E-commerce product searching method and device based on intelligent agent, medium and program product

Abstract

The invention relates to the technical field of intelligent search, and discloses an intelligent agent-based e-commerce product search method and device, a medium and a program product, the method comprises the following steps: receiving a current round of search input of a user, and analyzing and constructing a search condition; comparing historical search conditions, collecting change characteristics, and judging the stable state of the search intention; calling a search engine to output a commodity result, collecting user interaction behaviors, and judging a trust state of a search result; and adjusting a subsequent search condition construction strategy and / or a commodity sorting strategy according to the combination linkage of the two, and finally outputting a commodity search result. According to the method, the stability of the search intention and the credibility of the result are jointly evaluated and dynamically adjusted, negative circulation formed by the stability and the credibility of the search intention is broken, repeated search and invalid recommendation are reduced, the matching accuracy of the search result and the credibility of the user are improved, and e-commerce search experience and conversion efficiency are optimized.

Description

Technical Field

[0001] This application relates to the field of intelligent search technology, specifically to an e-commerce product search method, device, medium, and program product based on intelligent agents. Background Technology

[0002] With the rapid development of the e-commerce industry, the number of products on platforms is growing exponentially, and users' demand for accurate and efficient product searches is becoming increasingly urgent. Traditional e-commerce search systems are mostly built on keyword matching logic, relying on a linear process of semantic understanding, condition construction, and result ranking, assuming that the user's search intent gradually becomes clear and stable during the interaction.

[0003] However, in real-world multi-turn dialogue search scenarios, existing technologies have significant limitations. Users frequently negate and modify their established search criteria, causing their search intent to fluctuate or regress. Meanwhile, the system continues to use historical search paths for condition convergence and result reinforcement, leading to search results that deviate from the user's actual needs. When users continuously receive unexpected search results, their trust in the system gradually decreases, resulting in behaviors such as quickly skipping or repeatedly triggering searches, further exacerbating the instability of search intent and creating a negative cycle.

[0004] Existing technologies mostly optimize search intent understanding or search result ranking separately, lacking joint modeling and linkage adjustment mechanisms for the stability of search intent and the credibility of search results. They cannot distinguish between two types of result mismatch scenarios: "unclear intent" and "path failure," making it difficult to effectively solve the above problems. This leads to frequent duplicate searches and invalid recommendations, limiting the user search experience and platform conversion efficiency. Summary of the Invention

[0005] The purpose of this application is to provide an e-commerce product search method, device, medium, and program product based on intelligent agents to solve the problems mentioned in the background art.

[0006] According to one aspect of this application, an e-commerce product search method based on intelligent agents is provided, comprising the following steps: Receives the user's current search input in the search session; Semantic parsing is performed on the current round of search input, and search conditions for the current round are constructed based on the parsing results; The current round of search conditions is compared with the historical search conditions of the search session to collect the characteristics of search condition changes; Based on the characteristics of the search condition changes, calculate and determine the stable state of the search intent of the search session; Based on the current round of search conditions, the search engine is invoked to obtain a set of product search results and output to the user, and the user's interaction behavior with the set of product search results is collected; Based on the interaction behavior, calculate and determine the trust status of the search results in the search session; Based on the combination of the stable state of the search intent and the trust state of the search results, the search condition construction strategy and / or product search result ranking strategy in the subsequent search process are adjusted in a coordinated manner. Based on the search criteria construction strategy and / or product search result ranking strategy after linkage adjustment, output product search results.

[0007] Preferably, the search condition change features include: the negation behavior of the current round search condition on the historical search condition, the directional correction of the attribute values ​​in the historical search condition, the addition of new search condition attributes or constraints, and the removal or replacement of historical search condition attributes.

[0008] Preferably, calculating and determining the stable state of the search intent specifically includes: calculating the search condition change frequency and the search condition consistency degree based on the search condition change characteristics collected within a predetermined time window; comparing the search condition change frequency with a change frequency threshold, and comparing the search condition consistency degree with a consistency degree threshold; and determining the stable state of the search intent as a relatively stable state or an unstable state based on the comparison results.

[0009] Preferably, the interactive behaviors include: continuous skipping behavior of users in the search results list, re-triggering of the search without significantly modifying the search criteria, and explicit negative behavior of users towards the search results.

[0010] Preferably, calculating and determining the trust status of the search results specifically includes: calculating a trust score for the search results based on the interaction behavior collected within a predetermined time window; comparing the trust score for the search results with a trust score threshold; and determining the trust status of the search results as a relatively stable state or a declining state.

[0011] Preferably, the linkage adjustment specifically includes: when the search intent is in an unstable state, reducing the influence weight of historical search conditions during the subsequent search condition construction process; wherein, reducing the influence weight of historical search conditions specifically involves: dynamically adjusting the historical condition weight decay coefficient according to the degree of instability of the search intent; and weighting and fusing the historical search condition set and the current search condition set based on the historical condition weight decay coefficient to generate a new search condition set.

[0012] In another aspect, this application also provides an e-commerce product search device based on intelligent agents, comprising: The input receiving module receives the user's search input for the current round of the search session; The semantic parsing and condition construction module performs semantic parsing on the current round of search input and constructs the current round of search conditions based on the parsing results; The condition change feature acquisition module compares the current round of search conditions with the historical search conditions of the search session to acquire search condition change features; The intent stability determination module calculates and determines the search intent stability state of the search session based on the search condition change characteristics. The search execution and interaction collection module calls the search engine to obtain a set of product search results based on the current round of search conditions and outputs it to the user, and collects the user's interaction behavior with the set of product search results; The trust status determination module calculates and determines the trust status of the search results in the search session based on the interaction behavior. The linkage adjustment module adjusts the search condition construction strategy and / or product search result ranking strategy in the subsequent search process based on the combination of the stable state of the search intent and the trust state of the search results. The results output module outputs product search results based on the search criteria construction strategy and / or product search result ranking strategy after linkage adjustment.

