AI platform data recommendation method and device and computer equipment

Abstract

The invention provides an AI platform data recommendation method and device and computer equipment, and relates to the technical field of data processing, and the method comprises the steps: obtaining a historical consultation dialogue record set of a target user, carrying out progressive multi-round dialogue iterative memory analysis, and determining the understanding category of the target user; obtaining a real-time active reasoning dialogue sequence; when a backtracking detection module is called to carry out recommended data completeness detection on the real-time active reasoning dialogue sequence, a recommended data completeness detection result is obtained; and triggering a recommendation output instruction according to the recommendation data completeness coefficient, and pushing recommendation data based on the recommendation output instruction. The data recommendation method and device have the advantages that the technical problems that in the data recommendation process in the prior art, in-depth interaction with the user cannot be dynamically conducted, initiative questioning in accordance with the actual situation of the user cannot be achieved, the recommendation quality is low, and the real-time requirement of the user cannot be met easily are solved, and the technical effect of improving the data recommendation quality is achieved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to an AI platform data recommendation method, apparatus, and computer equipment. Background Technology

[0002] Current platform data recommendations primarily rely on predictive methods based on user behavior and preferences. However, when faced with complex recommendation tasks, especially when user needs and intentions are unclear or information is incomplete, they often depend on simple content-based or collaborative filtering algorithms, lacking proactive reasoning and dynamic adaptability. During the data recommendation process, misleading guidance or misinterpretation of user questions frequently leads to misjudging user direction and providing data that does not meet their needs. Furthermore, recommendation strategies often involve repeatedly asking templated questions, while user circumstances are frequently dynamic, significantly reducing the reliability of data recommendations.

[0003] Existing technologies suffer from the technical problem that, during the data recommendation process, it is impossible to dynamically engage in in-depth interaction with users and proactively ask questions that are tailored to the user's actual situation. This results in low recommendation quality and an inability to adapt to the user's real-time needs. Summary of the Invention

[0004] This application provides an AI platform data recommendation method, apparatus, and computer device to address the technical problem that existing technologies, in the data recommendation process, are unable to dynamically engage in in-depth interaction with users and proactively ask questions that are tailored to users' actual situations, resulting in low recommendation quality and difficulty in adapting to users' real-time needs.

[0005] In view of the above problems, this application provides an AI platform data recommendation method, apparatus and computer device.

[0006] The first aspect of this application provides an AI platform data recommendation method, the method comprising: acquiring a set of historical consultation dialogue records of a target user; performing progressive multi-turn dialogue iterative memory parsing on the set of historical consultation dialogue records to determine the target user's comprehension scope; acquiring real-time consultation information of the target user; using the target user's comprehension scope as a constraint, the AI ​​platform performs active reasoning based on the real-time consultation information to obtain a real-time active reasoning dialogue sequence, wherein the AI ​​platform is embedded with a deep learning model; when the number of dialogue turns in the real-time active reasoning dialogue sequence is greater than or equal to a preset dialogue turn threshold, calling a backtracking detection module to perform recommendation data completeness detection on the real-time active reasoning dialogue sequence to obtain a recommendation data completeness detection result, wherein the recommendation data completeness detection result includes a recommendation data completeness coefficient and recommendation data; triggering a recommendation output instruction based on the recommendation data completeness coefficient, and pushing the recommendation data based on the recommendation output instruction.

[0007] A second aspect of this application provides an AI platform data recommendation device, the device comprising: a target user comprehension category determination module, configured to acquire a set of historical consultation dialogue records of the target user, perform progressive multi-turn dialogue iterative memory parsing on the set of historical consultation dialogue records to determine the target user comprehension category; a real-time proactive reasoning dialogue sequence acquisition module, configured to acquire real-time consultation information of the target user, and, constrained by the target user comprehension category, the AI ​​platform performs proactive reasoning based on the real-time consultation information to obtain a real-time proactive reasoning dialogue sequence, wherein the AI ​​platform embeds a deep learning model; a recommendation data completeness detection result acquisition module, configured to, when the number of dialogue turns in the real-time proactive reasoning dialogue sequence is greater than or equal to a preset dialogue turn threshold, call a backtracking detection module to perform recommendation data completeness detection on the real-time proactive reasoning dialogue sequence to obtain a recommendation data completeness detection result, wherein the recommendation data completeness detection result includes a recommendation data completeness coefficient and recommendation data; and a recommendation information push module, configured to trigger a recommendation output instruction based on the recommendation data completeness coefficient, and push the recommendation data based on the recommendation output instruction. A third aspect of this application discloses a computer device comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement any step of the first aspect of this application.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application obtains a set of historical consultation dialogue records of the target user, performs progressive multi-turn dialogue iterative memory parsing on the historical consultation dialogue record set to determine the target user's comprehension scope, and then obtains the target user's real-time consultation information. Using the target user's comprehension scope as a constraint, the AI ​​platform performs proactive reasoning based on the real-time consultation information to obtain a real-time proactive reasoning dialogue sequence. The AI ​​platform embeds a deep learning model. When the number of dialogue turns in the real-time proactive reasoning dialogue sequence is greater than or equal to a preset dialogue turn threshold, a backtracking detection module is invoked to perform a recommendation data completeness check on the real-time proactive reasoning dialogue sequence, obtaining a recommendation data completeness check result. This result includes a recommendation data completeness coefficient and recommended data. Then, a recommendation output command is triggered based on the recommendation data completeness coefficient, and recommended data is pushed based on the recommendation output command. This achieves the technical effect of improving the reliability and accuracy of data recommendation. Attached Figure Description

[0009] Figure 1 This application provides a schematic diagram of an AI platform data recommendation method.

