Information recommendation method, electronic equipment and vehicle

Abstract

The invention relates to the field of data processing, and particularly provides an information recommendation method, electronic equipment and a vehicle. The method comprises the following steps: receiving user demand information, and sending the user demand information to an intention complexity classification model obtained by pre-training to obtain a complexity classification result; determining a corresponding target intention recognition model according to the complexity classification result, and performing analysis based on the user intention information by using the target intention recognition model to determine a target intention; and selecting a target recommendation model corresponding to the target intention, determining target recommendation content by using the target recommendation model, and feeding back the target recommendation content to the user. The user demand can be accurately understood according to the complexity of the user demand information, then the accurate target intention is determined, then the target recommendation model corresponding to the target intention is used for capturing the user demand and conducting intelligent analysis, and the most accurate target recommendation content is determined.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an information recommendation method, electronic equipment, and vehicle. Background Art

[0002] Currently, some information recommendation software directly recommends information based on the needs expressed by users (for example, voice needs or text needs).

[0003] However, current information recommendation software determines the corresponding keywords based on the needs expressed by users, and then intelligently makes corresponding information recommendations based on the keywords. This information recommendation based on keywords often fails to meet user needs, resulting in a poor user experience. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose an information recommendation method, electronic device and vehicle to solve the problem that the current method of directly recommending information based on specific keywords cannot meet user needs.

[0005] Based on the above objectives, this application provides an information recommendation method, including:

[0006] Receive user demand information, and send the user demand information to a pre-trained intent complexity classification model to obtain a complexity classification result;

[0007] Determine a corresponding target intent recognition model based on the complexity classification result, and use the target intent recognition model to analyze based on the user intent information to determine the target intent, wherein the target intent recognition model is a pre-trained model;

[0008] A target recommendation model corresponding to the target intent is selected, target recommendation content is determined using the target recommendation model, and the target recommendation content is fed back to the user, wherein the recommendation model corresponding to each intent is stored.

[0009] Based on the same inventive concept, the present application also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor implements the method described above when executing the computer program.

[0010] Based on the same inventive concept, the present application also provides a vehicle, comprising the electronic device as described above.

[0011] From the above, it can be seen that the information recommendation method, electronic device and vehicle provided in the present application can use the pre-trained intention complexity classification model to identify and process user demand information to determine the complexity classification result corresponding to the user demand information, and the complexity classification result can characterize the demand complexity of the user demand information; then, based on the complexity classification result, the corresponding target intention recognition model can be determined, and the target intention recognition model can accurately perform intent recognition analysis on the user intention information corresponding to the complexity classification result to obtain an accurate target intention; then, the target recommendation model corresponding to the target intention is selected, and the target recommendation model can process the content recommendation of the field corresponding to the target intention. In this way, the target recommendation model can be used to determine the target recommendation content according to user needs from the field corresponding to the target intention, and the obtained target recommendation content is more accurate, can better meet the actual needs of the user, and improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flowchart of an information recommendation method according to an embodiment of the present application;

[0013] Figure 2 This is a flowchart of an information recommendation method according to another embodiment of the present application;

[0014] Figure 3 This is a structural block diagram of an information recommendation device according to an embodiment of the present application;

[0015] Figure 4 This is a schematic structural diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] When a control system (e.g., a vehicle computer system) recommends information based on various user needs, the following problems generally arise:

[0017] First, the demand is not understood accurately.

[0018] It's difficult to accurately understand the user's true intentions when expressing complex needs. For example, if a user says, "I want to listen to some songs that are suitable for long-distance driving and to refresh myself," traditional control systems might make recommendations based solely on the keyword "song," failing to understand the underlying needs of "long-distance driving" and "refreshment," resulting in recommendations that don't match the user's expectations.

[0019] Second, poor scene adaptation capabilities

[0020] The system fails to effectively perceive the user's driving scenario and make targeted recommendations based on that scenario. For example, when a user feels tired after a long drive, the system recommends a large amount of driving-related content without considering that the user needs to rest and relax. Or, after the user arrives at the travel destination, the system fails to provide timely services such as local travel guides and food recommendations.

[0021] Third, the interactive experience is not good

[0022] General recommendation functions have a single interaction method, mostly button operations or simple voice command responses. They do not support natural and smooth multi-round conversations. Users have to perform cumbersome operations when expressing complex needs, which affects driving safety and experience.

[0023] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0024] The information recommendation method proposed in the embodiment of the present application is applied to a control system (for example, a vehicle control system) such as Figure 1 As shown, the method includes steps 101-103.

[0025] Step 101: Receive user demand information, and send the user demand information to a pre-trained intention complexity classification model to obtain a complexity classification result.

[0026] In a specific implementation, the user demand information includes: voice information, text information, posture information, or eye movement information, etc. If it is voice information, the voice information will be converted into text; if it is posture information, the meaning expressed by the posture information will be converted into text; if it is eye movement information, the meaning expressed by the eye movement information will be converted into text.

[0027] In this way, the text corresponding to the user demand information can be input into the intent complexity classification model to perform intent complexity classification and obtain a complexity classification result. The complexity classification result can be divided into multiple categories based on complexity. The specific number of categories can be set according to actual needs. Preferably, the classification is divided into two categories, namely simple category and complex category.

[0028] Among them, the corresponding intent complexity classification model is a binary classification model. A binary classification model will be built in advance, and user demand information of the simple class (for example, user demand information: "I really want to eat Sichuan cuisine now", its intention is clear and easy to identify, marked as a simple class, marked with a symbol of "0") and user demand information of the complex class (for example, user demand information: "We are finally about to arrive in Nanjing", its intention is vague and not easy to identify, marked as a complex class, marked with a symbol of "1") are collected as the first training samples to perform supervised training on the pre-built binary classification model, and the loss function is determined based on the difference between the training results output during training and the marked complexity classification (for example, simple class or complex class), and the parameters of the binary classification model are adjusted according to the loss function. After all the first training samples are trained, the final binary classification model is used as the intent complexity classification model. The intent complexity classification model can perform complexity classification on various user demand information. The corresponding marking symbols of the simple class and the complex class are different, and can be set according to actual needs.

[0029] Step 102: Determine a corresponding target intent recognition model based on the complexity classification result, and use the target intent recognition model to analyze based on the user intent information to determine the target intent, wherein the target intent recognition model is a pre-trained model.

