A guide intelligent agent intention-driven goods matching method

Abstract

The application discloses a kind of shopping guide intelligent agent intention driving goods matching method, it is related to artificial intelligence and shopping guide technical field, comprising: S1: receiving user natural language request and historical interaction information, generates user preference state information based on semantic analysis;S2: collection user real-time interaction behavior data to recommended goods, generate user cognitive state evaluation signal in combination with user preference state information, fusion signal generates cognitive convergence index;S3: according to cognitive convergence index, determine target matching mode from the matching strategy mode set of pre-setting;S4: according to target matching mode and user preference state information, execute recall, ordering and explanation generation to commodity database, obtain ordered recommended goods list.Through cognitive convergence index dynamic switching recommendation strategy, accurately capture user vague intention, and generate interpretable recommendation result, to significantly improve the intelligent level and decision efficiency of shopping guide interaction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and shopping guide technology, specifically to a shopping guide intelligent agent intent-driven product matching method. Background Technology

[0002] With the deep integration of e-commerce and artificial intelligence technologies, intelligent shopping guide systems have become an important tool for enhancing user shopping experience and promoting product conversion. Existing intelligent shopping guide solutions typically rely on users' historical behavior, real-time clicks, or specific keywords to search for and recommend products.

[0003] However, in practical applications, especially in shopping guide scenarios requiring multiple rounds of interaction and complex intents (such as clothing matching and gift selection), existing technical solutions still face several common challenges. First, at the intent understanding level, most systems rely on shallow semantic parsing of single-round user queries or fixed intent classifications, making it difficult to effectively capture and track the dynamic and often vaguely expressed complex preferences of users during interaction. Second, in terms of interaction patterns, systems often follow a fixed "query-response" process, lacking the ability to adaptively guide based on the user's decision-making state, resulting in low recommendation efficiency when the user's intent is unclear. Finally, in the recommendation generation stage, how to reliably match deep user intents with real-time, accurate product information and provide convincing recommendation reasons is also crucial to user experience. These factors, to some extent, constrain the evolution of intelligent shopping guide systems towards a more humanized and precise direction.

[0004] Therefore, the industry needs an intelligent matching method that can more deeply understand users' dynamic intentions and adaptively adjust recommendation strategies based on users' cognitive states in order to improve the accuracy and naturalness of shopping guide interactions. Summary of the Invention

[0005] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a shopping guide intelligent agent intent-driven product matching method to solve the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a product matching method driven by the intent of a shopping guide intelligent agent, comprising:

[0007] S1: Receive user natural language requests and historical interaction information, and generate user preference status information based on semantic parsing;

[0008] S2: Collect real-time interaction data of users with recommended products, combine user preference status information to generate user cognitive status evaluation signal, and fuse the signal to generate cognitive convergence index;

[0009] S3: Based on the cognitive convergence index, determine the target matching pattern from the preset matching strategy pattern set;

[0010] S4: Based on the target matching pattern and user preference status information, perform recall, sorting and interpretation generation on the product database to obtain a sorted list of recommended products.

[0011] The present invention is further configured such that S1 specifically includes:

[0012] Obtain user's historical interaction information and perform context fusion with the currently collected user natural language request to form complete context information of the current dialogue;

[0013] The complete contextual information is input into a pre-trained natural language understanding model to perform intent classification and semantic slot filling;

[0014] Extract structured data from the output of the natural language understanding model. The structured data includes at least: user core intent categories, newly added semantic slot key-value pairs, and confirmation or correction information of historical preference attributes.

[0015] Based on structured data, user preference state information is maintained through a state machine update mechanism. The user preference state information includes at least: a set of confirmed preference attributes, a set of preference attribute intent texts to be explored, and a set of excluded preference attributes.

[0016] The present invention is further configured such that S2 specifically includes: a step of collecting and normalizing interactive behavior data, a step of extracting and evaluating cognitive state, and a step of fusing and calculating cognitive convergence index.

[0017] The present invention is further configured such that the step of collecting and normalizing interaction behavior data specifically includes:

[0018] Collect real-time interactive behavior data of users in response to recommended products. The interactive behavior data includes at least: browsing, dwelling, clicking, adding to cart, removing, and feedback tags.

[0019] Based on a preset behavior weighting system, interactive behavior data is mapped to different behavior categories;

[0020] Based on the preset scaling rules, the behavioral data mapped to each category are aligned in terms of dimensions and normalized in terms of numerical range to form standardized behavioral feature vectors.

[0021] The present invention is further configured such that the step of extracting and evaluating cognitive states specifically includes:

[0022] By using weighted aggregation calculations to analyze the preference indices corresponding to each behavior category, we can obtain the user's real-time preference feedback for recommended products.