[0010] Figure 2 This application provides a schematic diagram of the structure of an AI platform data recommendation device.

[0011] Figure 3 This is a schematic diagram of a zero-knowledge proof document cloud transfer device provided in an embodiment of this application.

[0012] Explanation of reference numerals in the attached diagram: Target user understanding scope determination module 11, real-time proactive reasoning dialogue sequence acquisition module 12, recommendation data completeness detection result acquisition module 13, recommendation information push module 14, processor 21, memory 22, input device 23, output device 24. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0014] Example 1, as Figure 1 As shown, this application provides a data recommendation method for an AI platform, the method comprising: Step S100: Obtain the set of historical consultation dialogue records of the target user, and perform progressive multi-round dialogue iterative memory parsing on the set of historical consultation dialogue records to determine the target user's understanding scope; Furthermore, the historical consultation dialogue record set of the target user is obtained, and progressive multi-turn dialogue iterative memory parsing is performed on the historical consultation dialogue record set to determine the target user's understanding scope. Step S100 in this embodiment of the application also includes: Step S110: Perform progressive understanding and memory parsing on the historical consultation dialogue record set to obtain the target user's progressive understanding and memory set; Step S120: Perform multi-round iterative understanding granularity analysis based on the target user's progressive understanding memory set to determine the target user's understanding granularity feature vector; Step S130: Based on the target user's understanding granularity feature vector, perform understanding category identification to determine the target user's understanding category.

[0015] In one possible embodiment, the historical consultation dialogue record set is a collection of all dialogue records between the target user and the AI ​​platform within a historical period, including all questions raised by the target user in different consultation scenarios and the AI ​​platform's responses. By analyzing the understanding of each record in the historical consultation dialogue record set, and analyzing the user's overall depth and ability of understanding the technology, such as the degree of understanding of the questioner, and whether the AI ​​platform needs to explain in plain language or provide explanations of specialized terms, the target user's comprehension range can be obtained.

[0016] Preferably, for each historical consultation dialogue record in the historical consultation dialogue record set, the target user's questioning situation in the communication dialogue is analyzed from two aspects: the progression of user intent and the progression of user emotion. Redundant data is filtered out from the historical consultation dialogue record set, and key data is extracted to obtain the target user's progressive understanding memory set.

[0017] Furthermore, after obtaining the target user's progressive understanding memory set—that is, after obtaining the target user's key data in different consultation scenarios—multiple rounds of iterative understanding granularity analysis are performed on the target user's progressive understanding memory set. By analyzing the target user's understanding granularity in each communication with the AI ​​platform, and the iterative development of the overall understanding granularity after multiple communications, the target user's understanding granularity feature vector is determined. Preferably, the target user's understanding granularity feature vector reflects the target user's understanding level and depth of knowledge understanding. Optionally, the target user's understanding granularity feature vector includes domain understanding depth features, demand clarity features, and emotional feedback features. Preferably, emotional feedback features include features such as doubt, agreement, skepticism, and questioning.

[0018] In one embodiment, multiple sample user understanding granularity feature vectors are obtained, and corresponding multiple sample user understanding categories are obtained as training data. The training data is used to supervise the training of a framework built on a feedforward neural network. During training, constraints are analyzed using a cross-entropy loss function. When the loss meets the requirements, a trained understanding category recognizer is obtained. The target user understanding granularity feature vectors are input into the understanding category recognizer for parsing to obtain the target user's understanding category. Through progressive analysis, dynamic parsing, and granular recognition of user historical dialogues, the target user's needs and personal knowledge level are analyzed. Furthermore, even when subsequent information is insufficient, accurate recommendation data can be provided through proactive reasoning.

[0019] Furthermore, the historical consultation dialogue record set is parsed using a progressive understanding and memory analysis within the record to obtain the target user's progressive understanding and memory set. In this embodiment, step S110 further includes: Step S111: Extract the first historical consultation question from the first historical consultation dialogue record set; Step S112: Based on preset user intent keywords and preset user sentiment keywords, perform progressive understanding and memory analysis on the first historical consultation question to obtain the first initial target user's progressive understanding and memory; Step S113: Based on the progressive understanding memory of the first initial target user, perform progressive understanding memory analysis on the second historical consultation question extracted from the first historical consultation dialogue record, and update the progressive understanding memory of the first initial target user according to the analysis result to obtain the progressive understanding memory of the second initial target user, and so on, to obtain the progressive understanding memory of the first target user. Step S114: Add the first target user's progressive understanding memory to the target user's progressive understanding memory set.