[0013] This application also provides a computer device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the agent-based e-commerce product search method as described above.

[0014] In another aspect, this application provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor to implement the agent-based e-commerce product search method described above.

[0015] Another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the agent-based e-commerce product search method described above.

[0016] This application, during multi-round conversational search, accurately identifies whether the search intent is stable by collecting and analyzing the changing characteristics of search conditions, avoiding premature reinforcement of search conditions before the search needs are clear; based on the user's interaction with the search results, it effectively assesses the user's level of trust in the search results, dynamically adjusting the search result ranking strategy when trust decreases, reducing the repeated presentation of invalid results; by linking the stable state of search intent with the trust state of search results, it breaks the negative cycle formed by the two, improving the stability of the match between search results and the user's current real needs; without changing the overall architecture of the existing e-commerce search system, by optimizing and supplementing the search process, it can significantly reduce problems such as duplicate searches and invalid recommendations, improve the overall effect of converting conversational search into product lists, and enhance user trust and stickiness in the search system. Attached Figure Description

[0017] Figure 1 A schematic diagram of an e-commerce product search method based on intelligent agents provided in an embodiment of this application; Figure 2 This is a schematic diagram of the search intent stable state determination process provided in the embodiments of this application; Figure 3 A flowchart illustrating the calculation and determination of the trust status of search results provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of an e-commerce product search device based on an intelligent agent is provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It should be noted that all user information (including but not limited to user device information, user personal information, object information corresponding to device usage data, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, device usage data, etc.) involved in all embodiments of this application are information and data authorized by the user or fully authorized by all parties.

[0020] This method is applicable to e-commerce search scenarios with natural language interaction capabilities, specifically applications where users interact with the e-commerce platform's search system through text input, voice input, etc., in multiple rounds, and the system generates product search result lists multiple times during the interaction. This method can run on the e-commerce platform's server or cloud computing nodes, serving users accessing the platform via mobile terminals, web terminals, etc. E-commerce search systems typically include a semantic parsing module, a search condition construction module, a search state evaluation module, a search engine invocation module, and a search result ranking module. These modules work collaboratively to jointly evaluate and adjust the stable state of search intent and the trust state of search results. The system must have the ability to store and associate multi-round interaction data, effectively storing and retrieving user input data, search condition data, and user-search result interaction data in real time. The system should support real-time collection and analysis of search condition change characteristics and user interaction behavior characteristics to ensure the timeliness and accuracy of search state evaluation.

[0021] The implementation process of the agent-based e-commerce product search method described in this application will be explained in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions to each step made by those skilled in the art without departing from the concept of this application should be included in the protection scope of this application.

[0022] like Figure 1 As shown in the figure, this application discloses a schematic diagram 100 of an e-commerce product search method based on intelligent agents, which includes the following method steps: S1: Receive the user's current search input in the search session; S2: Perform semantic parsing on the current round of search input, and construct the current round of search conditions based on the parsing results; S3: Compare the current round of search conditions with the historical search conditions of the search session, and collect the characteristics of search condition changes; S4: Based on the search condition change characteristics, calculate and determine the stable state of the search intent of the search session; S5: Based on the current round of search conditions, call the search engine to obtain a set of product search results and output it to the user, and collect the user's interaction behavior with the set of product search results; S6: Based on the interaction behavior, calculate and determine the trust status of the search results of the search session; S7: Based on the combination of the stable state of the search intent and the trust state of the search results, adjust the search condition construction strategy and / or product search result ranking strategy in a coordinated manner during the subsequent search process. S8: Based on the search criteria construction strategy and / or product search result ranking strategy after linkage adjustment, output product search results.

[0023] In some embodiments, for step S1, the user's current round of search input in the search session is received. According to embodiments of this disclosure, after the search session is started, the e-commerce search system is in a real-time monitoring state, continuously receiving search requests submitted by the user through the terminal. The user terminal includes smartphones, tablets, desktop computers, etc., and the search input can be in the form of natural language text such as "recommendation of mid-length pure cotton trench coats suitable for working women", or text data processed by speech recognition. The speech recognition process is completed by the system's built-in speech processing module, which converts the user's speech signal into standard text format.

[0024] When the system receives the user's current search input, it first assigns a unique search session identifier to the search session. This identifier is in string form and, for example, can be generated by combining a timestamp, a random number, and the user's terminal device identifier. It is used to associate all data in the multi-round interaction process in subsequent steps, including user input data, search condition data, search result data, and user interaction behavior data, to ensure the relevance and traceability of multi-round interaction data.

[0025] Subsequently, the system associates and stores the current round of search input with the search session identifier in a distributed database (such as MongoDB) for later retrieval and analysis. Simultaneously, the system performs preliminary validation of the current round of search input, determining whether the input is empty or contains invalid characters. If an input anomaly is detected, the system returns an invalid input message to the user, guiding them to re-enter the information; if the input is valid, the system proceeds to step S2.

[0026] In some embodiments, for step S2, semantic parsing is performed on the current round of search input, and the current round of search conditions are constructed based on the parsing results.