[0023] The system compares the instant preference feedback expression with the user preference state information, performs pattern matching through state transition rules, and identifies the dynamic change pattern of preference attributes in the user preference state information. The dynamic change pattern includes: enhancement, maintenance, or transition.

[0024] Based on the dynamic change pattern, a multi-dimensional user cognitive state evaluation signal is calculated and output through a predefined signal extraction logic. The user cognitive state evaluation signal includes at least: a distribution signal reflecting the concentration of positive feedback, a consistency signal reflecting the temporal changes of preference attributes, and an efficiency signal reflecting the trend of decision delay.

[0025] The present invention is further configured such that the step of fusing and calculating the cognitive convergence index specifically includes:

[0026] The user's cognitive state assessment signals are weighted and fused according to preset weights to generate an initial cognitive state value;

[0027] The initial cognitive state values ​​are processed by temporal smoothing to output the final cognitive convergence index.

[0028] The present invention is further configured such that S3 specifically includes:

[0029] The cognitive convergence index is compared with a preset multi-level convergence threshold range to generate comparison results.

[0030] Based on the comparison results, the corresponding target matching pattern is determined from the preset matching strategy pattern set.

[0031] The present invention is further configured such that S4 specifically includes: a matching execution step and a result generation step.

[0032] The present invention is further configured such that the matching execution step specifically includes:

[0033] Based on target matching patterns and user preference status information, multi-path retrieval and recall are performed from a preset product database to form a candidate product set. The multi-path retrieval includes at least: precise matching retrieval based on confirmed preference attributes in user preference status information, vector similarity retrieval of intention text of preference attributes to be explored in user preference status information, and diversity retrieval based on product association relationships.

[0034] The present invention is further configured such that the result generation step specifically includes:

[0035] Based on the weights configured in the target matching mode, the products in the candidate product set are calculated for multi-dimensional matching degree. The multi-dimensional matching degree includes at least: attribute matching degree, semantic relevance and real-time popularity, and the products are sorted according to the calculation results.

[0036] Based on the ranking results, the calculation results of multi-dimensional matching degree, and the confirmed preference attributes in the user preference status information, the corresponding explanatory text is synthesized through natural language generation technology;

[0037] The sorted products are combined with explanatory text to form a list of recommended products.

[0038] This invention provides a product matching method driven by the intent of a shopping guide intelligent agent. The method comprises: S1: receiving user natural language requests and historical interaction information, and generating user preference state information based on semantic parsing; S2: collecting real-time interaction data of users with recommended products, combining the user preference state information to generate a user cognitive state evaluation signal, and fusing the signal to generate a cognitive convergence index; S3: determining a target matching mode from a preset matching strategy mode set based on the cognitive convergence index; and S4: performing recall, sorting, and interpretation on a product database based on the target matching mode and user preference state information to obtain a sorted list of recommended products. The beneficial effects include:

[0039] Dynamic strategy switching: By quantifying the user's decision-making status in real time through the "cognitive convergence index", the system can automatically switch the most suitable recommendation strategy at different stages such as user exploration, hesitation or clarity, just like a human shopping guide, which greatly improves the intelligence of interaction and guidance efficiency.

[0040] Fuzzy Intent Capture: By independently retrieving "intention text of preference attributes to be explored", the system can directly understand and respond to users' vague and unformed intentions, realizing in-depth mining and immediate satisfaction of potential needs.

[0041] Explainable and accurate recommendations: The "multi-path retrieval" mechanism, which integrates precise matching, semantic expansion and diversity exploration, combined with ranking and explanation generation based on multi-dimensional matching degree, ensures that the recommendation results are highly relevant, rich and clear, and effectively improve user trust and decision-making efficiency.

[0042] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0044] Figure 1 The flowchart illustrates an intention-driven product matching method for a shopping guide intelligent agent, as an exemplary embodiment of the present invention. Detailed Implementation

[0045] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0046] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0047] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0048] Example:

[0049] A shopping guide intelligent agent intent-driven product matching method, such as Figure 1 As shown, it includes:

[0050] S1: Receive user natural language requests and historical interaction information, and generate user preference status information based on semantic parsing;

[0051] S2: Collect real-time interaction data of users with recommended products, combine user preference status information to generate user cognitive status evaluation signal, and fuse the signal to generate cognitive convergence index;

[0052] S3: Based on the cognitive convergence index, determine the target matching pattern from the preset matching strategy pattern set;

[0053] S4: Based on the target matching pattern and user preference status information, perform recall, sorting and interpretation generation on the product database to obtain a sorted list of recommended products.

