Product recommendation method, device, equipment, system and program product

Abstract

The embodiment of the invention provides a product recommendation method. The method comprises the steps of obtaining corresponding dialogue data when a target object selects a product; performing semantic recognition based on the dialogue data to determine dominant features corresponding to the target object; calling associated information corresponding to the target object, and performing feature extraction according to the associated information to obtain hidden features; and fusing the dominant features and the recessive features to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information. Therefore, a product recommendation process based on dominant features and recessive features is realized, dialogue data is fully mined, multi-dimensional information fusion is performed, the accuracy of dialogue understanding is improved, information supplementation is performed through the recessive features, the comprehensiveness of object representation is further improved, and the accuracy of product recommendation is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to product recommendation methods, devices, equipment, systems and program products. Background Art

[0002] With the continuous development of the Internet and AI technology, companies have significantly improved their customer insights and personalized marketing technologies, especially in the field of outbound marketing. How to better utilize conversation information and implement personalized precision marketing has become a current research hotspot.

[0003] Generally, product recommendations in the outbound marketing process use fixed keyword rules to extract key information, and then make product recommendations based on the key information.

[0004] However, with the richness of products and the rapid increase in user scale, keyword rules may not be applicable to all recommendation scenarios, which may cause product recommendations to be inappropriate and affect the accuracy of product recommendations. Summary of the invention

[0005] In view of this, the present application provides a product recommendation method, which can improve the accuracy of product recommendations.

[0006] According to a first aspect of an embodiment of the present application, a method for product recommendation is provided, which can be applied to a system or program including a product recommendation function in a terminal device, and specifically includes:

[0007] Acquire conversation data of a target object when selecting a product, wherein the conversation data includes object information and product information corresponding to the target object;

[0008] Performing semantic recognition based on the conversation data to determine explicit features corresponding to the target object, wherein the explicit features include object features and product features;

[0009] Calling the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features;

[0010] The explicit features and the implicit features are fused to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information.

[0011] Optionally, in some possible implementations of the present application, the performing semantic recognition based on the conversation data to determine the explicit features corresponding to the target object includes:

[0012] Performing semantic recognition based on the conversation data, summarizing key information of the sentences involved in the conversation process to obtain semantic representation information;

[0013] Determine a speech node in the conversation data, and determine corresponding intent information according to the speech node;

[0014] Extracting key elements from the conversation data according to the configured key information categories to obtain object portrait information and product description information;

[0015] Performing emotion recognition on the conversation data to obtain emotion analysis information;

[0016] generating the object feature based on the semantic representation information, the intention information, the object portrait information and the sentiment analysis information, and generating the product feature based on the product description information;

[0017] The object characteristics and the product characteristics are combined to determine the dominant characteristics corresponding to the target object.

[0018] Optionally, in some possible implementations of the present application, determining a speech node in the conversation data, and determining corresponding intent information according to the speech node includes:

[0019] Obtaining scene information corresponding to the conversation data;

[0020] Calling the speech node diagram associated with the scene information;

[0021] Detecting the conversation data based on the speech node graph to determine the speech nodes in the conversation data;

[0022] Determine the intent information corresponding to the speech node.

[0023] Optionally, in some possible implementations of the present application, extracting key elements from the conversation data according to the configured key information categories to obtain object portrait information and product description information includes:

[0024] Performing corpus extraction on the conversation data according to the configured key information categories to obtain corpus extraction elements;

[0025] Logically associating the extracted elements of the corpus to obtain condensed information;

[0026] Key elements are extracted based on the condensed information to obtain the object portrait information and the product description information.

[0027] Optionally, in some possible implementations of the present application, the performing emotion recognition on the conversation data to obtain emotion analysis information includes:

[0028] Determining feedback information in the conversation data;

[0029] Identifying emotional keywords in each corpus in the feedback information to obtain an emotional keyword set;

[0030] Emotion recognition is performed based on the identified emotion keyword set to obtain the emotion analysis information.

[0031] Optionally, in some possible implementations of the present application, combining the object feature and the product feature to determine the explicit feature corresponding to the target object includes:

[0032] Determining feature types corresponding to the object feature and the product feature;

[0033] Determine a corresponding feature processing method based on the feature type;

[0034] According to the preset dimensions, the feature processing method is used to regularize the feature dimensions of the object features and the product features;

[0035] The processed object features and product features are combined to determine the dominant features corresponding to the target object.

[0036] Optionally, in some possible implementations of the present application, calling the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features includes:

[0037] Calling the associated information corresponding to the target object;

[0038] Determining the action of the target object on the product indicated in the association information to obtain behavior information;

[0039] Acquire a knowledge database corresponding to the associated information, perform knowledge call based on the knowledge database, and obtain open domain information corresponding to the target object;

[0040] Feature extraction is performed based on the behavior information and the open domain information respectively to obtain the latent feature.

[0041] Optionally, in some possible implementations of the present application, the fusing of the explicit feature and the implicit feature to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information, includes:

[0042] Determine a mapping relationship between the target object indicated in the explicit feature and each product, so as to restore the explicit feature to first knowledge representation data based on the mapping relationship;

[0043] Determine a mapping relationship between the target object indicated in the implicit feature and each product, so as to restore the implicit feature to second knowledge representation data based on the mapping relationship;

[0044] Performing semantic summarization based on the first knowledge representation data and the second knowledge representation data to obtain summary features;

[0045] Determine the cross-domain information corresponding to the target object according to the summary feature;

[0046] A target product is recommended to the target object based on the cross-domain information.

[0047] Optionally, in some possible implementations of the present application, the recommending a target product to the target object based on the cross-domain information includes:

[0048] Obtaining a candidate list corresponding to the target object, wherein the candidate list includes multiple candidate products;

[0049] Performing word segmentation processing on the cross-domain information, and performing vector representation on the cross-domain information after the word segmentation processing to obtain a cross-domain vector;

[0050] Calculating the similarity between the cross-domain vector and each of the candidate products respectively to obtain a target score for each of the candidate products;

[0051] Product screening is performed based on the target score of each of the candidate products to obtain the target product recommended to the target object.

[0052] Optionally, in some possible implementations of the present application, respectively calculating the similarity between the cross-domain vector and each of the candidate products to obtain a target score for each of the candidate products includes:

[0053] Respectively calculating the similarity between the cross-domain vector and each of the candidate products in a preset scoring dimension to obtain a semantic score, an object feature score, and a product feature score, wherein the preset scoring dimension includes a semantic dimension, an object feature dimension, and a product feature dimension;

[0054] The semantic score, the object feature score and the product feature score are weightedly calculated to obtain the target score.

[0055] Optionally, in some possible implementations of the present application, the method further includes:

[0056] Determine the speech nodes in the conversation data;

[0057] Obtain the recommended speech template corresponding to the speech node;

[0058] Determine a recommendation speech template corresponding to the target product, and rewrite the content of the recommendation speech template to obtain the target recommendation speech;

[0059] The target product is recommended to the target object based on the target recommendation words.

[0060] Optionally, in some possible implementations of the present application, the method further includes:

[0061] Collect feedback from target objects on target products;

[0062] Adjusting the feature extraction dimension corresponding to the target object based on the feedback information;

[0063] The open domain information is updated according to the knowledge information indicated by the feedback information.

[0064] According to a second aspect of an embodiment of the present application, a product recommendation device is provided, including:

[0065] An acquisition unit, used to acquire dialogue data of a target object when selecting a product, wherein the dialogue data includes object information and product information corresponding to the target object;

[0066] A processing unit, configured to perform semantic recognition based on the conversation data to determine explicit features corresponding to the target object, wherein the explicit features include object features and product features;

[0067] The processing unit is further used to call the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features;

[0068] A recommendation unit is used to fuse the explicit features and the implicit features to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information.

[0069] Optionally, in some possible implementations of the present application, the processing unit is specifically used to perform semantic recognition based on the dialogue data, and summarize key information of the sentences involved in the dialogue process to obtain semantic representation information;

[0070] The processing unit is specifically used to determine the speech nodes in the conversation data, and determine the corresponding intention information according to the speech nodes;

[0071] The processing unit is specifically used to extract key elements from the conversation data according to the configured key information categories to obtain object portrait information and product description information;

[0072] The processing unit is specifically used to perform emotion recognition on the conversation data to obtain emotion analysis information;

[0073] The processing unit is specifically configured to generate the object features based on the semantic representation information, the intention information, the object portrait information and the sentiment analysis information, and to generate the product features based on the product description information;

[0074] The processing unit is specifically used to combine the object characteristics and the product characteristics to determine the dominant characteristics corresponding to the target object.