[0020] Furthermore, based on preset user intent keywords and preset user sentiment keywords, a progressive understanding and memory analysis is performed on the first historical consultation question to obtain the progressive understanding and memory of the first initial target user. In this embodiment, step S112 further includes: Step S112-1: Extract keywords from the first historical consultation question to obtain the first historical consultation question keyword set; According to step S112-2: extract historical consultation question keywords in the first set of historical consultation question keywords whose similarity meets the preset similarity threshold, and obtain the first set of user intent keywords; Step S112-3: Extract historical consultation question keywords from the first set of historical consultation question keywords whose similarity meets the preset similarity threshold according to preset user sentiment keywords, and obtain the first set of user sentiment keywords; Step S112-4: Add the preset user intent keywords and the first user intent keyword set to the first user intent memory unit, and add the preset user emotion keywords and the first user emotion keywords to the first user emotion memory unit. Use the first user intent memory unit and the first user emotion memory unit as the first initial target user progressive understanding memory.

[0021] In one possible embodiment, the first historical consultation dialogue record is any one of the historical consultation dialogue records in the set of historical consultation dialogue records, where "first" is a general term. Taking the first historical consultation dialogue record as the analysis object, the first historical consultation question is extracted from it. This first historical consultation question reflects the target user's consultation objective. Furthermore, preset user intent keywords and preset user sentiment keywords, pre-defined by those skilled in the art, are obtained. The preset user intent keywords are keywords conveyed by the user through questioning in the dialogue, such as "check price" or "product function." The preset user sentiment keywords refer to keywords reflecting emotional tendencies in the user's questions or responses, such as "like," "dissatisfied," "affirmative," "no," or "yes."

[0022] In one embodiment, the first historical consultation question is segmented into individual words or phrases using NLP tools. Then, Stanford NLP is used to perform part-of-speech tagging on the segmented words, and the Word2Vec model is used to generate vector representations for each segmented word, converting the words into low-dimensional word vectors. The most frequent keywords in the segmented word vectors are extracted by word frequency analysis to obtain the keyword set for the first historical consultation question.

[0023] Then, the similarity between the preset user intent keywords and each historical consultation question keyword in the first historical consultation question keyword set is calculated using the cosine similarity function. Historical consultation question keywords that meet the similarity threshold preset by those skilled in the art are extracted from the calculation results to obtain the first user intent keyword set. Based on the same principle, the preset user sentiment keywords are used to extract sentiment-related keywords from the first historical consultation question keyword set to obtain the first user sentiment keyword set.

[0024] The preset user intent keywords and the first set of user intent keywords are added to an initially empty vector to obtain the first user intent memory unit. Simultaneously, the preset user emotion keywords and the first set of user emotion keywords are added to an initially empty vector to obtain the first user emotion memory unit. The first user intent memory unit and the first user emotion memory unit are used as the first initial target user for progressive understanding and memorization. This achieves the goal of effectively storing user intent and emotion memories.

[0025] Furthermore, based on the same principle as obtaining the progressive understanding memory of the first initial target user, the second historical consultation question is parsed using the progressive understanding memory of the first initial target user as a foundation. The obtained second user intent keyword set and second user sentiment keyword set are then used to populate and update the first user intent memory unit and first user sentiment memory unit in the progressive understanding memory of the first initial target user, respectively, to obtain the second initial target user progressive understanding memory. This process is repeated to obtain the first target user progressive understanding memory. The first target user progressive understanding memory reflects the target user's understanding of intent and acceptance of information provided by the AI ​​platform during a consultation. The updated first target user progressive understanding memory is added to the target user progressive understanding memory set, thereby storing the results of progressive understanding memory parsing for use in subsequent dialogues and recommendations.

[0026] Furthermore, based on the target user's progressive understanding memory set, multiple rounds of iterative understanding granularity parsing are performed to determine the target user's understanding granularity feature vector. In this embodiment, step S120 further includes: Step S121: Sort the target user's progressive understanding memory set in chronological order to obtain the target user's progressive understanding memory sequence; Step S122: According to the preset intent similarity bandwidth, the target user progressive understanding memory sequence is split into multiple target user progressive understanding memory sub-sequences by the same intent. Step S123: Traverse the multiple target user progressive understanding memory subsequences to perform multi-scale understanding granularity analysis and determine multiple sets of multi-scale understanding granularity feature vectors; Step S124: Perform cross-guided enhancement on the multiple sets of multi-scale understanding granularity feature vectors respectively to obtain multiple multi-scale understanding granularity cross-guided enhancement feature vectors; Step S125: Calculate the mean of the multiple multi-scale understanding granularity cross-guided enhancement feature vectors to obtain the target user understanding granularity feature vector.

[0027] Furthermore, the understanding granularity cross-guided enhancement is performed on the multiple sets of multi-scale understanding granularity feature vectors respectively to obtain multiple multi-scale understanding granularity cross-guided enhanced feature vectors. Step S124 of this embodiment also includes: Step S124-1: Randomly select two multi-scale understanding granularity feature vectors from the multiple multi-scale understanding granularity feature vector sets without replacement, and use them as multiple multi-scale understanding granularity feature vector groups. Step S124-2: Traverse and identify the feature vector consistency factors of multiple multi-scale understanding granularity feature vector groups, and construct multiple cross-guided enhancement matrices according to the identification results. Use the multiple cross-guided enhancement matrices to perform cross-guided enhancement on the multiple multi-scale understanding granularity feature vector groups to obtain multiple first initial multi-scale understanding granularity cross-guided feature vectors. Step S124-3: Randomly select a multi-scale understanding granularity feature vector from the multiple sets of multi-scale understanding granularity feature vectors again without replacement, and enhance the understanding granularity through multiple first initial multi-scale understanding granularity cross-guided feature vectors until all multi-scale understanding granularity feature vectors in the multiple sets of multi-scale understanding granularity feature vectors are enhanced, thereby obtaining multiple multi-scale understanding granularity cross-guided enhanced feature vectors.

