent deep learning analysis. User intent analysis: Analyzing the vocabulary, sentence structure, and context in user dialogue using natural language processing technology. For example, using word vector models to convert words into vector representations, and calculating the similarity between vectors to determine the semantics of words; combining contextual information, using sequence labeling models to identify key information inquired by users, such as the scope of car insurance coverage, claims process, etc., thereby determining the user's potential demand for insurance products and purchase intention. Agent tone analysis: Using sentiment analysis technology to analyze the agent's tone in the dialogue. By establishing a tone lexicon, identifying positive, negative, or neutral words in agent speech, and combining information such as tone and speed of speech, determining whether the agent's tone is enthusiastic, professional, patient, or other types. User emotion change observation: Observing the user's emotional fluctuations during the dialogue using emotion recognition models. By analyzing users' language expressions (such as the emotional tone of their word choice) and changes in tone of voice (such as pitch and speaking speed), we can determine the user's emotional shift from initial doubt and hesitation to later interest and acceptance. Product matching analysis: Based on the user's needs and circumstances, and combined with the characteristics of the recommended product, we analyze the degree of matching between the product and the user. We can establish a product feature library and a user demand model, and evaluate the matching degree by calculating the similarity between the two. For example, for users with high-value vehicles who frequently drive long distances, we recommend car insurance products that include comprehensive coverage and high compensation. By comparing the product's coverage with the user's needs, we can determine the degree of matching. Furthermore, we integrate the user intent, agent tone, user emotional changes, and product matching degree obtained from the analysis to obtain corresponding text feature data.

[0028] Step S203: Based on a preset multimodal feature extraction strategy, feature extraction is performed on the real-time call data to obtain the corresponding feature data.

[0029] In this embodiment, the specific implementation process of extracting features from the real-time call data based on the preset multimodal feature extraction strategy to obtain the corresponding feature data will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0030] Step S204: The text feature data and the feature data are fused to obtain the corresponding comprehensive feature vector.

[0031] In this embodiment, the specific implementation process of fusing the text feature data and the feature data to obtain the corresponding comprehensive feature vector will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0032] Step S205: Based on the preset intelligent agent, the comprehensive feature vector is compared with the pre-built knowledge base to obtain the corresponding case matching result.

[0033] In this embodiment, the aforementioned intelligent agent is a comprehensive entity based on a pre-trained target large language model, possessing multiple functions such as receiving call text input (through an interface) and performing matching reasoning based on a knowledge base. Specifically, after obtaining the comprehensive feature vector, the intelligent agent performs a matching operation based on a pre-built knowledge base. The knowledge base stores a large amount of structured feature information, covering multiple dimensions such as user intent, agent tone, user emotional changes, and product matching degree. For example, the knowledge base may record cases and judgment criteria such as "when a user frequently inquires about product details and the agent's tone is patient and professional, it may be a suitable time to make a recommendation." Then, the intelligent agent compares the aforementioned comprehensive feature vector with the feature information in the knowledge base one by one. It searches the knowledge base for cases similar to the user's current intent, agent tone, and user emotions. For example, if a user frequently inquires about the coverage of a certain insurance product in the real-time call text, and the agent's tone is professional and patient, the intelligent agent will search the knowledge base for judgment criteria for recommending products in similar scenarios, to see if there are successful cases of recommending products in this situation. Through this matching, the agent can find the knowledge base entry that is closest to the current dialogue situation, and extract the key features of the corresponding successful or failed cases that are most similar to the current scenario, as well as the recommendation timing judgment results, as the case matching results, providing a reference for subsequent reasoning.

[0034] Furthermore, the specific construction process of the aforementioned knowledge base will be described in more detail in subsequent specific embodiments of this application, and will not be elaborated upon here.

[0035] Step S206: Based on the preset target large language model, perform reasoning processing on the comprehensive feature vector and the case matching result to generate the corresponding recommendation timing recognition result.

[0036] In this embodiment, the specific implementation process of reasoning on the comprehensive feature vector and the case matching result based on the preset target large language model to generate the corresponding recommendation timing identification result will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0037] Step S207: Output the recommendation timing identification result.

[0038] In this embodiment, after obtaining the recommendation timing identification result, the agent returns the result (in a simple and clear way, such as "Y" indicating a suitable recommendation timing and "N" indicating an unsuitable recommendation timing) to the system. The system then decides on subsequent actions based on the agent's response. If the response is "Y," the system promptly prompts the agent, for example, by displaying a small window showing recommended product information and a sales pitch. The agent can then recommend products to the user based on this prompt. This ensures that the right product is recommended to the user at the right time, improving the accuracy and effectiveness of the recommendation. If the response is "N," the system does not interfere with the communication between the agent and the user but continues to monitor the conversation. It continuously monitors the content of the user and agent's interaction, analyzes changes in user intent and emotions, and waits for the next suitable recommendation timing to occur. Once a new situation is found that meets the conditions for a recommendation timing, the agent performs reasoning again and returns a result, and the system then makes a corresponding decision.