[0075] Optionally, in some possible implementations of the present application, the processing unit is specifically used to obtain scene information corresponding to the conversation data;

[0076] The processing unit is specifically used to call the speech node graph associated with the scene information;

[0077] The processing unit is specifically configured to detect the conversation data based on the speech node graph to determine the speech nodes in the conversation data;

[0078] The processing unit is specifically used to determine the intention information corresponding to the speech node.

[0079] Optionally, in some possible implementations of the present application, the processing unit is specifically configured to perform corpus extraction on the conversation data according to the configured key information category to obtain corpus extraction elements;

[0080] The processing unit is specifically used to logically associate the corpus extraction elements to obtain condensed information;

[0081] The processing unit is specifically used to extract key elements based on the condensed information to obtain the object portrait information and the product description information.

[0082] Optionally, in some possible implementations of the present application, the processing unit is specifically configured to determine feedback information in the conversation data;

[0083] The processing unit is specifically used to identify the emotion keywords in each corpus in the feedback information to obtain an emotion keyword set;

[0084] The processing unit is specifically configured to perform emotion recognition based on the identified emotion keyword set to obtain the emotion analysis information.

[0085] Optionally, in some possible implementations of the present application, the processing unit is specifically used to determine the feature type corresponding to the object feature and the product feature;

[0086] The processing unit is specifically used to determine a corresponding feature processing method based on the feature type;

[0087] The processing unit is specifically used to perform feature dimension regularization processing on the object feature and the product feature using the feature processing method according to the preset dimension;

[0088] The processing unit is specifically used to combine the processed object features and product features to determine the dominant features corresponding to the target object.

[0089] Optionally, in some possible implementations of the present application, the processing unit is specifically used to call the associated information corresponding to the target object;

[0090] The processing unit is specifically configured to determine the action of the target object on the product indicated in the association information to obtain behavior information;

[0091] The processing unit is specifically used to obtain a knowledge database corresponding to the associated information, so as to perform knowledge call based on the knowledge database to obtain the open domain information corresponding to the target object;

[0092] The processing unit is specifically configured to perform feature extraction based on the behavior information and the open domain information respectively to obtain the latent feature.

[0093] Optionally, in some possible implementations of the present application, the processing unit is specifically configured to determine a mapping relationship between the target object indicated in the explicit feature and each product, so as to restore the explicit feature to the first knowledge representation data based on the mapping relationship;

[0094] The processing unit is specifically configured to determine a mapping relationship between the target object indicated in the implicit feature and each product, so as to restore the implicit feature to second knowledge representation data based on the mapping relationship;

[0095] The processing unit is specifically used to perform semantic summarization based on the first knowledge representation data and the second knowledge representation data to obtain summary features;

[0096] The processing unit is specifically configured to determine the cross-domain information corresponding to the target object according to the summary feature;

[0097] The processing unit is specifically configured to recommend a target product to the target object based on the cross-domain information.

[0098] Optionally, in some possible implementations of the present application, the recommendation unit is specifically configured to obtain a candidate list corresponding to the target object, wherein the candidate list includes a plurality of candidate products;

[0099] The recommendation unit is specifically used to perform word segmentation processing on the cross-domain information, and perform vector representation on the cross-domain information after the word segmentation processing to obtain a cross-domain vector;

[0100] The recommendation unit is specifically configured to respectively calculate the similarity between the cross-domain vector and each of the candidate products to obtain a target score for each of the candidate products;

[0101] The recommendation unit is specifically configured to perform product screening based on the target scores of the candidate products to obtain the target product recommended to the target object.

[0102] Optionally, in some possible implementations of the present application, the recommendation unit is specifically used to respectively calculate the similarity between the cross-domain vector and each of the candidate products in a preset scoring dimension to obtain a semantic score, an object feature score, and a product feature score, wherein the preset scoring dimension includes a semantic dimension, an object feature dimension, and a product feature dimension;

[0103] The recommendation unit is specifically used to perform weighted calculation on the semantic score, the object feature score and the product feature score to obtain the target score.

[0104] Optionally, in some possible implementations of the present application, the recommendation unit is specifically used to determine a speech node in the conversation data;

[0105] The recommendation unit is specifically used to obtain the recommended speech template corresponding to the speech node;

[0106] The recommendation unit is specifically used to determine a recommendation speech template corresponding to a target product, and rewrite the content of the recommendation speech template to obtain a target recommendation speech;

[0107] The recommendation unit is specifically used to recommend the target product to the target object based on the target recommendation words.

[0108] Optionally, in some possible implementations of the present application, the recommendation unit is specifically used to collect feedback information of the target object on the target product;

[0109] The recommendation unit is specifically configured to adjust the feature extraction dimension corresponding to the target object based on the feedback information;

[0110] The recommendation unit is specifically configured to update the open domain information according to the knowledge information indicated by the feedback information.

[0111] According to a third aspect of an embodiment of the present application, there is provided a product recommendation device, including an input-output component and a processor;

[0112] The input and output components are used to obtain conversation data;

[0113] The processor is used to recommend the target product corresponding to the dialogue data obtained by the input-output component by executing the product recommendation method as described in the first aspect or any one of the first aspects.

[0114] According to a fourth aspect of an embodiment of the present application, there is provided a product recommendation system, including an interactive client and a server;

[0115] The interactive client is used to obtain the conversation data and send the conversation data to the server, and to display the product recommendation results output by the server;

[0116] The server is used to recommend the target product corresponding to the conversation data obtained by the interactive client by executing the product recommendation method described in the first aspect or any one of the first aspects.

[0117] According to a fifth aspect of an embodiment of the present application, a computer program product is provided, comprising: a computer program, which, when executed by a processor, implements the product recommendation method described in the first aspect or any one of the first aspects.

[0118] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0119] By obtaining the dialogue data corresponding to the target object when selecting a product, the dialogue data includes the object information and product information of the target object; then performing semantic recognition based on the dialogue data to determine the explicit features corresponding to the target object, which include object features and product features; and calling the associated information corresponding to the target object to extract features based on the associated information to obtain implicit features; and then fusing the explicit features and implicit features to obtain cross-domain information, so as to recommend the target product to the target object based on the cross-domain information. In this way, the product recommendation process based on explicit features and implicit features is realized, and the accuracy of dialogue understanding is improved by fully mining the dialogue data and performing multi-dimensional information fusion. In addition, information is supplemented by implicit features to further improve the comprehensiveness of object representation, thereby improving the accuracy of product recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0121] Figure 1 A diagram of the network architecture that runs the product recommendation system;

[0122] Figure 2A process architecture diagram of a product recommendation provided in an embodiment of the present application;

[0123] Figure 3 A flowchart of a product recommendation method provided in an embodiment of the present application;

[0124] Figure 4 A schematic diagram of a scenario of a product recommendation method provided in an embodiment of the present application;

[0125] Figure 5 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0126] Figure 6 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0127] Figure 7 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0128] Figure 8 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0129] Fig. 9 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0130] Fig.10 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0131] Fig.11 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0132] Fig.12 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0133] Fig.13 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0134] Fig.14 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0135] Fig.15 A schematic diagram of another scenario of a product recommendation method provided in an embodiment of the present application;

[0136] Fig.16 A schematic diagram of the structure of a product recommendation device provided in an embodiment of the present application;

[0137] Fig.17 A schematic diagram of the structure of a terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0138] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0139] It should be understood that the product recommendation method provided in the present application can be applied to a system or program including a product recommendation function in a terminal device, such as an online shopping application. Specifically, the product recommendation system can be run on a computer such as a computer. Figure 1 In the network architecture shown in Figure 1 As shown in the figure, it is a network architecture diagram of the product recommendation system. As can be seen from the figure, the product recommendation system can provide a product recommendation process with multiple information sources, that is, through the conversation data of the product on the terminal side, the explicit features and implicit features of the conversation data are processed and integrated on the server side, so as to make product recommendations; it can be understood that Figure 1 A variety of terminal devices are shown in FIG. 1 . The terminal devices may be computer devices. In actual scenarios, more or fewer types of terminal devices may participate in the product recommendation process. The specific number and type depend on the actual scenario and are not limited here. In addition, Figure 1 One server is shown in the figure, but in actual scenarios, multiple servers may be involved, especially in multi-disciplinary recommendation scenarios, and the specific number of servers depends on the actual scenario.

[0140] In this embodiment, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected through wired or wireless communication, and the terminal and the server can be connected to form a blockchain network, which is not limited in this application.

[0141] It is understandable that the above-mentioned product recommendation system can run on a personal mobile terminal, for example: as an application such as customer service management, it can also run on a server, and it can also run on a third-party device to provide product recommendations to obtain the product recommendation processing results of the information source; the specific product recommendation system can be run in the above-mentioned device in the form of a program, and it can also be run as a system component in the above-mentioned device, and it can also be run as a cloud service program. The specific operation mode depends on the actual scenario and is not limited here.

