Large language model-based item comparison method and apparatus, electronic device, and medium

Abstract

The present disclosure provides a large language model-based item comparison method and apparatus, an electronic device, and a medium, applied to the technical fields of artificial intelligence, intelligent customer service, intelligent robots, deep learning, and large language models. The large language model-based item comparison method comprises: in response to receiving item comparison information inputted by a user, determining a target category to which a first item and a second item belong in the item comparison information; on the basis of the target type, determining a plurality of attributes to be compared matching the item comparison information, and ordering information among said plurality of attributes, wherein said attributes are used for describing the first item and the second item; inputting said plurality of attributes, first item information of the first item, and second item information of the second item into an item comparison large language model to generate a first item comparison sub-result; and generating an item comparison result on the basis of the ordering information and the first item comparison sub-result.

Description

Item comparison methods, devices, electronic equipment, and media based on large models

[0001] Cross-references to related applications

[0002] This application claims priority to Chinese Patent Application No. 202411441460.8, filed on October 15, 2024, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to the fields of artificial intelligence, intelligent customer service, intelligent robots, deep learning, and large model technology, and more specifically, to a method, apparatus, electronic device, and medium for comparing items based on a large model. Background Technology

[0004] With the rapid development of artificial intelligence, intelligent customer service, intelligent robots, and deep learning, users can use item comparison functions before shopping to compare the similarities and differences between items they wish to purchase. However, most item comparisons on existing e-commerce platforms compare information on all attributes of an item in a fixed order from basic attributes to additional attributes, especially item comparisons implemented using Large Language Models (LLM).

[0005] In realizing the present invention, the inventors discovered at least the following problems in the related technology: In the case of item comparison, the importance of different types of attributes is different and will change over time. Therefore, the order of attributes in the existing item comparison scheme is unreasonable, resulting in a poor user experience of the item comparison function. Summary of the Invention

[0006] In view of this, the present disclosure provides a method, apparatus, electronic device and medium for comparing items based on a large model.

[0007] One aspect of this disclosure provides a large-scale model-based item comparison method, comprising: in response to receiving item comparison information input by a user, determining the target categories to which a first item and a second item belong in the item comparison information; determining, based on the target categories, multiple comparison attributes matching the item comparison information, and ranking information among the multiple comparison attributes, wherein the comparison attributes are used to describe the first item and the second item; inputting the multiple comparison attributes, the first item information of the first item, and the second item information of the second item into a large-scale item comparison model to generate a first item comparison sub-result; and generating an item comparison result based on the ranking information and the first item comparison sub-result.

[0008] According to embodiments of this disclosure, determining multiple attributes to be compared that match the item comparison information, and sorting information among the multiple attributes to be compared, based on the target category, includes: determining multiple original attributes of the target category and initial sorting information among the multiple original attributes based on the target category; filtering multiple attributes to be compared from the multiple original attributes based on the attention information of the original attributes; and determining sorting information among the multiple attributes to be compared from the initial sorting information.

[0009] According to embodiments of this disclosure, determining multiple original attributes of a target category and initial ranking information among the multiple original attributes based on the target category includes: querying multiple original attributes and initial ranking information from a database based on the target category; updating the attention information of each original attribute in response to the failure to find initial ranking information; and determining initial ranking information based on the updated attention information.

[0010] According to embodiments of this disclosure, updating the attention information of each original attribute includes: for each original attribute, obtaining a first parameter and a second parameter of the original attribute, wherein the first parameter includes a user profile parameter representing the elimination of user bias and a position parameter representing the elimination of attribute display bias; and inputting the first parameter and the second parameter into an attribute ranking model to obtain the output attention information.

[0011] According to embodiments of this disclosure, a large-scale item comparison model is used to perform an item comparison task. The first item comparison sub-result includes item similarity and difference information and item recommendation information. The process involves inputting multiple attributes to be compared, first item information of a first item, and second item information of a second item into the large-scale item comparison model to generate the first item comparison sub-result. This includes inputting multiple attributes to be compared, first item information, and second item information into the large-scale item comparison model, causing the large-scale item comparison model to perform the item comparison task and generate item similarity and difference information and item recommendation information.

[0012] According to embodiments of this disclosure, the large-scale item comparison model is also used to perform an attribute interpretation task, and the first item comparison sub-result further includes: attribute interpretation information; the method further includes: inputting multiple attributes to be compared into the large-scale item comparison model, so that the large-scale item comparison model performs an attribute interpretation task, interprets each attribute to be compared, and generates attribute interpretation information.

[0013] According to embodiments of this disclosure, generating an item comparison result based on sorting information and a first item comparison sub-result includes: generating a plurality of second item comparison sub-results with their attributes arranged in order based on the sorting information; and combining the first item comparison sub-result and the second item comparison sub-result into an item comparison result.

[0014] According to embodiments of this disclosure, the large-scale item comparison model is trained as follows: negative sample question information and a first positive sample pair similar to the negative sample question information are obtained, wherein the first positive sample pair includes the first positive sample question information and the first positive sample answer information output by the first initial large-scale item comparison model in response to the first positive sample question information; the negative sample question information and the first positive sample answer information are input into the large-scale model to generate training sample pairs; the first initial large-scale item comparison model is trained using the training sample pairs and the first positive sample pairs to obtain a second initial large-scale item comparison model; the second initial large-scale item comparison model is optimized for preferences using the second positive sample pairs and the second negative sample pairs to obtain a trained large-scale item comparison model, wherein the second positive sample pairs and the second negative sample pairs are determined based on the second initial large-scale item comparison model.

[0015] Another aspect of this disclosure provides an item comparison device based on a large model, comprising: a first determining module, configured to determine the target category to which a first item and a second item belong in the item comparison information in response to receiving item comparison information input by a user; a second determining module, configured to determine, according to the target category, a plurality of comparison attributes matching the item comparison information, and sorting information among the plurality of comparison attributes, wherein the comparison attributes are used to describe the first item and the second item; a first generating module, configured to input the plurality of comparison attributes, the first item information of the first item, and the second item information of the second item into the item comparison large model to generate a first item comparison sub-result; and a second generating module, configured to generate an item comparison result according to the sorting information and the first item comparison sub-result.

[0016] Another aspect of this disclosure provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the methods described above.

[0017] Another aspect of this disclosure provides a computer-readable storage medium storing computer-executable instructions that, when executed, are used to implement the methods described above.