[0039] Based on the above automated processing flow, this application collects real-time call data corresponding to the user during a call, then transforms the real-time call data into call text, extracts features from the call text to obtain text feature data, and uses a multimodal feature extraction strategy to extract features from the real-time call data to obtain feature data. Subsequently, the text feature data and feature data are fused to obtain a comprehensive feature vector. Then, based on the use of an intelligent agent, the comprehensive feature vector is compared with a pre-built knowledge base to obtain a case matching result. Following this, a target large language model is used to perform inference processing on the comprehensive feature vector and the case matching result to generate a recommendation timing recognition result. Finally, the recommendation timing recognition result is output. Thus, unlike existing recommendation timing recognition methods, this application, by using a target large language model, performing inference processing on the comprehensive feature vector fused with text feature data and feature data, and comparing it with a knowledge base to obtain the case matching result, can effectively improve the recognition efficiency and accuracy of recommendation timing recognition, ensuring the precision of the obtained recommendation timing recognition result.

[0040] In some alternative implementations, step S206 includes the following steps: Based on the target large language model, the comprehensive feature vector and the case matching result are integrated to obtain the corresponding input features.

[0041] In this embodiment, the target large language model (hereinafter referred to as the model) receives a fused comprehensive feature vector and relevant information from the case matching results at the input layer. The comprehensive feature vector includes features from various aspects such as user intent, agent tone, user emotional changes, product matching degree, voice intonation, pause frequency, and user reaction characteristics in the current scenario. The case matching results provide key features of the most similar successful or unsuccessful cases in the current scenario, as well as the recommendation timing judgment results. The target large language model integrates this information into a more comprehensive input representation (input features). For example, it can attach the key features from the case matching results to the comprehensive feature vector in the form of weights, or add the recommendation timing judgment results of the matched cases as an additional feature dimension. Furthermore, the input layer performs a preliminary transformation on the integrated input features, converting them into a numerical form that the model can process internally. For example, for textual features (such as user intent descriptions), word embedding technology is used to convert them into vector representations; for numerical features (such as product matching degree ratings), normalization processing is performed to adapt their range to the model's input requirements.

[0042] The input features are subjected to multi-layer feature transformation and deep analysis to obtain the corresponding target features.

[0043] In this embodiment, the aforementioned multi-layer feature transformation and deep analysis processing includes: shallow feature extraction: In the shallow hidden layer, the model mainly performs basic combinations and transformations on the input features to extract some intuitive and simple feature patterns. For example, it may be found that when the features of a user's positive emotion and a friendly agent's tone appear simultaneously, it is somewhat associated with a certain type of recommendation success. Through the computation of shallow neurons, the model can capture these basic and easily identifiable feature relationships.

[0044] Mid-level Feature Association: As the layers deepen, the mid-level hidden layers further uncover the complex relationships between features. It comprehensively considers the combined effects of multiple features, not just individual features. For example, in addition to considering user sentiment and agent tone, it also combines product relevance and the clarity of user intent to analyze the impact of different feature combinations on recommendation timing. For instance, when user sentiment is positive, agent tone is friendly, product relevance is high, and user intent is clear, the model will identify a strong correlation between this combination and successful recommendation.

[0045] Deep Feature Abstraction: In deep hidden layers, the model can capture more abstract and complex feature patterns. It learns potential, hard-to-observe patterns from massive amounts of data. For example, the model might discover seemingly unrelated feature combinations that actually have a significant impact on recommendation timing in specific business scenarios. Through deep nonlinear transformations, the model can map input features to a higher-dimensional feature space where different recommendation timings can be more clearly distinguished.

[0046] The target features are subjected to comprehensive reasoning processing to obtain the corresponding recommendation probability value.

[0047] In this embodiment, at the output layer, the target large language model uses the knowledge and patterns learned from previous layers, combined with knowledge base experience (case matching results), to perform comprehensive reasoning on the input target features and calculate a recommendation probability value. This probability value is derived from the model's analysis of the current scene features and knowledge base case features, as well as the complex relationship between various feature combinations learned internally by the model and the timing of recommendations. For example, the model comprehensively considers the weights and interrelationships of various features in the current scene and calculates the recommendation probability value through a complex function. This value is typically between 0 and 1; the closer the value is to 1, the more appropriate the recommendation timing.

[0048] The reasoning based on knowledge base experience includes: Case pattern matching: The model will conduct a more in-depth comparison and analysis of the features of the current scenario with the features of similar cases in the knowledge base. If a successful recommendation case is matched, the model will analyze the similarity and differences between the current scenario and the successful case in various features. For example, if it is found that the user's emotional level in the current scenario is slightly lower than that in a successful case, but the agent's tone is more professional, the model will comprehensively consider these factors to determine whether the current scenario still meets the conditions for a successful recommendation.