[0027] Specifically, the system performs semantic parsing on the text input for the current round of search based on a semantic parsing module. This semantic parsing module uses a deep learning-based NLP technology stack, and its parsing process is as follows: First, the semantic parsing module performs word segmentation on the search text, using a BERT pre-trained model as the base model. This BERT pre-trained model's network structure includes a 12-layer Transformer encoder, with a hidden layer dimension of 768 and 12 heads for the multi-head attention mechanism. Using this model to segment the search text, it can accurately identify key semantic elements such as product names, product attributes, usage scenarios, and price ranges. For example, for the search text "Recommendation of mid-length pure cotton trench coats suitable for working women, priced between 1000-1500 yuan", the segmentation results include "recommendation", "suitable", "working women", "mid-length", "pure cotton", "trench coat", "price", and "1000-1500 yuan".

[0028] Secondly, after word segmentation, the semantic parsing module performs part-of-speech tagging and named entity recognition on each segmentation result. Part-of-speech tagging uses a CRF (Conditional Random Field)-based tagging algorithm to label each segmentation result as a noun, verb, adjective, quantifier, etc. Named entity recognition is based on a BERT pre-trained model combined with an e-commerce domain entity dictionary to identify product name entities such as "trench coat", product attribute entities such as "mid-length" and "pure cotton", usage scenario entities such as "working women", and price range entities such as "1000-1500 yuan" in the search text.

[0029] Subsequently, the semantic parsing module maps the identified semantic elements into structured search condition fields. These preset search condition fields include product category, product attribute, usage scenario, price range, and brand. Each semantic element is mapped to its corresponding search condition field based on its type. For example, "trench coat" is mapped to the product category field, "mid-length" and "pure cotton" to the product attribute field, "working woman" to the usage scenario field, and "1000-1500 yuan" to the price range field.

[0030] After the semantic elements are mapped, the system constructs the search conditions for the current round based on the semantic elements parsed in the current round. These search conditions are stored in key-value pairs, where the key is the name of the search condition field and the value is the content of the corresponding semantic element. For example, the search conditions for the current round can be represented as {"Product Category": "Windbreaker", "Product Attributes": ["Mid-length", "Pure Cotton"], "Usage Scenarios": "Working Women", "Price Range": [1000, 1500]}.

[0031] Finally, the system associates the constructed search criteria for the current round with the search session identifier and stores them in a distributed database for comparison and analysis with historical search criteria in subsequent steps.

[0032] Optionally, the specific structure of the BERT pre-trained model is as follows: The input layer uses the WordPiece word segmentation method to divide the search text into sub-word units, and introduces position encoding and segment encoding. Position encoding is used to represent the position information of sub-words in the text, and segment encoding is used to distinguish different text segments (this method only involves single-segment search text, and the segment encoding is uniformly 0); The encoder layer contains 12 Transformer encoder layers, each Transformer encoder layer consists of a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism has 12 heads, and each head has a dimension of 64. It captures the semantic associations of different dimensions in the text through the attention calculation of multiple heads. The feedforward neural network contains two fully connected layers, the activation function is ReLU, and the hidden layer dimension is 3072; The output layer uses the Softmax activation function for word segmentation tasks to output the probability distribution of each sub-word unit. For part-of-speech tagging and named entity recognition tasks, a CRF layer is used to tag the part-of-speech or entity type of each sub-word unit.

[0033] The training process of the model is as follows: The training dataset uses a large-scale search text dataset from the e-commerce domain, containing more than 10 million user search texts, each labeled with word segmentation results, part-of-speech tags, and named entity tags; In the pre-training stage, based on a general BERT pre-trained model such as BERT-Base, further pre-training is performed using the e-commerce domain search text dataset. The optimization objectives include Masked Language Model (MLM) and Next Sentence Prediction (NSP). The pre-training epochs are 10, the batch size is 32, and the learning rate is 2e-5; In the fine-tuning stage, for the three specific tasks of word segmentation, part-of-speech tagging, and named entity recognition, corresponding output layers are added to the pre-trained model, and fine-tuning is performed using labeled task datasets. The fine-tuning epochs are 5, the batch size is 16, the learning rate is 1e-5, and the optimizer is AdamW. In some embodiments, for step S3, the search conditions of the current round are compared with the historical search conditions of the search session to collect the characteristics of search condition changes. During the duration of the search session, the system stores the search conditions constructed in each round as historical search conditions. After the construction of the search conditions for the current round is completed, the system retrieves the historical search conditions corresponding to the search session from the distributed database, compares the current round search conditions with the historical search conditions one by one, and collects the characteristics of search condition changes.

[0034] Specifically, the characteristics of search condition changes can be categorized into four types: The first type is the negation of historical search conditions by current search conditions. The system determines whether there is a negation of existing search conditions by identifying whether the current search text contains negative words such as "no," "don't want," or "exclude," and the corresponding historical search condition elements. For example, if the product attribute in the historical search conditions includes "pure cotton," and the current search text is "don't want a pure cotton trench coat, want a polyester one," then the system identifies that the current search condition negates the historical product attribute condition of "pure cotton." The second category involves directional adjustments to attribute values ​​in historical search criteria. The system compares the values ​​of the same search condition field in the current round with those in historical search criteria to determine if there has been an adjustment in the direction of the value. For example, if the price range in historical search criteria was "1000-1500 yuan", and the price range in the current round of search criteria is adjusted to "1200-1800 yuan", then this is a directional adjustment to the price range attribute value. The third category consists of newly added search condition attributes or constraints. The system compares the search condition fields in the current round with those in the historical search condition set to determine whether there are new search condition fields not included in the historical search conditions, or constraints added to existing search condition fields. For example, if the historical search condition set does not contain a brand field, but a brand field is added to the current round's search condition set with the value "a well-known brand," then this is considered a newly added search condition attribute. The fourth category involves the removal or replacement of historical search condition attributes. The system compares the current round with the historical search condition set to determine if there are any search condition fields included in historical search conditions but not included in the current round, or if the value of a certain search condition field has been completely replaced. For example, if historical search conditions include the usage scenario field "working women," and this field is removed from the current round's search conditions, this falls under the category of historical search condition attribute removal.