[0142] With the continuous development of the Internet and AI technology, companies have significantly improved their customer insights and personalized marketing technologies, especially in the field of outbound marketing. How to better utilize conversation information and implement personalized precision marketing has become a current research hotspot.

[0143] Generally, product recommendations in the outbound marketing process use fixed keyword rules to extract key information, and then make product recommendations based on the key information.

[0144] However, with the richness of products and the rapid increase in user scale, keyword rules may not be applicable to all recommendation scenarios, which may cause product recommendations to be inappropriate and affect the accuracy of product recommendations.

[0145] In order to solve the above problems, this application proposes a product recommendation method, which is applied to Figure 2 In the process framework of product recommendation shown in Figure 2 As shown, it is a process architecture diagram of a product recommendation provided in an embodiment of the present application, in which conversation data is generated through the interactive operation of the terminal based on the product, thereby triggering the server to perform data analysis on the conversation data to obtain explicit features; then, implicit features are determined in combination with external related information; and then, the fused features are analyzed to carry out the product recommendation process.

[0146] It is understandable that the product recommendation method provided by the present application can be written as a program to serve as a processing logic in a hardware system, or as a product recommendation device to implement the above processing logic in an integrated or external manner. As an implementation method, the product recommendation device obtains the corresponding dialogue data of the target object when selecting a product, and the dialogue data includes the object information and product information of the target object; then performs semantic recognition based on the dialogue data to determine the explicit features corresponding to the target object, and the explicit features include object features and product features; and calls the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features, and the associated information includes the behavior information and open domain information corresponding to the target object; and then fuses the explicit features and implicit features to obtain cross-domain information, so as to recommend the target product to the target object based on the cross-domain information. Thereby, a product recommendation process based on explicit features and implicit features is realized, and the accuracy of dialogue understanding is improved by fully mining the dialogue data and performing multi-dimensional information fusion, and the information is supplemented by implicit features to further improve the comprehensiveness of object representation, thereby improving the accuracy of product recommendation.

[0147] Combined with the above process architecture, the following will introduce the product recommendation method in this application. Please refer to Figure 3 , Figure 3 A flowchart of a product recommendation method provided in an embodiment of the present application, the embodiment of the present application at least includes the following steps:

[0148] 301. Acquire conversation data of a target object when selecting a product, where the conversation data includes object information and product information corresponding to the target object.

[0149] In this embodiment, the target object is the subject that interacts with the product, and its subject form can be a user, a terminal or other indicative identifier. Here, the user is used as an example for explanation. Correspondingly, the product can be a virtual product, such as a financial product, a software product, etc.; the product can also be a physical product, such as food, clothing, etc. Here, the financial product is used as an example for explanation. The specific object form and product form can be combined according to the actual scenario and are not limited here.

[0150] It is understandable that in the process of product selection, users may have a conversation with customer service. The customer service here can be a manual customer service or a virtual assistant. Based on the process of this conversation, conversation data can be generated, which shows the user's demand for the product, product selection and the user's own preferences, so that object information (also called user information) and product information can be obtained from the conversation data.

[0151] Specifically, the product recommendation process of this embodiment is based on a large language model (LLM), hereinafter referred to as a large model, which is a natural language processing model based on deep learning, designed to understand and generate human language. These models are trained on large amounts of text data and can perform a wide range of natural language processing tasks, such as text classification, question answering, dialogue systems, and machine translation.

[0152] In one possible scenario, the large model in this embodiment adopts Figure 4 The model architecture shown, Figure 4 A scenario diagram of a product recommendation method provided in an embodiment of the present application; that is, the large model of the present application starts from the conversation data collection and analysis and personalized marketing part to realize the product recommendation (also called marketing) process based on cross-domain information fusion of conversations. Figure 4 In the figure, all the large models are large models that can be applied in different scenarios. For different application scenarios, it includes a dialogue management and analysis module and a personalized marketing module. The dialogue management and analysis module is mainly composed of a large model and an external knowledge source. The large model has many parameters and strong capabilities, and is suitable for multi-scenario tasks. This module mainly analyzes user intentions in multiple dimensions of user dialogue content, extracts user portraits and product information, and obtains user emotions and potential points of interest.

[0153] In addition, the personalized marketing module mainly includes a feature understanding module, an interest interaction and recommendation module, and a sorting and scoring module. The feature understanding module is mainly used to decompose the results of the dialogue management and analysis module to gain a deeper understanding of user representation and product representation. These features are equivalent to the "explicit" features that appear during the dialogue process. In addition, this module will also introduce additional data sources, which are equivalent to "invisible features", such as customers' historical purchases of products and historical search information; the interest interaction and recommendation module mainly uses the common semantic information and rich recommendation features of the big model to obtain a product list; the sorting and scoring module mainly scores and sorts the recommended products, making full use of the prompt capability of the big model to realize the recommendation function.

[0154] Combined with the above-mentioned module configuration based on the big model, the process of executing product recommendation includes: first, analyzing the user's dialogue query, extracting the user portrait tags and product names that appear in the dialogue process, and fully understanding the user's intention and emotional perception. Secondly, the above-extracted user features and product features are integrated with external source features such as user historical purchase features or search records, and the capabilities of the big model recommendation and recommendation system are used to form a product recommendation list. Next, feature scores are formed for the candidate list and users, including semantic features, user feature co-occurrence scores, and product feature co-occurrence scores, and each product is scored using a weighted average method. Then, the best product is recommended based on the obtained product list and the corresponding score, and a personalized marketing speech plan is generated using prompt, and user feedback is collected. Furthermore, the correct product and user information recommended in the above process is updated to the user feature library and product feature library to form a product marketing plan based on cross-domain fusion of dialogue.

[0155] It is understandable that the existing product recommendation process cannot cover newly added products when sorting out the rules between keywords and products, and has timeliness and cold start problems; and because of the use of small models, there is a problem with the number of parameters, and it cannot understand the deep semantics and recommendation intentions. For recommendation tasks, common "explicit" features include: products mentioned by customers during the conversation, and "invisible" features include user historical interaction information, customer group characteristics, etc. Existing solutions do not fully utilize global information.

[0156] Therefore, compared with the existing solutions, this embodiment has cold start capability through the configuration process of the above-mentioned large model, and has a certain baseline effect in multiple scenarios, with better cold start effect. Experiments show that the baseline effect of large model recommendation is at least 10% higher than the baseline of traditional recommendation system, and can reach 20% or even higher in scenarios such as factor extraction. It also has great advantages in speech generation and data construction, reducing customization costs such as data labeling and model optimization.

[0157] Next, combine Figure 4 The model architecture shown in the figure explains the execution process of each module.

[0158] 302. Perform semantic recognition based on the conversation data to determine explicit features corresponding to the target object, where the explicit features include object features and product features.

[0159] In this embodiment, explicit features are used to directly display the user's intention or product demand, so features can be extracted from object features (also called user features) and product features respectively.

[0160] Specifically, since the conversation data is the user's request (query) during the conversation process, the semantic recognition process for the conversation data first needs to understand the customer's intention and speech nodes, that is, whether the current customer's query is consultation, suggestion, rejection or other, and then the user portrait label and product information that appear in the conversation process, and then the customer's emotions. Figure 5 In the scenario shown, Figure 5 A scenario diagram of another product recommendation method provided in an embodiment of the present application; the figure shows the marketing dialogue data in a financial management scenario, and then the implementation process of each part of the above-mentioned semantic recognition is gradually disassembled based on the example scenario.

[0161] Based on the above example description, semantic recognition can include semantic representation, that is, understanding the content of the conversation, summarizing the key information of the sentences involved in the conversation process, and aggregating them to form a semantic representation of the conversation, as well as the process of intent recognition, element extraction and emotion recognition, that is, firstly, semantic recognition is performed based on the conversation data, and the key information of the sentences involved in the conversation process is summarized to obtain semantic representation information; the speech nodes in the conversation data are determined, and the corresponding intent information is determined according to the speech nodes; then the key elements of the conversation data are extracted according to the configured key information categories to obtain object portrait information and product description information; and emotion recognition is performed on the conversation data to obtain emotion analysis information; then object features are generated based on the semantic representation information, intent information, object portrait information and emotion analysis information, and product features are generated based on the product description information; thereby combining the object features and product features to determine the explicit features corresponding to the target object.

