Information recommendation method and electronic equipment

Abstract

The invention discloses an information recommendation method and electronic equipment, and relates to the field of information processing, and the method comprises the steps: recognizing a first user intention in first session information; based on the first user intention, determining first candidate recommendation information and a first confidence coefficient of the first user intention; the first confidence is used for representing a target parameter of a target object indicated by the first user intention in at least one dimension; and in response to the condition that the confidence is greater than a preset confidence threshold, determining target recommendation information corresponding to the target object from the first candidate recommendation information.

Description

Technical Field

[0001] This application relates to the field of information processing, and more particularly to an information recommendation method and related apparatus. Background Technology

[0002] Current recommendation systems, including dialogue systems with rule engines, can analyze user conversations. However, because users typically express their needs very briefly, vaguely, and subjectively, current recommendation systems struggle to identify clear intents.

[0003] Because existing recommendation systems analyze and make recommendations based on each user's input independently, they struggle to deeply analyze ambiguous intents in dynamic scenarios and to create a continuous, collaborative intelligent interaction between the user's dialogue and the recommendation process. This results in too many interaction rounds to achieve a recommendation, leading to inefficiency and a poor user experience. Summary of the Invention

[0004] The first aspect of this application provides an information recommendation method, including:

[0005] Identify the first user intent in the first session information;

[0006] Based on the first user intent, a first candidate recommendation information and a first confidence level of the first user intent are determined; the first confidence level is used to characterize the target object indicated by the first user intent in at least one dimension as a target parameter.

[0007] In response to the confidence level being greater than a preset confidence threshold, target recommendation information corresponding to the target object is determined from the first candidate recommendation information.

[0008] In one possible implementation, in response to the first confidence level being less than or equal to the preset confidence threshold, the user is prompted to provide second session information;

[0009] Identify the second user intent in the second session information;

[0010] Based on the second user intent, a second confidence level of the second user intent is determined; the second confidence level is used to characterize the target object indicated by the second user intent in at least one dimension as a target parameter.

[0011] Based on the second confidence level, target recommendation information corresponding to the target object is determined.

[0012] In one possible implementation, determining the target recommendation information corresponding to the target object based on the second confidence level includes:

[0013] If the dimensions included in the second confidence level are of the same type as the dimensions included in the first confidence level, second candidate recommendation information is obtained from the first candidate recommendation information based on the second user intent; in response to the second confidence level being greater than the preset confidence threshold, target recommendation information corresponding to the target object is determined from the second candidate recommendation information; or

[0014] If the dimensions included in the second confidence level are of a different type than those included in the first confidence level, a third candidate recommendation information is obtained by filtering from a preset knowledge graph based on the first user intent and the second user intent; in response to the second confidence level being greater than a preset confidence threshold, the target recommendation information corresponding to the target object is determined from the third candidate recommendation information.

[0015] In one possible implementation, determining the first candidate recommendation information based on the first user intent includes:

[0016] Extract the corresponding demand keywords from the first user intent;

[0017] Based on the required keywords, the first candidate recommendation information is obtained by filtering in a preset knowledge graph; the preset knowledge graph includes a set of candidate objects in the target scenario and dimensional information and user evaluation information associated with the set of candidate objects; the first candidate recommendation information includes at least one candidate object in the target scenario and dimensional information associated with the at least one candidate object.

[0018] In one possible implementation, determining a first confidence level of the first user intent based on the first user intent includes:

[0019] Extract the corresponding demand keywords from the first user intent;

[0020] Using a confidence level detection model, determine the confidence level of the required keywords in at least one dimension;

[0021] The first confidence level is calculated based on the confidence level corresponding to each demand keyword.

[0022] In one possible implementation, calculating the first confidence level based on the confidence level corresponding to each demand keyword includes at least one of the following:

[0023] The frequency of keywords corresponding to at least one dimension in the first conversation information is determined; based on the frequency of keywords corresponding to the at least one dimension, the weight of the demand keyword in the at least one dimension is determined; based on the confidence and weight of each demand keyword, the first confidence is calculated; the frequency is positively correlated with the weight; the keyword is the word corresponding to the demand keyword.

[0024] The target parameter and session round corresponding to at least one dimension in the first session information are determined; based on the target parameter and the session round, the weight corresponding to the demand keyword in at least one dimension is determined; based on the confidence and weight corresponding to each demand keyword, the first confidence score is calculated; the target parameter is negatively correlated with the weight, and the session round is negatively correlated with the weight.

[0025] Based on the type of dimension in the first user intent, a preset weight value is assigned, and the preset weight value is determined as the weight of the demand keyword in at least one dimension; based on the confidence and weight corresponding to each demand keyword, the first confidence is calculated.

[0026] In one possible implementation, prompting the user to provide second session information includes at least one of the following:

[0027] If the first dimension of the preset dimension type is missing from the first confidence level, the user is prompted to provide the requirement information corresponding to the first dimension.

[0028] If the confidence level of the second dimension in the first confidence level is less than the target confidence level threshold, the user is prompted to provide the required information corresponding to the second dimension.

[0029] In one possible implementation, the method further includes:

[0030] Identify the current recommendation scenario; based on the current recommendation scenario, determine the target confidence threshold corresponding to at least one dimension; or

[0031] In response to the absence of a target object in the preset knowledge graph that corresponds to the first user intent, a third user intent is generated, and the recommended object and / or related information corresponding to the third user intent are output; the third user intent is generated after adjusting the variable values ​​of the target object in at least one dimension.

[0032] A second aspect of this application provides an electronic device, including at least one processor and a target application running on the processor, wherein the target application is capable of independently executing or invoking at least one artificial intelligence model to perform the following operations: identifying a first user intent in first session information; determining candidate recommendation information and a confidence level of the first user intent based on the first user intent; the confidence level being used to characterize a target parameter of a target object indicated by the first user intent in at least one dimension; and determining target recommendation information corresponding to the target object from the candidate recommendation information in response to the confidence level being greater than a preset confidence threshold.

[0033] A display device is used to display the target recommendation information.

[0034] In one possible implementation, the first session information belongs to the session information in the product recommendation scenario;

[0035] The target application is used to determine a candidate product recommendation list in a preset knowledge graph and a first confidence level of the first user intent based on a first user intent; the first confidence level is used to characterize the target product indicated by the first user intent in at least one dimension; the preset knowledge graph includes a set of candidate products in the product recommendation scenario and dimension information and user evaluation information associated with the set of candidate products; the candidate product recommendation list includes at least one candidate product in the product recommendation scenario and dimension information associated with the at least one candidate product.

[0036] In response to the first confidence level being greater than the preset confidence threshold, target recommendation information corresponding to the target product is determined from the candidate product recommendation list.

[0037] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the information recommendation method described in the first aspect or any implementation thereof.

[0038] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs that, when executed by an electronic device, enable the electronic device to perform the information recommendation method described in the first aspect or any implementation thereof. Attached Figure Description

[0039] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0040] Figure 1 This is a flowchart illustrating an information recommendation method provided in an embodiment of this application;

[0041] Figure 2 This is another flowchart illustrating an information recommendation method provided in an embodiment of this application;

[0042] Figure 3 This is a flowchart illustrating the process of determining first candidate recommendation information based on a first user intent, provided in an embodiment of this application.

[0043] Figure 4 This is a flowchart illustrating the process of determining a first confidence level of a first user intent based on a first user intent, as provided in an embodiment of this application.

