Commodity search result display method and device, equipment and medium

Abstract

The invention relates to a commodity search result display method and device, equipment and a medium, and the method comprises the steps: calling a preset intention extraction model when a commodity query request submitted by a user cannot query effective commodity data from an own commodity library of a current independent station, outputting a corresponding intention description message body corresponding to the original query statement, wherein the intention description message body comprises a target language, the original query statement and a scene tag and an attribute tag to which a commodity expected by a user belongs; calling a preset query generation model, and generating an extended query statement expressed in a target language according to the intention description message body and the consumption style preference characteristics of the user to embody the complete intention and the consumption style of the user; and constructing a placeholder interaction information block according to the extended query statement, and pushing the placeholder interaction information block to a search result page of the user for display. Therefore, the user residence rate and the commodity conversion rate of the independent station during user search can be improved, and the user experience is comprehensively improved.

Description

Technical Field

[0001] This application relates to the field of e-commerce technology, and in particular to a method, apparatus, and medium for displaying product search results. Background Technology

[0002] Independent websites, built independently by businesses or individuals, serve as e-commerce sites or brand websites with complete control. Each independent website operates independently and typically provides online transaction services to users in multiple countries worldwide. However, independent websites often face challenges such as limited product inventory, complex multi-language compatibility, and fragmented user search intent. These issues frequently result in users encountering zero-result pages when searching for products on independent websites.

[0003] Specifically, the average number of products on independent websites operated by small and medium-sized brands is only a few thousand, or even less, far lower than that of traditional e-commerce platforms. This results in many user searches triggering zero-result pages. Furthermore, user queries involve dozens of languages ​​globally and contain slang, spelling errors, and other issues. Traditional product recommendation algorithms, such as keyword matching and semantic matching, are prone to failure in these situations. Cross-border users' searches often include contextual needs, such as "beach weddingguest dress," which traditional semantic matching techniques simply cannot capture.

[0004] Existing optimization solutions for zero-results pages on independent websites mainly include static suggestion solutions, keyword expansion solutions, and third-party recommendation solutions. Static suggestion solutions typically display fixed text, such as "No related products found," along with a list of popular products sorted by sales volume across the entire site. This solution doesn't consider the user's current query intent, leading to ineffective intent matching. Keyword expansion solutions mechanically recommend similar search terms in isolation using a thesaurus, such as expanding "pants" to "trousers," but they cannot handle contextualized queries. Third-party recommendation solutions integrate with affiliate marketing product libraries, but this can lead to user churn, which is not the intention of independent websites.

[0005] This demonstrates the significant limitations of existing technologies in handling user queries. Firstly, when traditional recommendation algorithms fail to match valid product data for a user query, conventional solutions fail to address the user's true needs, especially when the query contains spelling errors, multiple languages, or contextual requirements. Secondly, they fail to fully explore and utilize the semantic relationships between user queries and product information in the actual product database, leading to user churn on independent websites when no effective solutions are available. Furthermore, they fail to provide effective solutions that address the unique characteristics of independent websites, such as single-site operation and a limited number of products. All of these factors contribute to a poor user experience when encountering zero-result pages on independent websites, necessitating urgent improvement. Summary of the Invention

[0006] The purpose of this application is to solve at least one of the above-mentioned problems by providing a method for displaying product search results and corresponding apparatus, devices, non-volatile readable storage media, and computer program products.

[0007] According to one aspect of this application, a method for displaying product search results is provided, comprising: In response to a user's product query request, based on the original query statement carried in the request, the system checks whether there is any matching valid product data in the current independent website's own product database. When no valid product data exists, a preset intent extraction model is invoked to identify the target intent contained in the original query statement in the target language corresponding to the original query statement, and output a corresponding intent description message body. The intent description message body includes the target language, the original query statement, and the user's expected product scene tag and attribute tag. The system invokes a preset query generation model to generate an extended query statement in the target language based on the intent description message body and the user's consumption style preference characteristics. The extended query statement reflects the complete intent expressed by the intent description message body and the user's consumption style. A placeholder interactive information block is constructed based on the extended query statement, and the placeholder interactive information block is pushed to the user's search results page for display. The placeholder interactive information block is suitable for the user to access product data retrieved from the proprietary product library by referring to the semantics of the extended query statement.

[0008] According to another aspect of this application, a product search result display device is provided, comprising: The regular query module is set to respond to the product query request submitted by the user and, based on the original query statement carried in the request, query whether there is any matching valid product data in the current independent website's own product database; The intent-based recommendation module is configured to call a preset intent extraction model when no valid product data exists. This model identifies the target intent contained in the original query statement in the target language of the original query statement and outputs a corresponding intent description message body. The intent description message body includes the target language, the original query statement, and the user's desired product scenario tag and attribute tag. The query extension module is configured to call a preset query generation model to generate an extended query statement in the target language based on the intent description message body and the user's consumption style preference characteristics. The extended query statement reflects the complete intent expressed by the intent description message body and the user's consumption style. The results display module is configured to construct a placeholder interactive information block based on the extended query statement, and push the placeholder interactive information block to the user's search results page for display. The placeholder interactive information block is suitable for the user to access product data retrieved from the proprietary product library by referring to the semantics of the extended query statement.

[0009] According to another aspect of this application, a product search result display device is provided, including a central processing unit and a memory, wherein the central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method described in this application.

[0010] According to another aspect of this application, a non-volatile readable storage medium is provided, which stores a computer program implemented according to the product search result display method in the form of computer-readable instructions, wherein the computer program, when invoked by a computer, executes the steps included in the method.

[0011] According to another aspect of this application, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the method.

[0012] Compared to traditional technologies, this application effectively solves the zero-results page problem encountered by users searching for products on independent websites through refined intent extraction and extended query generation, significantly improving user experience and the relevance of search results. First, an intent extraction model deeply analyzes user queries, generating an intent description message body containing target language, scenario tags, and attribute tags. This ensures accurate extraction of the user's true intent even when the query contains spelling errors, multiple languages, or contextual requirements. Second, a query generation model generates extended query statements based on the intent description message body and user consumption style preferences. This fully explores the deep semantic relationship between the original query statement and the independent website's product database, thereby recalling more indirectly related product data within the independent website, enriching user personalization features, and improving product conversion rates. Furthermore, the entire business logic is implemented within the independent website, making full use of limited resources to maximize the matching of user needs. This ensures a comprehensive improvement in user experience, reduces user churn, and enhances user retention even with a limited product selection on the independent website. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating one embodiment of the product search result display method of this application; Figure 2 This is a schematic block diagram of the product search result display device of this application; Figure 3 This is a schematic diagram of the structure of a product search result display device used in this application. Detailed Implementation

[0014] Taking independent websites on cross-border e-commerce platforms as an example, these websites, deployed by merchants, operate in multiple regions globally. Technical support is provided by independent website service providers, allowing merchants to configure their site content using simple and efficient editing tools, such as layout design and product data, thereby building their independent website page information and their own product library. Service providers offer various editing and operational tools for merchants to use.

[0015] Product searches are a frequent occurrence on independent websites. Independent website service providers offer standardized search architectures. Typically, when a user performs a product search on an independent website, the corresponding search request is submitted to a standardized search interface for execution. The search interface uses pre-defined product recommendation algorithms, such as keyword matching and collaborative filtering algorithms, to attempt to retrieve relevant products from its own product library and provide a list of corresponding product search results to the independent website. The independent website then constructs a search results page based on this list and returns it to the user for browsing.

[0016] When the product search results list does not contain any valid product data, if the current independent website only returns a blank page to the user, it will not meet the expectations of this application. Therefore, a product search results display method of this application can be implemented to provide the user with a corresponding results page.