[0162] It should be noted that compared to the existing feature splitting process, such as Figure 6 As shown, Figure 6A scenario diagram of another product recommendation method provided for an embodiment of the present application; In the existing dialogue analysis scheme, each model basically implements a scenario, and then uses a task flow orchestrator to string together the capabilities of each small model, and each model is responsible for its own task, that is, the current small model can only process 1 to 2 natural language processing tasks alone due to its own parameter problems, and it is impossible to realize multiple capabilities of a model, so when doing scenario tasks, the cold start effect is poor, and additional data optimization is required. For large models, these tasks can be unified into a set of standard paradigms, that is, unified into "text to text" tasks, and these information can be used at the same time, which is faster, and the information obtained by a single model is richer, which is equivalent to merging multiple tasks into one model for processing, thereby reducing the number of underlying models. For example, in a traditional NLP model, an NLP model is generally responsible for one task. If contrastive learning is used, it can be responsible for two tasks, but a large model can be responsible for multiple tasks, realizing a large model to realize multiple NLP tasks, and can make full use of the information feedback from each task to realize information fusion.

[0163] Specifically, the process of intent recognition can be based on the speech node graph. First, the scene information corresponding to the conversation data is obtained; then the speech node graph associated with the scene information is called; then the conversation data is detected based on the speech node graph to determine the speech nodes in the conversation data; and then the intent information corresponding to the speech nodes is determined. This is because the conversation process in the marketing field will have corresponding speech nodes and intents. For example, Figure 7 In the scenario shown, Figure 7 A scenario diagram of another product recommendation method provided in an embodiment of the present application; the figure shows an example of a speech node in a certain scenario of a bank, which includes a speech node diagram for an "activity promotion type" of a certain bank. For example, for the speech node "whether it is a wrong number", the intent recognition is performed as "wrong number / not answered by myself", or for the speech node of the opening remarks, the intent recognition is performed to obtain the user's attitude towards the product; therefore, each user query and historical conversation can be sent to the big model to analyze the specific speech nodes and related intentions.

[0164] In one possible scenario, for a financial management scenario, Figure 8 As shown, Figure 8 A scenario diagram of another product recommendation method provided in an embodiment of the present application; the figure shows relevant speech nodes and intent recognition examples when a user inquires about financial products, that is, for the speech node "financial product purchase", the corresponding intent recognition is "confirm", that is, the user is interested in the financial product, thereby guiding subsequent product recommendations.

[0165] In addition, for the process of extracting key elements, since some key element labels will appear during the conversation, including user portrait category and product name category, these key information are generally explicit features, and making full use of these features can achieve better marketing goals. Therefore, you can first extract the conversation data according to the configured key information categories (such as user portrait category, product name category) to obtain corpus extraction elements; then logically associate the corpus extraction elements to obtain condensed information; and extract key elements based on the condensed information to obtain object portrait information and product description information. Among them, for the process of logically associating the corpus extraction elements, that is, using the preset logical process to extract elements, for example, user A has a positive attitude towards financial product C of product type B (condensed information), such as Fig. 9 As shown, Fig. 9 A scenario diagram of another product recommendation method provided for an embodiment of the present application; the figure shows a sample of key element extraction, which reflects the conversation-level element extraction capability of the large model, that is, compared with the traditional element extraction model, which can only extract based on the elements appearing in the text and cannot perform in-depth understanding, such as "I hope to have stable returns, but I don't want to take too much risk", the traditional element extraction model such as BERT_CRF can only extract its key information but cannot perform logical normalization or summary on it. The large model can further condense such elements, that is, based on the semantic recognition process after logical association, thereby improving the accuracy of object portrait information and product description information.

[0166] In addition, regarding the process of emotion recognition, it is very important to be able to sense the emotional changes of customers and change marketing strategies in a timely manner during the marketing process. That is, the more positive the customer's emotions are, the greater the probability of marketing success. Therefore, we can first determine the feedback information in the conversation data; then identify the emotional keywords in each corpus in the feedback information to obtain an emotional keyword set; and perform emotion recognition based on the identified emotional keyword set to obtain emotional analysis information. For example, in Fig.10 In the described scenario, Fig.10 A scenario diagram of another product recommendation method provided for an embodiment of the present application; the figure shows the process of the large model analyzing the user query sentiment, that is, by identifying the emotional keywords "OK, I think I'm more interested" and "Great, thank you for your help", the "positive" emotional label, that is, the emotional analysis information, can be obtained.

[0167] It is understandable that the big model's emotion perception ability is better than the traditional model, and it has cold start capability. In addition, the big model has a certain cold start effect for multiple tasks such as context inheritance and reference resolution. Therefore, the application of the big model in the dialogue management and analysis module is very effective.

[0168] In a possible scenario, although the above feature information is explicit, the data formats are often different and cannot be directly spliced ​​or processed by a network. In this case, different networks can be used to process each explicit feature separately, and the information domains can be aligned before being put into the network for processing. That is, first determine the feature type corresponding to the object feature and the product feature; then determine the corresponding feature processing method based on the feature type; and use the feature processing method to regularize the feature dimension of the object feature and the product feature according to the preset dimension; and then combine the processed object feature and product feature to determine the explicit feature corresponding to the target object. For example, multi-layer perceptron or BERT can be used to process discrete features, and convolutional neural network can be used to process image features and speech features. For large language models, the above features can all be expressed as the same dimensional features, and the large model encoding part can be directly used to complete feature regularization to obtain features in the conversation process.

[0169] It can be seen that this embodiment unifies multiple dimensions such as intent understanding, key element extraction, sentiment analysis, and context understanding in conversation analysis into a large model, and automatically has historical context information. Compared with traditional solutions, one model realizes multiple tasks, has lower complexity, and obtains richer information.

[0170] 303. Call the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features.

[0171] In this embodiment, the implicit feature is a supplementary description of the object using external information source data, such as user behavior information, open domain information, etc., correspondingly, the description association information includes the behavior information and open domain information corresponding to the target object; for the extraction process of the implicit feature, the association information corresponding to the target object can be first called; then the action of the target object on the product indicated in the association information can be determined to obtain the behavior information; and the knowledge database corresponding to the association information is obtained to perform knowledge call based on the knowledge database to obtain the open domain information corresponding to the target object; and then feature extraction is performed based on the behavior information and the open domain information respectively to obtain the implicit feature. For example, the user's behavior information includes the user's purchase behavior, collection behavior and other behaviors for product operations, while the open domain information refers to the extraction of useful information from texts (knowledge databases) from various sources without specific domain restrictions, and its source can include rule databases such as user feature databases or product feature databases based on big data collection.

[0172] Specifically, explicit features are user features and product features that appear during the conversation. These features are directly presented and can help understand the user's intentions or needs. For example, in the above conversation, it is mentioned that the customer is a "risk-averse customer" and he is more interested in "financial product A". Implicit features are presented as potential data relative to the conversation process. For example, the user has browsed financial product A on the APP, and "risk-averse" customers prefer stable financial products.

[0173] 304. The explicit features and the implicit features are integrated to obtain cross-domain information, so as to recommend a target product to a target object based on the cross-domain information.

[0174] In this embodiment, the fusion of cross-domain information is to fuse the above-mentioned explicit features and implicit features of the conversation to form cross-domain information, which facilitates accurate recommendation.

[0175] Specifically, the process of fusing explicit features and implicit features to obtain cross-domain information can be combined with Figure 4 The module configuration of the large model shown is explained, that is, the process of forming cross-domain information recommendation by utilizing the feature information (explicit features) and external source information (implicit features) in the dialogue process is mainly based on natural language processing technology (NLP), large language model technology (LLM), and recommendation model (CTR); that is, condensing key information from the dialogue process and combining external data sources to complete product recommendations. Among them, the combination of natural language processing technology and large language model technology can obtain contextual semantic information, and the recommendation model technology mainly solves how users can obtain information of interest in the case of complex information and recommend suitable products. Therefore, the large model of this embodiment can make full use of the semantic information, user ID features, product features and external source knowledge in the dialogue domain, and complete knowledge alignment and fusion of different domain information through the encoding network, so that learning is more sufficient, and the knowledge content supports implementation updates, making full use of the semantic understanding ability, logical reasoning ability, and generation ability of the large model, so that it can be used in multiple scenarios after one training. Relevant experiments have shown that the cold start effect is better when recommending, which is at least 10% higher, without the need to annotate a large amount of data, and its customization cost is lower.

[0176] It can be understood that the cross-domain information recommendation process of this embodiment is as follows Fig.11 As shown, Fig.11 A scenario diagram of another product recommendation method provided for an embodiment of the present application; the figure shows a comparison of the differences between a traditional recommendation system and a cross-domain information recommendation system. Compared with the traditional recommendation system, the cross-domain information recommendation system has richer semantic information, including explicit and implicit features. In addition, the knowledge can be updated in real time according to the update frequency of the external information source, and has a certain potential reasoning ability.