[0035] The system collects the aforementioned search condition change characteristics as search intent change data, associates them with the search session identifier in chronological order, and stores them in a distributed database. Each piece of search intent change data includes information such as change feature type, value before change, value after change, and change timestamp, providing data support for subsequent judgment of the stable state of search intent.

[0036] In some embodiments, for step S4, the stable state of the search intent in the search session is calculated and determined based on the characteristics of changes in search conditions. Its core purpose is to address the technical problem in existing technologies where it is impossible to effectively identify whether the search intent is in a stable state, leading to deviations in results caused by the system continuing to use historical search paths when search intent fluctuates. Through this evaluation step, it is possible to accurately determine whether the user's search needs are in a continuously changing state at the current stage, providing a basis for adjusting subsequent search paths and achieving the technical effect of avoiding premature reinforcement of incorrect search paths.

[0037] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the search intent stability determination process provided in an embodiment of this application. Specifically, the system calculates and determines the search intent stability state based on the collected search condition change characteristics as follows: First, in S201, a predetermined time window for evaluation is determined. The system determines a predetermined time window for evaluation, which is the current round and the previous N rounds of interaction. N can be configured to be 3-5 rounds, and for example, it is configured to be 4 rounds. That is, the evaluation is based on the search condition change characteristics of the most recent 4 rounds of interaction to ensure that the evaluation results can reflect the recent changing trend of user search intent.

[0038] Secondly, in S202, the search condition change frequency is calculated. The system counts the total number of times the search condition changes within a predetermined time window, C, and calculates the search condition change frequency F. The formula for calculating the search condition change frequency F is:

[0039] Where C is the total number of times the search conditions change within the predetermined time window, and T is the total number of interaction rounds within the predetermined time window.

[0040] For example, if a search session has 3 changes in search criteria within a 4-round interaction window, such as 1 negative, 1 attribute modification, and 1 new attribute, then the search criteria change frequency F = 3 / 4 = 0.75.

[0041] In S203, the consistency of the search conditions across consecutive rounds is calculated. The system calculates the consistency S of the search conditions across consecutive rounds, and the formula for calculating the consistency S is:

[0042] Where T is the total number of interaction rounds within the predetermined time window. This represents the number of search condition fields that are identical in the search condition set of round i and round i+1. is the total number of search condition fields in the search condition sets of the i-th round and the (i + 1)-th round (taking the larger value of the two).

[0043] Exemplarily, within a predetermined time window, there are 4 rounds of interactions. The number of identical fields in the search condition sets of the 1st round and the 2nd round is 3, and the total number of fields is 4. Among them, the 1st round has 4 fields and the 2nd round has 4 fields. Then ; the number of identical fields in the 2nd round and the 3rd round is 2, and the total number of fields is 4. Then ; the number of identical fields in the 3rd round and the 4th round is 1, and the total number of fields is 5. Among them, the 3rd round has 4 fields and the 4th round has 5 fields. Then ; then the search condition consistency degree S = (0.75 + 0.5 + 0.2) / (4 - 1) = 1.45 / 3 ≈ 0.483.

[0044] After calculating the search condition change frequency F and the search condition consistency degree S, the system can preset two thresholds: the change frequency threshold F0 and the consistency degree threshold S0. The change frequency threshold F0 can be configured to be 0.5 - 0.7. Exemplarily, it is configured to be 0.6; the consistency degree threshold S0 can be configured to be 0.4 - 0.6. Exemplarily, it is configured to be 0.5.

[0045] Subsequently, in S204, the search intent stable state is determined. The system determines the search intent stable state based on the comparison results of F and F0, S and S0: when F ≤ F0 and S ≥ S0, it indicates that the search conditions change less within the predetermined time window and the search condition consistency between consecutive rounds is relatively high. The system determines the search intent stable state of the current search session as a relatively stable state. At this time, the user's search requirements are basically clear, and the system can follow the current search path for condition convergence and result enhancement; when F > F0 and S < S0, it indicates that the search conditions change frequently within the predetermined time window and the search condition consistency between consecutive rounds is relatively low. The system determines the search intent stable state of the current search session as an unstable state. At this time, the user's search requirements are still changing continuously, and the system needs to reduce the weight of the historical search conditions to avoid premature convergence; when F > F0 and S ≥ S0 or F ≤ F0 and S < S0, the system further combines the type of search condition change for determination. If the change type is mainly new attribute constraints and there are no negation and major modification behaviors, it is determined as the demand exploration stage (a sub-type of the search intent unstable state); if the change type includes multiple negation and major modification behaviors, it is determined as the intent oscillation stage (also a sub-type of the search intent unstable state).

[0046] After the determination of the search intent stable state is completed, the system associates and stores the determination result with the current search session identifier and synchronizes it to the search state evaluation module in real time as the basis for subsequent linkage adjustment.