[0017] Please see Figure 1 According to a product search result display method provided in this application, in some embodiments, the method includes the following steps: Step S3100: Respond to the product query request submitted by the user, and query the existing independent website's own product database to see if there is any matching valid product data based on the original query statement carried in the request. When a user enters a query in the search box of an independent website, it is submitted to the server. The server parses the request and extracts the original query statement. The original query statement is the text information entered by the user to describe the desired product, such as "running shoes" or "beach wedding guest dress". The server first submits the original query statement to the product recommendation engine. According to the product recommendation engine's preset business logic, it queries the independent website's own product database to see if there is any matching valid product data. The own product database is a product information database maintained internally by the independent website, containing detailed information such as product titles, descriptions, attribute tags, and product images. Valid product data refers to product information that is highly relevant to the user's query intent. This product data can be included in the product search results list generated by the product recommendation engine, directly satisfying the user's search needs.

[0018] To implement this query process, the business logic of a product recommendation engine can be implemented in various ways. For example, in one embodiment, a keyword matching algorithm can be used to compare the keywords in the user's original query with the product titles and descriptions in the product database to find products with a high degree of matching. In another embodiment, semantic analysis technology can be combined to understand the semantic intent of the user's query, thereby retrieving products from the product database that meet the user's needs more accurately. In other embodiments of practical applications, keyword matching and semantic analysis can be used in combination to improve the accuracy and recall of the query. For example, when a user enters "running shoes," not only will products containing the keyword "running shoes" in their titles or descriptions be searched, but semantic analysis will also be used to understand that the user may be interested in sports, fitness, and other related products, thereby expanding the search scope and improving the relevance of the search results. In addition, the business logic of a product recommendation engine can also be implemented using various mature collaborative filtering algorithms.

[0019] Product search engines are typically deployed independently of independent websites. When a search is needed, the independent website submits a raw query by calling the product search engine's preset interface. The product search engine then searches its own product database according to its preset business logic to retrieve the user's desired valid product data and constructs a corresponding product search results list. After the user's terminal device receives this product search results list, if the number of valid product data in the product search results list is not zero, the valid product data in the product search results list can be displayed on the search results page; if the number of valid product data is zero, subsequent steps can be executed.

[0020] Step S3200: When there is no valid product data, call the preset intent extraction model, identify the target intent contained in the original query statement according to the target language of the original query statement, and output the corresponding intent description message body. The intent description message body includes the target language, the original query statement, and the user's expected product scene tag and attribute tag. When a product recommendation engine fails to find valid product data matching the user's original query from the product database of the current independent website, it is usually because the product data in the database, such as product titles and descriptions, does not match the original query in terms of keywords or semantics, or although there is some relevance, the relevance is extremely low to the point of being unbelievable. In this case, the product recommendation engine returns an empty list.

[0021] To address the issue of product search results lists lacking any valid product data, this application invokes a pre-defined intent extraction model. The main task of this model is to identify the user's target intent within the text, corresponding to the target language of the user's original query, and output an intent description message. This intent description message can contain important information such as the target language, the original query, and the scene and attribute tags associated with the desired product. The target language can be determined before invoking the pre-defined intent extraction model and provided to it. Scene and attribute tags refer to relevant semantic information that may be contained or indirectly reflected in the user's original query, which can be determined by the intent extraction model using its learned multilingual reasoning capabilities based on the semantic features of the original query.

[0022] Intent extraction models can be implemented in various ways. For example, in one embodiment, natural language processing techniques from deep learning can be used to train a multilingual model based on the Transformer architecture, such as mBERT, to identify query intents in different languages. When processing user-input query text, the model can understand and extract key semantic information from the text, corresponding to the specified target language, including the type of product the user expects, the usage scenario, product attributes, etc. Taking the user input "beach wedding guest dress" as an example, the intent extraction model can identify, based on the fact that the target language is English, that the product the user expects is a dress related to a beach wedding guest scenario, and may extract attribute tags such as "dress" and scenario tags such as "beach wedding," which will be included in the intent description message body of the output.

[0023] The intent extraction model is pre-trained using data pairs consisting of input text in a specified target language and a corresponding intent message body. The input text may contain minimal information, while the intent message body may contain pre-annotated intent information. The semantic relationship between the two is determined by those skilled in the art when constructing the data pairs. For example, to achieve the functionality required in this step, the input text could be a single product term such as "massage pillow," and the intent message body could be: Original query: massage pillow; Target language: Chinese; Scene tags: Bedding, office supplies, cervical spine therapy; Attribute tags: Neck massage, office care, middle-aged and elderly As the examples above demonstrate, the data pairs used to train the intent extraction model rely on extremely limited information provided by the input text. The input corpus can be the simplest words directly representing the product name, while the corresponding intent message body can broadly encompass multiple scene tags and attribute tags for the product referred to by those simple words. These scene tags and attribute tags may not be directly related to the simple words on a literal level, but they are indirectly related in terms of product function and purpose. In other words, the data pairs used to train the intent extraction model extensively expand the semantics of the corresponding input text in terms of both scene and attribute aspects. This extensive expansion capability is typically lacking in traditional product recommendation algorithms. The intent extraction model trained in this way, compared to traditional product recommendation algorithms, can help expand the original input text submitted by users to include richer semantics in terms of scene and attribute aspects.

[0024] Step S3300: Invoke the preset query generation model and generate an extended query statement in the target language based on the intent description message body and the user's consumption style preference characteristics. The extended query statement reflects the complete intent expressed by the intent description message body and the user's consumption style. The query generation model of this application can be trained to convergence beforehand using training samples corresponding to the functions required to achieve its purpose, ready for use in the future. The main function of this model is to receive the intent description message body generated in the previous step and consumption style preference features determined in advance based on historical behavioral data generated from user access to product data as input. Based on this input information, the query generation model, according to its learned reasoning ability, can generate an expanded query statement expressed in the target language. This expanded query statement not only reflects the complete intent expressed in the intent description message body but also incorporates the user's consumption style preferences, thereby enabling more accurate matching of the user's personalized needs.

[0025] The query generation model is trained using a large number of training samples. These samples include the intent description message body and the corresponding extended query statement, as well as the user's consumption style preference features. For example, for the user input "massage pillow", the intent description message body follows the previous example, including the original query statement "massage pillow", the target language "Chinese", the scene tags "bedding, office supplies, cervical spine treatment", and the attribute tags "massage, cervical spine, office, care, middle-aged and elderly". The entire intent description message body is combined with the user's consumption style preference features, and then an extended query statement is set for it to play a supervisory role, such as "cervical massage pillow suitable for middle-aged and elderly people to use in the office", which constitutes a training sample. The query generation model is trained using a large number of such training samples until it converges, and then it can be used online. For example, when the intent description message body of the previous example is input, the query generation model can output an extended query statement like "cervical massage pillow suitable for middle-aged and elderly people to use in the office", thereby realizing the semantic expansion of the original query statement submitted by the user.

[0026] The query generation model can be implemented in various specific ways. One implementation uses natural language processing techniques from deep learning, such as models based on the Transformer architecture. These models can be pre-trained models, such as LLAMA or GPT, so that fine-tuning with a small number of training samples can achieve the purpose of this application, namely, understanding the semantic information in the intent description message body and generating related extended query statements. For example, the model can be trained to generate extended query statements containing more details and user preference features after receiving the intent description message body and the user's consumption style preference features. The user's consumption style preference will inevitably affect the user intent expressed in the extended query statement. For example, taking the "massage pillow" example, if the user's historical behavior data shows that they prefer high-end, high-tech products, the model may generate an extended query statement such as "high-end high-tech neck massage pillow". Therefore, introducing consumption style preference features to constrain the extended query statements generated by the query generation model can control the model's generated results to be more likely to match the user's intent.