[0018] Another aspect of this disclosure provides a computer program product including computer-executable instructions that, when executed, are used to implement the methods described above.

[0019] The embodiments of this disclosure, in response to receiving item comparison information input by a user, determine the target categories to which the first and second items in the item comparison information belong; based on the target categories, determine multiple comparison attributes that match the item comparison information, as well as the ranking information among the multiple comparison attributes; input the multiple comparison attributes, the first item information, and the second item information into a large item comparison model to generate a first item comparison sub-result; and generate an item comparison result based on the ranking information and the first item comparison sub-result. Compared to existing fixed-rank item comparisons, the embodiments of this disclosure add multiple comparison attributes that match the item comparison information and their ranking information, and combine the ranking information with the output of the large item comparison model. Therefore, the comparison attributes in the final output item comparison result have a dynamic ranking, thereby improving the user experience. Attached Figure Description

[0020] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 schematically illustrates an exemplary system architecture to which the large-model-based item comparison method and apparatus of this disclosure can be applied.

[0022] Figure 2 schematically illustrates a flowchart of a large-model-based item comparison method according to an embodiment of the present disclosure.

[0023] Figure 3 schematically illustrates a flowchart of a method for determining the attribute to be compared and the ranking information of the attribute to be compared according to an embodiment of the present disclosure.

[0024] Figure 4 schematically illustrates an application scenario of the article comparison function according to an embodiment of the present disclosure.

[0025] Figure 5 schematically illustrates a scenario in which user-inputted item comparison information is used to generate item comparison results according to an embodiment of the present disclosure.

[0026] Figure 6 schematically illustrates a scenario of training a large-scale item comparison model according to an embodiment of the present disclosure and implementing an item comparison function based on the large-scale item comparison model;

[0027] Figure 7 schematically illustrates a data flow diagram of an article comparison function according to an embodiment of the present disclosure;

[0028] Figure 8 schematically illustrates a block diagram of a large-model-based article comparison device according to an embodiment of the present disclosure; and

[0029] Figure 9 schematically illustrates a block diagram of an electronic device suitable for implementing a large-model-based item comparison method according to an embodiment of the present disclosure. Detailed Implementation

[0030] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0031] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0032] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0033] When using expressions such as "at least one of A, B, and C", they should generally be interpreted in accordance with the meaning that is commonly understood by a person skilled in the art (e.g., "a system having at least one of A, B, and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B, and C, etc.).

[0034] In the embodiments disclosed herein, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.

[0035] In the embodiments of this disclosure, user authorization or consent is obtained before acquiring or collecting user personal information. For example, user authorization or consent is obtained before acquiring user behavior data.

[0036] Existing item comparison schemes mostly use fixed, static sorting, which easily leads to useless or uncritical attributes being ranked higher than crucial attributes. For example, in the mobile phone industry, user A mainly wants to buy a phone with good battery life and focuses primarily on the "battery capacity" attribute. However, fixed, static sorting generally doesn't prioritize "battery capacity." Therefore, existing schemes require users to spend significant time searching for specific attributes, resulting in inconvenience and a poor user experience.

[0037] Furthermore, the importance of attributes varies across different categories, and users' level of concern for item attributes differs, also changing with time and market fluctuations. For example, for a newly released mobile phone with significantly improved attribute A, users will show increased concern and importance for attribute A when using the item comparison function. Therefore, existing attribute sorting based on a fixed order struggles to dynamically reflect changes in attribute importance and user concern.

[0038] Therefore, embodiments of this disclosure provide a method for item comparison based on a large model, comprising: in response to receiving item comparison information input by a user, determining the target categories to which the first item and the second item in the item comparison information belong; determining, based on the target categories, multiple attributes to be compared that match the item comparison information, and sorting information among the multiple attributes to be compared; inputting the multiple attributes to be compared, the first item information of the first item, and the second item information of the second item into the item comparison large model to generate a first item comparison sub-result; and generating an item comparison result based on the sorting information and the first item comparison sub-result.

[0039] Figure 1 schematically illustrates an exemplary system architecture to which the large-model-based item comparison method and apparatus of this disclosure can be applied.

[0040] As shown in Figure 1, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0041] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, query applications, instant messaging tools, email clients, and / or social media platform software, etc. (for example only).

[0042] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0043] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests and feed the processing results back to the terminal devices.

[0044] It should be noted that the item comparison method based on a large model provided in this disclosure can generally be executed by server 105. Correspondingly, the item comparison device based on a large model provided in this disclosure can generally be located in server 105. The item comparison method based on a large model provided in this disclosure can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105; or, it can also be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103. Correspondingly, the item comparison device based on a large model provided in this disclosure can also be located in the aforementioned server or server cluster; or, it can be located in the first terminal device 101, the second terminal device 102, or the third terminal device 103.

[0045] For example, a user can use an application from one of the e-commerce platforms, such as the first terminal device 101, the second terminal device 102, or the third terminal device 103, to input item comparison information. The server 105, such as the backend server of the e-commerce platform, executes the aforementioned item comparison method based on a large model and generates item comparison results. Then, the backend server of the e-commerce platform sends the item comparison results to the first terminal device 101, the second terminal device 102, or the third terminal device 103, and displays them within its application.

[0046] It should be understood that the number of terminal devices, networks, and servers shown in Figure 1 is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0047] Figure 2 schematically illustrates a flowchart of a large-model-based item comparison method according to an embodiment of the present disclosure.

[0048] As shown in Figure 2, the method includes operations S210 to S240.

[0049] In operation S210, in response to receiving the item comparison information input by the user, the target category to which the first item and the second item belong in the item comparison information is determined.

[0050] According to embodiments of this disclosure, the item comparison information includes a first item and a second item to be compared. The first and second items in the item comparison information can be in the form of item titles, item links, etc. The item comparison information can be entered by the user in a single input operation, or it can be entered through multiple sessions via a dialog window.

[0051] For example, a user can input item comparison information in a single operation: "Please compare phone A and phone B." In this case, the item comparison information can be represented by the item titles "Item A" and "Item B" to indicate the first and second items. Alternatively, the user can input multiple operations: "Link 1," "Link 2," and "Please help me compare the items in the above two links." In this case, the item comparison information will be the content of the three inputs, and the first and second items will be represented by item links.