[0047] In some embodiments, for step S5, the system invokes the search engine based on the current round of search conditions to obtain a set of product search results and outputs it to the user, while also collecting user interaction with the set of product search results. The system invokes the search engine based on the search conditions constructed for the current round. The search engine constructs a search query statement according to the search conditions, retrieves matching product data from the product index, and obtains a set of product search results. The product index adopts a hierarchical index design, including basic information indexes (name, brand, price), attribute indexes (size, color, material), and description indexes (product details, user reviews), etc. The retrieval process supports multiple retrieval methods such as Boolean search, fuzzy search, and range search to ensure the comprehensiveness and accuracy of the retrieval results.

[0048] After obtaining the product search results set, the system performs a preliminary sorting of the product search results set through the search results sorting module. The sorting criteria include the semantic similarity score between the product and the search conditions, the product popularity score (calculated based on sales volume, positive review rate, page views, etc.), etc., and generates a search results list. Then, the search results list is transmitted to the user terminal and displayed to the user through the API gateway.

[0049] After the search results are output, the system initiates a user interaction behavior collection program. This program combines front-end tracking and back-end logging to collect user interactions with the search results list. The collected interactions fall into three categories: The first category is continuous skipping behavior in the search results list. The system judges this by recording data such as the user's scrolling speed, dwell time, and clicks. For example, a preset scrolling speed threshold is 3 seconds per screen. If a user scrolls more than 3 screens within 3 seconds without clicking on any product, it is considered continuous skipping behavior. The second category is re-triggering the search without significantly modifying the search criteria. The system monitors user actions after the search results are displayed. If a user clicks the "Research" or "Regenerate" button more than twice without entering new search text or modifying search criteria, it is considered this type of behavior. The third category is explicit negative feedback on the search results. The system identifies user feedback after the search results are displayed. If a user enters negative text such as "does not match," "doesn't look good," or "not what I wanted," or clicks the "Dislike" or "Exclude" buttons in the search results list, it is considered explicit negative behavior.

[0050] The system collects the above-mentioned interactive behaviors as search result interaction data, and stores them in a distributed database in chronological order along with the search session identifier. Each interactive behavior data includes information such as behavior type, behavior occurrence time, and corresponding product identifier (if any), providing data support for the subsequent judgment of the trust status of search results.

[0051] In some embodiments, for step S6, the trust status of the search results in the search session is calculated and determined based on the interaction behavior. Its core purpose is to address the technical problem in the prior art of lacking an effective assessment of the user's trust status in search results, which leads to the inability to adjust search strategies in a timely manner when the user's trust in search results declines. Through this assessment step, the user's level of trust in the current search results can be accurately determined, providing a basis for subsequent coordinated adjustments and achieving the technical effect of proactively adjusting the search strategy to restore user trust when the trust in search results declines.

[0052] Please see Figure 3 , Figure 3 This is a flowchart illustrating the process of calculating and determining the trust status of search results, as provided in an embodiment of this application.

[0053] Specifically, the system calculates and determines the trust status of search results based on the collected interaction behavior as follows: First, in S301, the predetermined time window for the evaluation is determined. The system determines the predetermined time window for the evaluation, which is consistent with the predetermined time window for the evaluation of the stable state of the search intent, i.e., the current round and the previous N rounds of interaction, such as N configured as 3-5 rounds, to ensure that the evaluation results are consistent with the time dimension of the evaluation of the stable state of the search intent.

[0054] Secondly, in S302, key behavioral indicators are statistically analyzed. The system statistically analyzes three types of key user behaviors within a predetermined time window: The first type of metric is the skip frequency P, defined as the ratio of the number of consecutive skips by a user within a predetermined time window to the number of times search results are displayed. The calculation formula is as follows:

[0055] in, The number of consecutive skips within a predetermined time window. The total number of times search results are displayed within the scheduled time window.

[0056] For example, if the search results are displayed 4 times within a predetermined time window and the user skips them 3 times consecutively, then the skip frequency P = 3 / 4 = 0.75.

[0057] The second metric is the number of repeated search triggers, Q, defined as the total number of times a user repeatedly triggers a search or requests to regenerate search results within a predetermined time window without significantly modifying the search criteria. For example, if a user clicks the "Regenerate" button 3 times within the predetermined time window without modifying the search criteria, then Q=3.

[0058] The third type of metric is the number of negative actions R, which is defined as the total number of explicit negative actions of the user on the displayed results within a predetermined time window. Exemplarily, if the user enters negative text 2 times and clicks the "dislike" button 1 time within the predetermined time window, then R = 3.

[0059] In S303, a search result trust score model is constructed. After statistically obtaining the above three types of key behavior metrics, the system constructs a search result trust score model to calculate the trust score T of the user for the current search result. The score model is as follows:

[0060] where are weight coefficients, satisfying and The value ranges of both are 0 - 1, and they are obtained through machine learning algorithms (such as logistic regression) training based on the historical data of the e-commerce platform. Exemplarily, after training ; is the total number of interaction rounds within the predetermined time window; P is the skip frequency; Q is the number of repeated search triggers; R is the number of negative actions.

[0061] In S304, the trust state of the result is determined based on a preset threshold. The system presets a trust score threshold T0, and T0 can be configured to be 0.4 - 0.6. Exemplarily, it is configured to be 0.5. According to the comparison result of the trust score T and the threshold T0, the trust state of the search result is determined: when T ≥ T0, it indicates that the user has a relatively high trust in the search result, and the system determines the trust state of the search result in the current search session as a relatively stable state. At this time, the system can continue to use the current search strategy; when T < T0, it indicates that the user has a relatively low trust in the search result, and the system determines the trust state of the search result in the current search session as a decreasing state. At this time, the system needs to adjust the search strategy in a timely manner to restore the user's trust.