[0044] Figure 5 This is another flowchart illustrating the information recommendation method provided in the embodiments of this application;

[0045] Figure 6 This is a schematic diagram of an information recommendation method provided in an embodiment of this application;

[0046] Figure 7 This is a schematic diagram of the structure of an electronic device that applies an information recommendation method according to an embodiment of this application. Detailed Implementation

[0047] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0048] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0049] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0050] In relevant recommendation systems, the following technical solutions can be adopted: one is a rule-based dialogue system, which suffers from high manual maintenance costs and poor scalability; another is a static knowledge graph matching system, which, due to the use of a fixed database (static knowledge graph), cannot integrate dynamic data such as real-time user reviews; and a collaborative filtering recommendation algorithm, which, in its separate architecture for intent understanding and recommendation engine, requires an average of 7.2 rounds of dialogue and suffers from low accuracy in cold start scenarios. Furthermore, the traditional single-input mode struggles to understand ambiguous requests, resulting in insufficient intent ambiguity handling capabilities in the recommendation system.

[0051] To improve the accuracy of recommendations, existing recommendation systems need to be improved to understand fuzzy needs, incorporate real-time user feedback, and improve the accuracy of understanding fuzzy intents through incremental understanding.

[0052] Based on this, this application provides an information recommendation method that sets a confidence level parameter, which characterizes the target parameters of the target object indicated by the user's intent in at least one set dimension. The target parameters can represent the clarity of the user's intent in the corresponding dimension. First, a first user intent in the first session information is identified. Based on the first user intent, a first candidate recommendation and the confidence level of the first user intent are determined. This confidence level can represent the target parameters of the target object indicated by the user intent in the first session information in multiple dimensions. In response to the confidence level being greater than a preset confidence threshold, the target recommendation information corresponding to the target object is determined from the first candidate recommendation information. By quantifying the user intent in the user's session information with confidence, a quantitative representation of ambiguous intent is achieved. Then, the first candidate recommendation information is initially determined based on this user intent. Finally, when the confidence level is greater than the preset confidence threshold, the target recommendation information is determined from the first candidate recommendation information. Through these two screening steps, the range of recommendation information can be gradually narrowed. Moreover, by using a clearer user intent to determine the target recommendation information, the purpose of understanding ambiguous needs is achieved, enabling accurate recommendations.

[0053] Reference Figure 1 , Figure 1 This is a flowchart illustrating an information recommendation method provided in an embodiment of this application, such as... Figure 1 As shown in the embodiment of this application, an information recommendation method may include steps 101 to 103, which are described in detail below.

[0054] 101. Identify the first user intent in the first session information;

[0055] During a conversation, the conversation information carries the user's needs for the target object that they want to recommend. Accordingly, the conversation information can reflect the user's intent, which represents the user's needs for the target object.

[0056] In this application, the session information can be a conversation that a user inputs into an electronic device according to their needs. This session information can be text information entered via a keyboard and mouse, or voice information entered via a microphone, etc. This application does not limit the specific form of the session information.

[0057] In one possible implementation, the user intent contained in the input session information is obtained by using a pre-defined intent recognition model to identify the intent of the session information.

[0058] Accordingly, the first session information is input into the intent recognition model to identify the corresponding first user intent.

[0059] As an example, the first conversation message is "I want to buy a computer, can you recommend one with a larger screen?" The user intent that can be identified is: recommend a computer with a large screen.

[0060] In one possible implementation, semantic error correction can be performed on user session information. This semantic error correction may include identification of referential errors and quantification of descriptive representations.

[0061] In this context, the identification of referential errors involves determining the specific referential content by combining the context of the session information. For example, if the user session information is "want big", the specific referential content can be determined by combining the context (such as the previously discussed screen or storage).

[0062] This quantitative fuzzy expression quantifies a vague notion. For example, it converts "medium budget" into a specific range (5000-8000 yuan).

[0063] Subsequently, the first user's intent is determined using the first user's session information after semantic error correction.

[0064] 102. Based on the first user intent, determine the first candidate recommendation information and the first confidence level of the first user intent; the first confidence level is used to characterize the target parameters of the target object indicated by the first user intent in at least one dimension;

[0065] After determining the first user intent, the first candidate recommendation information can be determined based on the first user intent. The first candidate recommendation information is a number of recommendation information related to the first user intent.

[0066] In one possible implementation, a confidence detection model can be set up to determine the first confidence level of the first user intent.

[0067] The first session information can be the current round session or multiple round sessions in this exchange, including the current round session and historical multiple round sessions.

[0068] To improve the accuracy of confidence determination, this confidence detection model can perform confidence detection on the entire multi-round conversation to obtain the overall confidence of the conversation in this exchange after the current round.

[0069] This confidence detection model can employ a 3-layer Bi-LSTM (Bidirectional Long Short-Term Memory) + Attention mechanism network structure. The Bi-LSTM captures bidirectional semantic information of the conversation sequence, while the Attention mechanism focuses on key contextual information to combine multi-turn conversation information for a more accurate determination of user intent.

[0070] In one possible implementation, based on the first user intent, a portion of the information can be filtered from a massive amount of data to serve as the first candidate recommendation information.

[0071] The target object can be something the user wants to be recommended to, such as electronic products, travel routes, decoration companies, real estate developments, applications, or clothing combinations.

[0072] As an example, when the target of the recommendation is electronic products, it can include five dimensions: scenario, specifications, price, appearance, and upgrades.

[0073] As an example, when the target of the recommendation is furniture, it can include three dimensions: size, material, and style; when the target of the recommendation is travel services, it can include three dimensions: destination, budget, and number of trip days.

[0074] As an example, when the target object is a computer, based on the first user intent, 100 items can be filtered out from a massive amount of information related to computers as the first candidate recommendation information.

[0075] The number of dimensions can be set according to the actual situation; it can be one dimension or multiple dimensions. To improve the accuracy of recommendations, multiple dimensions are generally set, such as three, five, or even more dimensions.

[0076] Confidence can be used to represent the target parameter of the identified user intent in at least one dimension. This target parameter can represent the clarity (or explicitness) of the user intent. The higher the clarity of the user intent across multiple dimensions, the higher the confidence level of that user intent.

[0077] The level of clarity can be determined by whether the corresponding dimension provides an accurate numerical range. If only a wide range of values ​​is provided, the clarity is considered low. If a more precise range of values ​​is provided, the clarity is considered high.

[0078] As an example, regarding the price dimension, ranges like "expensive" and "high-priced" are unclear and can be considered to have low clarity, while "200 to 300 yuan" can be considered to have high clarity. Similarly, regarding the price dimension, "100 to 10,000 yuan" is too broad a range and can also be considered unclear, thus having low clarity, while "200 to 300 yuan" is relatively precise and can be considered to have high clarity.

[0079] Moreover, the confidence level of the user intent is determined by the clarity of the user intent from multiple dimensions. If the clarity of each dimension is high, the confidence level of the user intent will be higher. If the clarity of each dimension is low, the confidence level of the user intent will be lower. If the clarity of one or more dimensions is low, it will also lead to a low overall confidence level of the user intent.

[0080] 103. In response to a confidence level greater than a preset confidence threshold, determine the target recommendation information corresponding to the target object from the first candidate recommendation information.

[0081] A pre-set reliability threshold can determine whether the user's session information clearly represents their needs for the target object.

[0082] If the confidence level of the first user intent is greater than the preset confidence threshold, indicating that the first session information can clearly represent the user's need for the target object, then the target recommendation information corresponding to the target object is determined from the first candidate recommendation information.

[0083] The recommended information for this target can be one piece of information, or two or more, depending on the actual settings.

[0084] As an example, if the confidence level of the first user intent is 0.85 and the preset confidence threshold is 0.8, then the target recommendation information can be determined from the first candidate information.

[0085] Since the first candidate recommendation information is determined based on the first user's intent, it can clearly represent the user's needs for the target object. Therefore, each candidate recommendation information in the first candidate recommendation information is close to the user's needs for the target object. Thus, determining the target recommendation information corresponding to the target object from the first candidate recommendation information can achieve a high recommendation accuracy.

[0086] Different dimensions can be defined for different target objects to achieve targeted recommendations and improve recommendation accuracy.