[0052] The target category is the same category to which the first and second items belong, such as electronic items, daily necessities, clothing, etc. Furthermore, item categories typically include multiple levels; for example, electronic items include computers, mobile phones, watches, etc., and computers include desktop computers, laptops, etc. In one illustrative embodiment, a level can be pre-defined, and the same category to which the first and second items belong can be used as the target category.

[0053] In operation S220, based on the target type, multiple attributes to be compared with the item comparison information are determined, as well as the sorting information among the multiple attributes to be compared.

[0054] According to embodiments of this disclosure, each category typically includes multiple attributes to describe an item, and the attributes may differ for each category. For example, for a mobile phone, these attributes include multiple properties such as memory, color, charging power, and screen size; for clothing, these attributes include multiple properties such as texture, size, and color.

[0055] In one illustrative embodiment, when comparing items, all attributes under the target category can be used as comparison attributes. Similar to the original attributes, the comparison attributes are used to describe the first item and the second item.

[0056] In another illustrative embodiment, multiple attributes to be compared can be selected from multiple attributes under the target category for subsequent item comparison. For the currently input item comparison information, multiple attributes to be compared that match the item comparison information can be filtered from multiple attributes under the target category.

[0057] For example, user behavior data for a target category can be filtered within a preset time period. This could include the number of times a user clicks on certain attributes within the preset time period, or attributes mentioned in user reviews after ordering the first and / or second item. The user behavior data involved in this embodiment includes user browsing, clicking, and commenting activities on web pages.

[0058] According to embodiments of this disclosure, sorting information can be understood as the display order of multiple attributes to be compared when presenting item comparison results to a user. Sorting information can reflect the importance of multiple objects to be compared; for example, the earlier an attribute is listed, the more important it is. In one illustrative embodiment, the sorting information among multiple attributes to be compared can be determined based on the number of clicks on the attribute and the total number of times the attribute is mentioned in user reviews after ordering the first and / or second item.

[0059] In operation S230, multiple attributes to be compared, the first item information of the first item, and the second item information of the second item are input into the large item comparison model to generate the first item comparison sub-result.

[0060] According to embodiments of this disclosure, the first item information includes information about the first item, such as: item title, price, parameters under each attribute, etc. The second item information is similar to the first item information and may also include item title, price, parameters under each attribute, etc. For example, for the material attribute, the parameters for clothing A and clothing B under this attribute could be cotton and polyester fiber, respectively.

[0061] According to embodiments of this disclosure, a large-scale item comparison model is constructed by leveraging the powerful analytical and summarizing capabilities of large models to perform item comparison tasks. The large-scale item comparison model can be one or more of the following: Generative Pre-Trained Transformer (GPT), Chat Generative Pre-Trained Transformer (ChatGPT), General Language Model (GLM), etc., but is not limited to these.

[0062] In operation S240, an item comparison result is generated based on the sorting information and the first item comparison sub-result.

[0063] According to embodiments of this disclosure, after obtaining the first item comparison sub-result output by the large item comparison model, the first item comparison sub-result is not used as the output to be displayed to the user. Instead, the item comparison result combining the sorting information and the first item comparison result is used as the final output.

[0064] For example, the order of the first item comparison sub-results can be adjusted using sorting information. For instance, attributes ranked higher can be moved to a higher position. Alternatively, multiple second item comparison sub-results can be generated based on the sorting information, with the attributes to be compared arranged in order, and the item comparison result can be determined based on the first and second item comparison sub-results.

[0065] The embodiments of this disclosure, in response to receiving item comparison information input by a user, determine the target categories to which the first and second items in the item comparison information belong; based on the target categories, determine multiple attributes to be compared that match the item comparison information, and the ranking information among the multiple attributes to be compared; input the multiple attributes to be compared, the information of the first item, and the information of the second item into a large item comparison model to generate a first item comparison sub-result; and generate an item comparison result based on the ranking information and the first item comparison sub-result. Compared to existing fixed-rank item comparisons, the embodiments of this disclosure add multiple attributes to be compared that match the item comparison information and their ranking information, and combine the ranking information with the output of the large item comparison model. Therefore, the attributes to be compared in the final output item comparison result have a dynamic ranking, thereby improving the user experience.

[0066] Figure 3 schematically illustrates a flowchart of a method for determining the attribute to be compared and the ranking information of the attribute to be compared according to an embodiment of the present disclosure.

[0067] As shown in Figure 3, Embodiment 300 includes operations S321 to S323, which can be used as a specific embodiment of operation S220.

[0068] S321, Based on the target type, determine multiple original attributes of the target type and the initial sorting information among the multiple original attributes.

[0069] S322, Based on the attention information of the original attributes, select multiple attributes to be compared from multiple original attributes.

[0070] S323, determine the sorting information among multiple attributes to be compared from the initial sorting information.

[0071] According to embodiments of this disclosure, the original attributes are the attributes under a category, and multiple original data are all attributes under the target category. Here, to distinguish between attributes before and after filtering, the original attributes and the attributes to be compared are used.

[0072] In the e-commerce field, product categories and their attributes are usually relatively fixed and pre-stored in a database. Initial sorting information among multiple original attributes can also be pre-calculated and stored in the database. Therefore, embodiments of this disclosure can retrieve multiple original attributes and initial sorting information of a target product category from a pre-stored address based on the target product category identifier.

[0073] According to embodiments of this disclosure, the attention information of each original attribute can reflect the importance of the original attribute. The attention information can be represented numerically; the larger the value, the higher the importance of the original attribute. Therefore, determining the ranking information based on the attention information can also reflect the importance among multiple attributes to be compared.

[0074] In one illustrative embodiment, attention information can be calculated based on multiple dimensions of user behavior data to comprehensively evaluate the original attributes from multiple dimensions.

[0075] For example, multiple attributes to be compared can be selected from multiple raw attributes whose attention information meets predetermined conditions. These predetermined conditions could be: attention information greater than a preset threshold, or the attributes could be sorted from largest to smallest attention information, and k raw attributes with the highest attention information can be selected as the comparison attributes. Therefore, after obtaining multiple comparison attributes, ranking information among the multiple comparison attributes can be selected from the initial ranking information among the multiple raw attributes based on attention information.

[0076] The embodiments of this disclosure filter out multiple attributes to be compared from multiple original attributes based on the attention information of the original attributes, and filter out the ranking information between multiple attributes to be compared from the initial ranking information. This allows for the selective filtering of multiple attributes to be compared and their ranking information that match the current item comparison information, without having to display all original attributes in a fixed manner. This enables users to see attributes with higher attention when comparing target categories, thereby improving the user experience.