[0030] In practice, the control system pre-stores intent processing models corresponding to various complexity categories, with the number of stored intent processing models matching the number of complexity categories. This allows the target intent recognition model corresponding to the complexity classification results to process user intent information of the corresponding complexity, enabling more targeted intent recognition and more accurate determination of the target intent.

[0031] Among them, each intent processing model uses the user intent information of the corresponding complexity classification, and forms a second training sample after marking the corresponding intent for each user intent information, and performs supervised training on the pre-built network model (for example, LLM, Large Language Model, large language model; or binary classification model), and determines the loss function based on the difference between the training results output during training and the marked intent, and adjusts the parameters of the network model according to the loss function. Make sure that all the second training samples are trained, and use the final network model as the model for intent recognition. The intent recognition model is a model that can perform intent recognition on the user intent information of the complexity classification. For example, intents include: music intent, film and television intent, dining intent, and scenic spot intent, etc.

[0032] Step 103 : Select a target recommendation model corresponding to the target intent, use the target recommendation model to determine target recommendation content, and feed back the target recommendation content to the user.

[0033] In specific implementations, recommendation models corresponding to various intents are pre-stored. For example, a music recommendation model corresponds to the music intent; a movie recommendation model corresponds to the movie intent; a restaurant recommendation model corresponds to the restaurant intent; and a scenic spot recommendation model corresponds to the scenic spot intent. These different recommendation models can be used to recommend content corresponding to the user's intent based on their needs.

[0034] Among them, each recommendation model can be a functional module that can recommend corresponding content data according to the user's interests and preferences.

[0035] The recommendation model is obtained through training using a neural network model. Specifically, for each intent, the specific information of the intent (such as the interaction questions corresponding to the intent) will be collected first, and the corresponding recommended content will be marked for each specific information to form a third training sample. The pre-built neural network model is supervised by using the third training sample, and the loss function is determined based on the difference between the training results output during training and the marked recommended content, and the parameters of the neural network model are adjusted according to the loss function. After all the third training samples are trained, the final neural network model is used as the recommendation model for the intent. The intent is associated with the corresponding recommendation model and stored in the database of the control system.

[0036] In this way, after the target intent recognition model outputs the target intent, the target recommendation model corresponding to the target intent will be called up, and the target recommendation model will be used to determine the corresponding target recommendation content based on the user's demand characteristics and preference characteristics, and the target recommendation content will be fed back to the user. The specific feedback method can be voice broadcast and / or screen display, preferably voice broadcast and screen display at the same time. The corresponding target recommendation content can be one or more, so that the user can determine whether there is the user's demand content in at least one target recommendation content. If so, the user can trigger the demand content and execute the corresponding function.

[0037] The number of target recommended contents cannot be too many (for example, no more than 10), so as to avoid too many target recommended contents and the user taking too long to confirm the target recommended contents, which makes it impossible for the user to quickly select the content he needs.

[0038] Through the above scheme, the pre-trained intention complexity classification model can be used to identify and process user demand information to determine the complexity classification result corresponding to the user demand information. The complexity classification result can characterize the demand complexity of the user demand information; then, based on the complexity classification result, the corresponding target intention recognition model can be determined. The target intention recognition model can accurately perform intent recognition analysis on the user intention information corresponding to the complexity classification result to obtain an accurate target intention; then, the target recommendation model corresponding to the target intention is selected. The target recommendation model can process content recommendations in the field corresponding to the target intention. In this way, the target recommendation model can be used to determine the target recommendation content from the field corresponding to the target intention according to user needs. The obtained target recommendation content is more accurate, can better meet the actual needs of users, and improve user experience.

[0039] In some embodiments, if the complexity classification result is a simple class or a complex class, step 102 includes:

[0040] Step A: In response to determining that the complexity classification result is a complex class, determining that the corresponding target intent recognition model is a first intent recognition model, receiving multiple rounds of conversation results with the user, and using the first intent recognition model to recognize the multiple rounds of conversation results to determine the target intent.

[0041] In specific implementation, for user needs with complex intentions, it is necessary to use the first intent recognition model obtained through training of the large language model (LLM) to determine the target intention of the user needs through multiple rounds of dialogue.

[0042] Among them, various conversation contents are collected in advance, and the corresponding intents are marked for the conversation contents to obtain a fourth training sample. The fourth training sample is used to train the large language model. The large language model outputs the training intent result. The training intent result is compared with the marked intent. If the results are the same, the next fourth training sample is used for training. If they are different, the parameters of the large language model are adjusted (for example, through backpropagation) so that after the parameters are adjusted, the large language model can output the same training intent result as the marked intent, and then continue to use the next fourth training sample for training. After the fourth training sample is fully trained, the first intent recognition model can be obtained.

[0043] For example, through the following conversation, we can determine the target intention:

[0044] User: We’ll be at the hotel soon.

[0045] First intention recognition model: Dear car owner, I heard that you are about to arrive at the hotel. Do you need me to recommend some fun places nearby?

[0046] User: I’m a little hungry now. Please recommend a restaurant nearby.

[0047] First intent recognition model: OK, I’ll recommend some restaurants to you right away.

[0048] The first intent recognition model analyzes these conversation results and determines that the target intent is dining intention.

[0049] Alternatively, in step B, in response to determining that the complexity classification result is a simple class, the corresponding target intent recognition model is determined to be a second intent recognition model, and the second intent recognition model is used to identify the user intent information to determine the target intent.

[0050] In specific implementation, for user needs with simple intentions, a second intent recognition model obtained after training a relatively simple two-classification model is used to classify and identify user intent information to determine the target intent.

[0051] A large amount of user intent information (e.g., query) is collected in advance and annotated with the actual intent to obtain a fifth training sample. The fifth training sample is then used to train a binary classification model to obtain a second intent recognition model. The classification level of this binary classification model is determined based on the number of corresponding intent categories.

[0052] For example, if there are four intent categories—music, film, and television intent, dining intent, and scenic spot intent—the corresponding binary classification model has four classification levels, each used to identify a corresponding intent category. Specifically, the four classification levels include: a music classification level for identifying music intent, a film classification level for identifying film and television intent, a dining classification level for identifying dining intent, and a scenic spot classification level for identifying scenic spot intent.