[0054] The present invention is further configured such that S1 specifically includes:

[0055] Obtain user's historical interaction information and perform context fusion with the currently collected user natural language request to form complete context information of the current dialogue;

[0056] The complete contextual information is input into a pre-trained natural language understanding model to perform intent classification and semantic slot filling;

[0057] Extract structured data from the output of the natural language understanding model. The structured data includes at least: user core intent categories, newly added semantic slot key-value pairs, and confirmation or correction information of historical preference attributes.

[0058] Based on structured data, user preference state information is maintained through a state machine update mechanism. This user preference state information includes at least: a set of confirmed preference attributes, a set of intention texts for preference attributes to be explored, and a set of excluded preference attributes. Specifically, firstly, the system retrieves user historical interaction records stored in the database and concatenates these historical records with the currently collected user natural language requests in chronological order to form a context text containing the complete dialogue process. If the query result historical records are empty, the currently collected user natural language requests are used as the complete context text. Next, the system uses the context text as input and passes it to a pre-trained natural language understanding model fine-tuned by joint intent and slot filling tasks. This model can be implemented using a pre-trained model based on the Transformer architecture, such as a BERT model, ALBERT model, or RoBERTa model fine-tuned for a specific dialogue dataset. These are existing technology models in the field of natural language processing. This natural language understanding model understands the user's overall purpose and identifies key attribute information in the sentence by simultaneously performing two tasks: intent classification and semantic slot filling. Then, the system parses the output of the natural language understanding model and converts the original label sequence into clear structured data through sequence labeling parsing methods. Specifically, the system uses BIOES or BIO labeling system to parse each word tag in the model output, identifies and merges consecutive words belonging to the same semantic slot to form a complete slot value, and combines it with the slot type to form a key-value pair. At the same time, the intent category is obtained from the specific classification vector output by the model. This part of the transformed structured data clearly includes: the determined user core intent category, the newly added semantic slot key-value pairs identified from the current sentence, and the confirmation or correction instructions for the preference attributes mentioned in the historical interactions. Finally, based on the structured data, the system drives a rule-based state machine to update and maintain the core data structure of user preference state information. The rule-based state machine is a programming paradigm for managing and transitioning system states; it is an existing technology in software engineering. It is constructed by defining a finite set of states, events that trigger state transitions, and state update rules corresponding to each event. Specifically, the system moves specific attributes into or out of the excluded preference attribute set according to correction instructions, adds newly added explicit key-value pairs to the confirmed preference attribute set, and directly stores newly added ambiguous descriptive text into the preference attribute intent text set to be explored. After these operations, the final output of user preference state information consists of three parts: the confirmed preference attribute set, the preference attribute intent text set to be explored, and the excluded preference attribute set.

[0059] The present invention is further configured such that S2 specifically includes: a step of collecting and normalizing interactive behavior data, a step of extracting and evaluating cognitive states, and a step of fusing and calculating a cognitive convergence index. Specifically, the step of collecting and normalizing interactive behavior data is used to acquire the real-time interactive behavior of users during the product recommendation display process, and to transform different types of behavioral signals into standardized behavioral features with a unified metric. Its function is to enable different interactive behaviors to be compared and analyzed in the same representation space through dimensional alignment and range normalization, providing a quantifiable behavioral basis for the subsequent extraction of user cognitive states. The step of extracting and evaluating cognitive states is used to identify the trend of user preference changes based on standardized behavioral features, determine the user's intention to the current recommendation direction, and output a multi-dimensional evaluation signal that can characterize the fluctuation of user cognitive decisions. Its function is to map user behavioral features into a cognitive state expression that can reflect the dynamics of preferences, the degree of attention concentration, and the degree of decision hesitation, laying the foundation for generating intermediate signals that can be used to measure the stability of user cognition. The calculation step of integrating and calculating the cognitive convergence index is used to perform weighted fusion and temporal smoothing of multi-dimensional cognitive state evaluation signals to form a single indicator that can represent the degree to which user cognition tends to be stable or continuously dispersed. Its function is to compress the dynamic and multi-source cognitive state expression into a convergent indicator that can be used for matching strategy switching and recommendation rhythm regulation, so that the recommendation process can adjust the pattern according to the user's cognitive trend.

[0060] The present invention is further configured such that the step of collecting and normalizing interaction behavior data specifically includes:

[0061] Collect real-time interactive behavior data of users in response to recommended products. The interactive behavior data includes at least: browsing, dwelling, clicking, adding to cart, removing, and feedback tags.