[0049] Failure Case Avoidance: When a failed recommendation case is encountered, the model focuses on the similarities between the current scenario and the failed case, and analyzes whether the factors that might have led to the recommendation failure exist in the current scenario or whether they can be mitigated. For example, in a failed case, the user is in a negative mood and has a misunderstanding of the product, while in the current scenario, the user is only slightly dissatisfied, and the agent has already provided a detailed explanation of the product. The model will determine that the current scenario is fundamentally different from the failed case, and thus more cautiously evaluate the timing of the recommendation.

[0050] Experience-based weighting: The model dynamically adjusts the weights of different features based on the success or failure experiences of cases in the knowledge base. Features that frequently appear in successful cases and significantly impact recommendation success are assigned higher weights; conversely, features that appear in failed cases and lead to recommendation failures are weighted lower. For example, if the knowledge base indicates that a user's understanding of the product is a key factor for successful recommendation, then in the current scenario, the model will focus on the user's product knowledge and increase its weight in the reasoning process.

[0051] The recommended probability value is compared with a preset recommended threshold.

[0052] In this embodiment, the system pre-sets a recommended threshold, which is determined based on a large amount of experimental data and business experience.

[0053] If the recommendation probability value is greater than or equal to the recommendation threshold, a first recommendation timing identification result is generated, indicating that the current time is a suitable time for recommendation.

[0054] In this embodiment, the recommendation probability value calculated by the target large language model is compared with a preset recommendation threshold. If the recommendation probability value is greater than or equal to the threshold, it is determined that the current time is a suitable time for recommendation, and a first recommendation time identification result is generated, such as returning the judgment result "Y" and prompting the agent to make a product recommendation.

[0055] If the recommendation probability value is less than the recommendation threshold, a second recommendation timing identification result is generated, indicating that the current timing is not suitable for recommendation.

[0056] In this embodiment, if the recommendation probability value is less than the recommendation threshold, it is determined that the current time is not a suitable time for recommendation, and a second recommendation time identification result is generated, which is not a suitable time for recommendation. For example, the judgment result "N" can be returned. At the same time, some suggestions can be provided based on the reasoning result, such as guiding users to further clarify their needs or adjust their communication strategies.

[0057] Based on the above processing flow, this application integrates the comprehensive feature vector and case matching results using a target large language model to obtain input features; then, it performs multi-layer feature transformation and deep analysis on the input features to obtain target features; subsequently, it performs comprehensive reasoning on the target features to obtain a recommendation probability value; and then compares the recommendation probability value with a preset recommendation threshold. If the recommendation probability value is detected to be greater than or equal to the recommendation threshold, a first recommendation timing identification result is generated indicating that the current recommendation timing is appropriate; otherwise, a second recommendation timing identification result is generated indicating that the current recommendation timing is not appropriate. Thus, this application obtains recommendation probability values ​​by reasoning from the fused comprehensive feature vector and knowledge base matching results using a target large language model, and then compares the recommendation probability values ​​based on a preset recommendation threshold. Based on the comparison results, corresponding recommendation timing identification results can be generated, effectively improving the generation efficiency of recommendation timing identification results and ensuring the accuracy of the obtained recommendation timing identification results.

[0058] In some optional implementations of this embodiment, step S203 includes the following steps: Invoke the preset speech analysis tool.

[0059] In this embodiment, the aforementioned voice analysis tool is a tool that has the function of analyzing the audio signal of call data to obtain the voice intonation, and analyzing the time series of call data to obtain the pause frequency.

[0060] The voice analysis tool extracts and processes the voice features from the real-time call data to obtain the corresponding voice intonation.

[0061] In this embodiment, the corresponding voice intonation can be obtained by analyzing the audio signals of the aforementioned real-time call data using voice analysis tools. Specifically, by utilizing audio processing techniques, such as extracting features like the fundamental frequency and amplitude of the audio, the pitch and intensity of the voice can be determined. For example, a high-pitched voice usually corresponds to emotions such as excitement and approval, while a low-pitched voice may indicate dissatisfaction or hesitation. This differs from semantic understanding, which focuses on the meaning expressed by the utterance, while voice intonation analysis focuses on the emotional state conveyed by the physical characteristics of the sound. Voice intonation can help determine the user's emotional state; different intonations may indicate different levels of acceptance of agent recommendations, providing an emotional basis for determining the timing of recommendations.

[0062] The real-time call data is processed by the voice analysis tool to extract time features and obtain the corresponding pause frequencies.

[0063] In this embodiment, the pause frequency can be obtained by analyzing the time series of the aforementioned real-time call data using voice analysis tools. Specifically, the number and duration of user pauses within a conversation are statistically analyzed, for example, calculating the number of pauses per minute. Longer pauses may indicate that the user is carefully considering the agent's recommendations, while frequent short pauses may reflect hesitation or uncertainty. This does not rely on semantic understanding but is based on observation of call duration characteristics. The pause frequency reflects the user's thought process and acceptance of the recommended content, helping to determine whether the user is in a suitable state of mind to accept the recommendations.