[0077] According to embodiments of this disclosure, determining multiple original attributes of a target category and initial ranking information among the multiple original attributes based on the target category includes: querying multiple original attributes and initial ranking information from a database based on the target category; updating the attention information of each original attribute in response to the failure to find initial ranking information; and determining initial ranking information based on the updated attention information.

[0078] In this embodiment, the attention information for each original attribute under each category can be calculated in advance based on user behavior data and stored in a database, such as a Redis cache. Similarly, the initial ranking information for multiple original attributes under each category can be calculated based on the attention information and stored in a database. Therefore, when using the item comparison service, the initial ranking information and multiple original attributes under the target category can be queried online based on the target category.

[0079] If the initial ranking information cannot be found online, update the attention information for each original attribute, that is, recalculate the attention information. For example, the attention information can be calculated based on the following: attribute exposure (expo_pv), attribute clicks (click_pv), orders (order), search filter category attribute ranking (idx), etc., and the weighted average of the above numerical features can be calculated as the attention information.

[0080] The initial ranking information is determined based on the updated attention information. This involves sorting multiple original attributes according to the updated attention information to obtain the initial ranking information. For example, the initial ranking information could be a sequence obtained by chaining multiple original attributes.

[0081] The embodiments of this disclosure pre-write attention information into a database, allowing users to directly obtain attention information rather than calculate it when using the item comparison function online. This reduces the amount of online resources required and improves the efficiency of item comparison.

[0082] In one illustrative embodiment, attention information for each raw attribute can also be calculated by invoking an offline-deployed attribute ranking model, such as learning-to-rank.

[0083] Figure 4 schematically illustrates an application scenario of the article comparison function according to an embodiment of the present disclosure.

[0084] As shown in Figure 4, during the online service phase, users can use the item comparison function on the e-commerce platform. For example, by inputting item comparison information 401, the e-commerce platform generates an item comparison request based on item comparison information 401 and reads the initial sorting information 402 from the Redis cache based on the item comparison request. The part not shown in the figure also includes queries for multiple original attributes and attention information. Afterwards, attribute querying / filtering / truncating is performed based on the attention information to obtain the sorting information. Attribute querying can be understood as: querying the original attribute at a certain position in the initial sorting information based on the attention information; attribute filtering / truncating can be understood as: filtering out a segment or truncating the initial sorting information from multiple initial sorting information based on the attention information to obtain the sorting information of some original attributes, such as the sorting information among multiple attributes to be compared in this paper. Finally, the item comparison result generated based on the sorting information and the first item sub-result is fed back to the user.

[0085] For the offline processing stage, click behavior data of page attributes can be searched and filtered based on operation S403, that is, the attributes used to filter item results through click operations when searching for a certain item in user behavior data. Click behavior data is stored based on operation S404, such as by offline caching to a Hive table. Daily updates are performed based on operation S405, that is, the stored click behavior data is updated daily, for example, through a big data platform like SparkSQL. Then, the attention information of each original attribute under each category is calculated offline using the attribute ranking model M1, and the multiple original attributes under each category are ranked based on the attention information to obtain the initial ranking information among the multiple original attributes under each category. The initial ranking information is stored based on operation S406.

[0086] It should be noted that in the embodiments of this disclosure, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solutions of this disclosure. However, it does not mean that this disclosure has used or necessarily used such solutions.

[0087] According to embodiments of this disclosure, updating the attention information of each original attribute includes: for each original attribute, obtaining a first parameter and a second parameter of the original attribute, wherein the first parameter includes a user profile parameter representing the elimination of user bias and a position parameter representing the elimination of attribute display bias; and inputting the first parameter and the second parameter into an attribute ranking model to obtain the output attention information.

[0088] When using attribute ranking models to calculate attention information, multiple types of features can be used as input to comprehensively calculate attention information from multiple dimensions, such as ID-type features, numerical features, and embedded features.

[0089] For example, the input to an attribute ranking model can be multiple dimensions, including: item category (cid3), category name (cid3_name), item code (skuid), user profile (user_profile), search term (query), attribute ID (attr_id), attribute name (attr_name), attribute exposure (expo_pv), attribute clicks (click_pv), order placement (order), and the search filter category attribute ranking position (idx). Here, cid3, skuid, and attr_id are ID-type features; idx, expo_pv, click_pv, and order are numerical features; and cid3_name, user_profile, and query are embedding features, which can be pre-generated by a text encoding / decoding model (sentence_transformer).

[0090] Search filter category attribute ranking (idx) is the ranking of an attribute on the page when a user searches for a certain type of item; attribute exposure (expo_pv) is the number of times an attribute is displayed on the page when a user searches for a certain type of item; attribute clicks (click_pv) is the number of times a user clicks on an attribute to filter item results when searching for a certain type of item; and order placement (order) is the number of times a user ultimately places an order after clicking on an attribute when searching for a certain type of item.

[0091] The actual parameters of the above 11 dimensions can be determined based on user behavior data, such as browsing, clicking, and order history of items.

[0092] In practical applications, the ranking of search category attributes (idx) can exhibit positional biases, such as attributes ranking higher being more likely to be clicked by users; user profiles are also influenced by user personalization. Therefore, during training, the initial attribute ranking model is trained using actual data from the aforementioned dimensions. However, during actual prediction, predetermined user profile parameters are used to replace the actual obtained user profiles to eliminate user bias; predetermined positional parameters are used to replace the actual obtained idx, while other parameters remain unchanged.

[0093] The first parameter includes the pre-defined user profile parameter 'user_profile=user_profile' and the location parameter 'idx=0'. The second parameter is the actual parameter value of the remaining dimensions other than user_profile and 'idx', such as the actual parameters of item type (cid3), type name (cid3_name), item code (skuid), search term (query), attribute ID (attr_id), attribute name (attr_name), attribute exposure (expo_pv), attribute click (click_pv), and order placement (order).

[0094] For example, during the training phase, the actual parameter value of the attribute click for each original attribute i under a certain category can be divided by the sum of the actual parameter values ​​of all attribute clicks. For instance, the label attention_score for the original attribute i under a certain category can be calculated as follows: i attention_score i =click_pv i / ∑ i click_pv i (1)

[0095] Where i represents a primitive attribute i under this category, and click_pv i The actual parameters for the click feature of the original attribute i.