[0087] As an example, the target object is a laptop. The first candidate recommendation information is selected from multiple candidate laptops corresponding to the user's intent reflected in the first session information, and one or more of them are selected as the target recommendation information to recommend a laptop that meets the user's needs.

[0088] As an example, the target object is a travel guide. The first candidate recommendation information is selected from multiple candidate guides corresponding to the user's intent reflected in the first conversation information, and one or more of them are selected as the target recommendation information to recommend a travel guide that meets the user's needs.

[0089] In one possible implementation, when the confidence level is greater than a preset confidence threshold, values ​​can be assigned to the corresponding dimensions based on the content of the conversation information. For example, if the conversation information clearly states that the weight is within 2 kilograms and the scene is an office setting, parameters can be assigned to the corresponding weight and scene dimensions to obtain "weight = within 2 kilograms" and "scene = office scene". Using the dimensions with assigned parameters, the target recommendation information can be determined from the first candidate information.

[0090] The preset reliability threshold value can be adjusted according to the needs of the scenario.

[0091] In one possible implementation, filtering rules can be set to determine the target recommendation information corresponding to the target object from the first candidate recommendation information. This preset filtering rule can be the order of the filtering dimensions.

[0092] For example, following the order of dimensions 1 to N, first filter the first candidate recommendation information to obtain candidate information that meets the requirements of dimension 1, then further filter the candidate information that meets the requirements of dimension 1 to obtain candidate information that meets the requirements of dimension 2, and so on, until the target candidate information that meets dimensions 1 to N is finally obtained.

[0093] As an example, the three dimensions are price, performance, and portability. First, based on the user's input budget or the price range obtained through clarification, products that meet the price criteria are filtered out. Then, based on the user's required specifications (such as CPU model, graphics card performance, storage capacity, etc.), products that do not meet the performance threshold are excluded, resulting in an initial candidate set. The products that pass the first two steps of filtering (the initial candidate set) are sorted according to portability indicators (such as weight, thickness), and products with high portability are recommended first.

[0094] As an example, when this dimension involves price, users are generally more sensitive to price. For expensive items, users need a clearer intent in their conversation information. Therefore, a higher confidence threshold can be set for more expensive targets to ensure that the final recommended information better matches the user's needs, while a lower confidence threshold can be set for cheaper targets to ensure that the final recommended information matches the user's needs as closely as possible.

[0095] In this embodiment, a confidence level parameter is set, which characterizes the target parameters of the target object indicated by the user's intent in at least one set dimension. These target parameters can represent the clarity of the user's intent in the corresponding dimension. First, a first user intent in the first session information is identified. Based on this first user intent, a first candidate recommendation and the confidence level of the first user intent are determined. This confidence level can represent the target parameters of the target object indicated by the user intent in the first session information in multiple dimensions. In response to the confidence level being greater than a preset confidence threshold, the target recommendation information corresponding to the target object is determined from the first candidate recommendation information. By quantifying the user intent in the user's session information with confidence, the first candidate recommendation information is initially determined based on this user intent. When the confidence level is greater than the preset confidence threshold, the target recommendation information is determined from the first candidate recommendation information. Through this two-stage selection process, the range of recommendation information can be gradually narrowed. Moreover, by using a clearer user intent to determine the target recommendation information, accurate recommendations can be achieved.

[0096] Figure 2 This is another flowchart illustrating an information recommendation method provided in an embodiment of this application. The steps in this flowchart are parallel to step 103 and may include steps 201 to 204. These steps will be described in detail below.

[0097] 201. In response to the first confidence level being less than or equal to a preset confidence threshold, prompt the user to provide second session information;

[0098] If the first confidence level of the first user intent is less than or equal to the preset confidence threshold, then the clarity of the user intent expressed in the first session information is not high. If target recommendation information is selected based on the first user intent, it may be significantly different from the user's actual needs. Therefore, to more clearly determine the user's needs, the user is prompted to provide second session information so that the user can provide clearer information about their needs.

[0099] Prompting users to provide second-session information guides them to provide more specific and clearer information about their needs. This allows for a higher degree of confidence in determining the user's intent using this second-session information, leading to more accurate recommendations.

[0100] This prompts the user to provide a second set of conversational information. This could be providing more information on dimensions with lower confidence levels or providing information on other dimensions to improve the clarity of the user's intent in the conversational information.

[0101] 202. Identify the second user intent in the second session information;

[0102] Continue to identify the second session information to obtain the corresponding second user intent, which can increase the clarity of the content compared to the first user intent.

[0103] The added content could be either increased clarity for existing dimensions or added content for other dimensions, thereby improving the overall confidence in determining user intent based on conversation information.

[0104] In one possible implementation, the second user intent in the second session information is obtained by using a preset intent recognition model to identify the intent.

[0105] 203. Based on the second user intent, determine the second confidence level of the second user intent; the second confidence level is used to characterize the target parameters of the target object indicated by the second user intent in at least one dimension;

[0106] The second confidence level of the second user intent can be determined using a confidence level detection model.

[0107] The process of determining the second confidence level for the second user intent can be the same as the method described above for determining the first confidence level for the first user intent, for example, the same confidence level detection model can be used.

[0108] If the second session information provides clearer demand information for a certain dimension, then the dimension type corresponding to the second user intent can be the same as the dimension type corresponding to the first user intent; if the second session information provides demand information for a certain dimension not covered by the first session information, then the dimension type corresponding to the second user intent can be different from the dimension type corresponding to the first user intent.

[0109] In one possible implementation, the overall user intent encompassing both sessions can be obtained by combining the first user intent identified from the first session information and the second user intent identified from the second session information, thereby improving the clarity of the identified user intent.

[0110] 204. Based on the second confidence level, determine the target recommendation information corresponding to the target object.

[0111] After determining the second confidence level, the target recommendation information corresponding to the target object can be determined using this second confidence level.

[0112] The second confidence level is determined progressively based on the second session information. It has a higher confidence level than the first confidence level and is closer to the user's needs for the target object.

[0113] Compared to the first confidence level, this second confidence level identifies more dimensions of demand, or perhaps more detailed requirements within the same dimension. Accordingly, different processes for determining target recommendation information can be employed for different situations.

[0114] In one possible implementation, the target recommendation information corresponding to the target object is determined based on the second confidence level, including either of the following two cases:

[0115] As an exemplary embodiment, if the dimensions included in the second confidence level are of the same type as the dimensions included in the first confidence level, second candidate recommendation information is obtained by filtering from the first candidate recommendation information based on the second user intent; in response to the second confidence level being greater than a preset confidence threshold, target recommendation information corresponding to the target object is determined from the second candidate recommendation information.

[0116] If the dimensions included in the second confidence level are of the same type as those included in the first confidence level, since the second session information is a supplement to the requirement information already provided in the first session information, the second session information adds more detailed requirement information to the determined dimensions. Accordingly, the second confidence level may be higher than the first confidence level. Consequently, the second confidence level can be used to filter the second candidate information from the first candidate information to obtain the second candidate information.

[0117] Based on the second user intent, second candidate recommendations are first obtained from the first candidate recommendations. This process narrows down the selection by filtering the first candidate recommendations, which have relatively low demand clarity, according to the second user intent, which has higher demand clarity. Then, if the second confidence level is greater than a preset confidence threshold, the target recommendation is determined from the second candidate recommendations.

[0118] In one possible implementation, if the second confidence level is not greater than the preset confidence threshold, step 201 can be returned until the conditions for determining the target recommendation information are met.

[0119] In one possible implementation, the condition for achieving the target recommendation information could be that the confidence level of the determined user intent exceeds a preset confidence threshold, or that the number of conversation rounds reaches a set round threshold. Preset default recommendation parameters can be used to determine the target recommendation information. The goal is to collect demand information from various dimensions within a set number of conversation rounds, synthesize the demands from these dimensions to determine the user intent, thereby obtaining a higher confidence level for the user intent and achieving recommendation information that better matches the user's needs.