[0028] In one embodiment of this application, the target user's progressive understanding memory sequence is obtained by sorting the historical consultation dialogue records corresponding to the target user's progressive understanding memory set in chronological order according to the timestamps. Then, a preset intent similarity bandwidth pre-defined by those skilled in the art is obtained, wherein the preset intent similarity bandwidth is the intent similarity difference when multiple target user progressive understanding memories can be divided into the same subsequence.

[0029] Preferably, a first target user progressive understanding memory is extracted from the target user progressive understanding memory sequence. The intention similarity between the target user progressive understanding memory sequence and the first target user progressive understanding memory is calculated using a cosine similarity function within a preset intention similarity bandwidth. These are then arranged in chronological order to obtain a first target user progressive understanding memory subsequence. Next, the earliest target user progressive understanding memory (excluding the first target user progressive understanding memory subsequence) is extracted from the target user progressive understanding memory sequence, and subsequence identification is performed. After multiple identifications, multiple target user progressive understanding memory subsequences are obtained. By grouping users according to similar needs in their understanding, subsequent granular analysis of understanding can focus on needs in specific domains or topics, avoiding interference from content with low relevance, thus achieving the technical effect of improving the reliability of user understanding granularity analysis.

[0030] In one embodiment, multiple understanding granularity parsing branches are pre-constructed, each corresponding to a parsing scale. Preferably, multiple sample user progressive understanding memory subsequences and corresponding sample understanding granularity feature vectors obtained after feature recognition according to a parsing scale are acquired as branch training data. The framework based on a convolutional neural network is trained under supervision using the branch training data, and the framework parameters are updated according to the training output until convergence is achieved, thus obtaining a trained understanding granularity parsing branch. The same principle is used to train and obtain the trained multiple understanding granularity parsing branches.

[0031] Furthermore, the multiple understanding granularity analysis branches are used to analyze the multiple target user progressive understanding memory subsequences to obtain corresponding sets of multiple multi-scale understanding granularity feature vectors. Each set of multi-scale understanding granularity feature vectors corresponds to a target user progressive understanding memory subsequence. Each set of multi-scale understanding granularity feature vectors reflects the multi-scale consultation situation of the target user in a consultation scenario from different scales.

[0032] In one embodiment, to deeply fuse feature vectors at different scales based on the potential correlations between feature vectors at different scales within each multi-scale understanding granularity feature vector set, a cross-guided enhancement matrix is ​​constructed for feature vector fusion. Two multi-scale understanding granularity feature vectors are randomly selected without replacement from each of the multiple multi-scale understanding granularity feature vector sets. The cosine similarity function is used to calculate the intra-group feature vector consistency of each multi-scale understanding granularity feature vector group, resulting in multiple sets of intra-group feature vector consistency coefficients. Each set of intra-group feature vector consistency coefficients reflects the degree of similarity between different features within a multi-scale understanding granularity feature vector group. Furthermore, the multiple sets of intra-group feature vector consistency coefficients are normalized using a standardization formula, and the processed data is added to initially empty matrices to obtain the multiple cross-guided enhancement matrices.

[0033] Preferably, graph convolutional networks are used to enhance the feature vectors of the multiple cross-guided enhancement matrices and the multiple multi-scale understanding granularity feature vector groups, thereby obtaining multiple enhanced multi-scale understanding granularity feature vector groups. The mean of the two enhanced multi-scale understanding granularity feature vectors within each of the multiple enhanced multi-scale understanding granularity feature vector groups is calculated to obtain the multiple first initial multi-scale understanding granularity cross-guided feature vectors. Cross-guided enhancement effectively improves the expressive power of the feature vectors, achieving the technical effect of capturing details at different levels and dimensions of user needs, further improving the accuracy and relevance of recommendations.

[0034] Again, a multi-scale understanding granularity feature vector is randomly selected without replacement from the multiple sets of multi-scale understanding granularity feature vectors. Based on the same principle of cross-guided enhancement, multiple first initial multi-scale understanding granularity cross-guided feature vectors are used to perform understanding granularity cross-guided enhancement on them respectively, until all multi-scale understanding granularity feature vectors in the multiple sets of multi-scale understanding granularity feature vectors have been cross-guided enhanced, resulting in multiple multi-scale understanding granularity cross-guided enhanced feature vectors.

[0035] Preferably, the mean of the obtained multi-scale understanding granularity cross-guided enhancement feature vectors is calculated to obtain a target user understanding granularity feature vector that reflects the user understanding granularity after comprehensive analysis from different scales and consultation scenarios. Through multiple rounds of parsing and enhancement of the user's progressive understanding memory, the depth and detail of the user's understanding are accurately captured. Furthermore, through hierarchical analysis and feature enhancement of user needs, a more accurate user understanding granularity feature vector is generated, thereby achieving the technical effect of providing a more accurate basis for data recommendation.