[0064] The real-time call data is processed to extract reaction features, thereby obtaining the corresponding user reaction features.

[0065] In this embodiment, the acquisition of the aforementioned user response features may include: For video call scenarios, computer vision technologies, such as facial recognition and expression analysis, can be used to identify the user's nodding, smiling, frowning, and other facial expressions and actions. For example, a nod may indicate agreement with the agent's viewpoint, and a smile may indicate a positive attitude. For audio call scenarios, the user's verbal responses can be analyzed, such as positive feedback like "Hmm, I understand" or "That sounds good," or negative or hesitant feedback like "I'll think about it" or "I'm not sure." The identification of these verbal responses is not entirely equivalent to interpreting the complete meaning of a sentence in semantic understanding; rather, it focuses on capturing the user's immediate reaction to the recommended content. The user response features are used to intuitively reflect the user's attitude towards the agent's recommendations, assisting in determining whether the timing of the recommendation is appropriate.

[0066] The speech tone, the pause frequency, and the user response features are integrated to obtain the corresponding feature data.

[0067] In this embodiment, integrated features can be obtained by integrating the extracted speech intonation, pause frequency and user response features, and these integrated features can be used as the corresponding feature data.

[0068] Based on the above processing flow, this application calls a preset voice analysis tool; then, based on the use of the voice analysis tool, it extracts and processes sound features from real-time call data to obtain voice intonation; and based on the use of the voice analysis tool, it extracts and processes time features from real-time call data to obtain pause frequency; then, it extracts and processes reaction features from real-time call data to obtain user reaction features; and finally, it integrates voice intonation, pause frequency, and user reaction features to obtain the corresponding feature data. This enables efficient and accurate feature extraction and processing of real-time call data from multiple dimensions, improving the diversity and accuracy of the obtained feature data.

[0069] In some alternative implementations, step S204 includes the following steps: Obtain the preset feature fusion strategy.

[0070] In this embodiment, the selection of the aforementioned feature fusion strategy is not specifically limited and can be determined according to actual business needs. It can be implemented using weighted averaging, machine learning algorithms (such as neural networks), or attention mechanisms. For example, different features are assigned corresponding weights based on their importance in determining the timing of recommendations, and then a comprehensive feature vector is obtained by weighted summation of the different features based on these weights.

[0071] The text features and the feature data are fused based on the feature fusion strategy to obtain the corresponding fused features.

[0072] In this embodiment, since features of different modalities have different feature representations and time scales, a time alignment method can be used to match the features of different modalities in time. Then, according to the selected feature fusion strategy, the aligned text features and aligned feature data are fused, and the resulting fused features are used as the final comprehensive feature vector.

[0073] The fused features are used as the comprehensive feature vector.

[0074] In this embodiment, the aforementioned text features are obtained by analyzing the call text and understanding the meaning and intent of the user's words, including information such as user intent, agent tone, user emotional changes, and product suitability. Multimodal information such as voice tone, pause frequency, and user response characteristics are features obtained from sound features, time features, and the user's nonverbal behavior. By fusing multimodal features (feature data) with semantic understanding results (text features), more comprehensive and accurate information can be provided for determining the timing of recommendations. For example, if semantic understanding detects that a user inquires about the scope of car insurance coverage, indicating a certain purchase intention, and that the user's voice tone is positive, pause frequency is moderate, and the user exhibits positive responses such as smiling, then this is a more suitable time to make a recommendation.

[0075] Based on the above processing flow, this application obtains a preset feature fusion strategy; then, based on the use of the feature fusion strategy, it performs fusion processing on text features and feature data, and uses the obtained fused features as the required comprehensive feature vector. This allows for efficient and accurate fusion processing of text features and feature data, thereby providing more comprehensive and accurate information for recommendation timing judgment and effectively improving the comprehensiveness and accuracy of the generated comprehensive feature vector.

[0076] In some alternative implementations, prior to step S206, the electronic device may also perform the following steps: Obtain pre-collected call data.

[0077] In this embodiment, initial call data from experienced and skilled agents can be selected from telephone-based insurance business scenarios (such as auto insurance), with a focus on selecting successfully recommended product dialogue data. This initial call data contains rich information such as dialogue content, tone of voice, pause frequency, and user responses. Furthermore, the collected initial call data is preprocessed to obtain corresponding call data. Specifically, for text data, operations such as word segmentation, stop word removal, and stemming are performed to convert it into a numerical vector form that the model can process. For non-text data such as tone of voice and pause frequency, quantification is performed; for example, tone of voice is divided into high, medium, and low levels and assigned corresponding numerical values, and pause frequency is counted according to the number of pauses per minute.

[0078] Based on a preset feature extraction strategy, feature extraction is performed on the call data to obtain key features related to product recommendations.

[0079] In this embodiment, the above-mentioned process of extracting features from the call data based on a preset feature extraction strategy to obtain key features related to product recommendations can be referred to in detail in the aforementioned process of extracting features from the call text to obtain corresponding text feature data, which will not be elaborated on here.