[0027] It's easy to understand that expanded query statements, compared to the original query statements submitted by users, not only deepen the understanding of user intent but also add richer semantic information. This semantic information expands from various aspects such as the scenarios in which the products are used, their attributes, and the user consumption habits they are suited for. Although it includes components predicted by the query generation model based on its reasoning ability, since the query generation model itself is a product of pre-training to convergence, the expanded query statements it predicts are reliable, and the semantics provided by the expanded query statements are more likely to match the product data in the current independent website's own product library.

[0028] Step S3400: Construct a placeholder interactive information block according to the extended query statement, and push the placeholder interactive information block to the user's search results page for display. The placeholder interactive information block is suitable for the user to access product data retrieved from the proprietary product library by referring to the semantics of the extended query statement.

[0029] A placeholder interactive information block is constructed based on the extended query statement and pushed to the user's search results page. This aims to provide users with an interactive search results page, enabling them to retrieve relevant product data from the independent website's own product library based on the semantics of the extended query statement. The design of the placeholder interactive information block not only improves the user experience but also increases user interactivity on the search results page, thereby increasing user dwell time and product conversion rates on the independent website.

[0030] Specifically, the placeholder interactive information block can include various implementation methods. In one embodiment, the placeholder interactive information block directly contains product information of valid product data retrieved from the current independent website's own product database by a secondary call to the product search engine based on the extended query statement. For example, assuming the user's original query statement is "massage pillow," after processing by the intent extraction model and query generation model, the generated extended query statement might be "cervical massage pillow suitable for middle-aged and elderly people to use in the office." At this time, the placeholder interactive information block can directly display product information of valid product data matching the extended query statement, such as product images, prices, and detailed descriptions, which are retrieved from the current independent website's own product database. This approach allows users to immediately see products related to the extended query statement, thereby quickly satisfying the user's search needs.

[0031] In another embodiment, the placeholder interactive information block may only provide one interactive entry point. When the user touches this entry point, a secondary search request is triggered. The server responds to this request by re-invoking the product search engine, retrieving valid product data based on the semantics of the extended query, and displaying the corresponding product information on the terminal device. For example, the placeholder interactive information block can be implemented as a touchable image displayed on the page. After the user clicks on it, relevant product data is retrieved from the proprietary product library based on the extended query "cervical massage pillow suitable for middle-aged and elderly office use," and displayed to the user on the current page or after a redirect. This approach provides users with a dynamic search experience, allowing them to obtain more product information when needed, increasing the flexibility and interactivity of search results.

[0032] Placeholder interactive information blocks can also enhance the user experience by combining placeholder images generated by the text-based image model. For example, a placeholder image generated based on an extended query can be a product image or scene image related to the extended query. These placeholder images can be displayed along with product information or as part of an interactive entry point, guiding users to further explore related products. For example, for the extended query "cervical massage pillow suitable for middle-aged and elderly people to use in the office," the generated placeholder image could be an office scene image of middle-aged and elderly people using a massage pillow. After the user clicks on the image, relevant product data will be retrieved and displayed. Using self-generated images for display, compared to displaying product images from valid product data, can comprehensively represent multiple product data and utilize the richer semantics contained in the extended query to provide users with a more intuitive product scene display effect, which helps to attract users to continue to access valid product data.

[0033] Thus, there is a corresponding remedy for the unexpected event that the search results page fails to obtain valid product data from the current independent website's own product library using conventional product search engines. Through the technical solution of this application, it is possible to ensure that even under such unexpected events, the semantics of the user's original query can still be deeply mined, and valid product data that is as close as possible to the user's intent can be matched from the limited number of proprietary products on the current independent website with a broader semantic range. This not only improves user retention rate and product conversion rate, but also enhances user experience. Furthermore, since the entire process only needs to be implemented within the independent website's internal data system, the entire technical solution can be deployed and implemented in a lightweight manner within the independent website, making it more efficient.

[0034] As can be seen from the above embodiments, this application proposes an innovative solution to the problem of zero-result pages encountered by users when searching for products on independent websites. This solution effectively overcomes the limitations of existing technologies, significantly improves user experience and the relevance of search results, and its technical advantages include, but are not limited to: First, by refining the intent of the user's original query, an intent description message body containing target language, scene tags, and attribute tags is obtained. Then, the query generation model expands the intent based on this intent message body and the user's consumption style preference characteristics to obtain an expanded query. This query fully explores the deep semantic relationship between the original query and the product data in the current independent website's product library. Based on this, a placeholder interactive information block is constructed, allowing users to continue searching for product data on the current independent website based on the expanded query after the intent semantics have been expanded. This can prevent users from leaving due to a lack of effective product data after encountering a page with zero search results, thus helping to improve the current independent website's retention rate.

[0035] Secondly, because an intent extraction model is introduced into the intent recognition process, the generated intent description message body takes into account factors such as language, scenario, and attributes. The intent extraction model can handle multilingual queries, spelling errors, and scenario-based requirements, ensuring that even when user queries have these problems, the user's true intent can be accurately extracted. Based on this, the semantics of intent can be expanded, and when recalling products on the current independent website, more indirectly related product data can be recalled from the product library of the current independent website for users to access. Moreover, since this product data itself matches the user's consumption style and enriches the user's personalized characteristics, it is more easily accepted by users and can improve the product conversion rate.

[0036] Furthermore, the entire business logic of this application is implemented within the independent website. The products it utilizes focus on the internal product data of the current independent website, while its extended semantics are sufficiently detailed. It always confines user needs within the current independent website and maximizes the matching of user needs with the limited resources of the current independent website. It fully takes into account the characteristics of the current independent website being independent and having a limited number of products, ensuring that the user experience is maximized under limited resource conditions.

[0037] Based on any embodiment of the method in this application, before invoking a preset intent extraction model, the method includes: Step S2100: Determine whether the character length of the original query statement is lower than a preset value. If it is lower than the preset value, perform word segmentation on the original query statement and determine the target language based on the word segmentation. To accurately identify the target language of user query text and perform subsequent processing, the target language can be determined before invoking the intent extraction model. First, it's determined whether the character length of the original query submitted by the user is below a preset value. This preset value is a threshold, such as 5 to 8 characters, used to distinguish between short and long texts so that different processing strategies can be adopted. For example, short text, such as "running shoes," may have a shorter character length, while long text may contain more complex queries, such as "beach wedding guest dress." By determining the character length, different types of query text can be processed more effectively.

[0038] When the character length of the original query is less than a preset value, the query is segmented into words, typically using the NGRAM algorithm. This segmentation divides the continuous text string into meaningful units, such as words or phrases. For example, for the Chinese query "running shoes," segmentation and matching with a preset dictionary might yield two word units: "running" and "shoes." Based on the segmentation results, the target language of the original query can be further determined. This is because different languages ​​have different lexical structures and grammatical features; analyzing the segmentation results allows for a more accurate identification of the language of the short text. For instance, if all the word units obtained after segmentation match Chinese words in a preset Chinese dictionary, the target language can be determined to be Chinese.

[0039] Step S2200: When the value is not lower than a preset value, determine whether there is a historical record corresponding to the original query statement in the cache. If there is, determine the target language based on the historical record. When the character length of the original query submitted by the user is not less than the preset value, the system first checks whether there is a historical record corresponding to the original query in the cache, so as to quickly determine the target language of the original query and avoid repeatedly processing known query text, thereby improving the efficiency and response speed of the system.

[0040] Specifically, the cache stores previously processed query text and its corresponding target language information. When a new query text arrives, it is first checked whether this query text is already recorded in the cache. If it exists, the target language information is retrieved directly from the cache without performing a complex language identification process again. For example, if a user previously searched for "beach wedding guest dress" and the target language has already been determined to be English, then when the user submits the same query text again, this information can be retrieved directly from the cache, quickly responding to the user's request.