[0096] Upon completion of training, the actual parameter values ​​are input into the attribute ranking model, and the predicted attention information attention_score is output as: attention_score=σ(attr_id,attr_name|idx,user_profile,Ω) (2)

[0097] Here, σ() represents the model output, identifying the current original attribute through attribute ID (attr_id) and attribute name (attr_name), and Ω is the feature set excluding idx and user_profile.

[0098] The first parameter, which eliminates user bias, is the average value of all user profiles, 'user_profile'. The position parameter for eliminating attribute display bias is 0, i.e., idx = 0, in order to reduce the impact of position bias on attention_score.

[0099] Therefore, for each original attribute under each category, the attention information obtained after inputting the first and second parameters into the attribute ranking model is: attention_score=σ(attr_id,cid3_name|0,user_profile',Ω) (3)

[0100] Among them, cid3_name is used to identify the type; other parameters are explained above.

[0101] In one embodiment, for the sake of interpretability, the attribute ranking model can adopt an extreme gradient boosting (XGBoost) structure, such as a five-layer XGBoost structure.

[0102] The embodiments of this disclosure calculate the attention information of each original attribute by using user profile parameters that eliminate user bias, position parameters that eliminate attribute display bias, and a second parameter. This can eliminate position bias caused during model training, thereby obtaining more accurate attention information and helping to improve the accuracy of subsequent determination of the attributes to be compared and their ranking based on attention information.

[0103] According to embodiments of this disclosure, a large-scale item comparison model is used to perform an item comparison task. The first item comparison sub-result includes item similarity and difference information and item recommendation information. The process involves inputting multiple attributes to be compared, first item information of a first item, and second item information of a second item into the large-scale item comparison model to generate the first item comparison sub-result. This includes inputting multiple attributes to be compared, first item information, and second item information into the large-scale item comparison model, causing the large-scale item comparison model to perform the item comparison task and generate item similarity and difference information and item recommendation information.

[0104] In this embodiment, after inputting multiple attributes to be compared, the first item information of the first item, and the second item information of the second item into the large item comparison model, the multiple attributes to be compared, the first item information of the first item, and the second item information of the second item can be combined into the input of the large item comparison model according to the pre-set prompt template, so that it can perform the item comparison task.

[0105] For example, the prompt template includes "Please compare the similarities and differences between XX and XX in the following attributes: YY, recommend the applicable scenarios for XX".

[0106] The input for the large-scale item comparison model could be: [Please compare the similarities and differences of the following attributes of the first and second items: memory, charging power. Recommend suitable scenarios for the first item. The information for the first item is: Model X mobile phone with 128GB of memory and a large 6.1-inch screen..., and the information for the second item is: Model Y mobile phone with 128GB of memory and a super-long battery life of 5000mA...].

[0107] The output of the item comparison model could be something like, "[Both model X and model Y phones have 128GB of memory, but model X has a larger screen.... Model X with its larger screen is recommended for users aged 20-30]". Here, "Both model X and model Y phones have 128GB of memory, but model X has a larger screen..." represents the item's differences, while "Model X with its larger screen is recommended for users aged 20-30" represents the item recommendation.

[0108] The embodiments of this disclosure invoke a large-scale item comparison model to perform item comparison tasks, enabling the model to compare and recommend items based on their similarities and differences, providing auxiliary information for e-commerce users' choices and decisions, and helping to improve the user experience.

[0109] According to embodiments of this disclosure, the large-scale item comparison model is also used to perform an attribute interpretation task. The first item comparison sub-result further includes attribute interpretation information. The method further includes: inputting multiple attributes to be compared into the large-scale item comparison model, so that the large-scale item comparison model performs an attribute interpretation task, interprets each attribute to be compared, and generates attribute interpretation information.

[0110] Considering that many item attributes are technical terms and have a high cost of understanding for users, this embodiment of the disclosure can perform an attribute interpretation task while using the large item comparison model to perform the item comparison task, interpret each attribute to be compared, and generate attribute interpretation information.

[0111] The input to the large-scale item comparison model is always multiple attributes to be compared, information about the first item, and information about the second item. However, the inputs used for the item comparison task and the attribute interpretation task are different. As a result, the output of the large-scale item comparison model is different for the item comparison task and the attribute interpretation task.

[0112] For a large-scale item comparison model that can perform both item comparison and attribute interpretation tasks simultaneously, its prompt word template differs from the prompt word template that only performs item comparison tasks.

[0113] For example, the prompt template includes "Please compare the similarities and differences between XX and XX in the following attributes: YY, and interpret the attributes. Finally, recommend suitable scenarios for XX."

[0114] The input for the item comparison model could be: [Please compare the similarities and differences of the following attributes of the first and second items: memory, charging power, and interpret the attributes. Recommended applicable scenarios for the first item. First item information: Model X mobile phone, with 128G memory, 6.1-inch large screen, 90W charging power...; Second item information: Model Y mobile phone, with 128G memory and 5000mA super long battery life...].

[0115] The output of the item comparison model could be: [Both model X and model Y phones have 128GB of storage, but model X has a larger screen and a higher 90W charging power, resulting in faster charging speed… For the larger-screen model X, it is recommended for users aged 20-30]. The item difference information is "Both model X and model Y phones have 128GB of storage, but model X has a larger screen and a higher 90W charging power," and the item recommendation information is similar to the above, while the attribute interpretation information is "faster charging speed."

[0116] For users unfamiliar with charging power, the meaning of charging power may not be clear. However, by interpreting the attribute information "faster charging speed," they can understand that charging power is related to charging speed.

[0117] The embodiments of this disclosure enable the large model for item comparison to perform attribute interpretation tasks and output attribute interpretation information, helping users to better understand the meaning of each attribute in the attributes to be compared, and thus improving the user experience.

[0118] According to embodiments of this disclosure, generating an item comparison result based on sorting information and a first item comparison sub-result includes: generating a plurality of second item comparison sub-results with their attributes arranged in order based on the sorting information; and combining the first item comparison sub-result and the second item comparison sub-result into an item comparison result.

[0119] The second item comparison sub-result includes the attribute parameters of the first and second items for each attribute to be compared. The attribute parameters of the first or second item can be obtained from the information of the first and second items, respectively. For example, if the attribute to be compared is charging power, the attribute parameter of the first item can be 20W, and the attribute parameter of the second item can be 90W.