[0080] Call the preset pre-trained large language model.

[0081] In this embodiment, a suitable pre-trained large language model, such as GLM, Qwen, DeepSeek, or PingAnGPT, can be selected based on actual needs and resource availability. These pre-trained models have been trained on a large amount of general text data and possess a certain level of language understanding and generation capabilities, providing a solid foundation for subsequent fine-tuning in the auto insurance field. The selected pre-trained large language model can capture the semantic coherence and intent evolution path of customer dialogue through contextual modeling techniques, avoiding incorrect recommendation timing due to misjudgment of local information. Furthermore, the training objectives and metrics are defined: the training objective is to enable the model to accurately determine the recommendation timing of auto insurance products based on key features such as user intent, agent tone, user emotional changes, and product matching degree. Evaluation metrics, such as precision, recall, and F1 score, are determined to measure the model's performance in the recommendation timing judgment task.

[0082] Based on the call data and the key features, the pre-trained large language model is fine-tuned to obtain the corresponding first generative model.

[0083] In this embodiment, the fine-tuning training includes: integrating key features into the call data: extracting key features and incorporating them into the call data. For example, for each call data segment, in addition to the text content, it is labeled with feature tags such as user intent (explicit, ambiguous), agent tone (enthusiastic, professional, indifferent), user emotional changes (positive, negative, neutral), and product matching degree (high, medium, low). This allows the model to learn the relationship between these features and recommendation timing during training. Designing a loss function: designing a suitable loss function based on the training objective to measure the difference between the model's prediction and the true labels. For example, the cross-entropy loss function can be used to measure the difference between the model's predicted recommendation timing probability and the actual recommendation label. The model parameters are optimized by minimizing the loss function. Model training: training the pre-trained large language model using the fine-tuned call data to obtain a trained first generative model. Specifically, the backpropagation algorithm is used to update the model parameters based on the gradient information of the loss function. During training, appropriate training epochs and learning rates can be set to avoid overfitting or underfitting. For example, the initial learning rate can be set to 0.001, and as training progresses, a learning rate decay strategy can be adopted to gradually reduce the learning rate in order to improve the convergence of the model.

[0084] The first generative model is evaluated and iteratively optimized based on preset verification data to obtain the corresponding second generative model.

[0085] In this embodiment, a specified proportion of data can be randomly selected from the aforementioned call data to serve as the verification data. The specific value of this specified proportion is not limited and can be set according to actual business needs; for example, 0.6 could be used.

[0086] In addition, the above-mentioned performance evaluation and iterative optimization process includes: Evaluating model performance: Using pre-determined evaluation metrics, the trained first-generation model is evaluated on validation data. Precision, recall, F1 score, and other metrics are calculated to analyze the model's ability to determine recommendation timing under different feature combinations. For example, the proportion of times the model correctly determines the recommendation timing when the user intent is clear, the agent's tone is positive, the user's emotions are positive, and the product match is high is statistically analyzed. Analyzing error cases: In-depth analysis is conducted on cases where the model predicts incorrectly to identify the causes of the errors. For example, some features may be inaccurately extracted, or the model may have a bias in its understanding of certain feature combinations. Analyzing error cases provides direction for further model optimization. Iterative model optimization: Based on the evaluation results and error case analysis, the model is iteratively optimized. Specifically, the model structure can be adjusted, such as increasing or decreasing the number of neurons in the hidden layers, changing the activation function, etc.; the training data can also be further optimized, such as adding more diverse cases, correcting feature labeling errors, etc. By continuously iterating and optimizing, the accuracy of the model's recommendation timing judgment is improved until a second generative model with a recommendation timing judgment accuracy greater than the preset accuracy threshold is obtained, and the obtained second generative model is used as the final target large language model.

[0087] The second generative model is used as the target large language model.

[0088] Based on the above processing flow, this application acquires pre-collected call data; then, based on a preset feature extraction strategy, it extracts features from the call data to obtain key features related to product recommendations; subsequently, it calls a pre-trained large language model; and based on the combination of call data and key features, it fine-tunes and trains the pre-trained large language model to obtain a first generative model; subsequently, based on the use of validation data, it evaluates the effectiveness of the first generative model and iteratively optimizes it, and uses the obtained second generative model as the required target large language model. This enables the automatic and intelligent construction of a target large language model with the ability to accurately identify recommendation timing, improves the model construction efficiency of the target large language model, and ensures the model effect of the obtained target large language model.

[0089] In some optional implementations of this embodiment, after the step of extracting features from the call data based on a preset feature extraction strategy to obtain key features related to product recommendations, the electronic device may further perform the following steps: The key features are then structured to obtain the corresponding recommended feature data.

[0090] In this embodiment, the above-mentioned structured processing includes: using JSON format to organize key features such as user intent, agent tone, user emotional changes, and product matching degree in the form of key-value pairs, so that each feature has a clear identifier and value.