[0053] Through the above scheme, a more targeted target intent recognition model can be determined according to the complexity classification results. In this way, the complex class can use the first intent recognition model that can handle complex intents, obtain multiple rounds of dialogue results through multiple rounds of dialogue, and then identify the multiple rounds of dialogue results to accurately determine the target intent; if it is a simple class, the second intent recognition model that can handle simple intents can be used to determine the target intent through multiple rounds of binary classification; the accuracy of the determined target intent is guaranteed.

[0054] In some embodiments, in step A, receiving the results of multiple rounds of conversations with the user, and using the first intent recognition model to recognize the results of the multiple rounds of conversations and determine the target intent may include:

[0055] Step A1: Analyze the user intention information using the first intention recognition model to obtain a first analysis result and output it so that the user can make feedback based on the output first analysis result.

[0056] In specific implementations, the first intent recognition model is used to analyze and process the user's intent information, determine a first analysis result of a rough intent classification, and output feedback to the user (specifically, feedback can be provided to the user via voice and / or screen display). The user will provide feedback based on the first analysis result.

[0057] Step A2: receiving user feedback information, analyzing the user feedback information using the first intention recognition model to obtain a pending intention type and outputting it, so that the user can determine whether the outputted pending intention type is correct.

[0058] During implementation, the first intent recognition model analyzes the user's pending intent type based on the user feedback information, sends the pending intent type to the output module, and then feeds the pending intent type back to the user (specifically, it can be fed back to the user via voice and / or screen display). The user determines whether the pending intent type is correct.

[0059] Step A3, in response to receiving user feedback that the pending intent type is wrong, use the first intent recognition model to iteratively perform a correction process on the pending intent type until the corrected intent type is correct, and use the corrected intent type as the target intent.

[0060] In specific implementation, if the user feedback received by the first intent recognition model is that the pending intent type is incorrect, the pending intent type correction process will be performed, specifically:

[0061] Re-determine and correct the analysis results, output feedback to the user, receive new feedback information from the user, repeat the process of step A2 above, determine the corrected intention type, and output feedback to the user.

[0062] The above correction process is continuously iterated until the first intent recognition model receives user feedback that the corrected intent type is correct. The iteration is stopped and the corrected intent type is used as the target intent.

[0063] Alternatively, in step A4, in response to receiving user feedback that the pending intent type is correct, the pending intent type is used as the target intent.

[0064] In specific implementation, if the user feedback received by the first intent recognition model is that the pending intent type is correct, there is no need to perform a correction process, and the pending intent type can be directly used as the target intent.

[0065] For example, the first intention recognition model: based on the user intention information of "Finally almost at the destination", determines that the first analysis result is: "The car owner wants to eat", and issues the first analysis result of "Does the car owner want to confirm the restaurant at the destination?"

[0066] User: Based on the inquiry, respond with “I don’t want to go to the restaurant, I want to visit nearby attractions.”

[0067] The first intent recognition model: Based on the user feedback "I don't want to go to the restaurant, I want to visit nearby attractions", it is determined that the pending intent type is "attraction intent" and fed back to the user.

[0068] User: Make sure the pending intent type is correct.

[0069] The first intent recognition model takes the undetermined intent type as the target intent and sends the target intent to the target recommendation model corresponding to the target intent.

[0070] or,

[0071] User: Confirm that the pending intent type is incorrect.

[0072] The first intention recognition model: issues a corrected analysis result of "Does the car owner want to listen to music?" and feeds back to the user.

[0073] User: I want to listen to nostalgic music.

[0074] The first intent recognition model determines that the type of intent to be determined is "music intent" and feeds it back to the user.

[0075] User: Make sure the pending intent type is correct.

[0076] The first intent recognition model takes the undetermined intent type as the target intent and sends the target intent to the target recommendation model corresponding to the target intent.

[0077] Through the above scheme, even if the user's intention is complex, the first intention recognition model can be used to determine the target intention that best meets the user's needs through user polling interaction. The obtained target intention is more accurate, so the target recommendation content determined based on the target intention is also more in line with the user's needs, thereby improving the user experience.

[0078] In some embodiments, determining the target recommendation content required by the user using the target recommendation model in step 103 includes:

[0079] Step 1031: Obtain user behavior information using a target recommendation model, and determine a first predetermined number of corresponding content to be recommended based on the user behavior information.

[0080] In a specific implementation, the user's preferences can be determined based on the user's behavior information, and then a first predetermined number (e.g., 100) of recommended content that meets the preferences of the group of users can be determined. The first predetermined number of recommended content can be listed and integrated in order to obtain a recall list.

[0081] The first predetermined number may be a fixed number or a variable number.

[0082] The content to be recommended may be determined in the following ways: (1) content to be recommended determined by collaboration between the user and similar users; (2) content to be recommended similar to the content corresponding to the user's behavior; (3) content to be recommended found based on interest tags determined by the user's behavior.

[0083] Step 1032: Use the target recommendation model to sort and adjust the first predetermined number of contents to be recommended according to the degree of matching with the user's needs to obtain a sorting result.

[0084] During specific implementation, since the ranking of the first predetermined number of contents to be recommended obtained above is not accurate enough and it is impossible to determine which ones better meet the needs of users, the target recommendation model will be used to determine the degree of matching between each content to be recommended and the user needs in the first predetermined number of contents to be recommended, and then the first predetermined number of contents to be recommended will be re-sorted in order from high to low according to the degree of matching to obtain the ranking result.

[0085] Step 1033 : Using the target recommendation model, select a second predetermined number of to-be-recommended contents with the highest matching degree from the sorting results as the target recommended contents, wherein the second predetermined number is smaller than the first predetermined number.

[0086] In specific implementations, a value of the second predetermined number is pre-set, and is generally smaller than the first predetermined number. In this way, the target recommendation model selects the first second predetermined number of recommended content from the sorting results as target recommended content and feeds it back to the user for selection.

[0087] Through the above scheme, a first predetermined number of contents to be recommended can be preliminarily selected based on the user's behavioral information. Since the sorting of the first predetermined number of contents to be recommended is not accurate enough, in order to facilitate the screening of target recommended content that better meets the user's needs, it will be re-sorted according to a matching program with the user's needs, thereby obtaining a more accurate sorting result, so that the target recommended content selected therefrom will be more accurate.

[0088] In some embodiments, in step 1031, determining a first predetermined number of content to be recommended based on the user's behavior information includes:

[0089] (1) Recommended content determined collaboratively by the user and similar users.