[0120] Specifically, after determining the sorting information among multiple attributes to be compared, each attribute to be compared can be filled into the initial template based on the sorting information to obtain a layout in which multiple attributes to be compared are arranged in order. Then, the attribute parameters of the first item and the second item for each attribute to be compared are filled into the initial template to obtain the second item comparison sub-result.

[0121] Figure 5 schematically illustrates a scenario in which user-inputted item comparison information is used to generate item comparison results according to an embodiment of the present disclosure.

[0122] As shown in Figure 5, users can input item comparison information 501 through three operations, such as the link to the first item, the link to the second item, and "compare these two items for me." The item comparison result 502 obtained by the item comparison method based on the large model includes the first item comparison sub-result 5021 and the second item comparison sub-result 5022.

[0123] The first item comparison sub-result 5021 is the output of the large item comparison model, such as "Both items support 5G and 4G networks and have the same network access license. However, the first item does not support memory cards, and the CPU model of the first item is XXX... The second item..."

[0124] The second item comparison sub-result 5022 includes: images and prices of the first / second item, or links containing images and prices. The second item comparison sub-result 5022 also includes multiple comparison attributes and their parameters arranged according to sorting information, such as memory and charging power. For second item comparison sub-results that cannot be fully displayed, all comparison attributes and their parameters can be viewed using the "Expand All" control.

[0125] The embodiments of this disclosure generate the final item comparison result based on the first item comparison sub-result and the second item comparison sub-result, which can realize the item comparison function from both intuitive and textual summary aspects, thereby improving the user experience.

[0126] According to embodiments of this disclosure, the large-scale item comparison model is trained as follows: negative sample question information and a first positive sample pair similar to the negative sample question information are obtained, wherein the first positive sample pair includes the first positive sample question information and the first positive sample answer information output by the first initial large-scale item comparison model in response to the first positive sample question information; the negative sample question information and the first positive sample answer information are input into the large-scale model to generate training sample pairs; the first initial large-scale item comparison model is trained using the training sample pairs and the first positive sample pairs to obtain a second initial large-scale item comparison model; the second initial large-scale item comparison model is optimized for preferences using the second positive sample pairs and the second negative sample pairs to obtain a trained large-scale item comparison model, wherein the second positive sample pairs and the second negative sample pairs are determined based on the second initial large-scale item comparison model.

[0127] This disclosure allows for alignment training of a first initial large-scale object comparison model using supervised fine-tuning, also known as supervised learning (SFT), such as a pre-trained large-scale language model (Pre-trained LLM) for conversational purposes, to obtain a large-scale object comparison model.

[0128] In the first stage of the training process, a conversational pre-trained LLM using Retrieval Augmented Generation (RAG) can be used to generate corresponding answer information for the question information. The response information sysreturn of the open-source pre-trained LLM is: Sysreturn = Pre-trained LLM ([prompt, user_input, sku_info]) (4)

[0129] Here, user_input is the user input, sku_info is the item information, including the item title, attributes, price, etc., and prompt is a simple prompt, such as "Please answer according to the user's question and the item information:".

[0130] Based on online user feedback, the question and answer information from the pre-trained LLM was labeled as positive and negative sample pairs. Specifically, replies that were disliked, excessively long replies, replies with looping text, and replies that were stopped manually by the user before completion were labeled as negative samples, while replies that were liked, had follow-up questions and answers, or had an order record in the current round were labeled as positive sample answers (Spos).

[0131] For each negative sample pair mentioned above, a similar positive sample pair is determined based on the question information in the negative sample pair. This yields the aforementioned negative sample question information and the first positive sample pair; that is, the question information in the negative sample pair is the aforementioned negative sample question information, and the positive sample pair is the first positive sample pair. For example, the user input in the negative sample prompt information and the user input in the first positive sample question information can be converted into embedding form, and cosine similarity can be calculated to determine whether they are similar.

[0132] In the second stage of training, the negative sample question information and the first positive sample answer in the first positive sample pair are used as small samples. Through few-shot learning, a large sample generation model is used to generate more training sample pairs. This large sample generation model can be GPT4. The training sample pair Sysreturn_sample output by the large sample generation model can be: Sysreturn_sample = GPT4([prompt...neg ,user_input neg ,sku_info neg , Spos])(5)

[0133] Among them, prompt neg ,user_input neg ,sku_info neg The prompts, user input, and item information in the negative sample question information are represented by Spos, which represents the first positive sample answer information.

[0134] In one embodiment, the output of the large model for generating samples can also be manually modified to obtain training sample pairs.

[0135] In the third stage of training, the training sample pairs obtained in the second stage and the first positive sample response information are mixed into a training dataset. Supervised alignment training is then performed on the Pre-trained LLM to obtain a second initial item comparison large model, such as the SFT LLM. Alternatively, the training sample pairs, the first positive sample pairs, and the training data from the common dataset can be mixed in an 85:10:5 ratio to form a training dataset for supervised alignment training of the Pre-trained LLM. The common dataset can be Super-Natural Instructions.

[0136] In the fourth stage of training, the process of the second stage is repeated, labeling the SFT LLM with either a second positive sample pair or a second negative sample pair based on the online real-world question information and the generated answer information. At this point, the second positive sample pair includes the second positive sample question information and the second positive sample answer information, and the second negative sample pair includes the second negative sample question information and the second negative sample answer information. For the second negative sample question information, the model parameters are adjusted so that the SFT LLM regenerates the third positive sample answer information. Then, the second negative sample question information and the third positive sample answer information are combined into a positive-negative sample pair, and the SFT LLM is trained using Direct Preference Optimization (DPO). The resulting item comparison model is called the DPO LLM. Adjusting the model parameters can involve increasing parameters such as generation temperature, top-k, and top-p.

[0137] The embodiments of this disclosure obtain training sample pairs that better suit the item comparison scenario through the screening and continuous optimization of positive and negative samples. These training sample pairs are then used for direct preference optimization training, resulting in a large-scale item comparison model capable of item comparison and attribute interpretation. Therefore, the first item comparison sub-result output by the large-scale item comparison model can include specific item comparisons and attribute interpretations, helping users to better understand the meaning of attributes and the item comparison results, thus improving the user experience.

[0138] Figure 6 schematically illustrates a scenario of training a large-scale item comparison model according to an embodiment of the present disclosure and implementing an item comparison function based on the large-scale item comparison model.