[0091] The recommended feature data is supplemented with case information to obtain the corresponding target recommended feature data.

[0092] In this embodiment, the target recommendation feature data can be obtained by supplementing the obtained recommendation feature data with corresponding case information (such as successful cases or failed cases) and recommendation timing judgment results.

[0093] Call the preset database.

[0094] In this embodiment, the design of the database includes: designing a suitable database structure based on the structured target recommendation feature data. For example, a relational database (such as MySQL) or a non-relational database (such as MongoDB) can be selected, depending on the characteristics of the data and access requirements. When designing database tables, fields and data types are defined appropriately to ensure accurate data storage and efficient querying.

[0095] The target recommendation feature data is stored in the database to obtain the corresponding target database.

[0096] In this embodiment, the structured target recommendation feature data is stored in a pre-designed database to obtain the corresponding target database, which serves as the required knowledge base. During storage, attention must be paid to data integrity and consistency to avoid data loss or incorrect storage. Specifically, batch insertion or real-time insertion can be used, selected based on the data volume and real-time requirements.

[0097] The target database is used as the knowledge base.

[0098] In this embodiment, the knowledge base is regularly maintained and updated to ensure that the data in the knowledge base is always up-to-date and accurate. Furthermore, a data auditing mechanism is established to review newly stored data, promptly identifying and correcting erroneous data.

[0099] Based on the above processing flow, this application obtains recommended feature data by structuring key features; then, it supplements the recommended feature data with case information to obtain corresponding target recommended feature data; subsequently, it calls a preset database and stores the target recommended feature data in the database, using the resulting target database as the required knowledge base. This enables the automatic and intelligent construction of the knowledge base. The use of the knowledge base allows recommended feature data to be organized in a clear format, forming a real-time accessible knowledge base. This facilitates the rapid retrieval of knowledge from the knowledge base during actual user communication to support real-time recommendation timing judgments, thereby improving the system's response speed and recommendation quality.

[0100] In some optional implementations of this embodiment, after step S207, the electronic device may further perform the following steps: Get newly added call data collected periodically.

[0101] In this embodiment, as more successful referral cases from top agents are collected, the system periodically collects new data, namely new call data. Specifically, this can be achieved by monitoring insurance business, such as the operational scenarios of car insurance telephone lines, to promptly identify new successful referral cases and collect relevant call data and user information.

[0102] Based on a preset target training strategy, the newly added call data is used to retrain the target large language model to obtain the corresponding specified large language model.

[0103] In this embodiment, the target training strategy includes: integrating newly collected call data with existing call data to obtain integrated call data, and then retraining the target large language model using this integrated call data. During the retraining process, an incremental learning method is specifically employed to fine-tune the original target large language model, thereby improving training efficiency and effectiveness. Specifically, based on the strategy outlined above, the newly collected call data can be used to retrain the target large language model to obtain a well-trained specified large language model.

[0104] The specified large language model is evaluated based on a pre-defined evaluation dataset.

[0105] In this embodiment, a target proportion of data can be randomly selected from the integrated call data to serve as the evaluation dataset. The value of the target proportion is not specifically limited and can be set according to actual business needs; for example, 0.6 could be used. Then, the retrained specified large language model is evaluated using the evaluation dataset to test its performance in the new data environment, thereby obtaining a corresponding performance evaluation result. If the performance meets expectations, a performance evaluation result indicating that the specified large language model has passed the evaluation is generated; otherwise, a performance evaluation result indicating that the specified large language model has failed the evaluation is generated.

[0106] Based on the evaluation results, the model will be further adjusted and optimized, such as adjusting the model parameters and improving the feature fusion method, to improve the accuracy and adaptability of recommendation timing judgment.

[0107] If the specified large language model passes the evaluation, the specified large language model is used to update the target large language model.

[0108] In this embodiment, if the specified large language model passes the evaluation, it is directly used to replace the original target large language model. This allows for subsequent recommendation timing identification processing using the specified large language model, thus completing the model update process for the target large language model. Conversely, if the specified large language model fails the evaluation, the performance evaluation results are used to further adjust and optimize the model. This includes adjusting model parameters and improving feature fusion methods to enhance the accuracy and adaptability of recommendation timing judgment until a model with expected performance is obtained.

[0109] Based on the above processing flow, this application acquires periodically collected new call data; then, based on a preset target training strategy, it retrains the target large language model using the new call data to obtain a specified large language model; subsequently, it evaluates the specified large language model using a preset evaluation dataset; if the specified large language model passes the evaluation, it updates the target large language model using the specified large language model. This enables dynamic updating and continuous optimization of the target large language model, allowing the recommendation strategy to continuously adapt to new situations, improving the accuracy and adaptability of the recommendation strategy, thereby maintaining the system's competitiveness.

[0110] In some alternative implementations, the user information obtained is subject to user consent and complies with relevant laws and policies.