[0062] After the determination of the search result trust state is completed, the system associates and stores the determination result with the current search session identifier, and synchronizes it to the search state evaluation module in real time as the basis for subsequent linkage adjustment.

[0063] In some embodiments, for step S7, its core purpose is to solve the technical problem in the prior art that the understanding of search intent and the ranking of search results are independently optimized, lacking a linkage adjustment mechanism, resulting in a negative cycle of search intent oscillation and a decline in search result trust. Through the linkage adjustment in this step, the system can synergistically adjust the search path and result ranking strategy according to the combination of the stable state of the search intent and the trust state of the search result, inhibit the negative cycle, and improve the matching stability between the search result and the user's current true needs.

[0064] Specifically, based on the combination of the stable state of search intent and the trust state of search results, the system generates linkage adjustment instructions through the search state evaluation module, which are then sent to the search condition construction module and the search result ranking module to dynamically adjust the search condition construction strategy and / or the product search result ranking strategy. The specific adjustment logic is as follows: When the search intent and search result trust levels are relatively stable, the user's search needs are largely clear, and they have a high level of trust in the current search results. The system maintains its current search condition construction strategy and product search result ranking strategy. The search condition construction module continues to build a set of search conditions based on historical search conditions and the current search input, maintaining the weighting of historical search conditions. The search result ranking module continues to use the original ranking formula, combining semantic similarity scores, personalized recommendation scores, and product popularity scores to ensure the consistency and stability of the search results.

[0065] When the search intent is in an unstable state and the search result trust state is relatively stable, the user's search needs are still changing, but the user trusts the current search results. The system mainly adjusts the search condition construction strategy and keeps the product search result ranking strategy basically unchanged.

[0066] In terms of search condition construction, the search condition construction module reduces the influence weight of historical search conditions in the process, preventing premature convergence of search conditions. Specifically, a historical condition weight decay coefficient is introduced. , The value ranges from 0.3 to 0.7, and is dynamically adjusted according to the degree of instability of the search intent. The higher the frequency of search condition changes (F), the better. The smaller the value, the better. For example, when F=0.7, When F=0.6, .

[0067] The new search criteria are constructed using a weighted fusion method, which combines historical search criteria with the new search criteria obtained from parsing the current search input. The fusion formula is as follows:

[0068] in, This is the new set of search conditions generated after fusion. For the set of historical search criteria, The new set of search conditions is parsed from the current search input. This is the historical condition weight decay coefficient.

[0069] Regarding the ranking of product search results, the system maintains the original ranking strategy basically unchanged, but appropriately increases the ranking weight of products with new attributes to adapt to changes in user search needs. The ranking weight increase coefficient for products with new attributes is 1.2-1.5, for example, the increase coefficient is 1.3, to ensure that products with new attributes can be displayed to users first.

[0070] When the search intent is in a relatively stable state and the trust in the search results is in a decreasing state, the user's search needs are basically clear, but the trust in the current search results is low. The system mainly adjusts the product search result ranking strategy while keeping the search condition construction strategy basically unchanged.

[0071] Regarding the ranking of product search results, the search result ranking module has implemented the following two adjustments: First, the ranking weight of products that have been displayed multiple times is reduced. The system queries product display records to calculate the number of times each product is displayed (n) in the current search session, and introduces a display frequency decay coefficient. , The calculation formula is:

[0072] Where k is the attenuation coefficient, which can be configured to 0.2-0.5. For example, k=0.3, and n is the number of times the product is displayed in the current search session.

[0073] Second, increase the weight of products with attribute novelty in the ranking. Products with attribute novelty refer to those that have a moderate match with the current search criteria but contain new product attributes or styles. The system filters such products by calculating a product attribute novelty score N. The formula for calculating the product attribute novelty score N is:

[0074] Where m is the number of new attributes of the product that have not appeared in the current search session, and M is the total number of attributes of the product.

[0075] The system presets an attribute novelty threshold N0, which can be configured to 0.2-0.4. For example, N0=0.3. Products with N≥N0 are judged as having attribute novelty, and their ranking weight is increased in the ranking process. The weight increase coefficient is 1.5-2.0. For example, the increase coefficient is 1.8. At the same time, the proportion of such products in the search results list is controlled to be 30%-50%. For example, the proportion is 40%, to help users rebuild trust in the search.

[0076] Regarding search criteria construction, the system maintains the current search criteria construction strategy. However, the search criteria construction module needs to monitor users' interaction with the new ranking results in real time. If users click on products with novel attributes more frequently, relevant new attributes will be appropriately introduced in subsequent search criteria construction.

[0077] When the search intent is in an unstable state and the trust in the search results is decreasing, it means that the user's search needs are constantly changing and the trust in the current search results is low. The system needs to adjust the search criteria construction strategy and the product search result ranking strategy at the same time to break the negative cycle.

[0078] In terms of constructing search criteria, the system takes the following two measures: First, further reduce the influence weight of historical search conditions, and decrease the weight decay coefficient of historical conditions. The value ranges from 0.1 to 0.3, for example, This reduces the interference of historical search criteria on the current search.