[0120] As another exemplary embodiment, if the dimensions included in the second confidence level are of a different type than the dimensions included in the first confidence level, a third candidate recommendation information is obtained by filtering from a preset knowledge graph based on the first user intent and the second user intent; in response to the second confidence level being greater than a preset confidence threshold, the target recommendation information corresponding to the target object is determined from the third candidate recommendation information.

[0121] If the dimensions included in the second confidence level are of a different type than those included in the first confidence level, since the second session information is a supplement to the demand information already provided in the first session information, the second session information is for the dimensions for which demand information has not yet been collected in the first session information. If demand information has been collected, the second confidence level may be higher than the first confidence level.

[0122] Since the second user intent corresponding to the second session information contains different dimensions of confidence compared to the first user intent corresponding to the first session information, the third candidate recommendation information can be obtained by combining the first user intent and the second user intent in the preset knowledge graph.

[0123] This preset knowledge graph is a knowledge base designed to provide users with various recommendation information. The information in this knowledge base is stored in the form of a knowledge graph. For example, the preset knowledge graph may include relevant knowledge about electronic products, travel routes, decoration companies, real estate developments, applications, and clothing matching. Moreover, the preset knowledge graph records the parameter values ​​of each recommendation information in each dimension, which can be recorded in the form of key-value pairs.

[0124] For example, for tour route 1 in the recommended information, the recorded content in three dimensions includes "Destination = Forbidden City", "Budget = 500 yuan", and "Number of trip days = 2 days".

[0125] Correspondingly, the first candidate recommendation information can be obtained by filtering the preset knowledge graph based on the first user's intent.

[0126] Based on the second confidence level of the second user intent, the target recommendation information is further determined from the third candidate recommendation information. This achieves the goal of combining multi-round dialogue information, filtering candidate recommendation information in a preset knowledge graph, and then accurately locating the target recommendation information from the candidate recommendation information, so as to achieve recommendation information that better meets the user's needs in various dimensions.

[0127] In one possible implementation, the number of rounds of user-provided sessions can also be counted. If the number of rounds exceeds the set round threshold but still fails to reach the preset reliability threshold, the current application scenario can be determined based on the aforementioned multi-point session information, and the target recommendation information can be determined using the default recommendation parameters of the application scenario.

[0128] Regarding this conversation recommendation, multiple conversations have reached the round threshold, yet the established confidence level still does not exceed the preset confidence threshold. This failure to reach the preset confidence threshold could be due to various reasons, such as the user's conversation information being vague, or the user not explicitly selecting an option when presented with multiple choices. It can be determined that the user can only provide a vague answer and cannot provide more precise information about their needs. To improve recommendation efficiency, targeted recommendation information can be directly provided to the user.

[0129] As an example, for the current recommendation scenario (e.g., recommending laptops), if the threshold for this round is 4 times, and the number of sessions has reached 4, but the confidence level of the user's intent is still less than the preset confidence threshold, recommendations can be made using known user needs. The determined user need is a desire for a thin and light laptop. Based on the pre-stored correspondence of default recommendation parameters in various application scenarios, the default recommendation parameter for the thin and light laptop recommendation scenario is determined to be a weight of 1.5 kg. In the first candidate recommendation information, laptops weighing 1.5 kg or less are models A and B. Correspondingly, the generated target recommendation information is a 1.5 kg model A laptop and a 1.2 kg model B laptop. Alternatively, recommendations satisfying the default recommendation parameter (1.5 kg) can be directly filtered from the preset knowledge graph for this laptop recommendation scenario. This would result in laptops weighing 1.5 kg or less being models A, B, and C, and the generated target recommendation information would include a 1.2 kg model B laptop and a 1.4 kg model C laptop.

[0130] It should be noted that after receiving the second session information, it can be used as the new first session information, or the second session information and the previous first session information can be combined to form a new first session information. The confidence level of the user intent in this new first session information is then judged to be greater than a preset confidence threshold. Based on the judgment result, it is determined whether to select the target recommendation information from the candidate recommendation information or to prompt the user to provide more session information. This process is repeated every time new session information is received, calculating the confidence level of the user intent corresponding to the session information, generating candidate recommendation information, etc., until the target recommendation information is obtained. The process of processing the new first session information can be referred to the explanation in the subsequent embodiments, and will not be detailed here.

[0131] In this embodiment, when the first confidence level is less than or equal to a preset confidence threshold, the user is prompted to provide second session information to allow the user to provide more demand information. The second user intent in the second session information is identified. Based on the second user intent, a second confidence level is determined. The second confidence level is used to characterize the target parameters of the target object indicated by the second user intent in at least one dimension. Based on the second confidence level, target recommendation information corresponding to the target object is determined. If the first session information does not contain sufficient demand information, resulting in inaccurate recommendations, the user is prompted to provide second session information to allow the user to provide more demand information. The second confidence level of the second user intent is determined in conjunction with the provided second session information, and then target recommendation information is determined based on the second confidence level. This achieves multi-round user intent determination by combining the context of the session. Combining the user intent determined by multiple rounds of sessions enables incremental intent understanding, allowing for more accurate determination of user intent and achieving the goal of accurate recommendations.

[0132] Figure 3 This is a flowchart illustrating the process of determining first candidate recommendation information based on a first user intent, provided in an embodiment of this application. It may include steps 301 to 302, which are described in detail below.

[0133] 301. Extract the corresponding demand keywords from the first user intent;

[0134] After determining the first user intent based on the first session information, demand keywords that represent user needs are extracted from the first user intent.

[0135] The keyword for this requirement can be a word contained in the user's intent, or a word that corresponds to the user's intent.

[0136] In one possible implementation, the current scenario can be determined based on the user's intent, and the corresponding demand keywords can be determined using the scenario.

[0137] As an example, if the user's requirement is determined to be "playing 'Game A'", which could be a large online game, the corresponding requirement keyword extracted is "graphics card performance".

[0138] 302. Based on the keywords of the demand, filter in the preset knowledge graph to obtain the first candidate recommendation information;

[0139] The preset knowledge graph includes a set of candidate objects in the target scenario and dimensional information and user evaluation information associated with the set of candidate objects. The first candidate recommendation information includes at least one candidate object in the target scenario and dimensional information associated with at least one candidate object.

[0140] After determining the keywords for the requirement, the information is filtered in the preset knowledge graph to obtain the first candidate information corresponding to the keywords for the requirement.

[0141] The set of candidate objects in the preset knowledge graph can be a set of several candidate objects for the target object in the corresponding target scenario. Moreover, the set of candidate objects also stores the dimensional information and user evaluation information associated with each candidate object.

[0142] The dimensional information associated with the candidate object can be its static parameters, which can be inherent product attribute data. This data is updated in real time, and the preset knowledge graph is also updated in real time. User evaluation information can include dynamically updated information such as real-time user reviews and feedback, which is also updated in real time.

[0143] This pre-defined knowledge graph can be obtained by fusing the product's static parameters (metadata) and user evaluation information.

[0144] As an example, for electronic products, this static parameter may include attributes such as the product's CPU (central processing unit), weight, and price. As electronic products are updated, this static parameter will also be updated.

[0145] As an example, for after-sales service, the static parameters may include the model of the product being repaired and the fault code database; user evaluation information may include the user's historical repair records, evaluations of solutions for similar faults, etc.

[0146] In one possible implementation, keywords and parameter values ​​corresponding to each object in each dimension can be extracted from the database of the target objects according to the set dimension type. The parameter values ​​that can be standardized are standardized, and the importance of the keywords in the dimensions that cannot be standardized is calculated using the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to obtain the feature vector of the dimension. For the evaluation text of user reviews of products, the BERT model is used to convert the evaluation text into a semantic vector to capture the sentiment and semantic information in the evaluation. The keywords, feature vectors and semantic vectors of the dimensions after the above standardization are combined to obtain the preset knowledge graph.