[0090] Step C1 : determining similar users based on the user's behavior information, wherein similar users corresponding to each user are pre-stored, and the similar users are other users matching the user's behavior information.

[0091] In a specific implementation, the user's similarity with other users can be calculated based on the user's behavior information, and the other users can be sorted by similarity. A third predetermined number of other users with high similarity can be selected as similar users. Alternatively, other users whose similarity exceeds a similarity threshold can be selected as similar users. These similar users are then associated with the user's identity information (e.g., ID address) and stored in a database. In this way, the corresponding similar users can be directly retrieved based on the user's identity information.

[0092] Step C2: obtaining the operation behaviors of the similar users within a predetermined time period, and determining multiple trigger contents according to the operation behaviors.

[0093] In specific implementations, the actions include: clicking, collecting, or liking. Then, the trigger content corresponding to the user's action within a predetermined time period (e.g., one month or 10 days) is determined. For example, the music content that the user has collected or liked in the past month is determined.

[0094] Step C3 , filtering the plurality of triggering contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of triggering contents.

[0095] In specific implementation, since the multiple triggering contents obtained above may be repeated, they are filtered to remove duplicates, thereby obtaining multiple filtered triggering contents. The multiple triggering contents are sorted according to the operation frequency of the operation behavior, and a first predetermined number of contents to be recommended are selected.

[0096] Through the above scheme, we can determine the trigger content corresponding to the operation behavior of similar users who are closely matched with the user. Since they are similar users, their corresponding preferences are also relatively similar. Therefore, the multiple trigger contents obtained are also content that meets the user's preferences. From these, the first predetermined number of recommended content is also content that is closely matched with the user's needs. This makes the subsequent determination of target recommended content more accurate.

[0097] and / or, (2) determining content to be recommended that is similar to the content corresponding to the user's behavior based on the content.

[0098] Step D1: determining trigger content based on the user's behavior information.

[0099] During specific implementation, a behavior list corresponding to the user's behavior information (e.g., click, favorite, like) is obtained, and the corresponding trigger content is determined based on the user's behavior list within a predetermined time period.

[0100] Step D2: Acquire multiple similar contents corresponding to the triggering content, wherein similar contents corresponding to each triggering content are pre-stored.

[0101] In a specific implementation, the similarity between the contents is pre-calculated, and a fourth predetermined number of similar contents corresponding to each content (for example, the first 100, first 50, or first 10 sorted in descending order of similarity) is determined and stored in a database. In this way, the corresponding similar content can be directly searched from the database based on the triggering content.

[0102] If D1 obtains multiple triggering contents, it will determine multiple similar contents corresponding to each triggering content and integrate these similar contents together.

[0103] Step D3: filtering the plurality of similar contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of similar contents.

[0104] In specific implementation, since the multiple similar contents obtained above may be duplicated, they are filtered to remove duplicates, thereby obtaining multiple filtered similar contents. The multiple similar contents are sorted according to the operation frequency of the operation behavior, and a first predetermined number of contents to be recommended are selected.

[0105] Through the above scheme, trigger content for a predetermined time period can be determined based on user behavior. Based on pre-stored similar content corresponding to each content, multiple similar content corresponding to the trigger content can be selected. Since the content is relatively similar, these similar content obtained are all content that meets the user's preferences. The first predetermined number of recommended content selected from them is also content that is more in line with the user's needs. This makes the subsequent target recommendation content determined more accurate.

[0106] and / or, (3) determining interest tags based on user behavior and then finding content to be recommended.

[0107] Step E1: Determine interest tags based on the user's behavior information.

[0108] Step E2: searching for a plurality of matching contents corresponding to the interest tag, filtering the plurality of matching contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of matching contents.

[0109] In specific implementation, the user's interest tags can be determined based on the user's behavior information (for example, clicks, favorites, or likes), so that the corresponding matching content can be directly found based on the interest tags. These matching contents may be repeated, so they are deduplicated and filtered to obtain filtered matching content. The filtered matching content is then sorted according to popularity to select the first predetermined number of content to be recommended.

[0110] Through the above solution, interest tags are directly determined based on the user's behavior information, and then a first predetermined number of recommended content is filtered out based on the interest tags. This operation is simple and fast, and the recommended content found is more in line with the user's interests.

[0111] In a preferred embodiment, the contents to be recommended obtained in the above three ways can be integrated and deduplicated to obtain a final recall list of contents to be recommended, which can further ensure the accuracy of the contents to be recommended.

[0112] In some embodiments, step 1032 includes:

[0113] Step 10321, obtain the user's characteristic data, input the user's characteristic data and a first predetermined number of contents to be recommended into a click-through rate prediction model, determine the predicted click-through rate of each content to be recommended, and use the predicted click-through rate as the degree of matching between the content to be recommended and the user's needs.

[0114] In specific implementations, user characteristic data includes: gender, age, health status, interests, and preferences. The corresponding content characteristics of each piece of content to be recommended are determined. The user's characteristic data and the content characteristics of each piece of content to be recommended are input into a click-through rate prediction model. The model then analyzes the user's characteristic data to determine the predicted click-through rate (CTR) for the piece of content to be recommended, which serves as a measure of the degree of match between the content and the user's needs. The predicted CTR represents the probability that the user will click on the recommended content.

[0115] Step 10322: sort and adjust the first predetermined number of contents to be recommended according to the matching degree to obtain a sorting result.

[0116] In specific implementation, the recommended contents are sorted and adjusted according to the matching procedure, specifically sorted in descending order according to the matching degree to obtain a sorting result, so that the second predetermined number of recommended contents can be selected as target recommended contents according to the sorting result.

[0117] Through the above scheme, since the ranking of the first predetermined number of contents to be recommended is not accurate enough, in order to accurately sort the recommended contents, a pre-trained click-through rate prediction model will be used to combine and analyze the user's characteristic data with the content characteristics corresponding to each content to be recommended, and accurately determine the predicted click-through rate of each content to be recommended. In this way, the ranking result obtained by accurate sorting based on the predicted click-through rate as the degree of matching is more in line with the user's needs, and it is more accurate to select the first second predetermined number of contents to be recommended as the target recommended content.

[0118] In some embodiments, the training process of the click-through rate prediction model includes:

[0119] Step F1, obtaining the user's feature data, historical content, and historical click data corresponding to the historical content, combining the feature data with the historical content, and marking with the historical click data to obtain a training sample.