[0041] Step S2300: When there is no corresponding historical record in the cache, the original query statement is input into a preset language recognition model to determine the target language. The language recognition model is a multilingual recognition model. When no historical record corresponding to the user's original query statement is found in the cache, the original query statement is input into a preset language recognition model. This model then uses its learned reasoning ability to determine the target language of the query text. In this embodiment, the language recognition model is a multilingual model, capable of handling text input in multiple languages ​​and accurately determining the language of the text. The language recognition model is implemented using deep learning techniques, such as the Transformer architecture, and is trained on a large amount of multilingual text data, thus possessing powerful language recognition capabilities.

[0042] For example, suppose a user submits a query text of "beach wedding guest dress". Because its character length exceeds a preset value and there are no historical records in the cache, the original query is input into the language identification model. The language identification model accurately identifies that the query text belongs to English by analyzing the lexical structure, grammatical features, and contextual information of the text.

[0043] There are various implementation methods for implementing language identification models. A common approach is to use pre-trained multilingual models, such as mBERT or XLM-R. These models have been pre-trained on texts in multiple languages ​​and can effectively identify the language of the text. For example, the mBERT model, through its multilingual pre-training, can encode the input text and use a classifier to predict the language of the text. Another approach is to use specialized language identification algorithms, which may determine the language of the text based on features such as character distribution and word frequency. For example, by analyzing the frequency of specific characters in the text, it is possible to distinguish between Latin-based and non-Latin-based languages.

[0044] Step S2400: Call the language correction model corresponding to the target language to perform error correction processing on the original query statement, and construct the corrected original query statement and the target language as a history record for caching.

[0045] Once the target language of the original query is determined, the corresponding language correction model is invoked to correct it. This ensures that the corrected query provides a more accurate linguistic representation when input into the subsequent intent extraction model. The main function of the language correction model is to identify and correct spelling and grammatical errors in the query text, thereby improving its quality and comprehensibility. For example, if the user inputs "runing shoes," the language correction model can correct it to "running shoes." The corrected text not only conforms better to linguistic norms but also improves the accuracy and efficiency of subsequent processing steps.

[0046] In one embodiment, the language error correction model can employ a deep learning-based error correction model, such as BERT-Corrector or GPT-3. These models are pre-trained on large amounts of text data and can effectively identify and correct errors in text. For example, BERT-Corrector can use its contextual understanding capabilities to identify misspelled words and provide correct replacement suggestions. Another embodiment can use traditional spell-checking algorithms, such as edit distance-based algorithms (Levenshtein Distance), which identify and correct spelling errors by calculating the similarity between words.

[0047] After the error correction process is completed, the corrected original query statement and the target language can be constructed into a history record and stored in the cache. This allows users to directly call the corresponding history record in the cache when they submit a new original query statement and execute step S2300, without having to call the language error correction model.

[0048] Since the language correction model is invoked after the target language is identified, and each language correction model has its own specialization for a specific language, it can more effectively and accurately correct the original query statement, ensuring that the intent extraction model can reason based on the accurate expression of user needs.

[0049] By implementing the above embodiments, this application has achieved significant technical advantages in many aspects, including but not limited to: First, by determining the character length of the original query statement and adopting different processing strategies, the use of word segmentation, caching, and language identification models is gradually advanced. This enables efficient processing of different types of query text. Whether it is short or long text, the target language can be quickly and accurately determined, which not only improves the system's response speed but also ensures the accuracy of language identification.

[0050] Secondly, accurate identification of the target language serves as the foundation for subsequent error correction, intent recognition, and semantic expansion. Throughout the entire technical solution process, the corresponding business logic is executed in relation to the language. Compared to traditional product recommendation algorithms that typically do not consider language conditions, this application can ensure the accuracy of intent recognition and semantic expansion. As a disaster recovery measure, its positive significance is more pronounced.

[0051] Furthermore, by calling the language correction model corresponding to the target language, since each language correction model can professionally and efficiently correct errors in the text of its corresponding language, the quality and comprehensibility of the text can be improved, providing more accurate input for subsequent intent extraction and query generation.

[0052] Based on any embodiment of the method in this application, before responding to a user-submitted product query request, the method includes: Step S1100: Obtain the training dataset, which includes intent description message body and consumption style preference features extracted from multiple product data for the corresponding model input, and product description for the corresponding model output. To train the query generation model, a training dataset is prepared. This dataset includes intent description messages and consumer style preference features extracted from multiple product data for the corresponding model input, as well as product descriptions for the corresponding model output. The query generation model is trained using this training dataset to ensure that the model can learn the mapping relationship between user query intent and product data, thereby generating more accurate expanded query statements.

[0053] First, intent description message bodies can be extracted from multiple product data sets. These message bodies contain semantic information about the assumed user query, including but not limited to the target language, original query statement, scene tags, and attribute tags. They can be extracted from product data and represented as structured data.

[0054] Then, the training dataset, associated with each intent description message body, also includes consumption style preference features. These features characterize the user's consumption style towards the target product they wish to purchase. Consumption style preference features can be statistically determined from corresponding historical behavioral data such as the user's purchase history, browsing behavior, and favorites preferences. For example, if a user frequently purchases high-end, technologically advanced products, their consumption style preference feature could be labeled as "high-end technology." These features help the model generate expanded query statements that better match the user's personalized needs.

[0055] Finally, the training dataset also includes product descriptions for each intent description message body, used for the corresponding model output. These product descriptions are extracted from an in-house product library and are highly relevant to the user's query intent. For example, for the user's original query "running shoes," the product description might include phrases like "lightweight and breathable running shoes, suitable for daily exercise." These product descriptions serve as the model's output targets, helping the model learn how to generate product descriptions that match the user's intent.

[0056] There are several specific implementation methods for constructing the training dataset. One approach is through data annotation, where human annotators label user query text to generate intent description messages, consumption style preference features, and corresponding product descriptions. Another approach is to utilize existing user behavior and product data, automatically extracting this information through data mining algorithms. For example, by analyzing historical behavioral data such as user purchase history and browsing behavior, information such as user intent description messages, consumption style preference features, and product descriptions can be automatically extracted.

[0057] The training dataset above not only contains semantic information of user queries, but also user personalized preferences and product descriptions. After training the query generation model with it, the model can learn the complex mapping relationship between user intent and product data, thereby generating more accurate and personalized extended query statements.

[0058] Step S1200: Based on the training dataset, construct positive samples to fine-tune the pre-trained large language model until the model converges; In this step, positive samples can be constructed based on the acquired training dataset. That is, the intention description message body and consumption style preference features with semantic correspondence in the training dataset are used as the model input required for model inference. The corresponding product descriptions are used to supervise the model output obtained after inference by the module. In this way, the pre-trained large language model is fine-tuned until the model converges, so that the model can better adapt to the current independent website's product data and user query intent, thereby generating more accurate extended query statements.

[0059] The positive samples in the training dataset contain product descriptions highly relevant to user query intent, which can be used to guide the model in learning the mapping relationship between user intent and product data. Through fine-tuning, the model can adjust its parameters to better understand and generate product descriptions that match user intent. Pre-trained large language models, such as those based on the Transformer architecture, already possess powerful language understanding and generation capabilities. By fine-tuning on the specific training dataset of this application, the model can be further adapted to the business needs of the current independent website.

[0060] Fine-tuning typically involves freezing some of the model's parameters, adjusting only the last few layers. This allows for improving the model's performance on specific tasks while maintaining its general capabilities. The fine-tuning process ends when the model's performance on the training dataset no longer improves, indicating convergence. At this point, the model has the ability to generate high-quality expanded query statements.

[0061] Step S1300: Based on the training dataset, construct positive samples and noise samples according to the noise ratio corresponding to different stages, and perform multi-stage training on the fine-tuned model until the model converges. After fine-tuning the query generation model, its ability to generate higher-quality expanded query statements can be further improved. To this end, based on the training dataset, positive and noisy samples are constructed according to the noise ratios corresponding to different stages. Noisy samples refer to data pairs obtained by semantically reducing or removing the association between the model input and output parts compared to positive samples; that is, the semantic relationship between the intent description message body, consumer style preference features, and their corresponding product descriptions is weak or absent in general understanding. The fine-tuned model is then trained in multiple stages until convergence. By gradually increasing the noise ratio during training, the model's robustness and generalization ability can be significantly improved, enabling it to better handle various complex situations and generate more accurate expanded query statements.