[0062] Based on a preset behavior weighting system, interactive behavior data is mapped to different behavior categories;

[0063] Based on preset scaling rules, the system performs dimensional alignment and numerical range normalization on the behavioral data mapped to each category, forming standardized behavioral feature vectors. Specifically, firstly, the system collects real-time interactive behavior data generated by users towards recommended products through front-end tracking and real-time data stream processing technology. This real-time interactive behavior data includes at least the following types: browsing, dwelling, clicking, adding to cart, removing, and feedback tags. Each record contains the behavior type, the target product identifier, and a timestamp. Subsequently, the system maps the raw real-time interactive behavior data to different behavior categories according to a preset behavior weighting system. This behavior weighting system is a predefined mapping table that assigns a weight value representing the intensity of preference and a positive or negative polarity label representing the direction of preference to each behavior type. For example, adding to cart is mapped to a strongly positive category and given a high weight, while removing is mapped to a strongly negative category and given a negative weight. Next, the system sums the weight values ​​of the same behavior category for each product within a defined recent time window to obtain the original aggregate value. Then, it performs dimensional alignment and numerical range normalization on the original aggregate value according to a preset scaling rule. Common scaling rules are min-max normalization or Z-score standardization. Min-max normalization requires dividing the original aggregate value by the predefined theoretical maximum value of the behavior category, so that the result falls within the range of 0 to 1. Finally, the system combines the normalized values ​​of each category corresponding to each product in a fixed order according to all predefined behavior categories, forming a fixed-length numerical array, which generates a unique standardized behavior feature vector. The system predefines three behavior categories: 'strongly positive', 'slightly positive', and 'strongly negative'. For product P123, an add-to-cart action and a click within the window are mapped to the 'strongly positive' and 'slightly positive' categories, respectively. After aggregation and normalization, the standardized behavioral feature vector of product P123 is [0.33, 0.30, 0.00], where the first value 0.33 represents the normalized strength of 'strongly positive' behavior, the second value 0.30 represents the normalized strength of 'slightly positive' behavior, and the third value 0.00 represents the strength of 'strongly negative' behavior, which is zero. This multidimensional vector provides a quantifiable and comparable behavioral basis for subsequent analysis.

[0064] The present invention is further configured such that the step of extracting and evaluating cognitive states specifically includes:

[0065] By using weighted aggregation calculations to analyze the preference indices corresponding to each behavior category, we can obtain the user's real-time preference feedback for recommended products.

[0066] The system compares the instant preference feedback expression with the user preference state information, performs pattern matching through state transition rules, and identifies the dynamic change pattern of preference attributes in the user preference state information. The dynamic change pattern includes: enhancement, maintenance, or transition.

[0067] Based on the dynamic change pattern, a multi-dimensional user cognitive state evaluation signal is calculated and output through a predefined signal extraction logic. This signal includes at least: a distribution signal reflecting the concentration of positive feedback, a consistency signal reflecting the temporal changes in preference attributes, and an efficiency signal reflecting the trend of decision delay. Specifically, firstly, the system performs weighted aggregation calculations on standardized behavioral feature vectors to obtain the user's immediate preference feedback for recommended items. The weighted aggregation calculation assigns an aggregation weight to each predefined behavioral category, multiplies the values ​​of each dimension in the behavioral feature vector corresponding to each item by the aggregation weight of the corresponding behavioral category, and then sums all the products to obtain the immediate preference score for that item. All items are then sorted in descending order of their immediate preference scores to form a list of immediate preference feedback expressions. Next, the system compares the instant preference feedback with the user's preference state information and performs pattern matching using state transition rules to identify the dynamic change patterns of preference attributes in the user's preference state information. This process first extracts the recently active focus attributes from the set of confirmed preference attributes in the user's preference state information. Then, it compares the attributes possessed by the high-ranked items in the instant preference feedback list with these focus attributes and applies preset state transition rules to make a judgment: if all high-ranked items possess a certain focus attribute and the preference score increases significantly, then the attribute is determined to be in an enhancement mode; if some high-ranked items possess a certain focus attribute and the preference score remains stable, then it is determined to be in a maintenance mode; if most high-ranked items no longer possess a certain focus attribute and shift to other attributes, then it is determined to be in a transition mode.Finally, based on the identified dynamic change patterns, the system calculates and outputs multi-dimensional user cognitive state evaluation signals through predefined signal extraction logic. The signal extraction logic comprises three independent calculation units. The first unit calculates a distribution signal reflecting the concentration of positive feedback. Specifically, it analyzes the evenness of the distribution of positive feedback scores across different items in the immediate preference feedback expression. If the scores are highly concentrated in a few items, a high distribution signal value is output; if the scores are evenly dispersed, a low distribution signal value is output. The calculation is based on the principle of information entropy. The second unit calculates a consistency signal reflecting the temporal changes in preference attributes. Specifically, it identifies the core preference attributes that are key to decision-making from the set of confirmed preference attributes in the user preference state information. The method for identifying the core preference attributes specifically includes: the system setting a filtering rule based on a time window and frequency; this rule... The system counts the number of times each preference attribute within the set is consistently mentioned or confirmed in the most recent five or ten rounds of dialogue. Then, it identifies and marks preference attributes whose number of occurrences exceeds a preset threshold (e.g., more than half the total number of rounds) as core preference attributes for the current conversation phase. Next, it calculates the proportion of rounds in which these core preference attributes remain unchanged in recent rounds of dialogue; a higher proportion results in a higher consistency signal value. The third unit calculates an efficiency signal reflecting the trend of decision delay. Specifically, it compares the average time interval between receiving a recommendation and generating the first positive feedback within the current time window and the previous time window. A positive efficiency signal value is output if the current interval is significantly shortened, and a negative efficiency signal value is output if it is significantly lengthened. These three signal values ​​together constitute the user's cognitive state evaluation signal. For example, in a conversation, the user's historical confirmation preferences include the color beige. The current window calculates and sorts the immediate preference scores of five items, with the top two items both being beige and having high scores. Based on this, the state transition rule determines that the beige color attribute is an enhancement mode. Further analysis reveals that the high scores are concentrated in the top two items, resulting in a high distribution signal value of 0.8. Examining the last five rounds of dialogue history, the beige color attribute has been stable in four rounds, resulting in a consistency signal value of 0.8. The current decision interval is 33% shorter than the previous round, resulting in a positive efficiency signal value of 0.3. The final output user cognitive state evaluation signal is a vector composed of 0.8, 0.8, and 0.3.