[0177] In a possible scenario, for the process of cross-domain information fusion, such as Fig.12 As shown, Fig.12 A scenario diagram of another product recommendation method provided in the embodiment of the present application; the figure shows a comparison diagram of feature fusion methods. For traditional recommendation learning application features, for users (USER) and products (item), all features of both need to be ID-encoded. These features generally include attribute features, behavioral features, interaction features, etc. For example, product attribute features IDp include product type, operation mode, interest rate, etc., and user behavioral features IDu include browsing records, operation records, etc. These ID-encoded features generally do not have semantic information and cannot be understood by the language model. They are often stored in the form of databases or tables, and the data used for recommendation is often before the current moment. In the cross-domain information fusion process of this embodiment, explicit features such as user tags, product tags, user intentions, and emotions are extracted in the dialogue analysis module, and these features need to be processed in a variety of ways. The first is dialogue content understanding. The key information of all current sentences in the dialogue process is summarized and summarized to form a dialogue semantic representation S. Each sentence of the user contains the intention Y. The intention feature can help the agent find the speech node, thereby assisting in the guidance and control of the dialogue information. User-related portrait features F will appear during the dialogue process u , the product-related feature F appears p . User’s emotional characteristics F s It is an auxiliary feature and can be used to guide the recognition of the dialogue intention. The above-mentioned explicit feature is the comprehensive feature A. ALL Its "explicitness" is reflected in the fact that it is explicitly mentioned in the conversation process and can be perceived by the outside world. In addition, considering that the essence of recommendation learning is to fit historical user behavior data and predict future user behavior, the general steps are to learn features in historical data and then predict the items that users like. The explicit features in the conversation process introduced above, the implicit features will be introduced next. Implicit features mainly include two parts, one is the traditional recommendation learning features, and the other is the external source features. Since the external source refers to some open information independent of the model, this information contains the real-time updated knowledge base information K now , Database Information K other , and even some external plug-ins. This information is very rich, but the representations are different and need to be organized before being connected to the recommendation system.

[0178] Therefore, for explicit features and implicit features, since these features belong to information in different domains, feature regularization is required, such as Fig.13 As shown, Fig.13A scenario schematic diagram of another product recommendation method provided for an embodiment of the present application shows a flow chart of cross-domain feature fusion, namely, first determining the mapping relationship between the target object indicated in the explicit feature and each product, so as to restore the explicit feature to the first knowledge representation data based on the mapping relationship, wherein the first knowledge representation data includes a user semantic representation and a product semantic representation, the user semantic representation is encoded / decoded based on the mapping between the user ID-type feature and the preference tag, and the product semantic representation is encoded / decoded based on the mapping between the product ID-type feature and the product description tag; in addition, the mapping relationship between the target object indicated in the implicit feature and each product (i.e., the preference settings or product descriptions contained in the knowledge base associated with the exogenous data) can also be determined, so as to restore the implicit feature to the second knowledge representation data based on the mapping relationship; then, semantic summarization is performed based on the first knowledge representation data and the second knowledge representation data to obtain the summary feature; and the cross-domain information corresponding to the target object is determined based on the summary feature; and then the target product is recommended to the target object based on the cross-domain information.

[0179] Specifically, for the process of user semantic representation and product semantic representation, the behavioral feature ID of the explicit feature user u and product attribute feature ID p According to the original mapping table, the original knowledge representation is restored, and then the content decoder is used to semantically summarize the features. The relevant expressions are as follows:

[0180] Content = LLM(f -1 (ID u ),f -1 (ID p ))

[0181] Furthermore, for the processing of implicit features, the feature K of the knowledge base now , other external information source characteristics K other And the comprehensive features in the dialogue analysis process are integrated. Because after the previous step of processing, both implicit features and explicit features in the dialogue process have been converted into representations containing semantic information, and then combined with the prompt instructions generated by the explicit features, they are input into the large language model to obtain a list of recommended products for the large model. Through this step, the purpose of information complementation can be achieved. For example, there is no risk tolerance feature of Zhang San in the original user feature library, but this feature appears in the dialogue process, and the two complement each other. For the process of semantically summarizing the first knowledge representation data and the second knowledge representation data to obtain the summary feature, the following expression can be used:

[0182] Reclist_llm=LLM(prompt,Content,K now ,K other )

[0183] Among them, prompt represents the prompt word generated based on the explicit features, content represents the semantic summary based on the explicit features, and K now Represents the characteristics of the knowledge base, K other Indicates other external information source characteristics.

[0184] Furthermore, for the process of recommending a target product to a target object based on cross-domain information, the candidate list corresponding to the target object can be first obtained, and the candidate list contains multiple candidate products; then the cross-domain information is segmented, and the cross-domain information after the segmentation is represented by a vector to obtain a cross-domain vector; and the similarity between the cross-domain vector and each candidate product is calculated to obtain the target score of each candidate product; and then the product is screened based on the target score of each candidate product to obtain the target product recommended to the target object. Among them, the target score can include the score using the recommendation system model CRM and the score using the cross-domain information; because the large model can make full use of the semantic feature domain to complete the recommendation of related products, for the ID domain features, the recommendation system model CRM is used to complete the product recommendation, and the relevant expressions are as follows:

[0185] Reclist_crm = CRM (ID u ,ID p )

[0186] By combining the two recommendation lists above, we can obtain a product recommendation list AL = {A1, A2, A3, A4…An}, which is expressed as follows:

[0187] Reclist=Reclist_llm+Reclist_crm

[0188] It can be seen that starting from the two dimensions of "implicit" features and "explicit" features, we first use the explicit features mined by the big model during the conversation process, such as user age, risk tolerance, and other features that appear during the conversation process, and then use features such as user ID, product ID, and external source features to regularize the features, and use the above multiple features to implement cross-domain information fusion technology to obtain more semantic information. The explicit features come from the conversation, and the implicit features come from two aspects. One is the user behavior feature data of the traditional recommendation system, which is mostly ID-based and does not have semantic information. It needs to be converted into a representation with semantic information using a big model to achieve information domain alignment. The other is external source knowledge, which converts open domain knowledge into semantic information and then integrates it into the big model. Compared with the traditional recommendation system, the information in this case is richer, including conversation semantic features, user ID features, product ID features, external source knowledge, etc. Using cross-domain fusion technology, the knowledge of different domains is aligned, making knowledge learning more sufficient, and the real-time update frequency of knowledge can be maintained, thereby achieving the purpose of precision marketing.

[0189] In one possible scenario, the product recommendation list is scored according to a certain priority order and the final recommendation list is output. That is, the candidate items are scored one by one using a large model, and finally the final ranked list is obtained by sorting according to the scores. The score is mainly composed of the weighted semantic score, user feature score, and product feature score. That is, the similarity between the cross-domain vector and each candidate product in the preset scoring dimensions is calculated separately to obtain the semantic score, object feature score, and product feature score. The preset scoring dimensions include the semantic dimension, object feature dimension, and product feature dimension; then the semantic score, object feature score, and product feature score are weighted to obtain the target score. The specific calculation steps for the scores of different parts are as follows: Fig.14 As shown, Fig.14 A scenario diagram of another product recommendation method provided in an embodiment of the present application; the recommendation scoring process shown in the figure includes the processes of semantic scoring, object feature scoring and product feature scoring.

[0190] Specifically, for the semantic score, its semantic feature is the semantic score between the user query and the product list. For example, during the conversation, the customer said "I hope to have a stable income, but I don't want to take too much risk", then the score between it and the recommended list is:

[0191] P sents = cos(A L ,A sents )

[0192] Among them, A L represents the candidate products in the candidate list, A sents Indicates user needs.

[0193] For user feature scores, user features refer to the scores between the products previously purchased by users with similar features and the current product list. For example, if the customer is a "risk-averse customer" and the similar customer B is also a "risk-averse customer", and the list of products previously purchased by the customer is BL{B1, B2, B3…Bn}, then the score between the customer and the product recommendation list is:

[0194] P user = cos(A L ,B L )

[0195] Among them, A L represents the candidate products in the candidate list, B L Represents a set of users with similar user characteristics.

[0196] For product feature score: The product feature score refers to the score between the product list Ap mentioned in the user conversation and the current recommended product list AL. For example, product A is mentioned in the conversation, and the agent recommends products A1 and A2. The product list mentioned in the conversation is Ap = {A, A1, A2}, then the score between it and the product recommendation list is:

[0197] P products = cos(A L ,A P )

[0198] Among them, A L represents the candidate products in the candidate list, A p Represents a list of products mentioned by a user conversation.

[0199] After obtaining the scores of the above dimensions, the ranking scores can be performed, that is, the above scores are weighted to obtain the final semantic score. The calculation formula is as follows:

[0200] P total =αP sents +βP user +(1-α-β)P products

[0201] Among them, α and β are adjustment coefficients. Generally, α is set to 0.3 and β is set to 0.7. They can also be obtained after training. After calculating the scores of all products in the recommended product list, they are sorted first, and then they can be filtered again by threshold parameters or by selecting the highest n products (TOPn).