[0079] Second, the search context of the current search session is pruned. The search context includes historical search inputs, historical search conditions, and historical search results. The pruning process is as follows: retain the search inputs and search conditions from the most recent two rounds in the current search session, and delete earlier search context data; at the same time, remove search condition elements related to the user's explicit negative behavior. For example, if the user negatively rejects the "pure cotton" attribute, then the "pure cotton" attribute constraint will be directly excluded in subsequent search condition construction.

[0080] After the search context is trimmed, the search condition building module reconstructs the search condition set based on the two most recent search inputs and the current search input. The new search condition set only contains the core semantic elements parsed from the current search input and the elements that were not negated in the two most recent search conditions, ensuring that the search conditions are highly relevant to the user's current needs.

[0081] Regarding the ranking of product search results, in addition to implementing the ranking adjustment measures mentioned above when "search intent is in a relatively stable state and search result trust is decreasing" (reducing the weight of products that have been displayed multiple times and increasing the proportion of products with novel attributes), the system also needs to further optimize the ranking formula by introducing search intent stability score and search result trust score as ranking factors. The new ranking formula is as follows:

[0082] in, For semantic similarity scores, To score personalized recommendations, To score the popularity of a product, The search intent stability score (when the search intent stability state is relatively stable) is used to calculate the search intent stability score. When unstable T represents the trust score of the search results; For the weighting coefficients, satisfying For example, .

[0083] By introducing search intent stability score and search result trust score, the ranking results can better adapt to the current search status and further improve the matching degree between search results and users' current real needs.

[0084] During the coordinated adjustment process, the system records various adjustment parameters in real time, such as the historical condition weight attenuation coefficient. The adjustment process, including parameters such as the display frequency decay coefficient k and the attribute novelty threshold N0, along with adjustment effect data such as adjusted user click frequency, conversion rate, and search session duration, provides a basis for subsequent parameter optimization. Simultaneously, the adjustment process must be completed before the existing semantic parsing module and search engine are invoked, without involving new hardware devices or dedicated sensors, ensuring that the adjustment mechanism can be seamlessly integrated into the existing e-commerce search system.

[0085] In some embodiments, for step S8, after the above-mentioned linkage adjustment is completed, the search condition construction module sends the adjusted search condition set (if the search condition construction strategy is adjusted) to the search engine calling module. The search engine calling module constructs a retrieval query statement based on the new search condition set, and calls the search engine again to retrieve matching product data from the product index library to obtain a new set of product search results.

[0086] After the new set of product search results is transmitted to the search result sorting module, the module sorts them using the adjusted sorting strategy (if the product search result sorting strategy has been adjusted), generating the final product search result list. Subsequently, the system transmits the final product search result list to the user terminal via the API gateway and displays it to the user, completing the output of search results after a coordinated adjustment.

[0087] After viewing the final product search results list, users can click, add to cart, or place an order. The system records the user's subsequent actions in real time, serving as a basis for future search optimization. When a user ends their search session (e.g., closes the search page or exits the app) or reaches predetermined termination conditions (e.g., the search session duration exceeds 30 minutes or the number of interaction rounds exceeds 10), the system releases temporary state data associated with that search session, including search session identifier, search intent change data, user interaction behavior data, and linked adjustment parameters. Only long-term user behavior data, such as purchase history and preference tags, is retained for subsequent personalized recommendation model training.

[0088] Therefore, during multi-round dialogue search, by collecting and analyzing the changing characteristics of search conditions, it is possible to accurately identify whether the search intent is stable, avoiding premature reinforcement of search conditions before the search needs are clear. Based on the user's interaction with the search results, it is possible to effectively assess the user's level of trust in the search results, and dynamically adjust the search result ranking strategy when trust decreases, reducing the repeated presentation of invalid results. By linking the stable state of search intent with the trust state of search results, the negative cycle formed by the two is broken, improving the stability of the match between search results and the user's current real needs. Without changing the overall architecture of the existing e-commerce search system, by optimizing and supplementing the search process, problems such as duplicate searches and invalid recommendations can be significantly reduced, improving the overall effect of converting conversational search into product lists, while enhancing user trust in the search system and user stickiness.

[0089] Please see Figure 4 , Figure 4 This application provides an embodiment of an e-commerce product search device 400 based on an intelligent agent. This device embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiment, this device can be specifically applied to various computer devices. The device specifically includes: The input receiving module 401 receives the user's current search input in the search session; The semantic parsing and condition construction module 402 performs semantic parsing on the current round of search input and constructs the current round of search conditions based on the parsing results; The condition change feature acquisition module 403 compares the current round of search conditions with the historical search conditions of the search session and acquires the search condition change features; The intent stability determination module 404 calculates and determines the search intent stability state of the search session based on the search condition change characteristics. The search execution and interaction acquisition module 405 calls the search engine to obtain a set of product search results based on the current round of search conditions and outputs it to the user, and collects the user's interaction behavior with the set of product search results; Trust status determination module 406 calculates and determines the trust status of the search results of the search session based on the interaction behavior. The linkage adjustment module 407 adjusts the search condition construction strategy and / or product search result ranking strategy in the subsequent search process based on the combination of the stable state of the search intent and the trust state of the search results. The result output module 408 outputs product search results based on the search condition construction strategy and / or product search result sorting strategy after linkage adjustment.

[0090] Based on the same inventive concept, this application also provides a computer device, the method corresponding to which can be the method in the foregoing embodiments, and the principle of solving the problem is similar to that method. The computer device provided in this application includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the methods and / or technical solutions of the foregoing embodiments of this application.