[0068] The present invention is further configured such that the step of fusing and calculating the cognitive convergence index specifically includes:

[0069] The user's cognitive state assessment signals are weighted and fused according to preset weights to generate an initial cognitive state value;

[0070] The system performs time-series smoothing on the initial cognitive state values ​​to output the final cognitive convergence index. Specifically, the system receives a multi-dimensional user cognitive state evaluation signal output from the cognitive state extraction and evaluation step. This signal includes specific values ​​for distribution, consistency, and efficiency signals. First, the system performs weighted fusion to generate the initial cognitive state value. This process is based on a predefined and fixed set of weights. For example, the default weight allocation scheme can be set as follows: 0.5 weight for the distribution signal, 0.3 weight for the consistency signal, and 0.2 weight for the efficiency signal. The system multiplies the signal value of each dimension in the user cognitive state evaluation signal by its corresponding preset weight, and then sums all the products. The sum is the initial cognitive state value. This essentially applies linear weighted fusion, a fundamental technique in multi-attribute decision-making. Subsequently, the system performs time-series smoothing on the initial cognitive state value to output the final cognitive convergence index. This process uses the exponentially weighted moving average method, commonly used in time series analysis. The system maintains a historical record of the cognitive convergence index from the previous round and presets a parameter called the smoothing factor, for example, 0.3. The smoothing calculation process is to multiply the historical value by 1 and subtract the smoothing factor, then add the current initial cognitive state value multiplied by the smoothing factor. The result is the final cognitive convergence index for this round. The new cognitive convergence index will be updated with the historical value for the next round of calculation. Through this mechanism, the cognitive convergence index can smoothly reflect the trend of the user's cognitive state. For example, assuming the current input user cognitive state evaluation signal is a distribution signal of 0.8, a consistency signal of 0.7, and an efficiency signal of 0.4, the initial cognitive state value after weighted fusion is 0.69. If the historical value of the cognitive convergence index from the previous round is 0.65 and the smoothing factor is 0.3, then after exponential weighted moving average calculation, the final cognitive convergence index output is approximately 0.66. This index serves as a stable scalar indicator, used in subsequent steps to determine the convergence trend of user cognition and drive the switching of matching strategies.

[0071] The present invention is further configured such that S3 specifically includes:

[0072] The cognitive convergence index is compared with a preset multi-level convergence threshold range to generate comparison results.