[0202] Furthermore, in the process of product recommendation, you can also use recommendation words for joint recommendation, that is, Fig.14The speech configuration process in is the marketing speech generation process. Since different speech nodes correspond to different templates, in the process of forming marketing speech, it is necessary to integrate the product list obtained in the previous step to form a new prompt instruction, so that the big model can perform polishing and rewriting operations to form the final marketing speech, that is, first determine the speech node in the dialogue data; then obtain the recommended speech template corresponding to the speech node; and determine the recommended speech template corresponding to the target product, and perform rewriting operations on the recommended speech template content to obtain the target recommended speech, where the rewriting operation is used to improve the readability of the recommended speech; and then recommend the target product to the target object based on the target recommended speech.

[0203] Furthermore, after completing the marketing product recommendation, the user's feedback on the recommendation results will be collected to achieve a closed-loop analysis of the overall marketing process. That is, first collect the target object's feedback on the target product; then adjust the feature extraction dimension corresponding to the target object based on the feedback information; and update the open domain information based on the knowledge information indicated by the feedback information. For example, if the customer's portrait during this call is relatively complete and the marketing results are relatively clear, this information will also be added to the corresponding user feature library.

[0204] The above embodiment is described below in conjunction with an outbound marketing scenario. Fig.15 As shown, Fig.15 A scenario diagram of another product recommendation method provided in the embodiment of the present application; product recommendation in the outbound marketing scenario includes five processes, namely, conversation information analysis, cross-domain information recommendation, recommendation result scoring and ranking, marketing speech generation, marketing result evaluation and information return. Among them, the input data is the query in the user conversation information and the external information source data; correspondingly, the output data is the corresponding product and the corresponding marketing speech recommended according to the user query, the user's feedback and characteristics.

[0205] Step 1: Conversation information analysis. For the input user query, combined with semantic information, first identify the intent of the conversation, extract the word slot information in the query, which generally includes user portraits (such as age, gender, income, etc.) and product names (xx financial products, fund products), and output the features. Because the big model itself has multi-task capabilities, it can fully understand the contextual semantic information and customer emotional perception.

[0206] Step 2: Cross-domain information recommendation. The product features or user features extracted in the previous step are sent to the explicit feature encoder for encoding, and combined with the implicit features in the external information source data (open domain information or user behavior features), they are sent to the collaborative recommendation system to form a product recommendation list. For example, during the conversation, the customer clearly mentioned the features of "A financial product" and "risk-averse customers" (risk-averse customers refer to those who tend to avoid risks when facing investment decisions and are more willing to choose investment options with lower expected returns but relatively safe). By querying the external source, it is found that A financial product belongs to "growth-oriented financial product". This product is not suitable for risk-averse customers. At this time, a financial product in the product library that is close to A financial product but more stable should be recommended.

[0207] Step 3: Score and sort the recommendation results. This step mainly uses the large model to score the candidate items one by one, and finally sorts them according to the scores to get the final sorting category. The score is obtained by weighting the semantic features, user feature co-occurrence scores, and product feature co-occurrence scores. The semantic feature is the semantic score between the user query and the product recommendation list. The user feature score refers to the score between the previous product purchase preferences of users with similar customer profiles and the current product list. The product feature score refers to the similarity score between the product name mentioned in the conversation and the current product list.

[0208] The fourth step is to generate marketing scripts. This step mainly combines the product recommendation list in the previous step with the prompt instruction to let the big model generate the golden scripts for the marketing scenario, so that customers can feel the personification and multi-source, and avoid the cold feeling brought by template scripts. For example, in the above recommendation process, it is learned that the customer is a risk-averse customer, and it is mentioned that product A is a "growth-oriented financial product", but it is not suitable to recommend to customers. At this time, it is obtained that products A1 and A2 are more suitable for customers. The above information is integrated into the prompt instruction to let the big model generate the corresponding marketing script: "Dear customer, based on your preference for stable and safe investment, I especially recommend products A1 and A2. These two products focus on protecting the principal and ensuring stable returns. They are ideal choices for your financial planning, allowing you to enjoy investment returns while minimizing risks."

[0209] The fifth step is marketing result evaluation and information reflux. There is a feedback module in each round of interaction. When the user forms effective feedback, such as approval of the recommended A1 and A2 products, or direct purchase records, the user's portrait information and the product will be mapped and updated to the user and product database to form data reflux.

[0210] In combination with the above embodiments, it can be known that by obtaining the dialogue data corresponding to the target object when selecting a product, the dialogue data includes the object information and product information of the target object; then semantic recognition is performed based on the dialogue data to determine the explicit features corresponding to the target object, the explicit features include object features and product features; and the associated information corresponding to the target object is called to extract features according to the associated information to obtain implicit features, the associated information includes the behavior information and open domain information corresponding to the target object; then the explicit features and the implicit features are fused to obtain cross-domain information, so as to recommend the target product to the target object based on the cross-domain information. Thus, the product recommendation process based on explicit features and implicit features is realized, and the accuracy of dialogue understanding is improved by fully mining the dialogue data and performing multi-dimensional information fusion, and the information is supplemented by implicit features to further improve the comprehensiveness of object representation, thereby improving the accuracy of product recommendation.

[0211] In order to better implement the above solution of the embodiment of the present application, the following also provides related devices for implementing the above solution. Fig.16 , Fig.16 This is a schematic diagram of the structure of a product recommendation device provided in an embodiment of the present application. The product recommendation device 1600 includes:

[0212] An acquisition unit 1601 is used to acquire the dialog data of the target object when selecting a product, where the dialog data includes the object information and product information corresponding to the target object;

[0213] The processing unit 1602 is used to perform semantic recognition based on the conversation data to determine the explicit features corresponding to the target object, where the explicit features include object features and product features;

[0214] The processing unit 1602 is further configured to call the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features, wherein the associated information includes behavior information and open domain information corresponding to the target object;

[0215] The recommendation unit 1603 is used to fuse the explicit feature and the implicit feature to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information.

[0216] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to perform semantic recognition based on the conversation data, and summarize key information of the sentences involved in the conversation process to obtain semantic representation information;

[0217] The processing unit 1602 is specifically configured to determine a speech node in the conversation data, and determine corresponding intention information according to the speech node;

[0218] The processing unit 1602 is specifically used to extract key elements from the conversation data according to the configured key information category to obtain object portrait information and product description information;

[0219] The processing unit 1602 is specifically configured to perform emotion recognition on the conversation data to obtain emotion analysis information;

[0220] The processing unit 1602 is specifically configured to generate the object feature based on the semantic representation information, the intention information, the object portrait information and the sentiment analysis information, and to generate the product feature based on the product description information;

[0221] The processing unit 1602 is specifically configured to combine the object feature and the product feature to determine the explicit feature corresponding to the target object.

[0222] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to obtain scene information corresponding to the conversation data;

[0223] The processing unit 1602 is specifically used to call the speech node graph associated with the scene information;

[0224] The processing unit 1602 is specifically configured to detect the conversation data based on the speech node graph to determine the speech nodes in the conversation data;

[0225] The processing unit 1602 is specifically used to determine the intention information corresponding to the speech node.

[0226] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to perform corpus extraction on the conversation data according to the configured key information category to obtain corpus extraction elements;

[0227] The processing unit 1602 is specifically used to logically associate the corpus extraction elements to obtain condensed information;

[0228] The processing unit 1602 is specifically used to extract key elements based on the condensed information to obtain the object portrait information and the product description information.

[0229] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to determine feedback information in the conversation data;

[0230] The processing unit 1602 is specifically used to identify the emotion keywords in each corpus in the feedback information to obtain an emotion keyword set;

[0231] The processing unit 1602 is specifically configured to perform emotion recognition based on the identified emotion keyword set to obtain the emotion analysis information.

[0232] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to determine a feature type corresponding to the object feature and the product feature;

[0233] The processing unit 1602 is specifically configured to determine a corresponding feature processing method based on the feature type;

[0234] The processing unit 1602 is specifically configured to perform feature dimension regularization processing on the object feature and the product feature using the feature processing method according to the preset dimension;

[0235] The processing unit 1602 is specifically configured to combine the processed object features and product features to determine the explicit features corresponding to the target object.

[0236] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to call the associated information corresponding to the target object;

[0237] The processing unit 1602 is specifically configured to determine the action of the target object on the product indicated in the association information to obtain behavior information;

[0238] The processing unit 1602 is specifically configured to obtain a knowledge database corresponding to the associated information, and perform knowledge call based on the knowledge database to obtain open domain information corresponding to the target object;

[0239] The processing unit 1602 is specifically configured to perform feature extraction based on the behavior information and the open domain information respectively to obtain the implicit feature.