[0120] In specific implementation, when the historical content corresponds to clicks, likes or collections, the historical click data is the first label (for example, 1); when the historical content does not correspond to clicks, likes or collections, the historical click data is the second label (for example, 0). In this way, the historical click data is used to label each combination of historical content and feature data, so that the training samples obtained are more accurate.

[0121] Step F2: Use the training samples to train the constructed initial model, output the click-through rate training results corresponding to the training samples, and adjust the parameters of the initial model according to the difference between the click-through rate training results and the marked historical click data.

[0122] In specific implementations, the initial model is a pre-built initial CTR (Click-Through-Rate) model. The user feature data in the training sample is converted into a user feature vector, and the content features of the corresponding historical content are converted into a content feature vector. The user feature vector and content feature vector are analyzed to determine the degree of overlap between the user feature vector and the content feature vector. This overlap is then used to determine the CTR training result.

[0123] The loss function is determined based on the difference between the click-through rate training results and the marked historical click data. The parameters of the initial model are adjusted according to the loss value calculated by the loss function to complete the training of a training sample.

[0124] Step F3: After all training samples are trained, the final adjusted initial model is used as the click-through rate prediction model.

[0125] In specific implementation, after confirming that all training samples have been trained, a new set of test samples will be determined according to the aforementioned training sample collection method. The final adjusted initial model will be tested using these test samples to determine its prediction accuracy. If the prediction accuracy is greater than or equal to the accuracy threshold, the final adjusted initial model will be used as the click-through rate prediction model. If the prediction accuracy is less than the accuracy threshold, the final adjusted initial model will be trained using the test samples as training samples until the prediction accuracy of the final adjusted initial model is greater than or equal to the accuracy threshold, at which point the final adjusted initial model will be used as the click-through rate prediction model.

[0126] Through the above scheme, it is possible to use relatively accurate training samples based on the initial model for learning and training, and then continuously adjust the parameters of the initial model, so that the final click-through rate prediction model can accurately identify the user's needs for each content, and then accurately determine the user's predicted click-through rate for each content, thereby ensuring the accuracy of the click-through rate prediction model.

[0127] In some embodiments, after step 103, the method further includes:

[0128] Step 104, in response to receiving user's unsatisfactory feedback on the target recommended content, using the target recommendation model to iteratively execute the interactive dialogue process with the user, and determine new target recommended content based on the dialogue content, until the user makes a choice based on the new target recommended content to stop the iteration.

[0129] In practice, the target recommendation model pre-stores target recommendation models corresponding to target intent. This target recommendation model is trained using a large language model (LLM). If the user is dissatisfied with the target recommendation content, the target recommendation model in the target recommendation model interacts with the user. Based on the interactive dialogue, new target recommendations are determined until the user selects the most desired content from the new target recommendations.

[0130] Step 105 : In response to receiving the execution content selected by the user from the target recommended content or the new target recommended content, executing a corresponding function according to the execution content.

[0131] During specific implementation, the user selects a content that he or she needs most from the target recommended content (or new target recommended content) as the execution content according to his or her needs, and the function will be executed according to the execution content.

[0132] Through the above scheme, even if the target recommended content obtained does not meet the user's needs, new target recommended content can be re-determined through interactive dialogue with the user, ensuring that the user can select the most required execution content from it to perform the function, providing the user with more accurate required functions and facilitating user use.

[0133] The following is a specific embodiment of the information recommendation method of the present application. Figure 2 As shown in the figure, the specific process is executed in the vehicle control system.

[0134] 1. Receive user demand information (e.g., user voice request).

[0135] 2. Preprocess user demand information (for example, converting voice requests into text content through a speech-to-text model).

[0136] 3. Input user demand information into the intent complexity classification model and output the intent complexity classification results.

[0137] The training process of the intent complexity classification model is as follows:

[0138] Training samples are obtained by collecting various user demand information (e.g., queries) and labeling them with corresponding complexity types. For example, the user demand information "I really want to eat Sichuan cuisine now" has a clear intention and is easy to identify, so it is labeled as a simple class with a marker symbol of "0". The user demand information "We are finally arriving in Nanjing" has a vague intention and is difficult to identify, so it is labeled as a complex class with a marker symbol of "1".

[0139] The training samples are then used to train a pre-built binary classification model (e.g., BERT, Bidirectional Encoder Representations from Transformers, a pre-trained language model) to obtain an intent complexity classification model.

[0140] 4. If the complexity classification result is complex, it is input to the first intent recognition model trained based on the large language model (LLM) to determine the target intent through multiple rounds of conversations with the user.

[0141] a) Using the first intention recognition model to obtain a first analysis result based on the user demand information and output feedback to the user.

[0142] b) The user provides feedback based on the first analysis result. The first intent recognition model receives the user feedback information, analyzes it, obtains the pending intent type, and outputs it to the user.

[0143] c) If the user accepts, feedback is given that the pending intent type is correct, and the pending intent type is used as the target intent;

[0144] If the user feedback is that the pending intent type is incorrect, the first intent recognition model will be used to correct the pending intent type until the user feedback is correct and it is used as the target intent.

[0145] For example, user: I will be at the hotel soon.

[0146] First intention recognition model: Dear car owner, I heard that you are about to arrive at the hotel. Do you need me to recommend some fun places nearby?

[0147] User: I’m a little hungry now. Please recommend a restaurant nearby.

[0148] First intent recognition model: OK, I’ll recommend some restaurants to you right away.

[0149] The first intent recognition model analyzes these conversation results and determines that the target intent is dining intention.

[0150] 5. If the complexity classification result is a simple class, it is input into the second intent recognition model trained based on the binary classification model to determine the target intent.

[0151] The training process of the second intent recognition model is:

[0152] a) Collect a large amount of user intent information (e.g., query) in advance and use the actual intent to annotate the user intent information to obtain training samples.

[0153] b) Train a pre-built binary classification model (e.g., BERT, Bidirectional Encoder Representations from Transformers, a pre-trained language model) to obtain a second intent recognition model.

[0154] 6. Select the corresponding target recommendation model based on the target intention.

[0155] a) If the target intent is music intent, select the music recommendation model;

[0156] b) If the target intent is film and television intent, select the film and television recommendation model;

[0157] c) If the target intent is dining intent, then select the dining recommendation model;

[0158] d) If the target intent is attraction intent, the attraction recommendation model is selected.