[0073] Based on the comparison results, the corresponding target matching pattern is determined from the preset matching strategy pattern set. Specifically, the system first presets multiple convergence threshold intervals. These intervals are continuous segmented ranges within the cognitive convergence index value range, used to map continuous index values ​​to discrete stages of user cognitive states. For example, the system can set intervals where the cognitive convergence index is less than or equal to 0.2 to correspond to the cognitive dispersion and exploration state, intervals where the cognitive convergence index is greater than 0.2 and less than or equal to 0.6 to correspond to the cognitive formation and guidance state, and intervals where the cognitive convergence index is greater than 0.6 to correspond to the cognitive clarity and convergence state. When the system receives the latest cognitive convergence index value calculated from the previous steps, it performs a threshold comparison operation, comparing the cognitive convergence index value sequentially with the preset interval boundaries to determine which specific interval it falls into. For example, when the cognitive convergence index is 0.75, since 0.75 is greater than 0.6, the system determines that it falls into the interval corresponding to the cognitive clarity and convergence state. This determination result is the generated comparison result. The system also presets a matching strategy pattern set, which contains multiple predefined and complete product matching strategy configurations. Each strategy clearly specifies the retrieval method, sorting rules, and interaction to be used in subsequent steps. The system employs various communication styles, each corresponding to a specific convergence threshold interval. For instance, a strategy bound to a cognitively dispersed state interval is called the "active exploration mode," which emphasizes broad semantic retrieval, introduces a high proportion of diverse products, de-emphasizes the weight of precise attributes during ranking, and uses guiding questioning techniques. A strategy bound to a cognitively formed state interval is called the "balanced guidance mode," which balances precise and semantic retrieval, balances attribute matching and user historical preferences during ranking, and provides comparison options. A strategy bound to a cognitively defined state interval is called the "precise cultivation mode," which prioritizes precise retrieval based on defined attributes, maximizes the weight of attribute matching and product popularity during ranking, and uses techniques to confirm preferences and provide pairing suggestions. After obtaining the comparison result, i.e., the specific interval identifier, the system selects the corresponding matching strategy configuration from the matching strategy mode set based on the pre-established binding relationship. For example, when the comparison result is a cognitively defined state interval, the system automatically selects the precise cultivation mode and all its parameter configurations as the target matching mode for this round. Finally, this target matching mode is directly output as the core control command.

[0074] The present invention is further configured such that S4 specifically includes a matching execution step and a result generation step. Specifically, the matching execution step is used to perform multi-way retrieval and recall on the product database based on the target matching pattern and user preference state information, forming a set of products that meet the constraints and have candidate value; its function is to improve the effectiveness and richness of subsequent ranking and interpretation generation by parallelly integrating precise retrieval of confirmed preference attributes, semantic vector retrieval of preference attributes to be explored, and diverse retrieval driven by product association relationships, so that the candidate product set can simultaneously cover both explicit and potential user needs. The result generation step, after obtaining the candidate product set, performs a comprehensive ranking of the candidate products based on the multi-dimensional matching weights configured according to the target matching mode. Simultaneously, it generates explanatory text that reflects the rationality of the recommendations, thus constructing the final recommended product list. Its function is to achieve priority stratification of candidate products through a comprehensive ranking of multi-dimensional features such as attribute matching degree, semantic relevance, and real-time popularity. Furthermore, it utilizes natural language generation technology to generate semantic explanations consistent with the ranking results, ensuring that the recommended list reflects structured matching logic while possessing user-friendly comprehensibility and perceptible justification, thereby improving recommendation credibility and the quality of the interaction loop.

[0075] The present invention is further configured such that the matching execution step specifically includes:

[0076] Based on the target matching pattern and user preference state information, the system performs multi-path retrieval and recall from a pre-set product database to form a candidate product set. The multi-path retrieval includes at least: precise matching retrieval based on confirmed preference attributes in the user preference state information, vector similarity retrieval based on the intent text of preference attributes to be explored in the user preference state information, and diversity retrieval based on product association relationships. Specifically, the system performs multi-path retrieval and recall from a pre-set product database based on the target matching pattern and user preference state information to form a candidate product set. The target matching pattern is a strategy configuration determined by previous steps, which dynamically specifies the activation priority and result ratio of different retrieval paths in this retrieval. User preference state information is the core input, containing the confirmed preference attribute set, the preference attribute intent text set to be explored, and the excluded preference attribute set. First, multi-path retrieval includes: exact matching retrieval based on a confirmed set of preference attributes. The system transforms explicit attribute key-value pairs (e.g., category: dress, scene: outdoor) into database query conditions and uses attribute retrieval technology to execute queries in the product database, thereby efficiently recalling a list of matching products. The attribute retrieval technology refers to existing technologies for efficiently processing structured attribute queries, and its implementation includes, but is not limited to, queries based on relational database indexes or queries based on dedicated search engines (such as Elasticsearch) using inverted indexes. Second, multi-path retrieval includes: vector similarity retrieval based on a set of intent texts for the preference attributes to be explored. The system uses a pre-built text embedding model, such as S... The entence-BERT model transforms ambiguous intent text, such as text with a strong design element, into high-dimensional semantic vectors. Then, it performs an approximate nearest neighbor search in a dedicated vector index of the product database to quickly find the top few products whose semantic vectors are most similar to the query vector. This process relies on existing technologies such as text embedding models and vector databases. Finally, multi-path retrieval includes diversity retrieval based on product associations. The system uses some products obtained from the first two retrieval paths as seeds. In a pre-built product knowledge graph, it traverses the graph along edges related to style similarity or matching recommendations to discover and recall other products strongly associated with the seed products, thus expanding the selection range. This process utilizes graph database query and traversal techniques. These three retrieval paths are executed in parallel or sequentially according to the target matching mode instructions. For example, in balanced guided mode, all three retrieval paths are activated simultaneously and the number of recalled products is allocated according to a preset ratio. All recalled products are then merged and deduplicated. Finally, filtering is performed based on the excluded preference attribute set, such as removing all products with the color black, thus forming the final, purified candidate product set.