[0240] Optionally, in some possible implementations of the present application, the processing unit 1602 is specifically configured to determine a mapping relationship between the target object indicated in the explicit feature and each product, so as to restore the explicit feature to the first knowledge representation data based on the mapping relationship;

[0241] The processing unit 1602 is specifically configured to determine a mapping relationship between the target object indicated in the implicit feature and each product, so as to restore the implicit feature to second knowledge representation data based on the mapping relationship;

[0242] The processing unit 1602 is specifically configured to perform semantic summarization based on the first knowledge representation data and the second knowledge representation data to obtain a summary feature;

[0243] The processing unit 1602 is specifically configured to determine the cross-domain information corresponding to the target object according to the summary feature;

[0244] The processing unit 1602 is specifically configured to recommend a target product to the target object based on the cross-domain information.

[0245] Optionally, in some possible implementations of the present application, the recommendation unit 1603 is specifically configured to obtain a candidate list corresponding to the target object, the candidate list including multiple candidate products;

[0246] The recommendation unit 1603 is specifically used to perform word segmentation processing on the cross-domain information, and perform vector representation on the cross-domain information after the word segmentation processing to obtain a cross-domain vector;

[0247] The recommendation unit 1603 is specifically used to calculate the similarity between the cross-domain vector and each of the candidate products to obtain a target score for each of the candidate products;

[0248] The recommendation unit 1603 is specifically configured to perform product screening based on the target score of each candidate product to obtain the target product recommended to the target object.

[0249] Optionally, in some possible implementations of the present application, the recommendation unit 1603 is specifically used to respectively calculate the similarity between the cross-domain vector and each of the candidate products in a preset scoring dimension to obtain a semantic score, an object feature score, and a product feature score, wherein the preset scoring dimension includes a semantic dimension, an object feature dimension, and a product feature dimension;

[0250] The recommendation unit 1603 is specifically configured to perform weighted calculation on the semantic score, the object feature score and the product feature score to obtain the target score.

[0251] Optionally, in some possible implementations of the present application, the recommendation unit 1603 is specifically used to determine a speech node in the conversation data;

[0252] The recommendation unit 1603 is specifically used to obtain a recommended speech template corresponding to the speech node;

[0253] The recommendation unit 1603 is specifically used to determine the recommendation speech template corresponding to the target product, and rewrite the content of the recommendation speech template to obtain the target recommendation speech;

[0254] The recommendation unit 1603 is specifically configured to recommend the target product to the target object based on the target recommendation words.

[0255] Optionally, in some possible implementations of the present application, the recommendation unit 1603 is specifically used to collect feedback information of the target object on the target product;

[0256] The recommendation unit 1603 is specifically configured to adjust the feature extraction dimension corresponding to the target object based on the feedback information;

[0257] The recommendation unit 1603 is specifically configured to update the open domain information according to the knowledge information indicated by the feedback information.

[0258] By obtaining the dialogue data corresponding to the target object when selecting a product, the dialogue data includes the object information and product information of the target object; then performing semantic recognition based on the dialogue data to determine the explicit features corresponding to the target object, which include object features and product features; and calling the associated information corresponding to the target object to extract features based on the associated information to obtain implicit features, which include the behavior information and open domain information corresponding to the target object; and then fusing the explicit features and implicit features to obtain cross-domain information, so as to recommend the target product to the target object based on the cross-domain information. In this way, the product recommendation process based on explicit and implicit features is realized, and the accuracy of dialogue understanding is improved by fully mining the dialogue data and performing multi-dimensional information fusion, and the information is supplemented by implicit features to further improve the comprehensiveness of object representation, thereby improving the accuracy of product recommendation.

[0259] The product recommendation device provided in this embodiment belongs to the same application concept as the method provided in the above embodiments of this application, and can execute the product recommendation method provided in any of the above embodiments of this application, and has the corresponding functional modules and beneficial effects of executing the product recommendation method. For technical details not fully described in this embodiment, please refer to the specific processing content of the product recommendation method provided in the above embodiments of this application, and will not be repeated here.

[0260] The functions implemented by the above acquisition unit 1601, processing unit 1602 and recommendation unit 1603 may be implemented by the same or different processors respectively, which is not limited in the embodiment of the present application.

[0261] It should be understood that each unit in the above device can be implemented in the form of a processor calling software. For example, the device includes a processor, the processor is connected to a memory, and instructions are stored in the memory. The processor calls the instructions stored in the memory to implement any of the above methods or realize the functions of each unit of the device, wherein the processor can be a general-purpose processor, such as a CPU or a microprocessor, etc., and the memory can be a memory in the device or a memory outside the device. Alternatively, the unit in the device can be implemented in the form of a hardware circuit, and the functions of some or all units can be realized by designing the hardware circuit. The hardware circuit can be understood as one or more processors; for example, in one implementation, the hardware circuit is an ASIC, and the functions of some or all units above are realized by designing the logical relationship of the components in the circuit; for another example, in another implementation, the hardware circuit can be implemented by PLD, taking FPGA as an example, which can include a large number of logic gate circuits, and the connection relationship between the logic gate circuits is configured by a configuration file, so as to realize the functions of some or all units above. All units of the above device can be implemented in the form of a processor calling software, or in the form of a hardware circuit, or in the form of a processor calling software, and the remaining part can be implemented in the form of a hardware circuit.

[0262] In an embodiment of the present application, a processor is a circuit with the ability to process signals. In one implementation, the processor may be a circuit with the ability to read and run instructions, such as a CPU, a microprocessor, a GPU, or a DSP; in another implementation, the processor may implement certain functions through the logical relationship of a hardware circuit, and the logical relationship of the hardware circuit is fixed or reconfigurable, such as a hardware circuit implemented by an ASIC or PLD, such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration document to implement the hardware circuit configuration can be understood as the process of the processor loading instructions to implement the functions of some or all of the above units. In addition, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as an NPU, TPU, DPU, etc.

[0263] It can be seen that each unit in the above device can be one or more processors (or processing circuits) configured to implement the above method, such as: CPU, GPU, NPU, TPU, DPU, microprocessor, DSP, ASIC, FPGA, or a combination of at least two of these processor forms.

[0264] In addition, all or part of the units in the above device can be integrated together, or can be implemented independently. In one implementation, these units are integrated together and implemented in the form of a SOC. The SOC may include at least one processor for implementing any of the above methods or implementing the functions of each unit of the device. The type of the at least one processor may be different, for example, including a CPU and an FPGA, a CPU and an artificial intelligence processor, a CPU and a GPU, etc.

[0265] The embodiment of the present application also provides a product recommendation device; the product recommendation device includes a processor and an input-output component. The input-output component is used to obtain conversation data;

[0266] The processor is used to recommend the target product corresponding to the dialogue data obtained by the input-output component by executing any one of the product recommendation methods in the above embodiments.

[0267] The above-mentioned interface circuit can be any interface circuit that can realize the data communication function, for example, it can be a USB interface circuit, a Type-C interface circuit, a serial port circuit, a PCIE circuit, etc.

[0268] Optionally, the embodiment of the present application further provides a system, wherein the interactive client is used to obtain the conversation data and send the conversation data to the server, and to display the product recommendation results output by the server;

[0269] The server is used to recommend the target product corresponding to the conversation data obtained by the interactive client by executing the product recommendation method described in any one of the above embodiments.

[0270] Another embodiment of the present application also provides a product recommendation device, see Fig.17 As shown, the device includes:

[0271] Memory 1700 and processor 1710;

[0272] The memory 1700 is connected to the processor 1710 and is used to store programs;

[0273] The processor 1710 is used to implement the product recommendation method disclosed in any of the above embodiments by running the program stored in the memory 1700.

[0274] Specifically, the above-mentioned product recommendation device may also include: a bus, a communication interface 1720 , an input device 1730 and an output device 1740 .

[0275] The processor 1710, the memory 1700, the communication interface 1720, the input device 1730 and the output device 1740 are connected to each other via a bus.

[0276] A bus may include a pathway that transfers information between components of a computer system.

[0277] Processor 1710 may be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the program of the scheme of the present invention. It may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0278] The processor 1710 may include a main processor, and may also include a baseband chip, a modem, and the like.

[0279] The memory 1700 stores a program for executing the technical solution of the present invention, and may also store an operating system and other key services. Specifically, the program may include a program code, and the program code includes a computer operation instruction. More specifically, the memory 1700 may include a read-only memory (ROM), other types of static storage devices that can store static information and instructions, a random access memory (RAM), other types of dynamic storage devices that can store information and instructions, a disk storage, a flash, and the like.

[0280] The input device 1730 may include a device for receiving data and information input by a user, such as a keyboard, a mouse, a camera, a scanner, a light pen, a voice input device, a touch screen, a pedometer, or a gravity sensor.