[0159] 7. Input user demand information into the target recommendation model, determine the first predetermined number of content to be recommended through the recall process, and then obtain the sorting results through fine sorting, and select the second predetermined number (for example, the top 5 in the sorting results) of target recommended content.

[0160] The recall process is divided into three situations.

[0161] (1) Based on the content to be recommended determined by the user and similar users (Usercf), a recall list is formed.

[0162] a) calculating the similarity between the user and other users based on the user's behavior information, sorting the other users according to the similarity, selecting a third predetermined number of other users with high similarity as similar users, and associating the similar users with the user's identity information (e.g., ID address) and storing them in a database.

[0163] b) Obtain similar users from the database based on the user's identity information, determine the operation behaviors (clicks, favorites, or likes) of the similar users within a predetermined time period, determine multiple triggering contents based on the operation behaviors, and filter out a first predetermined number of contents to be recommended after deduplication to form a recall list.

[0164] (2) Based on the content corresponding to the user's behavior, determine the content to be recommended that is similar to the content (Itemc) and form a recall list.

[0165] a) Calculate the similarity between contents and determine a fourth predetermined number (eg, the first 100, first 50, or first 10 sorted by similarity) of similar contents corresponding to each content, and store them in a database.

[0166] b) Obtain a behavior list corresponding to the user's behavior information (e.g., clicks, favorites, likes), determine the corresponding trigger content based on the user's behavior list within a predetermined time period, and retrieve similar content corresponding to the trigger content from the database, dedupe these similar contents and filter out a first predetermined number of content to be recommended to form a recall list.

[0167] (3) Determine the interest tags based on the user’s behavior and then find the content to be recommended.

[0168] a) Determine interest tags based on user behavior information.

[0169] b) searching for multiple contents corresponding to the interest tag, removing duplicates from the multiple contents, and selecting a first predetermined number of the most popular contents to be recommended, to form a recall list.

[0170] II refined process is:

[0171] a) Training a click-through rate prediction model: Obtain user feature data (including at least one of basic features, historical behavior features, and preference features), historical content, and historical click data corresponding to the historical content; determine content features based on the historical content (including at least one of content release time, content title, content tags, and other content features); combine the feature data with the content features, and label them using the historical click data to obtain training samples; and use the training samples to train an initial CTR (Click-Through-Rate) model to obtain a click-through rate prediction model.

[0172] b) Input the user's characteristic data and the first predetermined number of recommended content in the recall list into the click-through rate prediction model, output the predicted click-through rate corresponding to each recommended content (that is, the degree of match between the recommended content and the user's needs), and accurately sort the recall list according to the predicted click-through rate to obtain the sorting result.

[0173] 8. Get detailed information about the target recommended content.

[0174] a) If the target recommendation content is music, then obtain singer information, style information, album information, emotion information, language information, etc.

[0175] b) If the target recommendation content is film or television, then obtain actor information, director information, style information, subject information, language information, film review information, etc.

[0176] c) If the target recommendation content is catering, then obtain restaurant rating information, cuisine information, average price information, positive reviews information, special dish information, address information, etc.

[0177] d) If the target recommendation content is a tourist attraction, obtain check-in point information, travel route information, feature information, route information, etc.

[0178] 9. Send the target recommendation content and detailed information to the user.

[0179] 10. If the user is satisfied, he / she will select the execution content, and the corresponding execution system will be started to execute the corresponding function according to the execution content.

[0180] a) If the execution content is music, start the audio system and play the song corresponding to the execution content.

[0181] b) If the execution content is a video, start the video system and play the video corresponding to the execution content.

[0182] c) If the execution content is dining or attractions, open the navigation system to determine the destination and navigate to the destination.

[0183] 11. If the user is not satisfied, the interactive dialogue process with the user is carried out, and new target recommendation content is determined based on the dialogue content until the user selects the execution content.

[0184] The target recommendation model is used to iteratively execute an interactive dialogue process with the user, and new target recommendation content is determined according to the dialogue content, until the user makes a selection based on the new target recommendation content and stops the iteration.

[0185] For example, user: I will be at the hotel soon.

[0186] Target recommendation model: Car owner, we are about to arrive at the hotel. Do you need recommendations for nearby attractions?

[0187] User: I'm hungry, please recommend some restaurants.

[0188] Intent processing model: Redefine the target intent and determine a new target recommendation model.

[0189] New target recommendation model: determines restaurants based on destination hotels (i.e., target recommendation content) and feeds it back to users.

[0190] User: I don’t like these restaurants. I’ll choose a Hangzhou cuisine restaurant instead.

[0191] New target recommendation model: Determine restaurants serving Hangzhou cuisine (i.e., new target recommendation content) and provide feedback to users.

[0192] User: Select the first restaurant.

[0193] New target recommendation model: Send the first restaurant and its detailed information to the navigation system and navigate to the first restaurant.

[0194] To sum up, we can combine various models according to the complexity of user demand information and combine multiple rounds of dialogue interactions to accurately understand user needs, and then determine the accurate target intent. We can then use the target recommendation model corresponding to the target intent to capture user needs, and conduct intelligent analysis through multiple rounds of interactive feedback with users to determine the most accurate target recommendation content.

[0195] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario and performed by multiple devices working together. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the method.

[0196] It should be noted that the above description is limited to some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in an order different from that described in the above embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0197] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application also provides an information recommendation device.

[0198] refer to Figure 3 , the device comprises:

[0199] The intent complexity classification module 201 is configured to receive user demand information and send the user demand information to a pre-trained intent complexity classification model to obtain a complexity classification result;

[0200] The intent analysis module 202 is configured to determine a corresponding target intent recognition model based on the complexity classification result, and use the target intent recognition model to analyze the user intent information to determine the target intent, wherein the target intent recognition model is a pre-trained model;

[0201] The recommended content determination module 203 is configured to select a target recommendation model corresponding to the target intent, determine target recommended content using the target recommendation model, and feed back the target recommended content to the user, wherein the recommendation model corresponding to each intent is stored.

[0202] In some embodiments, the complexity classification result is a simple class or a complex class;

[0203] The intention analysis module 202 is specifically configured to:

[0204] In response to determining that the complexity classification result is a complex class, determining that the corresponding target intent recognition model is a first intent recognition model, receiving multiple rounds of conversation results with the user, and using the first intent recognition model to recognize the multiple rounds of conversation results to determine the target intent;

[0205] or,

[0206] In response to determining that the complexity classification result is a simple class, determining that the corresponding target intent recognition model is a second intent recognition model, and using the second intent recognition model to identify the user intent information to determine the target intent.