[0147] As an example, the dimensions of the target object include price, specifications, and reviews. Price data is obtained from the product database and standardized (normalized to the [0, 1] interval). Product specification descriptions (such as "16GB (gigabytes) memory" and "512GB SSD (Solid State Drives)") are extracted. User reviews of the product are collected and converted into semantic vectors using the BERT model to capture the sentiment and semantic information in the reviews. The keywords, feature vectors, and semantic vectors of the standardized dimensions are combined to obtain the pre-defined knowledge graph.

[0148] In one possible implementation, new user reviews can be retrieved from the evaluation platform at preset intervals to incrementally update the BERT vector and TF-IDF features.

[0149] In the process of filtering candidate recommendation information in a pre-defined knowledge graph based on user intent, a metadata semantic joint attention mechanism can be adopted to achieve cross-modal feature fusion.

[0150] For example, the weighted calculation of the information in the preset knowledge graph and the information representing the user's intent can be expressed by the following formula:

[0151] Fused_Feature =α×Meta_Embed+β×Dialog_Embed(1)

[0152] Here, Fused_Feature (fusion feature) integrates information from the metadata graph and information representing user intent. Products are filtered and recommendations are generated based on this fusion feature. Meta_Embed (metadata embedding vector) represents the static parameters of products and user reviews in the preset knowledge graph. Dialog_Embed (session embedding vector) represents the semantic analysis of the user session. α and β are weight coefficients. The weight of β is adjusted according to the clarity of the user session (corresponding to the confidence level); the clearer the session, the larger the value of β, and the lower the value of α.

[0153] The first candidate recommendation information obtained after filtering includes the candidate object and the dimension information associated with the candidate object. The dimension information associated with the candidate object corresponds to the keyword of the requirement.

[0154] In one possible implementation, based on the above formula (1), the information of each dimension of each candidate object in the preset knowledge graph can be weighted and calculated with the user's intention keywords to obtain the fusion feature. The fusion feature is obtained by weighting the vector values ​​and is also expressed in numerical form. Accordingly, the candidate objects can be sorted according to the numerical value of the fusion feature, and the preset number of candidate objects with larger values ​​can be used as candidate recommendation information.

[0155] For example, if the preset number of candidate recommendation information is 20, according to the above formula (1), the information of each dimension of each candidate object in the preset knowledge graph is weighted and calculated with the demand keywords in the determined first user intent to obtain the fusion feature of each candidate object. In descending order, the 20 candidate objects with the highest fusion feature values ​​are selected as the first candidate recommendation information, and then one or more of these 20 candidate objects are selected as the target recommendation information.

[0156] As an example, for candidate laptop model A in the preset knowledge graph, user reviews are "lightweight, suitable for office use, and cost-effective." The dimensions corresponding to this user review information are weight, application scenarios, and price. Combining the "lightweight" in the user review information with the weight "1.2 kg" in the product's static parameters yields the weight dimension feature of the laptop in the preset knowledge graph. Combining the "office scenarios" in the user review information with application scenarios such as "offline gaming and office work" in the product's static parameters yields the application scenario dimension feature of the laptop in the preset knowledge graph. Combining the "cost-effective" in the user review information with the price "2899 yuan" in the product's static parameters yields the price dimension feature of the laptop in the preset knowledge graph. Accordingly, after obtaining the first user intent's demand keywords including "lightweight" and "low price," the first candidate recommendation information filtered from the preset knowledge graph can include this laptop model A.

[0157] As an example, the filtered candidate recommendations include laptop model A and laptop model B. Laptop model A's dimensional information includes configuration, price, and weight, while laptop model B's dimensional information includes configuration, price, and weight. The dimensional information of the candidates obtained through this filtering process can be the same or different. For example, laptop models A and B might have the same weight but different prices and configurations.

[0158] Combining the incremental understanding of user intent provided in the aforementioned embodiments, this solution can improve recommendation accuracy and efficiency by using a collaborative mechanism of dynamic metadata fusion and incremental intent understanding, while maintaining the fluency of natural dialogue.

[0159] In this embodiment, corresponding demand keywords are extracted from the first user intent; based on these demand keywords, a first candidate recommendation information is obtained by filtering within a preset knowledge graph. The preset knowledge graph includes a set of candidate objects in the target scenario, dimensional information associated with the candidate object set, and user evaluation information. The first candidate recommendation information includes at least one candidate object in the target scenario and dimensional information associated with at least one candidate object. A preset knowledge graph is constructed for the set of candidate objects in the target scenario, along with the associated dimensional information and user evaluation information. This preset knowledge graph can be updated in real time based on the candidate object situation and can also be updated based on user evaluation information, thus achieving the construction of a dynamic knowledge graph. Therefore, by using the demand keywords determined from the user intent and filtering the first candidate recommendation information from this dynamically updated preset knowledge graph, it is possible to combine user evaluations for recommendation, thereby improving the recommendation effect.

[0160] Figure 4 This is a flowchart illustrating the determination of a first confidence level of a first user intent based on a first user intent, provided in an embodiment of this application. It may include steps 401 to 403, which are described in detail below.

[0161] 401. Extract the corresponding demand keywords from the first user intent;

[0162] The keyword for this requirement can be a word contained in the user's intent, or a word that corresponds to the user's intent.

[0163] For example, if the first user intent is "a thin and light laptop for office use", the corresponding key requirements are "weight" and "office use".

[0164] The specific implementation of step 401 can be the same as step 301. For a detailed explanation, please refer to the explanation of step 301. It will not be repeated here.

[0165] 402. Using a confidence level testing model, determine the confidence level of the demand keywords in at least one dimension;

[0166] This confidence level detection model can determine the confidence level of the corresponding dimension based on the input keywords.

[0167] The extracted demand keywords can be one or multiple, and each demand keyword corresponds to a dimension.

[0168] In one possible implementation, the confidence level of each dimension can be determined for the demand keywords in the first user intent. These determined demand keywords may only correspond to some dimensions among multiple preset dimension types; in this case, only the confidence level corresponding to that dimension is generated, while other dimensions are left empty or set to zero.

[0169] The confidence score for each dimension ranges from [0, 1].

[0170] As an example, the demand keywords include "lightweight" and "high graphics card performance". Inputting these demand keywords into the confidence detection model, we can obtain the confidence scores for the corresponding weight and performance dimensions. The confidence scores for the weight dimension of this user intent are 0.4, the confidence scores for the performance dimension are 0.3, and the confidence scores for the scenario dimension are 0.2.

[0171] As an example, the demand keywords include "weight under 3 catties", "S model graphics card", and "office scenario". Inputting these demand keywords into the confidence test model, we can obtain the confidence scores for the corresponding weight and performance dimensions. The confidence scores for the weight dimension of this user intent are 0.8, the confidence scores for the performance dimension are 0.9, and the confidence scores for the scenario dimension are 0.7.

[0172] As an example, the demand keywords include "80-liter capacity" and "energy saving". Inputting these demand keywords into the confidence test model, we can obtain the confidence scores for the corresponding power consumption and capacity dimensions. The confidence score for the power consumption dimension of this user intent is 0.4, and the confidence score for the capacity dimension is 0.8.

[0173] In one possible implementation, the confidence level of the demand keyword under the corresponding dimension can also be determined by combining the situation of the demand keyword in multiple rounds of the conversation.

[0174] As an example, we can count the frequency of the keyword in the several rounds that have occurred in this session to determine whether it is a word that users are more concerned about. The higher the frequency, the more concerned the user is, and the higher the confidence level.

[0175] In one possible implementation, the confidence level of the keyword in the corresponding dimension can also be determined by the behavioral parameters input by the keyword.