[0036] Step S200: Obtain real-time consultation information from the target user. Based on the target user's understanding scope, the AI ​​platform performs active reasoning based on the real-time consultation information to obtain a real-time active reasoning dialogue sequence. The AI ​​platform is embedded with a deep learning model. Furthermore, real-time consultation information of the target user is obtained. Constrained by the target user's understanding scope, the AI ​​platform performs proactive reasoning based on the real-time consultation information to obtain a real-time proactive reasoning dialogue sequence. The AI ​​platform embeds a deep learning model. Step S200 in this embodiment further includes: The deep learning model is invoked, and the target user's understanding scope is used as a constraint to actively reason about the real-time consultation information to obtain the first active reasoning information. Obtain the first active reasoning feedback information of the target user on the first active reasoning information; The deep learning model is invoked again, and the target user's understanding scope is used as a constraint to actively reason about the first active reasoning feedback information to obtain the second active reasoning information. The real-time consultation information and the first active reasoning information are used as the first round of active reasoning dialogue, and the first active reasoning feedback information and the second active reasoning information are used as the second round of active reasoning dialogue. After multiple rounds of dialogue, the real-time active reasoning dialogue sequence is obtained.

[0037] In one embodiment, real-time consultation information refers to the latest questions, needs, or information raised by the target user in the current conversation. This information can be the content the user is inquiring about or relevant data provided by the user during the conversation. Real-time proactive reasoning dialogue sequence refers to the sequence of dialogue information generated by the AI ​​platform through proactive reasoning in each round of multi-round dialogue with the target user. Each round of dialogue generates new reasoning content based on previous feedback information, thereby gradually obtaining a more complete reasoning result. A deep learning model is a neural network model embedded in the AI ​​platform, used to perform reasoning, prediction, and classification by learning from large amounts of data. This model can learn from real-time consultation information and generate reasoning results that meet the user's needs.

[0038] In one embodiment, a pre-trained deep learning model is invoked to analyze real-time consultation information by limiting the target user's understanding scope. Preferably, the input data of the deep learning model is the target user's understanding scope and the real-time consultation information, and the output data is first active reasoning information. The first active reasoning information is pushed to the target user through an AI platform, and the target user's feedback is obtained, which is the first active reasoning feedback information. The first active reasoning feedback information reflects the target user's affirmation, further inquiry, or negation of the first active reasoning information. Then, the deep learning model is invoked again, using the target user's understanding scope and the first active reasoning feedback information as input for active reasoning to obtain the second active reasoning information. This process continues, through multiple rounds of dialogue, using the real-time consultation information and the first active reasoning information as the first round of active reasoning dialogue, and using the first active reasoning feedback information and the second active reasoning information as the second round of active reasoning dialogue. That is, a consultation information or active reasoning feedback information is combined with the corresponding active reasoning information output by the deep learning model as a dialogue, and through multiple rounds of dialogue, the real-time active reasoning dialogue sequence is obtained. The user's feedback information can further confirm the direction of reasoning or correct errors, ensuring the accuracy and relevance of the reasoning results. Through multiple rounds of dialogue, the AI ​​platform's deep learning model can continuously improve its reasoning based on user feedback, gradually arriving at a recommendation or answer that meets the user's needs. This process ensures the accuracy and consistency of the reasoning, thereby enhancing the user experience.

[0039] Step S300: When the number of dialogue rounds in the real-time proactive reasoning dialogue sequence is greater than or equal to a preset dialogue round threshold, the backtracking detection module is invoked to perform a recommendation data completeness detection on the real-time proactive reasoning dialogue sequence to obtain a recommendation data completeness detection result, wherein the recommendation data completeness detection result includes a recommendation data completeness coefficient and recommendation data; Step S400: Trigger a recommendation output command based on the completeness coefficient of the recommendation data, and push the recommendation data based on the recommendation output command.

[0040] In the embodiments of this application, the preset dialogue turn threshold is a threshold for dialogue turns pre-set by those skilled in the art. If the number of turns in the real-time proactive reasoning dialogue sequence is greater than or equal to this threshold, it indicates that the dialogue has progressed to a certain extent and meets the conditions for triggering subsequent processing. At this time, the backtracking detection module is invoked to perform a retrospective check on the generated dialogue sequence to determine whether the information in the dialogue is sufficient, thereby performing a completeness check. Preferably, multiple sample real-time proactive reasoning dialogue sequences and corresponding multiple sample recommendation data completeness coefficients and multiple sample recommendation data are obtained. The multiple sample recommendation data completeness coefficients and multiple sample recommendation data are identified. The identified multiple sample recommendation data completeness coefficients and multiple sample recommendation data, as well as multiple sample real-time proactive reasoning dialogue sequences, are input into a framework constructed based on a convolutional neural network for training. The output multiple verification recommendation data completeness coefficients and multiple verification recommendation data are compared with the identified multiple sample recommendation data completeness coefficients and multiple sample recommendation data. If the comparison success rate is greater than the preset comparison success rate, the backtracking detection module that has been trained is obtained.

[0041] Then, the real-time proactive reasoning dialogue sequence is input into the backtracking detection module for analysis to obtain the recommendation data completeness coefficient and recommendation data. The recommendation data completeness coefficient reflects the completeness of the data in the real-time proactive reasoning dialogue sequence. When the recommendation data completeness coefficient is greater than or equal to a threshold preset by those skilled in the art, it indicates that reasoning has been completed, triggering a recommendation output command to push the recommendation data to the target user. When the recommendation data completeness coefficient is less than the threshold preset by those skilled in the art, dialogue reasoning needs to continue, triggering a dialogue reasoning continuation command, which calls the AI ​​platform to continue proactive reasoning for the target user.