[0111] Furthermore, any software tools or components not belonging to our company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0112] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0113] It should be emphasized that, in order to further ensure the privacy and security of the above-mentioned recommendation timing identification results, the above-mentioned recommendation timing identification results can also be stored in a node of a blockchain.

[0114] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0115] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by 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 accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0116] Further reference Figure 3 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of an artificial intelligence-based data recognition device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0117] like Figure 3As shown, the artificial intelligence-based data recognition device 300 described in this embodiment includes: a data acquisition module 301, a first processing module 302, a first extraction module 303, a fusion module 304, a comparison module 305, an inference module 306, and an output module 307. Wherein: The acquisition module 301 is used to acquire real-time call data corresponding to the user during a call with the user; The first processing module 302 is used to convert and process the real-time call data to obtain the corresponding call text, and to extract features from the call text to obtain the corresponding text feature data. The first extraction module 303 is used to extract features from the real-time call data based on a preset multimodal feature extraction strategy to obtain corresponding feature data. The fusion module 304 is used to fuse the text feature data and the feature data to obtain the corresponding comprehensive feature vector; The comparison module 305 is used to compare the comprehensive feature vector with a pre-built knowledge base based on a preset intelligent agent to obtain the corresponding case matching result; The reasoning module 306 is used to perform reasoning processing on the comprehensive feature vector and the case matching result based on the preset target large language model, and generate the corresponding recommendation timing recognition result; The output module 307 is used to output the recommendation timing identification result.

[0118] In some optional implementations of this embodiment, the inference module 306 includes: The first integration submodule is used to integrate the comprehensive feature vector and the case matching result based on the target large language model to obtain the corresponding input features; The processing submodule is used to perform multi-layer feature transformation and depth analysis on the input features to obtain the corresponding target features; The reasoning submodule is used to perform comprehensive reasoning processing on the target features to obtain the corresponding recommendation probability value; The comparison submodule is used to compare the recommendation probability value with a preset recommendation threshold. The first generation submodule is used to generate a first recommendation timing identification result if the recommendation probability value is greater than or equal to the recommendation threshold. The second generation submodule is used to generate a second recommendation timing identification result if the recommendation probability value is less than the recommendation threshold.

[0119] In some optional implementations of this embodiment, the first extraction module 303 includes: Calling submodules is used to invoke preset speech analysis tools; The first extraction submodule is used to extract sound features from the real-time call data based on the voice analysis tool to obtain the corresponding voice intonation. The second extraction submodule is used to extract time features from the real-time call data based on the voice analysis tool to obtain the corresponding pause frequency. The third extraction submodule is used to extract reaction features from the real-time call data to obtain the corresponding user reaction features. The second integration submodule is used to integrate the voice intonation, the pause frequency, and the user response features to obtain the corresponding feature data.

[0120] In some optional implementations of this embodiment, the fusion module 304 includes: The acquisition submodule is used to acquire preset feature fusion strategies; The fusion submodule is used to fuse the text features and the feature data based on the feature fusion strategy to obtain the corresponding fused features; A submodule is defined to use the fused features as the comprehensive feature vector.

[0121] In some optional implementations of this embodiment, the artificial intelligence-based data recognition device further includes: The first acquisition module is used to acquire pre-collected call data; The second extraction module is used to extract features from the call data based on a preset feature extraction strategy to obtain key features related to product recommendations. The first calling module is used to call the preset pre-trained large language model; The first training module is used to fine-tune the pre-trained large language model based on the call data and the key features to obtain the corresponding first generative model. The second processing module is used to evaluate and iteratively optimize the first generative model based on preset verification data to obtain the corresponding second generative model. The first determining module is used to use the second generated model as the target large language model.

[0122] In some optional implementations of this embodiment, the artificial intelligence-based data recognition device further includes: The third processing module is used to perform structured processing on the key features to obtain the corresponding recommendation feature data; The supplementary module is used to supplement the recommended feature data with case information to obtain the corresponding target recommended feature data; The second calling module is used to call the preset database; The storage module is used to store the target recommendation feature data into the database to obtain the corresponding target database; The second determining module is used to use the target database as the knowledge base.

[0123] In some optional implementations of this embodiment, the artificial intelligence-based data recognition device further includes: The second acquisition module is used to acquire newly added call data collected periodically; The second training module is used to retrain the target large language model using the newly added call data based on a preset target training strategy, so as to obtain the corresponding specified large language model. The evaluation module is used to evaluate the specified large language model based on a preset evaluation dataset; An update module is used to update the target large language model using the specified large language model if the specified large language model passes the evaluation.

[0124] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0125] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0126] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0127] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions based on artificial intelligence data recognition methods. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0128] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions of the artificial intelligence-based data recognition method.

[0129] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0130] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the artificial intelligence-based data recognition method described above.