[0207] In some embodiments, the intent analysis module 202 is further configured to:

[0208] Analyzing the user intention information using the first intention recognition model to obtain a first analysis result and output it, so that the user can provide feedback based on the output first analysis result;

[0209] receiving user feedback information, analyzing the user feedback information using the first intent recognition model to obtain a pending intent type and outputting the result, so that the user can determine whether the outputted pending intent type is correct;

[0210] In response to receiving user feedback that the pending intent type is incorrect, iteratively performing a correction process on the pending intent type using the first intent recognition model until the corrected intent type is correct, and using the corrected intent type as the target intent; or

[0211] In response to receiving user feedback that the pending intent type is correct, the pending intent type is used as the target intent.

[0212] In some embodiments, the recommended content determination module 203 is specifically configured to:

[0213] Obtaining user behavior information using a target recommendation model, and determining a first predetermined number of corresponding content to be recommended based on the user behavior information;

[0214] Using the target recommendation model, the first predetermined number of contents to be recommended are sorted and adjusted according to the degree of matching with the user's needs to obtain a sorting result;

[0215] A target recommendation model is used to select a second predetermined number of to-be-recommended contents with the highest matching degree from the ranking results as the target recommended contents, wherein the second predetermined number is smaller than the first predetermined number.

[0216] In some embodiments, the recommended content determination module 203 is further configured to:

[0217] Determining similar users based on the user's behavior information, wherein similar users corresponding to each user are pre-stored, and the similar users are other users matching the user's behavior information;

[0218] Obtaining operation behaviors of the similar users within a predetermined time period, and determining multiple trigger contents according to the operation behaviors;

[0219] Filtering the plurality of triggering contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of triggering contents;

[0220] and / or,

[0221] Determining trigger content based on the user's behavior information;

[0222] Acquire a plurality of similar contents corresponding to the triggering content, wherein similar contents corresponding to each triggering content are pre-stored;

[0223] Filtering the plurality of similar contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of similar contents;

[0224] and / or,

[0225] Determining interest tags based on the user's behavior information;

[0226] A plurality of matching contents corresponding to the interest tag is searched, the plurality of matching contents are filtered, and a first predetermined number of to-be-recommended contents are screened out from the filtered plurality of matching contents.

[0227] In some embodiments, the recommended content determination module 203 is further configured to:

[0228] Obtaining user feature data, inputting the user feature data and a first predetermined number of to-be-recommended content into a click-through rate prediction model, determining a predicted click-through rate for each to-be-recommended content, and using the predicted click-through rate as a degree of match between the to-be-recommended content and the user's needs;

[0229] The first predetermined number of contents to be recommended are sorted and adjusted according to the matching degree to obtain a sorting result.

[0230] In some embodiments, the apparatus further comprises: a model training module configured to:

[0231] Acquire user feature data, historical content, and historical click data corresponding to the historical content, combine the feature data with the historical content, and use the historical click data for labeling to obtain training samples;

[0232] Using the training samples to train the constructed initial model, outputting click-through rate training results corresponding to the training samples, and adjusting the parameters of the initial model according to the degree of difference between the click-through rate training results and the marked historical click data;

[0233] After all training samples are trained, the final adjusted initial model is used as the click-through rate prediction model.

[0234] In some embodiments, the apparatus further includes: a modification and adjustment module configured to:

[0235] After feeding back the target recommended content to the user, in response to receiving user feedback indicating dissatisfaction with the target recommended content, iteratively executing an interactive dialogue process with the user using the target recommendation model, and determining new target recommended content based on the dialogue content, until the user makes a selection based on the new target recommended content and the iteration stops;

[0236] After the target recommended content is fed back to the user, in response to receiving the execution content selected by the user from the target recommended content or from the new target recommended content, a corresponding function is executed according to the execution content.

[0237] For the convenience of description, the above devices are described as being divided into various modules according to their functions. Of course, when implementing this application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0238] The apparatus of the above embodiment is used to implement the corresponding method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0239] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any of the above embodiments when executing the computer program.

[0240] Figure 4 10 is a schematic diagram showing a more specific hardware structure of an electronic device provided in this embodiment. The device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are communicatively connected to each other within the device via the bus 1050.

[0241] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0242] The memory 1020 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, etc. The memory 1020 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0243] The input / output interface 1030 is used to connect input / output modules to implement information input and output. The input / output modules can be configured as components within the device (not shown in the figure) or can be externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, various sensors, etc., and output devices may include a display, speaker, vibrator, indicator light, etc.

[0244] The communication interface 1040 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via a wired method (such as USB, network cable, etc.) or a wireless method (such as mobile network, WiFi, Bluetooth, etc.).

[0245] The bus 1050 comprises a path for transmitting information between the various components of the device (eg, the processor 1010 , the memory 1020 , the input / output interface 1030 , and the communication interface 1040 ).

[0246] It should be noted that although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040, and the bus 1050, in a specific implementation, the device may also include other components necessary for normal operation. In addition, it will be understood by those skilled in the art that the above device may only include the components necessary to implement the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0247] The electronic device of the above embodiment is used to implement the corresponding method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0248] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method described in any of the above embodiments.

[0249] The computer-readable media of this embodiment include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM, Parameter Random Access Memory), static random access memory (SRAM, Static Random-Access Memory), dynamic random access memory (DRAM, Dynamic Random Access Memory), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM, Electrically Erasable Programmable read only memory), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM, Compact Disc Read-Only Memory), digital versatile disc (DVD, Digital Video Disc) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0250] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0251] Based on the same concept, corresponding to any of the above-mentioned embodiments, the present application also provides a computer program product, including computer program instructions. When the computer program instructions are run on a computer, the computer executes the method described in any of the above embodiments, which has the beneficial effects of the corresponding method embodiments and will not be repeated here.

[0252] Based on the same inventive concept, the present application further provides a vehicle including the vehicle-mounted charging device or the electronic device described in the above embodiments. The beneficial effects of the corresponding vehicle-mounted charging device or electronic device embodiments are not further described here.

[0253] It is understandable that before using the technical solutions of each embodiment of this application, the type, scope of use, usage scenarios, etc. of the personal information involved will be informed to the user in an appropriate manner, and the user's authorization will be obtained.