[0176] As an example, the behavioral parameters when the user inputs the keyword for their needs can be recorded. Based on these behavioral parameters, it can be analyzed whether the user is paying attention to the keyword for their needs. The more attention the user pays, the higher the confidence level.

[0177] For example, the behavioral parameter may include any one or more of the following: tone words, handwriting pressure, and keyboard typing pressure. Emphasizing tone words, handwriting pressure being greater than other words, and keyboard typing pressure being greater than other words can all indicate that the user is paying attention to the keyword in question.

[0178] In one possible implementation, labels can also be added to the user intent recognition model. These labels can be hidden labels that the model can recognize. By combining the frequency of the demand keywords in the user intent with the results obtained by the user intent recognition model, the demand keywords can be distinguished from other keywords, thereby highlighting the demand keywords that the user is more concerned about. This will improve the confidence of the demand keywords that the user is more concerned about when determining the confidence of the demand keywords in the corresponding dimension.

[0179] For example, if the keywords identified by the user intent recognition model include "weight" and "height", the user intent recognition model will ultimately identify "weight" as the demand keyword because a hidden label was added to "weight".

[0180] 403. Calculate the first confidence level based on the confidence level corresponding to each requirement keyword.

[0181] The confidence level of the first user intent corresponding to the first session information is calculated based on the confidence level of the corresponding demand keywords for each dimension.

[0182] In one possible implementation, for each requirement keyword in the same dimension, a weighted summation method can be used to obtain the confidence level of this dimension. The first confidence level of the first user intent can be calculated by weighted summation of the confidence levels of each dimension.

[0183] In conjunction with the confidence detection module, a dialogue state tracking module can also be set up. This dialogue state tracking module determines the session state, which can be used to determine the weighted weights. The process will be described in detail in subsequent embodiments.

[0184] In this embodiment, corresponding demand keywords are extracted from the first user intent; a confidence level detection model is used to determine the confidence level of the demand keywords in at least one dimension; and a first confidence level is calculated based on the confidence levels corresponding to each demand keyword. For the demand keywords extracted from the first user intent, the confidence levels corresponding to each dimension are determined, and the first confidence level of the first user intent is calculated using the confidence levels corresponding to each demand keyword. This process of determining the first confidence level for the first user intent is quantified, enabling highly accurate determination of the first confidence level.

[0185] In one possible implementation, a first confidence level is calculated based on the confidence level corresponding to each requirement keyword, including at least one of the following:

[0186] 1. Determine the frequency of keywords corresponding to at least one dimension in the first conversation information; based on the frequency of keywords corresponding to at least one dimension, determine the weight of the required keywords under at least one dimension; calculate the first confidence level based on the confidence level and weight of each required keyword; frequency and weight are positively correlated; keywords are words corresponding to the required keywords.

[0187] The first session information can be the current round session or multiple round sessions in this exchange, including the current round session and historical multiple round sessions.

[0188] To reduce data processing volume, the first session information can include three rounds of sessions: the current session and two previous rounds. These previous rounds include the previous two rounds of sessions. By combining context, the weights of key requirements can be determined. As the session progresses, the frequency of keywords in each dimension changes, enabling dynamic adjustment of the weights.

[0189] Each dimension's keywords correspond to the demand keywords extracted from the first user intent for that dimension; the two are semantically related or completely identical. Since the frequency of keyword occurrences in this first conversation information can characterize the user's emphasis on the content corresponding to that keyword, the frequency of keyword occurrences in any dimension of this first conversation information can be used to determine the importance of the corresponding demand keyword in that dimension. The importance corresponds to the user's level of attention.

[0190] Correspondingly, the higher the frequency of keywords appearing in the first conversation, the higher the weight of the corresponding demand keywords.

[0191] As an example, regarding the battery life dimension, the keyword "power saving" appeared twice and "battery life" appeared three times in the first session information. This frequency is relatively high, so the keyword "battery life" can be given a higher weight, such as 0.5. On the other hand, regarding the weight dimension, the keyword "lightweight" appeared once in the first session information. This frequency is relatively low, so the keyword "weight" can be given a lower weight, such as 0.2.

[0192] By using the confidence levels corresponding to the demand keywords of each dimension and the weights determined in the previous steps, the first confidence level of the first user intent can be obtained through weighted calculation.

[0193] 2. Determine the target parameter and session round corresponding to at least one dimension in the first session information; based on the target parameter and session round, determine the weight of the demand keyword in at least one dimension; calculate the first confidence level based on the confidence level and weight of each demand keyword; the target parameter is negatively correlated with the weight, and the session round is negatively correlated with the weight.

[0194] The target parameter can represent the clarity of the user's intent. This target parameter can be information entropy, which represents the ambiguity of the information. The more ambiguous the answers in the first conversation, the higher the information entropy. Consequently, the recommended information determined based on this conversation information may not meet the user's needs. Therefore, this target parameter is set to be negatively correlated with the weight.

[0195] The vague response can be an uncertain response such as "fast", "good", "light", "not heavy", "large screen" etc. This application does not limit the specific content of the vague response. The uncertainty of the response in the conversation information can be used to determine that it is a vague response.

[0196] The number of conversation rounds indicates how many conversations have been conducted for this dimension. The more conversation rounds for this dimension, the less likely the user is to quickly provide a clear need for that dimension. The number of conversation rounds can characterize the severity of the user's inability to clearly express their needs. Therefore, setting the number of conversation rounds is negatively correlated with the weight.

[0197] In one possible implementation, the weights for any dimension can be determined using the following formula:

[0198] w_k(t) = w_k(t-1)× e^(-0.5×E(t)×δ_k) (2)

[0199] Where t represents the dialogue round or time point, which reflects the change of weight as the dialogue round progresses or time passes; w_k(t) is the weight of the k-th dimension in the t-th round; w_k(t-1) is the weight of the k-th dimension in the t-1-th round; E(t) represents the information entropy; δ_k represents the number of sessions in the k-th dimension.

[0200] The algorithmic logic of formula (2) is that the weight of any dimension decays with the information entropy E(t) and the number of sessions δ_k.

[0201] Based on the target parameters and session rounds corresponding to each dimension, the first confidence level can be calculated according to the weight of the demand keywords in the corresponding dimension, and the confidence level and weight of the demand keywords in each dimension.

[0202] In one possible implementation, the conversation state tracking module determines the weight value of each dimension by tracking the state of the session, and uses the weight value to calculate the first confidence level.

[0203] Third, based on the type of dimension in the first user intent, assign preset weight values, and determine the preset weight values ​​as the weights corresponding to the demand keywords in at least one dimension; calculate the first confidence level based on the confidence level and weight corresponding to each demand keyword.

[0204] For each dimension of the first user intent, a preset weight value can be set. This preset weight value can be set by the user according to their own needs, or it can be configured by technicians based on the statistical results of a large number of users' usage.

[0205] In one possible implementation, by statistically analyzing the usage of a large number of users, the dimensions with high user attention can be identified. Accordingly, higher weights can be assigned to the dimensions with high user attention, and lower weights can be assigned to the dimensions with low user attention.

[0206] In one possible implementation, the preset weight values ​​can be the same for all dimensions.

[0207] In this embodiment, during the process of calculating the first confidence level corresponding to the first user intent based on the confidence level corresponding to each demand keyword, the weights of the demand keywords under each dimension can be determined, and then the first confidence level can be calculated based on the confidence level and weights of the demand keywords. The determination of the weights of the demand keywords under each dimension can be a preset weight; since the keywords in the first conversation information correspond to the demand keywords, the weights of the demand keywords can also be determined based on the frequency of the keywords corresponding to at least one dimension in the first conversation information, and these weights are positively correlated with the frequency; since the target parameter characterizes the clarity of the user's expression of the demand, and the conversation round characterizes the severity of the user's inability to clearly express the demand, the weights of the demand keywords are determined based on the target parameter and the conversation round, and these target parameter, conversation round, and weights are all negatively correlated. Through the above methods, the weights of the demand keywords under at least one dimension can be determined, and then the first confidence level can be calculated using these weights and the determined confidence level. This achieves multiple ways to calculate the first confidence level, providing high flexibility in calculation.