[0131] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A data recognition method based on artificial intelligence, characterized in that, Includes the following steps: During a call with a user, real-time call data corresponding to that user is collected. The real-time call data is transformed to obtain the corresponding call text, and the call text is then subjected to feature extraction to obtain the corresponding text feature data. Based on a preset multimodal feature extraction strategy, features are extracted from the real-time call data to obtain the corresponding feature data; The text feature data and the feature data are fused together to obtain the corresponding comprehensive feature vector; Based on the preset intelligent agent, the comprehensive feature vector is compared with the pre-built knowledge base to obtain the corresponding case matching result; Based on the preset target large language model, the comprehensive feature vector and the case matching result are subjected to reasoning processing to generate the corresponding recommendation timing recognition result; The recommendation timing identification results are then processed for output.

2. The data recognition method based on artificial intelligence according to claim 1, characterized in that, The step of performing reasoning processing on the comprehensive feature vector and the case matching result based on the preset target large language model to generate the corresponding recommendation timing recognition result specifically includes: Based on the target large language model, the comprehensive feature vector and the case matching result are integrated and processed to obtain the corresponding input features; The input features are subjected to multi-layer feature transformation and depth analysis to obtain the corresponding target features; The target features are subjected to comprehensive reasoning processing to obtain the corresponding recommendation probability value; The recommendation probability value is compared with a preset recommendation threshold; If the recommendation probability value is greater than or equal to the recommendation threshold, a first recommendation timing identification result is generated indicating that the current time is a suitable time for recommendation. If the recommendation probability value is less than the recommendation threshold, a second recommendation timing identification result is generated, indicating that the current timing is not suitable for recommendation.

3. The data recognition method based on artificial intelligence according to claim 1, characterized in that, The step of extracting features from the real-time call data based on a preset multimodal feature extraction strategy to obtain corresponding feature data specifically includes: Invoke the preset voice analysis tool; Based on the aforementioned voice analysis tool, the real-time call data is processed to extract sound features and obtain the corresponding voice intonation. The real-time call data is processed by the voice analysis tool to extract time features and obtain the corresponding pause frequency. The real-time call data is processed to extract reaction features to obtain the corresponding user reaction features; The speech tone, the pause frequency, and the user response features are integrated to obtain the corresponding feature data.

4. The data recognition method based on artificial intelligence according to claim 1, characterized in that, The step of fusing the text feature data with the feature data to obtain the corresponding comprehensive feature vector specifically includes: Obtain the preset feature fusion strategy; The text features and the feature data are fused based on the feature fusion strategy to obtain the corresponding fused features; The fused features are used as the comprehensive feature vector.

5. The data recognition method based on artificial intelligence according to claim 1, characterized in that, Before the step of performing inference processing on the comprehensive feature vector and the case matching result based on the preset target large language model to generate the corresponding recommendation timing recognition result, the method further includes: Obtain pre-collected call data; Based on a preset feature extraction strategy, feature extraction is performed on the call data to obtain key features related to product recommendations. Call the preset pre-trained large language model; Based on the call data and the key features, the pre-trained large language model is fine-tuned to obtain the corresponding first generative model. The first generative model is evaluated and iteratively optimized based on the preset verification data to obtain the corresponding second generative model. The second generative model is used as the target large language model.

6. The data recognition method based on artificial intelligence according to claim 5, characterized in that, After the step of extracting features from the call data based on a preset feature extraction strategy to obtain key features related to product recommendations, the method further includes: The key features are then processed in a structured manner to obtain the corresponding recommendation feature data; The recommended feature data is supplemented with case information to obtain the corresponding target recommended feature data; Call the preset database; The target recommendation feature data is stored in the database to obtain the corresponding target database; The target database is used as the knowledge base.

7. The data recognition method based on artificial intelligence according to claim 1, characterized in that, After the step of outputting the recommendation timing identification result, the method further includes: Obtain newly added call data collected periodically; Based on the preset target training strategy, the target large language model is retrained using the newly added call data to obtain the corresponding specified large language model. The specified large language model is evaluated based on a pre-defined evaluation dataset; If the specified large language model passes the evaluation, the target large language model is updated using the specified large language model.

8. A data recognition device based on artificial intelligence, characterized in that, include: The data acquisition module is used to collect real-time call data corresponding to the user during a call. The first processing module is used to transform and process the real-time call data to obtain the corresponding call text, and to extract features from the call text to obtain the corresponding text feature data. The first extraction module is used to extract features from the real-time call data based on a preset multimodal feature extraction strategy to obtain corresponding feature data. The fusion module is used to fuse the text feature data and the feature data to obtain the corresponding comprehensive feature vector; The comparison module is used to compare the comprehensive feature vector with a pre-built knowledge base based on a preset intelligent agent to obtain the corresponding case matching result; The reasoning module is used to perform reasoning processing on the comprehensive feature vector and the case matching result based on the preset target large language model, and generate the corresponding recommendation timing recognition result. The output module is used to process the recommendation timing identification results.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the data recognition method based on artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the data recognition method based on artificial intelligence as described in any one of claims 1 to 7.

Images