[0077] The present invention is further configured such that the result generation step specifically includes:

[0078] Based on the weights configured in the target matching mode, the products in the candidate product set are calculated for multi-dimensional matching degree. The multi-dimensional matching degree includes at least: attribute matching degree, semantic relevance and real-time popularity, and the products are sorted according to the calculation results.

[0079] Based on the ranking results, the calculation results of multi-dimensional matching degree, and the confirmed preference attributes in the user preference status information, the corresponding explanatory text is synthesized through natural language generation technology;

[0080] The sorted products are combined with explanatory text to form a recommended product list. Specifically, the system first calculates and sorts the multi-dimensional matching degree of each product in the candidate product set according to the weights configured in the target matching mode. The weights configured in the target matching mode refer to the importance coefficients pre-set for different matching dimensions in the currently effective matching strategy. For example, in the precise and in-depth cultivation mode, attribute matching degree has the highest weight. The multi-dimensional matching degree calculation includes three core dimensions: attribute matching degree, semantic relevance, and real-time popularity. Attribute matching degree is calculated by comparing the structured attributes of candidate products with the set of confirmed preference attributes in the user's preference status information. The specific method is to count the proportion of product attributes that meet the confirmed preference attributes. For example, if a user requires the color to be off-white and the price to be less than 500 yuan, 1 point is awarded if the preference is fully met, and 0.5 points are awarded if only one preference attribute is met. This process is implemented using attribute comparison logic. Semantic relevance is measured by calculating the similarity in semantic space between the text descriptions of candidate products and the intent text set of the desired preference attributes in the user preference state information. The specific process involves using a pre-trained text embedding model, such as the Sentence-BERT model, to transform the text to be explored (e.g., retro style) and the title description of each product into high-dimensional semantic vectors. Then, the cosine similarity between these vectors is calculated as a score. This process relies on semantic vector models and similarity calculation techniques. Real-time popularity is a standardized metric obtained from a real-time statistics module in the background, reflecting the recent dynamic popularity of products. For example, it is pre-calculated based on sales and clicks within 24 hours using a time decay weighted formula. After calculating the scores for each product in the above three dimensions, the system performs a weighted sum according to the weights configured in the target matching mode to obtain a comprehensive matching score for each product. Subsequently, all products are sorted in descending order based on the comprehensive matching score to generate a ranking result. Next, based on the ranking results, the calculation results of multi-dimensional matching scores, and the confirmed preference attributes in the user preference status information, the system synthesizes corresponding explanatory text using natural language generation technology. Specifically, the system analyzes the multi-dimensional scores of top-ranked items, identifying the one or two dimensions with the highest scores as the core reason for recommendation. Then, it selects a text template corresponding to the core reason from a pre-set explanatory template library. For example, for items with outstanding attribute matching, a template emphasizing attribute fit is selected; for items with high semantic relevance, a template emphasizing style intent matching is selected. The system fills the template with specific item attribute values, user preference values, and other variable information to form the original sentence. It can further refine and connect the sentence using a fine-tuned natural language generation model, such as the T5 model, ultimately generating a fluent, natural, and personalized explanatory text. Finally, the system binds the ranked item sequence with the explanatory text generated for each item, assembling them into a structured list containing complete display information, thus forming the final recommended item list.Taking a specific scenario as an example, assuming the target matching mode is the balanced guided mode, with the following weights: attribute matching degree 0.4, semantic relevance 0.4, and real-time popularity 0.2; the user has confirmed their preferred attribute as off-white, and the text of the intent to be explored is a retro style; candidate products include product X and product Y. In the calculation, product X is off-white and its title contains retro elements, with an attribute matching degree of 1, semantic relevance of 0.9, and real-time popularity of 0.7, resulting in a weighted overall score of 0.9; product Y is light yellow and its title is minimalist, with an attribute matching degree of 0, semantic relevance of 0.2, and real-time popularity of 0.95, resulting in an overall score of 0.27; therefore, product X ranks higher than product Y. For product X, based on its high attribute matching degree and high semantic relevance, the system generates the explanatory text "This dress not only matches your desired off-white color but also satisfies your favorite retro style elements"; for product Y, based on its high real-time popularity, the system generates the explanatory text "This dress is very popular recently; perhaps you could check out this different style."