[0062] In a more specific embodiment, the multi-stage training process includes any number of the following three stages, with the noise ratio gradually increasing in each stage, so that the model can be trained on samples of varying difficulty to gradually improve its performance.

[0063] In the first stage, positive and noisy samples with a noise ratio within a preset minimum range (e.g., 20% to 40%) are constructed. The model is trained to generate product descriptions with a relevance higher than a preset maximum threshold (e.g., 90%) to the user intent represented by the intent description message. The goal of this stage is to allow the model to learn on relatively simple samples and establish basic intent matching capabilities. For example, for the user input "running shoes," the model can generate product descriptions highly relevant to "running shoes," such as "lightweight and breathable running shoes, suitable for daily exercise."

[0064] In the second stage, positive and noisy samples with a noise ratio falling within a preset intermediate range (e.g., 40% to 60%) are constructed. The model is trained to generate product descriptions whose relevance to the user intent represented by the intent description message body is higher than a preset second-highest threshold (e.g., 80%). In this stage, the noisy samples are constrained to interfere with keywords using synonym substitution. For example, replacing "running shoes" with "sports shoes," the model needs to still be able to generate product descriptions relevant to the user intent under this interference, such as "lightweight shoes suitable for sports." The goal of this stage is to expand the model's generative capabilities, enabling it to handle a wider range of query intents.

[0065] In the third stage, positive and noisy samples with a noise ratio falling within a preset maximum range (e.g., 60% to 80%) are constructed. The model is then trained to generate product descriptions whose relevance to the user intent represented by the intent description message body is higher than a preset minimum threshold. In this stage, the noisy samples are constrained to use semantic association to perturb the scene information within the noisy samples. For example, replacing "running shoes" with "shoes suitable for outdoor activities" requires the model to generate product descriptions relevant to the user intent under this more complex perturbation, such as "high-performance shoes suitable for outdoor running." The goal of this stage is to improve the model's adaptability and robustness, enabling it to provide reasonable alternatives when faced with complex queries.

[0066] The above multi-stage training method can control the model to learn step by step on samples of different difficulties, thereby better handling various complex situations in practical applications. It can not only improve the robustness and generalization ability of the model, but also ensure the quality and relevance of the generated extended query statements.

[0067] Step S1400: For minor languages ​​with a total sample size below a preset threshold, fine-tune the model using the corresponding preset parallel corpus of minor languages ​​after training with samples from other languages ​​until the model converges. During the training process described above, the query generation model is trained on training samples from one or more typical languages ​​(e.g., Chinese and English) by default. These typical languages ​​typically have a sufficient number of samples, providing rich training data for the model. However, for less commonly taught languages ​​with a total sample size below a preset threshold, directly using samples from these typical languages ​​for training may not adequately cover the semantic features and linguistic habits of the less commonly taught languages, thus affecting the model's performance on those languages.

[0068] To overcome this problem, this application combines a two-step transfer learning method to improve the query generation model's ability to adapt to the needs of less commonly spoken languages. Building upon the model training using training samples from other languages ​​(such as Chinese and English), transfer learning is then applied to fine-tune the model using pre-set parallel corpora for less commonly spoken languages ​​with limited sample sizes. These parallel corpora are carefully selected and labeled beforehand to reflect the semantic features and linguistic habits of the less commonly spoken languages. In this way, the model can be specifically optimized for less commonly spoken languages, thereby improving its adaptability and generation quality in these languages.

[0069] Step S1500: Use the converged model as the query generation model, and use it to generate the corresponding language model output as the extended query statement in response to the model input.

[0070] After the query generation model has undergone systematic training through the above process and reached a convergence state, it is actually mature and can automatically generate extended query statements in the corresponding language based on the given model input. This enables semantic expansion of the user's original query statement, allowing it to be deployed online. Typically, it can be embedded as an internal service of an independent website for the website to call at any time.

[0071] By implementing the above embodiments, this application provides a query generation model with stronger reasoning capabilities, capable of generating extended query statements in a one-stop manner, demonstrating significant technical advantages. Specifically, by constructing a training dataset containing intent description messages, consumer style preference features, and product descriptions, the model can learn the complex mapping relationship between user query intent and product data, thereby generating more accurate and personalized extended query statements. This dataset construction method not only covers the semantic information of user queries but also incorporates users' personalized preferences, enabling the model to better understand and meet users' personalized needs. By fine-tuning the pre-trained large language model based on the training dataset, the model can further adapt to the current independent website's product data and user query intent, thereby improving its performance on specific tasks. On this basis, through multi-stage training, the model gradually learns on samples of different difficulty levels, progressively improving its robustness and generalization ability, enabling the model to generate more accurate extended query statements when facing various complex situations. Finally, through a two-step transfer learning method for minority languages, the model can effectively cover the semantic features and language habits of minority languages, improving its adaptability and generation quality in minority languages. These technological effects work together to significantly improve the reasoning capabilities of the query generation model. It can more accurately understand and generate product descriptions that match the user's intent as extended query statements, effectively achieving a deep understanding of the user's original query intent and precise semantic expansion. This ensures effective disaster recovery measures when a page of zero search results appears, and further matches users with product data from independent websites, thereby providing users with a better and more personalized search experience.

[0072] Based on any embodiment of the method in this application, a placeholder interactive information block is constructed according to the extended query statement, including: Step S4100: Call the preset text graph model and generate the first placeholder graph according to the extended query statement; A text-based graph model is an artificial intelligence model capable of generating corresponding images based on text descriptions. In this application, the expanded query statement contains rich semantic information, such as the type of product, usage scenario, and attributes. The text-based graph model utilizes this information to generate an image that matches the expanded query statement. For example, if the expanded query statement is "cervical massage pillow suitable for middle-aged and elderly people to use in the office," the text-based graph model can generate an image of a middle-aged or elderly person using a massage pillow in an office setting. This image not only visually demonstrates the usage scenario of the product but also attracts the user's attention and increases their interest.

[0073] When implementing text-to-image models, deep learning-based generative models such as DALL·E or StableDiffusion can be used. These models are trained on large amounts of text and image data and are able to generate high-quality images based on the input text description. For example, DALL·E can generate images of various scenes and objects based on the input text description through its powerful generative capabilities.

[0074] Step S4200: Based on the extended query statement, determine multiple target products from the proprietary product library that match the complete semantics of the extended query statement and construct a product display list; Based on the extended query statement, multiple target products that match the complete semantics of the query can be identified from the website's own product database. This can be achieved by calling the product search engine's preset interface again. After obtaining multiple target products, they are constructed into a product display list. Therefore, by leveraging the rich semantic information of the extended query statement, product data that better meets user needs can be retrieved from the website's own product database, thereby improving the relevance of search results and user experience.

[0075] Specifically, expanded query statements not only include the semantics of the user's original query, but also incorporate the user's consumption style preferences and detailed intent descriptions generated by the intent extraction model. For example, if the user's original query is "running shoes," after processing by the intent extraction and query generation model, the expanded query statement might be "lightweight and breathable running shoes, suitable for daily exercise." This expanded query statement includes more specific product attributes and usage scenarios, enabling a more accurate match to product data in the product database.

[0076] The default interface of a product search engine receives extended query statements as input and retrieves product data matching the extended query statements from its own product database according to its business logic. As revealed earlier, product search engines can employ various algorithms and technologies to achieve this function, such as keyword matching, semantic analysis, and collaborative filtering. These algorithms and technologies can be used individually or in combination to improve the accuracy and recall rate of search results.