[0091] The computer device can be a user device, or a device formed by integrating user devices and network devices through a network, or it can be an application running on the aforementioned devices. The user device includes, but is not limited to, various terminal devices such as computers, mobile phones, tablets, smartwatches, and smart bands. The network device includes, but is not limited to, network hosts, single network servers, multiple network server sets, or cloud computing-based computer sets, and can be used to implement some processing functions when setting an alarm clock. Here, the cloud consists of a large number of hosts or network servers based on cloud computing. Cloud computing is a type of distributed computing, consisting of a virtual computer composed of a group of loosely coupled computer sets.

[0092] Figure 5 The diagram illustrates the structure of an apparatus suitable for implementing the methods and / or technical solutions in the embodiments of this application. The apparatus 500 includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 502 or a program loaded from a storage portion 508 into a random access memory (RAM) 503. The RAM 503 also stores various programs and data required for system operation. The CPU 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0093] The following components are connected to I / O interface 505: input section 506 including keyboard, mouse, touch screen, microphone, infrared sensor, etc.; output section 507 including cathode ray tube (CRT), liquid crystal display (LCD), LED display, OLED display, etc., and speakers, etc.; storage section 508 including one or more computer-readable media such as hard disk, optical disk, magnetic disk, semiconductor memory, etc.; and communication section 509 including network interface card such as LAN (local area network) card, modem, etc. Communication section 509 performs communication processing via a network such as the Internet.

[0094] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by the central processing unit (CPU) 501, it performs the functions defined in the methods of this application.

[0095] Another embodiment of this application provides a computer-readable storage medium having computer program instructions stored thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0096] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0097] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0098] Furthermore, the inclusion of a single word does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device through software or hardware. Terms such as "first," "second," etc., are used to indicate names and do not indicate any particular order.

Claims

1. A method for searching e-commerce products based on intelligent agents, characterized in that, Includes the following steps: Receives the user's current search input in the search session; Semantic parsing is performed on the current round of search input, and search conditions for the current round are constructed based on the parsing results; The current round of search conditions is compared with the historical search conditions of the search session to collect the characteristics of search condition changes; Based on the characteristics of the search condition changes, calculate and determine the stable state of the search intent of the search session; Based on the current round of search conditions, the search engine is invoked to obtain a set of product search results and output to the user, and the user's interaction behavior with the set of product search results is collected; Based on the interaction behavior, calculate and determine the trust status of the search results in the search session; Based on the combination of the stable state of the search intent and the trust state of the search results, the search condition construction strategy and / or product search result ranking strategy in the subsequent search process are adjusted in a coordinated manner. Based on the search criteria construction strategy and / or product search result ranking strategy after linkage adjustment, output product search results.

2. The e-commerce product search method based on intelligent agents according to claim 1, characterized in that, The characteristics of the search condition changes include: the negation of historical search conditions by the current round of search conditions, the directional correction of attribute values ​​in historical search conditions, the addition of new search condition attributes or constraints, and the removal or replacement of historical search condition attributes.

3. The e-commerce product search method based on intelligent agents according to claim 1, characterized in that, The calculation and determination of the stable state of the search intent specifically includes: calculating the frequency of search condition changes and the degree of consistency of search conditions based on the search condition change characteristics collected within a predetermined time window; comparing the frequency of search condition changes with a change frequency threshold and comparing the degree of consistency of search conditions with a consistency degree threshold; and determining the stable state of the search intent as a relatively stable state or an unstable state based on the comparison results.

4. The e-commerce product search method based on intelligent agents according to claim 1, characterized in that, The interactive behaviors include: users continuously skipping search results, users re-triggering a search without significantly modifying their search criteria, and users explicitly rejecting search results.

5. The e-commerce product search method based on intelligent agents according to claim 1, characterized in that, The calculation and determination of the trust status of the search results specifically includes: calculating the trust score of the search results based on the interaction behavior collected within a predetermined time window; comparing the trust score of the search results with a trust score threshold, and determining the trust status of the search results as a relatively stable state or a declining state.

6. The e-commerce product search method based on intelligent agents according to claim 1, characterized in that, The linkage adjustment specifically includes: when the search intent is in an unstable state, reducing the influence weight of historical search conditions during the subsequent search condition construction process; wherein, reducing the influence weight of historical search conditions specifically involves: dynamically adjusting the historical condition weight decay coefficient according to the degree of instability of the search intent; and weighting and fusing the historical search condition set and the current search condition set based on the historical condition weight decay coefficient to generate a new search condition set.

7. An e-commerce product search device based on intelligent agents, characterized in that, include: The input receiving module receives the user's search input for the current round of the search session; The semantic parsing and condition construction module performs semantic parsing on the current round of search input and constructs the current round of search conditions based on the parsing results; The condition change feature acquisition module compares the current round of search conditions with the historical search conditions of the search session to acquire search condition change features; The intent stability determination module calculates and determines the search intent stability state of the search session based on the search condition change characteristics. The search execution and interaction collection module calls the search engine to obtain a set of product search results based on the current round of search conditions and outputs it to the user, and collects the user's interaction behavior with the set of product search results; The trust status determination module calculates and determines the trust status of the search results in the search session based on the interaction behavior. The linkage adjustment module adjusts the search condition construction strategy and / or product search result ranking strategy in the subsequent search process based on the combination of the stable state of the search intent and the trust state of the search results. The results output module outputs product search results based on the search criteria construction strategy and / or product search result ranking strategy after linkage adjustment.

8. A computer device, wherein the computer device is characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.

9. A computer-readable medium having computer program instructions stored thereon, characterized in that, The computer program instructions can be executed by a processor to implement the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images