[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A product matching method driven by the intent of a shopping guide intelligent agent, characterized in that, include: S1: Receive user natural language requests and historical interaction information, and generate user preference state information based on semantic parsing, including: acquiring user historical interaction information, fusing it with the currently collected user natural language requests to form complete context information for the current dialogue; inputting the complete context information into a pre-trained natural language understanding model to perform intent classification and semantic slot filling; extracting structured data output by the natural language understanding model, wherein the structured data includes at least: user core intent categories, newly added semantic slot key-value pairs, and confirmation or correction information of historical preference attributes; maintaining user preference state information through a state machine update mechanism based on the structured data, wherein the user preference state information includes at least: a set of confirmed preference attributes, a set of preference attribute intent texts to be explored, and a set of excluded preference attributes; S2: Collect real-time user interaction data on recommended products, combine it with user preference state information to generate a user cognitive state evaluation signal, and fuse the signal to generate a cognitive convergence index. Specifically, this includes: collecting and normalizing interaction data, extracting and evaluating cognitive state, and fusing and calculating the cognitive convergence index. The step of extracting and evaluating cognitive state specifically includes: analyzing the preference orientation corresponding to each behavior category through weighted aggregation calculation to obtain the user's immediate preference feedback expression for recommended products; comparing the immediate preference feedback expression with the user preference state information, performing pattern matching through state transition rules, and identifying the dynamic change pattern of preference attributes in the user preference state information. The dynamic change pattern includes: enhancement, maintenance, or transition; based on the dynamic change pattern, calculating and outputting a multi-dimensional user cognitive state evaluation signal through predefined signal extraction logic. The user cognitive state evaluation signal includes at least: a distribution signal reflecting the concentration of positive feedback, a consistency signal reflecting the temporal change of preference attributes, and an efficiency signal reflecting the trend of decision delay. S3: Based on the cognitive convergence index, determine the target matching pattern from the preset matching strategy pattern set; S4: Based on the target matching pattern and user preference status information, perform recall, sorting, and interpretation generation on the product database to obtain a sorted list of recommended products. This includes a matching execution step and a result generation step. The matching execution step specifically includes: performing multi-path retrieval and recall from a preset product database based on the target matching pattern and user preference status information to form a candidate product set. The multi-path retrieval includes at least: precise matching retrieval based on confirmed preference attributes in the user preference status information, vector similarity retrieval of intent text for preference attributes to be explored in the user preference status information, and diversity retrieval based on product association relationships. The result generation step specifically includes: calculating multi-dimensional matching degree for products in the candidate product set based on the weights configured by the target matching pattern. The multi-dimensional matching degree includes at least: attribute matching degree, semantic relevance, and real-time popularity, and sorting according to the calculation results; synthesizing corresponding explanatory text using natural language generation technology based on the sorting results, the calculation results of the multi-dimensional matching degree, and the confirmed preference attributes in the user preference status information; and combining the sorted products with the explanatory text to form a list of recommended products.

2. The method for matching goods driven by the intent of a shopping guide intelligent agent according to claim 1, characterized in that, The steps for collecting and normalizing interaction behavior data specifically include: Collect real-time interactive behavior data of users in response to recommended products. The interactive behavior data includes at least: browsing, dwelling, clicking, adding to cart, removing, and feedback tags. Based on a preset behavior weighting system, interactive behavior data is mapped to different behavior categories; Based on the preset scaling rules, the behavioral data mapped to each category are aligned in terms of dimensions and normalized in terms of numerical range to form standardized behavioral feature vectors.

3. The method for matching goods driven by the intent of a shopping guide intelligent agent according to claim 1, characterized in that, The specific steps for fusing and calculating the cognitive convergence index include: The user's cognitive state assessment signals are weighted and fused according to preset weights to generate an initial cognitive state value; The initial cognitive state values ​​are processed by temporal smoothing to output the final cognitive convergence index.

4. The method for matching goods driven by the intent of a shopping guide intelligent agent according to claim 1, characterized in that, S3 specifically includes: The cognitive convergence index is compared with a preset multi-level convergence threshold range to generate comparison results. Based on the comparison results, the corresponding target matching pattern is determined from the preset matching strategy pattern set.

Images