[0077] In practical applications, product search engines can retrieve multiple target products from a product database based on keywords and semantic information in extended queries. For example, for the extended query "lightweight and breathable running shoes, suitable for daily exercise," the product search engine can retrieve products with attributes such as "lightweight," "breathable," and "running shoes," and the usage scenario for these products may be labeled as "daily exercise." These target products will be constructed into a product display list for presentation to the user in subsequent steps.

[0078] This demonstrates that not only can the rich semantic information of expanded query statements be utilized to more accurately retrieve product data that meets user needs, but the relevance of search results and user experience can also be improved. This product retrieval method based on expanded query statements can effectively solve the problem of irrelevant search results caused by overly short or vague user query text in traditional search methods, thereby improving the search performance and user satisfaction of independent websites.

[0079] Step S4300: Construct the link between the first placeholder image and the product display list as a first placeholder interactive information block.

[0080] Constructing the first placeholder image and its link to the product display list as the first interactive information block means combining the generated image with the link to the product display list to form an interactive interface element. For example, a user can click on the first placeholder image to directly access the product display list related to that image, thereby quickly obtaining more product information. This interaction method not only increases user engagement but also allows users to more intuitively understand the usage scenarios and features of the products.

[0081] In one embodiment, the first placeholder image can be embedded in the search results page, and a hyperlink can be added to it, pointing to the product display list page. In another embodiment, JavaScript or other front-end technologies can be used to dynamically load the product display list upon clicking the first placeholder image. For example, when a user clicks the first placeholder image, the webpage can retrieve the product display list data from the server via an AJAX request and dynamically display this data on the page without reloading the entire page.

[0082] Accordingly, the above embodiments, through the first placeholder interactive information block, not only provide an intuitive visual display but also a convenient interactive entry point, enabling users to quickly access and expand product data related to their query statements. This design, combining visual display and interactive functionality, significantly enhances the user experience, preventing users from encountering zero-result pages when performing product searches and allowing them to continue browsing the current independent website's own product data based on their interaction with the first placeholder image, thereby increasing user dwell time and product conversion rates on the independent website.

[0083] Based on any embodiment of the method in this application, a placeholder interactive information block is constructed according to the extended query statement, including: Step S5100: Call the preset text graph model and generate a second placeholder graph according to the extended query statement; Similar to the previous embodiment, in this embodiment, a second placeholder image can be generated by calling the Wensheng image model; details will not be elaborated here.

[0084] Step S5200: Based on the extended query statement, after replacing the scene information with similar scene information, determine multiple target products from the proprietary product library that match the complete semantics of the replaced extended query statement to construct a product display list; Based on the extended query statement, the scenario information is replaced with similar scenario information. Then, multiple target products that match the complete semantics of the replaced extended query statement are determined from the proprietary product library and constructed into a product display list. The aim is to further enrich the search results and improve the relevance and diversity of the search results through semantic expansion and scenario replacement, thereby better meeting the user's search needs.

[0085] Specifically, expanded query statements contain rich semantic information, such as product type, usage scenario, and attributes. In this step, the scenario information in the expanded query statement can be replaced with similar scenario information. For example, if the expanded query statement is "cervical massage pillow suitable for middle-aged and elderly people to use in the office," "office use" can be replaced with similar scenario information such as "home use" or "travel use," generating new expanded query statements, such as "cervical massage pillow suitable for middle-aged and elderly people to use at home" or "cervical massage pillow suitable for middle-aged and elderly people to use while traveling." This scenario replacement can cover a wider range of user needs and provide more product choices.

[0086] Then, the product search engine's default interface is invoked again, using the replaced extended query as input. This retrieves multiple target products from its own product database that match the complete semantics of the extended query. These target products are not only related to the original query but also take into account the replaced scenario information, thus providing a more comprehensive selection of products. For example, for the replaced extended query "cervical massage pillow suitable for middle-aged and elderly families," the product search engine can retrieve products labeled with the "family use" scenario, which may be more suitable for the needs of use in a home environment.

[0087] Several specific implementation methods can be used to achieve this step. One embodiment is to use semantic analysis technology to automatically identify and replace similar scene information by analyzing keywords and semantic information in the extended query. For example, a pre-trained semantic model (such as BERT or Word2Vec) can be used to calculate the similarity between scene information and select the most similar scene for replacement. Another embodiment is to predefine a set of mapping relationships between similar scenes through manual annotation, and then automatically apply these mapping relationships to replace scenes. For example, similarity relationships between scenes such as "office use" and "home use" and "travel use" can be defined, and automatic replacement can be performed when processing extended queries.

[0088] Therefore, this approach not only provides products related to the original query, but also offers a wider range of product choices through scenario replacement, thereby improving the diversity and relevance of search results. This product retrieval method based on semantic expansion and scenario replacement effectively solves the problem of irrelevant search results caused by overly short or vague user query text in traditional search methods, thus improving the search performance and user satisfaction of independent websites.

[0089] Step S5300: Construct a second placeholder interactive information block by linking the second placeholder image with the product display list.

[0090] Similarly, constructing a second placeholder image and its link to the product display list as a second interactive information block means combining the generated image with the link to the product display list to form an interactive interface element. For example, a user can click on the second placeholder image to directly access the product display list related to that image, thereby quickly obtaining more product information. This interaction method not only increases user engagement but also allows users to more intuitively understand the usage scenarios and features of the products.

[0091] In one embodiment, the second placeholder image can be embedded in the search results page and a hyperlink can be added to it, pointing to the product display list page. In another embodiment, JavaScript or other front-end technologies can be used to dynamically load the product display list upon clicking the second placeholder image. For example, when a user clicks the second placeholder image, the webpage can retrieve the product display list data from the server via an AJAX request and dynamically display this data on the page without reloading the entire page.

[0092] The above embodiments further enrich product search capabilities by replacing the extended query statement with contextual information. The second placeholder interactive information block not only provides an intuitive visual display but also a convenient interactive entry point, enabling users to quickly access product data related to the extended query statement. This design, combining visual display and interactive functionality, significantly improves the user experience, preventing users from encountering zero-result pages when performing product searches. Furthermore, users can continue browsing the independent website's own product data based on their interaction with the second placeholder image, thereby increasing user dwell time and product conversion rates on the independent website.

[0093] Based on any embodiment of the method in this application, the first placeholder image and the second placeholder image can be inserted into different positions on the product search results page at the same time, enriching the types of user interaction entry points on the search results page, providing users with more interactive options, guiding users to place orders on the current independent site, thereby further improving the conversion rate of the current independent site's own products.

[0094] Based on any embodiment of the method in this application, before invoking a preset query generation model, the method includes: Step S6100: Obtain historical behavior data formed by the user's historical access to products, wherein the historical behavior data includes the price, category, feature description and selling points of multiple accessed products; Historical behavioral data, formed from users' past product visits, consists of records of a user's browsing and interactions on one or more independent websites. Specifically, it can include information such as the price, category, features, and selling points of multiple visited products. This data reflects users' shopping preferences and behavioral patterns, effectively constituting a description of their consumption style or tone, and providing a foundation for subsequently determining their consumption style and preference characteristics.

[0095] Specifically, historical behavioral data can be acquired in several ways. One approach is to utilize a user behavior log system on an independent website to record user actions each time they visit a product page, including clicks, browsing, and purchases. This log data can be stored in a server database for later analysis. For example, when a user browses a product page for a high-tech neck massage pillow, information such as the product's price, category (e.g., massage equipment), features (e.g., technological features), and selling points (e.g., intelligent massage function) will be recorded.

[0096] Another approach involves collecting all historical behavioral data of users on independent websites through a user account system. For example, if a user purchases a high-end, technologically advanced neck massage pillow on an independent website, relevant information about the product (such as price, category, features, and selling points) will be recorded in the user's purchase history. This data includes not only products actually purchased by the user but also products viewed but not purchased, thus providing a more comprehensive reflection of the user's interests and preferences.