[0208] In one possible implementation, the user is prompted to provide second session information, including at least one of the following:

[0209] 1. If the first dimension of the preset dimension type is missing in the first confidence level, prompt the user to provide the required information corresponding to the first dimension;

[0210] Multiple dimension types can be set for different application scenarios. In order to determine clearer user intent through conversation information, sufficiently clear requirement information needs to be provided for each dimension.

[0211] As an example, when the target product is a laptop, there are five preset dimension types: scenario, specifications, price, appearance, upgrade, etc.

[0212] In one possible implementation, the initial confidence level being less than or equal to a preset confidence threshold may be due to one or more of the preset dimension types not yet being collected, resulting in a low overall dimension level. Accordingly, the user is prompted to provide the required information corresponding to the missing dimension.

[0213] Through a dynamic slot-filling strategy, regardless of which dimension a user starts with, the system collects demand information for each dimension within a preset dialogue round. Furthermore, the collection order is not fixed and can be dynamically adjusted based on the dimensions whose demand information has already been collected.

[0214] In one possible implementation, for the first confidence level of the first user intent, after determining the included dimension types, the remaining unincluded first dimension can be identified. For the first dimension, a prompt can be generated to prompt the user to provide the required information for that first dimension, in order to achieve the goal of collecting the required information for all dimensions.

[0215] It determines whether the confidence level of each dimension in the preset dimension type has been confirmed in the first session information, so as to complete the demand statistics of each dimension in a limited number of sessions, so as to achieve more accurate recommendations with fewer exchanges.

[0216] For example, in the scenario of recommending note-taking products, the corresponding preset dimension types include price, performance, scenario, screen size and weight. The first confidence level includes performance, portability and price, with screen size and weight as the first dimension, prompting users to provide their needs for screen size and weight.

[0217] As an example, if the user mentions a price first, the generated prompt could be, "With this budget, do you value gaming performance or office battery life more?"

[0218] Second, if the confidence level of the second dimension in the first confidence level is less than the target confidence level threshold, prompt the user to provide the corresponding requirement information for the second dimension.

[0219] This target confidence threshold can define whether the user's intent clearly represents their need for the target object in a single dimension. For the same target object, the target confidence threshold for each dimension type in the corresponding preset dimension type can be the same or different, and can be set according to the actual situation; this application does not impose any restrictions.

[0220] The preset intent recognition model can generate corresponding confidence scores for each of the preset dimensions during the process of recognizing user intent from the first session information.

[0221] Accordingly, the confidence level of each dimension in the first user intent is compared with the target confidence level threshold. If the confidence level of the second dimension is less than the target confidence level threshold, it indicates that the demand for the second dimension is unclear, and the user is prompted to provide the demand information corresponding to the second dimension.

[0222] In one possible implementation, the scope can be further narrowed based on the content identified in the first session information, in order to make the requirements for the second dimension clearer and improve the confidence of the second dimension.

[0223] As an example, if the confidence level for the price dimension is less than the target confidence threshold, the generated prompt message could be "Is a price of around 5000 yuan acceptable?".

[0224] The above-mentioned prompts can achieve a confidence weight gain of at least 0.3 for each round of dialogue compared to the previous round, enabling a clear understanding of user needs and accurate recommendations to be made to users in a smaller number of rounds of conversation.

[0225] In one possible implementation, a three-stage clarification funnel approach can be used to determine the prompt message, which is used to prompt the user to provide information for the second session.

[0226] This three-stage clarification funnel can be configured with multiple questions ranging from large to small, and one of them can be selected as the prompt message according to the actual situation.

[0227] The three-stage clarification funnel includes: Stage 1: Open-ended guidance; Stage 2: Option-limited; Stage 3: Data-supported.

[0228] For example, if the first session message is "I need a fast computer," the corresponding prompt for the first level of performance could be "What exactly do you mean by 'fast'?"; the corresponding prompt for the second level could be "Do you mean ① boot speed ② software responsiveness ③ file transfer?"; and the corresponding prompt for the third level could be "An i7 processor is 40% faster than an i5 for file processing, do you need this level?".

[0229] In one possible implementation, the selection of prompts in the three-level clarification funnel can be achieved using reinforcement learning. Multiple question templates can be set in the action space for each prompt, and the specific content of each template can be determined based on the information from the first session. Alternatively, the optimal question can be selected from the action space based on the session state to improve the efficiency of demand resolution and recommendation conversion rate.

[0230] This reinforcement learning method allows for the creation of a state space and an action space. The state space is a vector that encodes and maps session information; the action space predefines multiple question templates as optional actions. These question templates include open-ended guided questions, option-limited questions, and data-supported questions. The strategy is updated by selecting an action (a question) from the action space based on the state mapped to the state space, and then incorporating user feedback (rewards). The reward can be the user's response to the recommended object.

[0231] For example, when recommending a product, the reward function can be calculated using the following formula:

[0232] Reward = 2 × purchase conversion + 1 × click - 0.5 × session rounds (3)

[0233] Purchase conversion earns a reward of +2 when a user ultimately purchases the recommended product; click earns a reward of +1 when a user clicks on the details of the recommended product; and session rounds are the number of sessions from the beginning to the current round, with each additional session deducting -0.5 from the reward to encourage shorter session durations.

[0234] As an example, the state space can be 256-dimensional. By encoding the first session information into a 256-dimensional vector, the session state is mapped to the state space. These 256 dimensions can include the current turn, collected information, confidence scores for each dimension, and features of the first candidate recommendation. Based on this 256-dimensional vector, the optimal question is selected from the action space to shorten the dialogue process.

[0235] Of course, the dimensions of this state space are not limited to 256, and the dimensions in the state space are not limited to those in the example above.

[0236] The reinforcement learning algorithm can be Q-learning, PPO (Proximal Policy Optimization Algorithms, Deep Reinforcement Learning), etc. This application does not restrict the algorithm used for reinforcement learning.

[0237] As an application example, the user conversation trajectory can be as follows: When a user enters "I want a laptop for office work", the corresponding prompt message is "Do you value battery life or performance more?"; when a user enters "I need long battery life", the corresponding prompt message is "What is your acceptable weight limit?"; when a user enters "under 1.5kg", the target recommendation message is "Laptop A (1.13kg / 15h)".

[0238] The dialogue status tracking module can also execute the scheme for generating prompt information in this embodiment, which can prompt the user to provide second dialogue information based on the overall dialogue status.

[0239] In this embodiment, when prompting the user to provide second session information, the user can be prompted to provide the corresponding demand information for the first dimension if the confidence level of the first dimension is missing, in order to obtain the confidence level of the first dimension; or the user can be prompted to provide the corresponding demand information for the second dimension if the confidence level of the second dimension is less than the target confidence level threshold, in order to improve the confidence level of the second dimension. This improves the confidence level of the session from two directions, thereby improving the overall confidence level of the session information and providing a basis for accurate recommendation of objects in the future.

[0240] In one possible implementation, the information recommendation method also includes:

[0241] Identify the current recommendation scenario; based on the current recommendation scenario, determine the target confidence threshold for at least one dimension.

[0242] After receiving the first round of session information, the current recommendation scenario can be identified based on the session information. For different recommendation scenarios, the target confidence thresholds for each type of dimension can be preset.

[0243] In one possible implementation, multiple recommendation scenarios can be pre-set, such as electronic product recommendations (which can be further divided into various more detailed recommendation scenarios such as washing machine recommendations, computer recommendations, etc.), travel route recommendations, recipe recommendations, clothing recommendations, etc.