[0042] By performing backtracking checks on real-time proactive reasoning dialogue sequences, it can be ensured that the recommended data meets the completeness requirements of user needs after multiple rounds of interaction. The introduction of the backtracking check module not only improves the accuracy of recommendations but also optimizes the user experience, avoiding invalid recommendations due to insufficient information.

[0043] In summary, the AI ​​platform data recommendation method provided in this application has the following technical effects: This application employs a progressive, multi-turn dialogue iterative memory parsing and proactive reasoning mechanism to accurately capture and dynamically adjust user needs, ensuring the generation of high-quality recommendations even with incomplete information. Furthermore, by acquiring user consultation information in real time and combining it with the target user's understanding, it can perform effective proactive reasoning to compensate for information gaps and improve recommendation accuracy. Simultaneously, a backtracking detection module is used to check the completeness of the recommendation data, ensuring that the recommended content is sufficient and meets user needs. This achieves the technical effect of providing accurate data recommendations and improving data recommendation efficiency.

[0044] Example 2, based on the same inventive concept as the AI ​​platform data recommendation method in the foregoing examples, such as... Figure 2 As shown, this application provides an AI platform data recommendation device, the device comprising: The target user's understanding scope determination module 11 is used to obtain a set of historical consultation dialogue records of the target user, and to perform progressive multi-round dialogue iterative memory parsing on the set of historical consultation dialogue records to determine the target user's understanding scope. The real-time proactive reasoning dialogue sequence acquisition module 12 is used to acquire real-time consultation information of the target user. With the target user's understanding scope as a constraint, the AI ​​platform performs proactive reasoning based on the real-time consultation information to obtain a real-time proactive reasoning dialogue sequence. The AI ​​platform is embedded with a deep learning model. The module 13 for obtaining the recommendation data completeness detection result is used to call the backtracking detection module to perform recommendation data completeness detection on the real-time active reasoning dialogue sequence when the number of dialogue rounds of the real-time active reasoning dialogue sequence is greater than or equal to a preset dialogue round threshold, and obtain the recommendation data completeness detection result. The recommendation data completeness detection result includes the recommendation data completeness coefficient and the recommendation data. The recommendation information push module 14 is used to trigger a recommendation output instruction based on the completeness coefficient of the recommendation data, and push the recommendation data based on the recommendation output instruction.

[0045] Furthermore, the target user understanding scope determination module 11 is used to perform the following steps: The historical consultation dialogue record set is analyzed using a progressive understanding and memory parsing method within the record to obtain the progressive understanding and memory set of the target user; Based on the target user's progressive understanding memory set, perform multiple rounds of iterative understanding granularity analysis to determine the target user's understanding granularity feature vector; Based on the target user's understanding granularity feature vector, the understanding category is identified to determine the target user's understanding category.

[0046] Furthermore, the target user understanding scope determination module 11 is used to perform the following steps: Extract the first historical consultation question from the first historical consultation dialogue record from the set of historical consultation dialogue records; Based on preset user intent keywords and preset user sentiment keywords, the first historical consultation question is analyzed in a progressive manner to obtain the progressive understanding and memory of the first initial target user; Based on the progressive understanding memory of the first initial target user, the second historical consultation question extracted from the first historical consultation dialogue record is analyzed using progressive understanding memory, and the progressive understanding memory of the first initial target user is updated according to the analysis result to obtain the progressive understanding memory of the second initial target user, and so on, to obtain the progressive understanding memory of the first target user. Add the first target user's progressive understanding memory to the target user's progressive understanding memory set.

[0047] Furthermore, the target user understanding scope determination module 11 is used to perform the following steps: Extract keywords from the first historical consultation question to obtain the keyword set of the first historical consultation question; According to the preset user intent keywords, extract the historical consultation question keywords in the first set of historical consultation question keywords whose similarity meets the preset similarity threshold to obtain the first set of user intent keywords; According to preset user sentiment keywords, extract the historical consultation question keywords in the first set of historical consultation question keywords whose similarity meets the preset similarity threshold to obtain the first set of user sentiment keywords; The preset user intent keywords and the first user intent keyword set are added to the first user intent memory unit, and the preset user emotion keywords and the first user emotion keywords are added to the first user emotion memory unit. The first user intent memory unit and the first user emotion memory unit are used as the first initial target user progressive understanding memory.

[0048] Furthermore, the target user understanding scope determination module 11 is used to perform the following steps: The target user's progressive understanding memory set is sorted in chronological order from front to back to obtain the target user's progressive understanding memory sequence; According to the preset intent similarity bandwidth, the target user progressive understanding memory sequence is split into multiple target user progressive understanding memory sub-sequences by the same intent. The multiple target user progressive understanding memory subsequences are traversed to perform multi-scale understanding granularity analysis, and multiple sets of multi-scale understanding granularity feature vectors are determined. The understanding granularity cross-guided enhancement is performed on the multiple sets of multi-scale understanding granularity feature vectors respectively to obtain multiple multi-scale understanding granularity cross-guided enhanced feature vectors; The mean of the multiple multi-scale understanding granularity cross-guided enhancement feature vectors is calculated to obtain the target user understanding granularity feature vector.