[0254] For example, in response to receiving a user's active request, a prompt message is sent to the user to clearly inform the user that the operation requested will require the acquisition and use of the user's personal information. Thus, the user can independently choose whether to provide personal information to the electronic device, application, server, storage medium, or other software or hardware that performs the operation of the technical solution of this application based on the prompt message.

[0255] As an optional but non-limiting implementation, in response to a user's active request, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. Furthermore, the pop-up window may also contain a selection control for the user to select "agree" or "disagree" to provide personal information to the electronic device.

[0256] It is understandable that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0257] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. Within the scope of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0258] In addition, for simplicity of description and discussion, and in order not to make the embodiment of the application difficult to understand, the known power supply / ground connection with integrated circuit (IC) chip and other components may or may not be shown in the accompanying drawings provided. In addition, the device can be shown in the form of a block diagram to avoid making the embodiment of the application difficult to understand, and this also takes into account the following fact, that is, the details of the embodiment of these block diagram devices are highly dependent on the platform to be implemented in the embodiment of the application (that is, these details should be fully within the scope of understanding of those skilled in the art). When specific details (for example, circuit) are set forth to describe exemplary embodiments of the application, it will be apparent to those skilled in the art that the embodiment of the application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered to be illustrative rather than restrictive.

[0259] Although the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may utilize the embodiments discussed.

[0260] The embodiments of the present application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of this application.

Claims

1. An information recommendation method, characterized in that: include: Receiving user demand information, and sending the user demand information to a pre-trained intent complexity classification model to obtain a complexity classification result; Determine a corresponding target intent recognition model based on the complexity classification result, and use the target intent recognition model to analyze the user intent information to determine the target intent, wherein the target intent recognition model is a pre-trained model; A target recommendation model corresponding to the target intention is selected, target recommendation content is determined using the target recommendation model, and the target recommendation content is fed back to the user.

2. The method according to claim 1, characterized in that The complexity classification result is a simple class or a complex class; Determining a corresponding target intent recognition model according to the complexity classification result, and analyzing the target intent based on the user intent information using the target intent recognition model to determine the target intent includes: In response to determining that the complexity classification result is a complex class, determining that the corresponding target intent recognition model is a first intent recognition model, receiving multiple rounds of conversation results with the user, and using the first intent recognition model to recognize the multiple rounds of conversation results to determine the target intent; or, In response to determining that the complexity classification result is a simple class, determining that the corresponding target intent recognition model is a second intent recognition model, and using the second intent recognition model to identify the user intent information to determine the target intent.

3. The method according to claim 2, characterized in that The receiving of multiple rounds of conversation results with the user, and identifying the multiple rounds of conversation results using the first intent recognition model to determine the target intent includes: Analyzing the user intention information using the first intention recognition model to obtain a first analysis result and output it, so that the user can provide feedback based on the output first analysis result; Receiving user feedback information, analyzing the user feedback information using the first intent recognition model to obtain a pending intent type and outputting the result, so that the user can determine whether the outputted pending intent type is correct; In response to receiving user feedback that the pending intent type is incorrect, iteratively performing a correction process on the pending intent type using the first intent recognition model until the corrected intent type is correct, and using the corrected intent type as the target intent; or In response to receiving user feedback that the pending intent type is correct, the pending intent type is used as the target intent.

4. The method according to claim 1, wherein The method of using the target recommendation model to determine the target recommendation content required by the user includes: Using the target recommendation model to perform: Acquiring user behavior information, and determining a first predetermined number of corresponding content to be recommended based on the user behavior information; Sorting and adjusting the first predetermined number of contents to be recommended according to their matching degree with the user's needs to obtain a ranking result; A second predetermined number of to-be-recommended contents with the highest matching degree is selected from the sorting results as the target recommended contents, wherein the second predetermined number is smaller than the first predetermined number.

5. The method according to claim 4, characterized in that The determining of a first predetermined number of contents to be recommended corresponding to the user's behavior information includes: Determining similar users based on the user's behavior information, wherein similar users corresponding to each user are pre-stored, and the similar users are other users matching the user's behavior information; Obtaining operation behaviors of the similar users within a predetermined time period, and determining multiple trigger contents according to the operation behaviors; Filtering the plurality of triggering contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of triggering contents; and / or, Determining trigger content based on the user's behavior information; Acquire a plurality of similar contents corresponding to the triggering content, wherein similar contents corresponding to each triggering content are pre-stored; Filtering the plurality of similar contents, and selecting a first predetermined number of contents to be recommended from the filtered plurality of similar contents; and / or, Determining interest tags based on the user's behavior information; A plurality of matching contents corresponding to the interest tag is searched, the plurality of matching contents are filtered, and a first predetermined number of to-be-recommended contents are screened out from the filtered plurality of matching contents.

6. The method according to claim 4, characterized in that The step of sorting and adjusting the first predetermined number of contents to be recommended according to the degree of matching with the user's needs to obtain a sorting result includes: Obtaining user feature data, inputting the user feature data and a first predetermined number of to-be-recommended content into a click-through rate prediction model, determining a predicted click-through rate for each to-be-recommended content, and using the predicted click-through rate as a measure of the degree of match between the to-be-recommended content and the user's needs; The first predetermined number of contents to be recommended are sorted and adjusted according to the matching degree to obtain a sorting result.

7. The method according to claim 6, characterized in that The training process of the click-through rate prediction model includes: Acquire user feature data, historical content, and historical click data corresponding to the historical content, combine the feature data with the historical content, and use the historical click data for labeling to obtain training samples; Using the training samples to train the constructed initial model, outputting click-through rate training results corresponding to the training samples, and adjusting the parameters of the initial model according to the degree of difference between the click-through rate training results and the marked historical click data; After all training samples are trained, the final adjusted initial model is used as the click-through rate prediction model.

8. The method according to claim 1, characterized in that After feeding back the target recommended content to the user, the method further includes: In response to receiving user feedback indicating dissatisfaction with the target recommended content, iteratively executing an interactive dialogue process with the user using the target recommendation model, and determining new target recommended content based on the dialogue content, until the user makes a selection based on the new target recommended content and the iteration stops; In response to receiving the execution content selected by the user from the target recommended content or from the new target recommended content, a corresponding function is executed according to the execution content.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

10. A vehicle, characterized in that: The electronic device comprising claim 9.

Images