[0139] As shown in Figure 6, during the offline alignment training phase of the large-scale item comparison model, the user input 602 entered by the user within application 601 is recorded offline. Then, background information 603 related to the question is retrieved using RAG. The background information (B) can be information related to the question, such as item information or a simple prompt. Based on the generation path of the pre-trained LLM, QA pairs, including question information (Q) and answer information (A), are generated according to the user input 602 and background information 603. After labeling with positive and negative samples, the QA pairs and QB pairs are sent together to the large-scale sample generation model M61 for few-shot learning. Then, the answer 604 generated by the large-scale sample generation model M61 is combined with the user input 602 to form a training sample pair 605 for alignment training of the first initial large-scale item comparison model M62, also known as supervised fine-tuning (SFT), resulting in the second initial large-scale item comparison model M63. Finally, the large-scale item comparison model M64 is obtained through direct preference optimization training.

[0140] During the online service phase of the large-scale item comparison model, users can input item comparison information through application 601. Application 601 generates an item comparison request based on this information, identifying multiple attributes to be compared and ranking information 606 that match the item comparison information. Subsequently, the trained large-scale item alignment model M64 generates a first item comparison sub-result 607, such as attribute interpretation information, item similarities and differences information, and item recommendation information. The item comparison result generated from the first item comparison sub-result 607 and the ranking information 606 is then fed back to the user through application 601.

[0141] To facilitate understanding of the technical solutions of this application, a specific data flow diagram is provided below to describe an embodiment of this disclosure. Figure 7 schematically illustrates a data flow diagram of an item comparison function according to an embodiment of this disclosure.

[0142] As shown in Figure 7, when a user uses application 701, user behavior data 702 is obtained. This data is stored as a user behavior log 703 and updated daily. Then, for multiple original attributes 704 under different categories for the user, attribute ranking model M71 is used to calculate the initial ranking information between these original attributes for each category. Next, when the user uses the item comparison function, the large item comparison model M72 is used to compare and interpret multiple attributes under the target category, obtaining the first item comparison sub-result, and determining the ranking information between the multiple attributes from the calculated initial ranking information. Finally, the item comparison result is generated based on the ranking information and the first item comparison sub-result, and fed back to the user through application 701. The large item comparison model M72 is obtained by aligning and training the first initial large item comparison model M73 using user behavior data 703.

[0143] Figure 8 schematically illustrates a block diagram of a large-model-based item comparison device according to an embodiment of the present disclosure.

[0144] As shown in Figure 8, the item comparison device 800 includes a first determining module 810, a second determining module 820, a first generating module 830, and a second generating module 840.

[0145] The first determining module 810 is used to determine the target category to which the first item and the second item belong in the item comparison information in response to receiving the item comparison information input by the user.

[0146] The second determining module 820 is used to determine, based on the target type, multiple attributes to be compared that match the item comparison information, and the sorting information among the multiple attributes to be compared, wherein the attributes to be compared are used to describe the first item and the second item.

[0147] The first generation module 830 is used to input multiple attributes to be compared, the first item information of the first item, and the second item information of the second item into the large item comparison model to generate the first item comparison sub-result.

[0148] The second generation module 840 is used to generate item comparison results based on the sorting information and the first item comparison sub-result.

[0149] According to embodiments of this disclosure, the second determining module 820 includes: a first determining submodule, a second determining submodule, and a third determining submodule.

[0150] The first determination submodule is used to determine multiple original attributes of the target type and the initial sorting information among the multiple original attributes, based on the target type.

[0151] The second determination submodule is used to filter out multiple attributes to be compared from multiple original attributes based on the attention information of the original attributes.

[0152] The third determination submodule is used to determine the sorting information among multiple attributes to be compared from the initial sorting information.

[0153] According to embodiments of this disclosure, the first determining submodule includes: a query unit, an update unit, and a determining unit.

[0154] The query unit is used to retrieve multiple raw attributes and initial sorting information from the database based on the target type.

[0155] The update unit is used to update the attention information for each original attribute in response to the failure to find the initial ranking information.

[0156] The determination unit is used to determine the initial ranking information based on the updated attention information.

[0157] According to embodiments of this disclosure, the updating unit includes an acquisition subunit and a calculation subunit. For each original attribute, the acquisition subunit is used to acquire a first parameter and a second parameter of the original attribute, wherein the first parameter includes a user profile parameter representing the elimination of user bias and a position parameter representing the elimination of attribute display bias. The calculation subunit is used to input the first parameter and the second parameter into an attribute ranking model to obtain output attention information.

[0158] According to embodiments of this disclosure, a large-scale item comparison model is used to perform an item comparison task, and the first item comparison sub-result includes item similarities and differences information and item recommendation information.

[0159] The first generation module 830 includes a first generation submodule, which is used to input multiple attributes to be compared, first item information and second item information into the item comparison big model, so that the item comparison big model performs the item comparison task and generates item similarity and difference information and item recommendation information.

[0160] According to embodiments of this disclosure, the large item comparison model is also used to perform an attribute interpretation task, and the first item comparison sub-result further includes attribute interpretation information.

[0161] The first generation module 830 also includes a second generation submodule, which is used to input multiple attributes to be compared into the large item comparison model, so that the large item comparison model performs an attribute interpretation task, interprets each attribute to be compared, and generates attribute interpretation information.

[0162] According to embodiments of this disclosure, the second generation module 840 includes a third generation submodule and a combination submodule.

[0163] The third generation submodule is used to generate second item comparison sub-results with multiple attributes to be compared arranged in order, based on the sorting information.

[0164] The combination submodule is used to combine the first item comparison sub-result and the second item comparison sub-result to form the item comparison result.

[0165] According to embodiments of this disclosure, the large-scale item comparison model is trained as follows: negative sample question information and a first positive sample pair similar to the negative sample question information are obtained, wherein the first positive sample pair includes the first positive sample question information and the first positive sample answer information output by the first initial large-scale item comparison model in response to the first positive sample question information; the negative sample question information and the first positive sample answer information are input into the large-scale model to generate training sample pairs; the first initial large-scale item comparison model is trained using the training sample pairs and the first positive sample pairs to obtain a second initial large-scale item comparison model; the second initial large-scale item comparison model is optimized for preferences using the second positive sample pairs and the second negative sample pairs to obtain a trained large-scale item comparison model, wherein the second positive sample pairs and the second negative sample pairs are determined based on the second initial large-scale item comparison model.