[0049] Furthermore, the target user understanding scope determination module 11 is used to perform the following steps: Two multi-scale understanding granularity feature vectors are randomly selected without replacement from the multiple sets of multi-scale understanding granularity feature vectors to form multiple sets of multi-scale understanding granularity feature vectors. The feature vector consistency factors of multiple multi-scale understanding granularity feature vector groups are identified by traversal, and multiple cross-guided enhancement matrices are constructed according to the identification results. The multiple cross-guided enhancement matrices are used to perform cross-guided enhancement on the multiple multi-scale understanding granularity feature vector groups to obtain multiple first initial multi-scale understanding granularity cross-guided feature vectors. Again, a multi-scale understanding granularity feature vector is randomly selected without replacement from the multiple sets of multi-scale understanding granularity feature vectors. Understanding granularity is enhanced by multiple first initial multi-scale understanding granularity cross-guided feature vectors until all multi-scale understanding granularity feature vectors in the multiple sets of multi-scale understanding granularity feature vectors are enhanced, thus obtaining multiple multi-scale understanding granularity cross-guided enhanced feature vectors.

[0050] Furthermore, the real-time proactive reasoning dialogue sequence acquisition module 12 is used to perform the following steps: The deep learning model is invoked, and the target user's understanding scope is used as a constraint to actively reason about the real-time consultation information to obtain the first active reasoning information. Obtain the first active reasoning feedback information of the target user on the first active reasoning information; The deep learning model is invoked again, and the target user's understanding scope is used as a constraint to actively reason about the first active reasoning feedback information to obtain the second active reasoning information. The real-time consultation information and the first active reasoning information are used as the first round of active reasoning dialogue, and the first active reasoning feedback information and the second active reasoning information are used as the second round of active reasoning dialogue. After multiple rounds of dialogue, the real-time active reasoning dialogue sequence is obtained.

[0051] Example 3, Figure 3 This is a schematic diagram of the structure of a computer device provided in Embodiment 3 of this application, showing a block diagram of an exemplary computer device suitable for implementing embodiments of the present invention. Figure 3 The computer device shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments of the present invention. Figure 3 As shown, the computer device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the computer device can be one or more. Figure 3 Taking a processor 21 as an example, the processor 21, memory 22, input device 23, and output device 24 in a computer device can be connected via a bus or other means. Figure 3 The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0052] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A data recommendation method for an AI platform, characterized in that, The method includes: Obtain a set of historical consultation dialogue records of the target user, and perform progressive multi-round dialogue iterative memory parsing on the set of historical consultation dialogue records to determine the target user's comprehension scope; The AI ​​platform acquires real-time consultation information from the target user, and, constrained by the target user's understanding scope, performs proactive reasoning based on the real-time consultation information to obtain a real-time proactive reasoning dialogue sequence. The AI ​​platform is embedded with a deep learning model. When the number of dialogue rounds in the real-time proactive reasoning dialogue sequence is greater than or equal to a preset dialogue round threshold, the backtracking detection module is invoked to perform a recommendation data completeness detection on the real-time proactive reasoning dialogue sequence to obtain a recommendation data completeness detection result, wherein the recommendation data completeness detection result includes a recommendation data completeness coefficient and recommendation data. The recommendation output instruction is triggered based on the completeness coefficient of the recommendation data, and the recommendation data is pushed based on the recommendation output instruction.

2. The AI ​​platform data recommendation method as described in claim 1, characterized in that, Obtain a set of historical consultation dialogue records of the target user, and perform progressive multi-turn iterative memory parsing on the set of historical consultation dialogue records to determine the target user's comprehension scope, including: The historical consultation dialogue record set is analyzed using a progressive understanding and memory parsing method within the record to obtain the progressive understanding and memory set of the target user; Based on the target user's progressive understanding memory set, perform multiple rounds of iterative understanding granularity analysis to determine the target user's understanding granularity feature vector; Based on the target user's understanding granularity feature vector, the understanding category is identified to determine the target user's understanding category.

3. The AI ​​platform data recommendation method as described in claim 2, characterized in that, The historical consultation dialogue record set is analyzed using a progressive understanding and memory parsing method to obtain the target user's progressive understanding and memory set, including: Extract the first historical consultation question from the first historical consultation dialogue record from the set of historical consultation dialogue records; Based on preset user intent keywords and preset user sentiment keywords, the first historical consultation question is analyzed in a progressive manner to obtain the progressive understanding and memory of the first initial target user; Based on the progressive understanding memory of the first initial target user, the second historical consultation question extracted from the first historical consultation dialogue record is analyzed using progressive understanding memory, and the progressive understanding memory of the first initial target user is updated according to the analysis result to obtain the progressive understanding memory of the second initial target user, and so on, to obtain the progressive understanding memory of the first target user. Add the first target user's progressive understanding memory to the target user's progressive understanding memory set.