[0097] Step S6200: Call the preset style classification model and determine the consumer style preference features mapped to each visited product based on the price, category, feature description and selling point features of each visited product. The system invokes a pre-defined style classification model to determine the corresponding consumer style preference features for each accessed product based on its price, category, features, and selling points. This aims to transform the specific product's feature information into a type label for the user's consumption style, thereby providing a basis for subsequent personalized recommendations.

[0098] Specifically, the main function of a style classification model is to transform product characteristics (such as price, category, feature description, and selling points) into consumer style preference characteristics. This characteristic information can be obtained from users' historical behavioral data, such as records of product pages viewed or purchased by users. By analyzing this characteristic information, the style classification model identifies the corresponding consumer style type. For example, if a product has a high price, is categorized as "massage equipment," has a "technological" feature description, and its selling point is "intelligent massage function," then this product might be mapped to a "high-end technology" consumer style preference characteristic.

[0099] There are several specific implementation methods for implementing style classification models. One embodiment is to use a rule-based classification method, pre-defining a set of rules to map product features to consumer style preference features. For example, a rule could be defined: if a product's price exceeds a certain threshold and it has a "technological" description, it is mapped to a "high-end technology" consumer style preference feature. Another embodiment is to use machine learning models, such as decision trees, random forests, or neural networks, to learn the relationship between product features and consumer style preference features through training data. For example, historical user behavior data can be used as a training set to train a classification model that can automatically identify product consumer style preference features.

[0100] In this way, the style classification model can transform the specific features of each accessed product into consumption style preference features, specifically manifested as a consumption style type, providing a foundation for subsequent statistical analysis of users' consumption style preference features. This method of determining consumption style preference features based on product characteristics can more accurately reflect users' personalized needs, thereby improving the accuracy and relevance of personalized recommendations.

[0101] Step S6300: Calculate the consumption style preference feature with the largest number of visited products, and use it as the consumption style preference feature representing the user.

[0102] By statistically analyzing the most frequently accessed product consumption style preferences of a user, we can identify the characteristics that best represent that user's consumption style and extract the features that best represent that user's consumption style, thus providing an accurate basis for subsequent personalized recommendations.

[0103] Specifically, in step S6200, each visited product is mapped to a consumer style preference feature. These features reflect the type of preferences a user has when browsing and purchasing products. For example, if a user repeatedly visits and purchases products with the "high-end technology" feature, then this feature will appear multiple times in the user's historical behavior data.

[0104] In this step, the number of each consumption style preference feature among the products visited by the user is counted. For example, if a user visits 10 products, 5 of which are mapped to "high-end technology," 3 to "fashionable and casual," and 2 to "affordable," then the feature "high-end technology" has the highest number of visits. Therefore, "high-end technology" is identified as the consumption style preference feature representing this user.

[0105] This statistical method can effectively extract the most representative features of a user's consumption style from their historical behavioral data. In this way, we can more accurately understand users' personalized needs, thereby providing content that better matches their preferences in subsequent query generation and product recommendations.

[0106] Through the above embodiments, users' consumption style preferences can be accurately determined, thereby providing users with more personalized and accurate product recommendations and services. This method of determining consumption style preferences based on users' historical behavior data not only improves the accuracy and relevance of personalized recommendations but also enhances users' shopping experience and satisfaction.

[0107] Please see Figure 2 According to one aspect of this application, a product search result display device includes a regular query module 3100, an intent-based recommendation module 3200, a query expansion module 3300, and a result display module 3400. The regular query module 3100 is configured to respond to a user-submitted product query request and, based on the original query statement carried in the request, query whether matching valid product data exists in the current independent website's proprietary product database. The intent-based recommendation module 3200, when no valid product data exists, invokes a preset intent extraction model to identify the target intent contained in the original query statement according to the target language of the original query statement, and outputs a corresponding intent description message body, wherein the intent description message body contains... The target language, the original query statement, and the user's desired product's scenario tag and attribute tag; the query extension module 3300 is configured to call a preset query generation model to generate an extended query statement in the target language based on the intent description message body and the user's consumption style preference characteristics. The extended query statement reflects the complete intent expressed by the intent description message body and the user's consumption style; the result display module 3400 is configured to construct a placeholder interactive information block based on the extended query statement and push the placeholder interactive information block to the user's search results page for display. The placeholder interactive information block is suitable for the user to access and refer to the semantics of the extended query statement to retrieve product data from the proprietary product library.

[0108] Based on any embodiment of the device in this application, the device further includes: a word segmentation inference module, configured to determine whether the character length of the original query statement is lower than a preset value; if it is lower than the preset value, the original query statement is segmented into words, and the target language is determined based on the word segmentation; a cache reuse module, configured to determine whether there is a historical record corresponding to the original query statement in the cache if it is not lower than the preset value; if it exists, the target language is determined based on the historical record; a model inference module, configured to input the original query statement into a preset language recognition model to determine the target language if there is no corresponding historical record in the cache, wherein the language recognition model is a multilingual recognition model; and a text correction module, configured to call the language correction model corresponding to the target language to perform error correction processing on the original query statement, and construct a historical record by combining the corrected original query statement with the target language for caching.

[0109] Based on any embodiment of the device in this application, the device further includes: a data acquisition module, configured to acquire a training dataset, the training dataset including intent description message bodies and consumer style preference features extracted from multiple product data for corresponding model input, and product descriptions for corresponding model output; an initial training module, configured to construct positive samples based on the training dataset to fine-tune a pre-trained large language model until the model converges; a gradient interference module, configured to construct positive samples and noise samples according to the noise ratio corresponding to different stages based on the training dataset, and perform multi-stage training on the fine-tuned model until the model converges; a language transfer module, configured to, for minor languages ​​with a total sample volume below a preset threshold, fine-tune the model using a preset parallel corpus of minor languages ​​after training with samples from other languages ​​until the model converges; and a model application module, configured to use the converged model as a query generation model, and use it to generate model output in the corresponding language as an extended query statement in response to model input.

[0110] Based on any embodiment of the device in this application, the multi-stage training in the initial training module includes: a first stage of training, which constructs positive samples and noise samples with a noise ratio belonging to a preset minimum ratio range, and trains the model to generate product descriptions with a relevance to the user intent represented by the intent description message body that is higher than a preset maximum threshold; a second stage of training, which constructs positive samples and noise samples with a noise ratio belonging to a preset middle ratio range, and trains the model to generate product descriptions with a relevance to the user intent represented by the intent description message body that is higher than a preset second-highest threshold, wherein the noise samples are constrained to use synonym replacement to interfere with keywords in the noise samples; and a third stage of training, which constructs positive samples and noise samples with a noise ratio belonging to a preset maximum ratio range, and trains the model to generate product descriptions with a relevance to the user intent represented by the intent description message body that is higher than a preset minimum threshold, wherein the noise samples are constrained to use semantic association expressions to interfere with scene information in the noise samples.

[0111] Based on any embodiment of the device in this application, the result display module 3400 includes: a first image generation module, configured to call a preset text-to-image model to generate a first placeholder image according to the extended query statement; a complete recall module, configured to determine multiple target products from the proprietary product library that match the complete semantics of the extended query statement based on the extended query statement and construct a product display list; and a first construction module, configured to construct a first placeholder interactive information block by linking the first placeholder image with the product display list.

[0112] Based on any embodiment of the device in this application, the result display module 3400 includes: a second image generation module, configured to call a preset text-to-image model to generate a second placeholder image according to the extended query statement; a scene optimization module, configured to replace the scene information in the extended query statement with similar scene information, and then determine multiple target products from the proprietary product library that match the complete semantics of the replaced extended query statement to construct a product display list; and a second construction module, configured to construct a second placeholder interactive information block by linking the second placeholder image with the product display list.