[0166] Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure, or at least part of the functions of any one or more of them, can be implemented in one module. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be implemented by dividing them into multiple modules. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as hardware circuitry, such as Field Programmable Gate Arrays (FPGAs), Programmable Logic Arrays (PLAs), Systems-on-Chip, Systems-on-Substrate, Systems-on-Package, Application-Specific Integrated Circuits (ASICs), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, one or more of the modules, submodules, units, and subunits according to embodiments of the present disclosure can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0167] It should be noted that the apparatus in the embodiments of this disclosure corresponds to the method in the embodiments of this disclosure. The description of the apparatus is specifically referred to in the method section, and will not be repeated here.

[0168] Figure 9 schematically illustrates a block diagram of an electronic device suitable for implementing a large-model-based item comparison method according to an embodiment of the present disclosure.

[0169] The electronic device shown in Figure 9 is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0170] As shown in FIG9, an electronic device 900 according to an embodiment of the present disclosure includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage portion 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0171] RAM 903 stores various programs and data required for the operation of electronic device 900. Processor 901, ROM 902, and RAM 903 are interconnected via bus 904. Processor 901 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 902 and / or RAM 903. It should be noted that the programs may also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in said one or more memories.

[0172] According to embodiments of this disclosure, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to a bus 904. The electronic device 900 may also include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 910 as needed so that computer programs read from it can be installed into the storage section 908 as needed.

[0173] According to embodiments of this disclosure, the method flow according to embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 909, and / or installed from removable medium 911. When the computer program is executed by processor 901, it performs the functions defined in the system of embodiments of this disclosure. According to embodiments of this disclosure, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0174] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0175] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0176] For example, according to embodiments of this disclosure, a computer-readable storage medium may include the ROM 902 and / or RAM 903 described above and / or one or more memories other than ROM 902 and RAM 903.

[0177] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this disclosure. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the large-model-based item comparison method provided in the embodiments of this disclosure.

[0178] When the computer program is executed by the processor 901, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0179] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 909, and / or installed from a removable medium 911. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0180] According to embodiments of this disclosure, program code for executing the computer programs provided in embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can execute entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0181] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present disclosure can be combined and / or combined in various ways, even if such combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments of this disclosure may be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0182] The embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A large model-based item comparison method, wherein, The method includes: In response to receiving item comparison information input by the user, determine the target category to which the first item and the second item in the item comparison information belong; Based on the target category, determine multiple attributes to be compared that match the item comparison information, and sorting information among the multiple attributes to be compared, wherein the attributes to be compared are used to describe the first item and the second item; Input multiple attributes to be compared, the first item information of the first item, and the second item information of the second item into the large item comparison model to generate a first item comparison sub-result; and Based on the sorting information and the first item comparison sub-result, an item comparison result is generated.

2. The method of claim 1, wherein, The step of determining, based on the target category, multiple attributes to be compared that match the item comparison information, and the sorting information among the multiple attributes to be compared, includes: Based on the target type, determine multiple original attributes of the target type and initial sorting information among the multiple original attributes; Based on the attention information of the original attributes, multiple attributes to be compared are selected from the multiple original attributes; and The sorting information among the multiple attributes to be compared is determined from the initial sorting information.

3. The method of claim 2, wherein, The step of determining multiple original attributes of the target type and initial sorting information among the multiple original attributes based on the target type includes: Based on the target type, query multiple original attributes and the initial sorting information from the database; In response to the failure to find the initial ranking information, update the attention information for each of the original attributes; and The initial ranking information is determined based on the updated attention information.

4. The method of claim 3, wherein, The update of the attention information for each of the original attributes includes: for each of the original attributes, Obtain a first parameter and a second parameter of the original attribute, wherein the first parameter includes a user profile parameter representing the elimination of user bias and a position parameter representing the elimination of attribute display bias; and The first parameter and the second parameter are input into the attribute sorting model to obtain the output attention information.

5. The method of claim 1, wherein, The item comparison model is used to perform item comparison tasks. The first item comparison sub-result includes item similarity / difference information and item recommendation information. The step of inputting multiple attributes to be compared, the first item information of the first item, and the second item information of the second item into the item comparison model to generate the first item comparison sub-result includes: The multiple attributes to be compared, the first item information, and the second item information are input into the item comparison model, so that the item comparison model performs the item comparison task and generates the item similarity and difference information and the item recommendation information.

6. The method of claim 5, wherein, The item comparison model is also used to perform attribute interpretation tasks, and the first item comparison sub-result further includes: attribute interpretation information; the method further includes: Multiple attributes to be compared are input into the item comparison model, causing the item comparison model to perform the attribute interpretation task, interpret each attribute to be compared, and generate the attribute interpretation information.

7. The method of claim 5 or 6, wherein, The step of generating an item comparison result based on the sorting information and the first item comparison sub-result includes: Based on the sorting information, generate multiple second item comparison sub-results arranged in the order of the attributes to be compared; and The first item comparison result and the second item comparison result are combined to form the item comparison result.

8. The method of claim 1, wherein, The large-scale model for comparing the items was trained in the following way: Obtain negative sample question information and a first positive sample pair similar to the negative sample question information, wherein the first positive sample pair includes first positive sample question information and first positive sample answer information output by the first initial item comparison model for the first positive sample question information; The negative sample question information and the first positive sample answer information are input into the large model to generate training sample pairs. Using the training sample pairs and the first positive sample pairs, the first initial item comparison model is trained to obtain the second initial item comparison model; Using the second positive sample pair and the second negative sample pair, the second initial item comparison model is optimized to obtain the trained item comparison model, wherein the second positive sample pair and the second negative sample pair are determined based on the second initial item comparison model.

9. An item comparison device based on a large model, wherein, The device includes: The first determining module is used to determine the target category to which the first item and the second item belong in the item comparison information in response to receiving item comparison information input by the user; The second determining module is used to determine, based on the target type, multiple attributes to be compared that match the item comparison information, and sorting information among the multiple attributes to be compared, wherein the attributes to be compared are used to describe the first item and the second item; The first generation module is used to input multiple attributes to be compared, the first item information of the first item, and the second item information of the second item into the item comparison model, and generate a first item comparison sub-result; and The second generation module is used to generate an item comparison result based on the sorting information and the first item comparison sub-result.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 8.

11. A computer-readable storage medium having stored thereon executable instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 8.

12. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 8.

Images