4. The AI ​​platform data recommendation method as described in claim 3, characterized in that, Based on preset user intent keywords and preset user sentiment keywords, a progressive understanding and memory analysis is performed on the first historical consultation question to obtain the first initial target user's progressive understanding and memory, including: Extract keywords from the first historical consultation question to obtain the keyword set of the first historical consultation question; According to the preset user intent keywords, extract the historical consultation question keywords in the first set of historical consultation question keywords whose similarity meets the preset similarity threshold to obtain the first set of user intent keywords; According to preset user sentiment keywords, extract the historical consultation question keywords in the first set of historical consultation question keywords whose similarity meets the preset similarity threshold to obtain the first set of user sentiment keywords; The preset user intent keywords and the first user intent keyword set are added to the first user intent memory unit, and the preset user emotion keywords and the first user emotion keywords are added to the first user emotion memory unit. The first user intent memory unit and the first user emotion memory unit are used as the first initial target user progressive understanding memory.

5. The AI ​​platform data recommendation method as described in claim 3, characterized in that, Based on the target user's progressive understanding memory set, multiple rounds of iterative understanding granularity analysis are performed to determine the target user's understanding granularity feature vector, including: The target user's progressive understanding memory set is sorted in chronological order from front to back to obtain the target user's progressive understanding memory sequence; According to the preset intent similarity bandwidth, the target user progressive understanding memory sequence is split into multiple target user progressive understanding memory sub-sequences by the same intent. The multiple target user progressive understanding memory subsequences are traversed to perform multi-scale understanding granularity analysis, and multiple sets of multi-scale understanding granularity feature vectors are determined. The understanding granularity cross-guided enhancement is performed on the multiple sets of multi-scale understanding granularity feature vectors respectively to obtain multiple multi-scale understanding granularity cross-guided enhanced feature vectors; The mean of the multiple multi-scale understanding granularity cross-guided enhancement feature vectors is calculated to obtain the target user understanding granularity feature vector.

6. The AI ​​platform data recommendation method as described in claim 5, characterized in that, The multiple sets of multi-scale understanding granularity feature vectors are respectively subjected to understanding granularity cross-guided enhancement to obtain multiple multi-scale understanding granularity cross-guided enhanced feature vectors, including: Two multi-scale understanding granularity feature vectors are randomly selected without replacement from the multiple sets of multi-scale understanding granularity feature vectors to form multiple sets of multi-scale understanding granularity feature vectors. The feature vector consistency factors of multiple multi-scale understanding granularity feature vector groups are identified by traversal, and multiple cross-guided enhancement matrices are constructed according to the identification results. The multiple cross-guided enhancement matrices are used to perform cross-guided enhancement on the multiple multi-scale understanding granularity feature vector groups to obtain multiple first initial multi-scale understanding granularity cross-guided feature vectors. Again, a multi-scale understanding granularity feature vector is randomly selected without replacement from the multiple sets of multi-scale understanding granularity feature vectors. Understanding granularity is enhanced by multiple first initial multi-scale understanding granularity cross-guided feature vectors until all multi-scale understanding granularity feature vectors in the multiple sets of multi-scale understanding granularity feature vectors are enhanced, thus obtaining multiple multi-scale understanding granularity cross-guided enhanced feature vectors.

7. The AI ​​platform data recommendation method as described in claim 1, characterized in that, The AI ​​platform acquires real-time consultation information from a target user, and, constrained by the target user's understanding, performs proactive reasoning based on this information to obtain a real-time proactive reasoning dialogue sequence. The AI ​​platform embeds a deep learning model, including: The deep learning model is invoked, and the target user's understanding scope is used as a constraint to actively reason about the real-time consultation information to obtain the first active reasoning information. Obtain the first active reasoning feedback information of the target user on the first active reasoning information; The deep learning model is invoked again, and the target user's understanding scope is used as a constraint to actively reason about the first active reasoning feedback information to obtain the second active reasoning information. The real-time consultation information and the first active reasoning information are used as the first round of active reasoning dialogue, and the first active reasoning feedback information and the second active reasoning information are used as the second round of active reasoning dialogue. After multiple rounds of dialogue, the real-time active reasoning dialogue sequence is obtained.

8. An AI platform data recommendation device, characterized in that, The apparatus for implementing the AI ​​platform data recommendation method according to any one of claims 1 to 7, the apparatus comprising: The target user's understanding scope determination module is used to obtain a set of historical consultation dialogue records of the target user, and to perform progressive multi-round dialogue iterative memory parsing on the set of historical consultation dialogue records to determine the target user's understanding scope. The real-time proactive reasoning dialogue sequence acquisition module is used to acquire real-time consultation information of the target user. With the target user's understanding scope as a constraint, the AI ​​platform performs proactive reasoning based on the real-time consultation information to obtain a real-time proactive reasoning dialogue sequence. The AI ​​platform is embedded with a deep learning model. The module for obtaining the recommendation data completeness detection result is used to call the backtracking detection module to perform recommendation data completeness detection on the real-time active reasoning dialogue sequence when the number of dialogue rounds of the real-time active reasoning dialogue sequence is greater than or equal to a preset dialogue round threshold, and obtain the recommendation data completeness detection result. The recommendation data completeness detection result includes the recommendation data completeness coefficient and the recommendation data. The recommendation information push module is used to trigger a recommendation output instruction based on the completeness coefficient of the recommendation data, and push the recommendation data based on the recommendation output instruction.

9. A computer device, characterized in that, The computer device includes: processor; Memory used to store the processor's executable instructions; The processor is used to execute the steps of the AI ​​platform data recommendation method according to any one of claims 1 to 7.

Images