[0113] Based on any embodiment of the device in this application, the device further includes: a data engineering module, configured to acquire historical behavior data formed by the user's historical access to product data, the historical behavior data including the price, category, feature description and selling point features of multiple accessed products; a style reasoning module, configured to call a preset style classification model to determine the consumer style preference feature mapped to each accessed product based on the price, category, feature description and selling point features of each accessed product; and a statistical calibration module, configured to statistically identify the consumer style preference feature with the largest number of accessed products covered, as the representative consumer style preference feature of the user.

[0114] Another embodiment of this application provides a product search result display device. For example... Figure 3 The diagram shows the internal structure of a product search result display device. This device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. The computer-readable, non-volatile storage medium stores an operating system, a database, and computer-readable instructions. The database stores information sequences, and when executed by the processor, these computer-readable instructions enable the processor to implement a product search result display method.

[0115] The processor of the product search results display device provides computing and control capabilities to support the operation of the entire device. The device's memory can store computer-readable instructions, which, when executed by the processor, cause the processor to perform the product search results display method of this application. The network interface of the device is used for communication with a terminal.

[0116] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the product search result display device to which the present application is applied. A specific product search result display device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0117] In this embodiment, the processor is used to execute... Figure 2 The specific functions of each module are described, and the memory stores the program code and various data required to execute the aforementioned modules or sub-modules. A network interface is used to enable data transmission between user terminals and the server. In this embodiment, the non-volatile readable storage medium stores the program code and data required to execute all modules in the product search result display device of this application. The server can call the server's program code and data to execute the functions of all modules.

[0118] This application also provides a non-volatile readable storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the product search result display method of any embodiment of this application.

[0119] This application also provides a computer program product, including a computer program / instructions that, when executed by one or more processors, implement the steps of the method described in any embodiment of this application.

[0120] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a computer-readable storage medium such as a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM).

Claims

1. A method for displaying product search results, characterized in that, include: In response to a user's product query request, based on the original query statement carried in the request, the system checks whether there is any matching valid product data in the current independent website's own product database. When no valid product data exists, a preset intent extraction model is invoked to identify the target intent contained in the original query statement in the target language corresponding to the original query statement, and output a corresponding intent description message body. The intent description message body includes the target language, the original query statement, and the user's expected product scene tag and attribute tag. The system invokes a preset query generation model to generate an extended query statement in the target language based on the intent description message body and the user's consumption style preference characteristics. The extended query statement reflects the complete intent expressed by the intent description message body and the user's consumption style. A placeholder interactive information block is constructed based on the extended query statement, and the placeholder interactive information block is pushed to the user's search results page for display. The placeholder interactive information block is suitable for the user to access product data retrieved from the proprietary product library by referring to the semantics of the extended query statement.

2. The method for displaying product search results according to claim 1, characterized in that, Before invoking the preset intent extraction model, the following steps are included: Determine whether the character length of the original query statement is lower than a preset value. If it is lower than the preset value, perform word segmentation on the original query statement and determine the target language based on the word segmentation. When the value is not lower than a preset value, it is determined whether there is a historical record corresponding to the original query statement in the cache. If there is, the target language is determined according to the historical record. When the corresponding historical record is not found in the cache, the original query statement is input into a preset language recognition model to determine the target language. The original query statement is corrected by calling the language correction model corresponding to the target language. The corrected original query statement and the target language are then used to construct a history record for caching.

3. The method for displaying product search results according to claim 1, characterized in that, Before responding to a user's product query request, the following steps are included: Obtain a training dataset, which includes intent description message bodies and consumption style preference features extracted from multiple product data for the corresponding model input, and product descriptions for the corresponding model output; Based on the training dataset, positive samples are constructed to fine-tune the pre-trained large language model until the model converges. Based on the training dataset, positive samples and noise samples are constructed according to the noise ratio corresponding to different stages, and multi-stage training is performed on the fine-tuned model until the model converges. For minority languages ​​with a total sample size below a preset threshold, after training using samples from other languages, fine-tuning is performed using the corresponding preset parallel corpus of minority languages ​​until the model converges. The converged model is used as the query generation model, and it is used to generate model outputs in the corresponding language in response to the model input as extended query statements.

4. The method for displaying product search results according to claim 3, characterized in that, Positive and noisy samples are constructed according to the noise ratio corresponding to different stages, and multi-stage training is performed on the fine-tuned model, including: In the first stage of training, positive samples and noise samples with a noise ratio within the preset minimum ratio range are constructed, and the product descriptions generated by the training model have a correlation with the user intent represented by the intent description message body that is higher than the preset maximum threshold. In the second stage of training, positive samples and noise samples with a noise ratio belonging to a preset middle ratio range are constructed. The training model generates product descriptions with a correlation higher than a preset second-highest threshold between the user intent represented by the intent description message body and the noise samples are constrained to use synonym replacement to interfere with the keywords in the noise samples. In the third stage of training, positive samples and noise samples with a noise ratio belonging to a preset maximum ratio range are constructed. The training model generates product descriptions with a correlation higher than a preset minimum threshold between the product description and the user intent represented by the intent description message body. The noise samples are constrained to use semantic association expressions to interfere with the scene information in the noise samples.

5. The method for displaying product search results according to any one of claims 1 to 4, characterized in that, Constructing a placeholder interactive information block based on the extended query statement includes: Invoke the preset text graph model and generate a first placeholder graph based on the extended query statement; Based on the extended query statement, multiple target products that match the complete semantics of the extended query statement are determined from the proprietary product library and constructed into a product display list; The link between the first placeholder image and the product display list is constructed as the first placeholder interactive information block.

6. The method for displaying product search results according to any one of claims 1 to 4, characterized in that, Constructing a placeholder interactive information block based on the extended query statement includes: The preset text graph model is invoked to generate a second placeholder graph based on the extended query statement; Based on the extended query statement, after replacing the scene information with similar scene information, multiple target products that match the complete semantics of the replaced extended query statement are determined from the proprietary product library and constructed into a product display list. The link between the second placeholder image and the product display list is constructed as a second placeholder interactive information block.

7. The method for displaying product search results according to any one of claims 1 to 4, characterized in that, Before invoking the preset query generation model, the following steps are included: Historical behavior data is obtained by acquiring the user's historical access to product data. The historical behavior data includes the price, category, feature description, and selling points of multiple accessed products. The system calls a preset style classification model to determine the consumer style preference characteristics mapped to each visited product based on its price, category, feature description, and selling points. The consumption style preference feature with the largest number of accessed products is identified and used as the representative consumption style preference feature of the user.

8. A product search result display device, characterized in that, include: The regular query module is set to respond to the product query request submitted by the user and, based on the original query statement carried in the request, query whether there is any matching valid product data in the current independent website's own product database; The intent-based recommendation module is configured to call a preset intent extraction model when no valid product data exists. This model identifies the target intent contained in the original query statement in the target language of the original query statement and outputs a corresponding intent description message body. The intent description message body includes the target language, the original query statement, and the user's desired product scenario tag and attribute tag. The query extension module is configured to call a preset query generation model to generate an extended query statement in the target language based on the intent description message body and the user's consumption style preference characteristics. The extended query statement reflects the complete intent expressed by the intent description message body and the user's consumption style. The results display module is configured to construct a placeholder interactive information block based on the extended query statement, and push the placeholder interactive information block to the user's search results page for display. The placeholder interactive information block is suitable for the user to access product data retrieved from the proprietary product library by referring to the semantics of the extended query statement.

9. A product search result display device, comprising a central processing unit and a memory, characterized in that, The central processing unit is used to invoke and run a computer program stored in the memory to perform the steps of the method as described in any one of claims 1 to 7.

10. A non-volatile readable storage medium, characterized in that, It stores, in the form of computer-readable instructions, a computer program implemented according to any one of claims 1 to 7, which, when invoked by a computer, executes the steps included in the corresponding method.

Images