[0281] Output device 1740 may include devices that allow information to be output to a user, such as a display screen, printer, speakers, etc.

[0282] The communication interface 1720 may include any transceiver or the like to communicate with other devices or communication networks, such as Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc.

[0283] The processor 1710 executes the program stored in the memory 1700 and calls other devices, which can be used to implement each step of any product recommendation method provided in the above embodiments of the present application.

[0284] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the product recommendation method according to various embodiments of the present application described in any of the above embodiments of this specification.

[0285] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0286] In addition, the embodiment of the present application may also be a storage medium on which a computer program is stored. The computer program is executed by a processor to execute the steps of the product recommendation method according to various embodiments of the present application described in any of the above embodiments of this specification, and specifically the following steps may be implemented:

[0287] Acquire conversation data of a target object when selecting a product, wherein the conversation data includes object information and product information corresponding to the target object;

[0288] Performing semantic recognition based on the conversation data to determine explicit features corresponding to the target object, wherein the explicit features include object features and product features;

[0289] Calling the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features;

[0290] The explicit features and the implicit features are fused to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information.

[0291] For the aforementioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0292] It should be noted that each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0293] The steps in the methods of each embodiment of the present application can be adjusted in order, combined and deleted according to actual needs, and the technical features recorded in each embodiment can be replaced or combined.

[0294] The modules and sub-modules in the devices and terminals of the various embodiments of the present application can be combined, divided and deleted according to actual needs.

[0295] In the several embodiments provided in the present application, it should be understood that the disclosed terminals, devices and methods can be implemented in other ways. For example, the terminal embodiments described above are only schematic, for example, the division of modules or submodules is only a logical function division, and there may be other division methods in actual implementation, for example, multiple submodules or modules can be combined or integrated into another module, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0296] The modules or submodules described as separate components may or may not be physically separated, and the components of the modules or submodules may or may not be physical modules or submodules, that is, they may be located in one place, or they may be distributed on multiple network modules or submodules. Some or all of the modules or submodules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0297] In addition, each functional module or submodule in each embodiment of the present application may be integrated into one processing module, or each module or submodule may exist physically separately, or two or more modules or submodules may be integrated into one module. The above-mentioned integrated modules or submodules may be implemented in the form of hardware or in the form of software functional modules or submodules.

[0298] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0299] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly by hardware, software units executed by a processor, or a combination of the two. The software units may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0300] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0301] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A product recommendation method, characterized in that: include: Acquire conversation data of a target object when selecting a product, wherein the conversation data includes object information and product information corresponding to the target object; Performing semantic recognition based on the conversation data to determine explicit features corresponding to the target object, wherein the explicit features include object features and product features; Calling the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features; The explicit features and the implicit features are fused to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information.

2. The method according to claim 1, characterized in that The performing semantic recognition based on the conversation data to determine the explicit features corresponding to the target object includes: Performing semantic recognition based on the conversation data, summarizing key information of the sentences involved in the conversation process to obtain semantic representation information; Determine a speech node in the conversation data, and determine corresponding intent information according to the speech node; Extracting key elements from the conversation data according to the configured key information categories to obtain object portrait information and product description information; Performing emotion recognition on the conversation data to obtain emotion analysis information; generating the object feature based on the semantic representation information, the intention information, the object portrait information and the sentiment analysis information, and generating the product feature based on the product description information; The object characteristics and the product characteristics are combined to determine the dominant characteristics corresponding to the target object.

3. The method according to claim 2, characterized in that The determining of the speech nodes in the conversation data, and determining corresponding intention information according to the speech nodes, includes: Obtaining scene information corresponding to the conversation data; Calling the speech node diagram associated with the scene information; Detecting the conversation data based on the speech node graph to determine the speech nodes in the conversation data; Determine the intent information corresponding to the speech node.

4. The method according to claim 2, characterized in that: The extracting of key elements from the conversation data according to the configured key information categories to obtain object portrait information and product description information includes: Performing corpus extraction on the conversation data according to the configured key information categories to obtain corpus extraction elements; Logically associating the extracted elements of the corpus to obtain condensed information; Key elements are extracted based on the condensed information to obtain the object portrait information and the product description information.

5. The method according to claim 2, characterized in that: The performing emotion recognition on the conversation data to obtain emotion analysis information includes: Determining feedback information in the conversation data; Identifying emotional keywords in each corpus in the feedback information to obtain an emotional keyword set; Emotion recognition is performed based on the identified emotion keyword set to obtain the emotion analysis information.

6. The method according to claim 2, characterized in that The combining the object feature and the product feature to determine the dominant feature corresponding to the target object includes: Determining feature types corresponding to the object feature and the product feature; Determine a corresponding feature processing method based on the feature type; According to the preset dimensions, the feature processing method is used to regularize the feature dimensions of the object features and the product features; The processed object features and product features are combined to determine the dominant features corresponding to the target object.

7. The method according to claim 1, characterized in that The calling of the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features includes: Calling the associated information corresponding to the target object; Determining the action of the target object on the product indicated in the association information to obtain behavior information; Acquire a knowledge database corresponding to the associated information, perform knowledge call based on the knowledge database, and obtain open domain information corresponding to the target object; Feature extraction is performed based on the behavior information and the open domain information respectively to obtain the latent feature.

8. The method according to claim 1, characterized in that The fusing the explicit feature and the implicit feature to obtain cross-domain information, and recommending a target product to the target object based on the cross-domain information, includes: Determine a mapping relationship between the target object indicated in the explicit feature and each product, so as to restore the explicit feature to first knowledge representation data based on the mapping relationship; Determine a mapping relationship between the target object indicated in the implicit feature and each product, so as to restore the implicit feature to second knowledge representation data based on the mapping relationship; Performing semantic summarization based on the first knowledge representation data and the second knowledge representation data to obtain summary features; Determine the cross-domain information corresponding to the target object according to the summary feature; A target product is recommended to the target object based on the cross-domain information.

9. The method according to claim 8, characterized in that The recommending a target product to the target object based on the cross-domain information includes: Obtaining a candidate list corresponding to the target object, wherein the candidate list includes multiple candidate products; Performing word segmentation processing on the cross-domain information, and performing vector representation on the cross-domain information after the word segmentation processing to obtain a cross-domain vector; Calculating the similarity between the cross-domain vector and each of the candidate products respectively to obtain a target score for each of the candidate products; Product screening is performed based on the target score of each of the candidate products to obtain the target product recommended to the target object.

10. The method according to claim 9, characterized in that The respectively calculating the similarity between the cross-domain vector and each of the candidate products to obtain a target score for each of the candidate products includes: Respectively calculating the similarity between the cross-domain vector and each of the candidate products in a preset scoring dimension to obtain a semantic score, an object feature score, and a product feature score, wherein the preset scoring dimension includes a semantic dimension, an object feature dimension, and a product feature dimension; The semantic score, the object feature score and the product feature score are weightedly calculated to obtain the target score.

11. The method according to any one of claims 1 to 10, characterized in that: The method further comprises: Determine the speech nodes in the conversation data; Obtain the recommended speech template corresponding to the speech node; Determine a recommendation speech template corresponding to the target product, and rewrite the content of the recommendation speech template to obtain the target recommendation speech; The target product is recommended to the target object based on the target recommendation words.

12. The method according to any one of claims 1 to 10, characterized in that: The method further comprises: Collect feedback from target objects on target products; Adjusting the feature extraction dimension corresponding to the target object based on the feedback information; The open domain information is updated according to the knowledge information indicated by the feedback information.

13. A product recommendation device, characterized in that: include: An acquisition unit, used to acquire dialogue data of a target object when selecting a product, wherein the dialogue data includes object information and product information corresponding to the target object; A processing unit, configured to perform semantic recognition based on the conversation data to determine explicit features corresponding to the target object, wherein the explicit features include object features and product features; The processing unit is further used to call the associated information corresponding to the target object to extract features according to the associated information to obtain implicit features; A recommendation unit is used to fuse the explicit features and the implicit features to obtain cross-domain information, so as to recommend a target product to the target object based on the cross-domain information.

14. A product recommendation device comprising an input and output component and a processor; The input and output components are used to obtain conversation data; The processor is used to recommend a target product corresponding to the conversation data acquired by the input-output component by executing the product recommendation method as described in any one of claims 1 to 12.

15. A product recommendation system, comprising an interactive client and a server; The interactive client is used to obtain the conversation data and send the conversation data to the server, and to display the product recommendation results output by the server; The server is used to recommend the target product corresponding to the conversation data obtained by the interactive client by executing the product recommendation method according to any one of claims 1 to 12.

16. A computer program product, characterized in that include: A computer program, which, when executed by a processor, implements the product recommendation method according to any one of claims 1 to 12.

Images