[0244] Since different confidence levels can be determined using different types of dimensions for different recommendation scenarios, the overall confidence level of the session intent corresponding to the session information can be determined using the confidence levels of each dimension. Accordingly, target confidence thresholds for each dimension type and each dimension can be set in advance for different recommendation scenarios.

[0245] As an example, if the received conversation message is "I want a laptop for office work", the current recommendation scenario can be identified as a "laptop" recommendation scenario. Accordingly, the laptop recommendation scenario (a branch of electronic product recommendation scenarios) is pre-defined with three dimensions, including specifications, price, and appearance, as well as the target confidence threshold for each dimension. For example, the confidence threshold for price is 0.8, the confidence threshold for specifications is 0.7, and the confidence threshold for appearance is 0.5.

[0246] As an example, if the received conversation message is "Recommend a Beijing travel guide", it can be identified that the current recommendation scenario is a "travel route" recommendation scenario. Accordingly, the travel route recommendation scenario is obtained by pre-setting three dimensions, including destination, budget, and number of travel days, as well as the target confidence threshold for each dimension. For example, the confidence threshold for destination is 0.8, the confidence threshold for budget is 0.9, and the confidence threshold for number of travel days is 0.5.

[0247] In this embodiment, the current recommendation scenario can be identified first; then, based on the current recommendation scenario, the target confidence threshold corresponding to at least one dimension can be determined, thus realizing the determination of the target confidence threshold corresponding to each dimension for different recommendation scenarios, which can be compatible with various recommendation scenarios.

[0248] In one possible implementation, the information recommendation method also includes:

[0249] In response to the absence of a target object in the preset knowledge graph that corresponds to the first user intent, a third user intent is generated, and the recommended object and / or related information corresponding to the third user intent are output; the third user intent is generated after adjusting the variable values ​​of the target object in at least one dimension.

[0250] If, during the filtering process in the preset knowledge graph based on the first user intent, no target object indicating the first user intent is found, then the parameter values ​​of the user intent in at least one dimension are adjusted to obtain the third user intent.

[0251] The first user intent is the range of the target object in one or more dimensions. However, there is no corresponding object in the preset knowledge graph for this range. The value range of this dimension can be adjusted to obtain the third user intent.

[0252] In one possible implementation, the Pareto optimal solution algorithm can be used to determine the recommended object from the candidate objects in the pre-defined knowledge graph using a compromise recommendation method.

[0253] The system can be adjusted based on the variable values ​​of each dimension. The adjustment can be made on one or more dimensions to obtain candidate objects that can balance the relationship between the dimensions as recommended objects.

[0254] Based on the Pareto optimal solution algorithm, the variable values ​​of the first user intent dimension are adjusted. The basis for adjusting the variable values ​​and the results obtained can also be output as related information, and the comparison and determination process of the recommended object can be visualized.

[0255] The parameter values ​​of at least one dimension in the third user intent are different from the parameter values ​​of the corresponding dimension in the first user intent.

[0256] After determining the third user's intent, objects that match the third user's intent and are selected from the preset knowledge graph for the target object can be used as recommended objects and output.

[0257] The Pareto optimal solution algorithm aims to balance the dimensions that users value, and filters from a pre-defined knowledge graph.

[0258] As an example, in a recommendation scenario for laptops, it's impossible to select a hard drive that perfectly meets both the criteria of lowest price and largest storage space from the preset knowledge graph. The first user intent might be a price range of 100 to 500 yuan and a storage range of more than 500GB. Since no hard drive meeting these two requirements is found in the preset knowledge graph, adjusting the price to 100 to 600 yuan and the storage space to 400 to 500GB can filter for one or more candidate hard drives, which are then used as recommendations.

[0259] For example, the output information is: Model A is recommended. Model A is 15% more expensive than Model B, but its storage space performance is 30% greater.

[0260] The output contains relevant information to enhance its persuasiveness to the user.

[0261] This cognitive guidance information can consist of three parts: reconstructing needs, revealing costs, and providing empirical persuasion.

[0262] For example, a restructuring requirement could be presented as "Actually, 90% of office users would find this product sufficient," convincing users with specific office needs that this approach helps adjust their expectations when they prioritize high performance but have limited budgets. Cost disclosure could be as "Every 10% performance improvement requires an additional 500 yuan," persuading users that a significant price increase is necessary for a small performance boost, arguing that simply pursuing performance is not cost-effective and helping users weigh the value proposition. Empirical persuasion could be as "This configuration supports all Adobe certifications," providing empirical evidence that the recommended product is adequate for office software, enhancing credibility. Through these three parts of cognitive guidance, users can understand the differences between products, allowing them to adjust their expectations and choose a product that better suits their needs.

[0263] Furthermore, since the recommended objects generally involve price, the price and performance of candidate objects in the preset knowledge graph under this recommendation scenario can be used to calculate the cost-effectiveness value of each product according to the confidence level, forming the Pareto cost-effectiveness curve, so as to select the target recommendation information (price and performance) from it.

[0264] This related information can serve as cognitive guidance, automatically generating comparative dialogue to flexibly respond to various questioning styles while subtly guiding the conversation toward the desired outcome. For example, if a user inputs "a cheap computer that can play Cyberpunk," the system first confirms the graphics card requirement, then explains that "the C model has twice the performance of the D model, but is 25% more expensive," and finally offers a compromise solution.

[0265] In this embodiment, in response to the absence of a target object in the preset knowledge graph that matches the first user intent, a third user intent is generated, and the recommended object and / or related information corresponding to the third user intent are output. The third user intent is generated after adjusting the variable values ​​of the target object in at least one dimension. When the target object does not exist in the preset knowledge graph, the variable values ​​of one or more dimensions of the first user intent can be adjusted to generate a third user intent. This third user intent is then used to filter through the preset knowledge graph to obtain the corresponding recommended object. This recommended object balances the relationships across various dimensions, resolves conflicting needs, reduces session rounds, and achieves a rapid recommendation process.

[0266] Figure 5 This is another flowchart illustrating the information recommendation method provided in this application embodiment. This flowchart is for a specific product recommendation scenario. In this scenario, the information recommendation method may include steps 501 to 503, which will be described in detail below.

[0267] 501. Identify the third user intent in the third session information; the third session information is session information in the context of product recommendation.

[0268] In the context of product recommendation, the system receives user-inputted conversation information, which is recorded as the third conversation information.

[0269] The goods can be any tradable object, such as electronic products, clothing, commercial housing, jewelry, etc. This application does not restrict the specific form of the goods.

[0270] In one possible implementation, a pre-defined intent recognition model can be used to identify the intent contained in the third session information.

[0271] In one possible implementation, semantic error correction can be performed on the third-party session information. This semantic error correction may include identifying referential errors and quantifying descriptive representations, and then using the corrected third-party session information to identify user intent.

[0272] 502. Based on the third user intent, determine the candidate product recommendation list in the preset knowledge graph and the third confidence level of the third user intent; the third confidence level is used to characterize the target parameters of the target product indicated by the third user intent in at least one dimension; the preset knowledge graph includes the candidate product set in the product recommendation scenario and the dimension information and user evaluation information associated with the candidate product set, and the candidate product recommendation list includes at least one candidate product in the product recommendation scenario and the dimension information associated with at least one candidate product;

[0273] In this product recommendation scenario, a preset knowledge graph can be set up for the target recommended product. This preset knowledge graph consists of a set of candidate products for the product recommendation scenario, the dimensional information associated with the candidate product set, and user review information. Through the process described in the previous embodiment, the candidate product set, the associated dimensional information, and the user review information are integrated to obtain the preset knowledge graph. In this preset knowledge graph, for each candidate product, corresponding information can be assigned to its associated dimensions based on the information of the candidate product. Moreover, corresponding information is also assigned to the review dimension based on the user review information.

[0274] After determining the third user intent, the third confidence level of the third user intent can be determined. The process of determining the third confidence level can refer to the process of determining the first confidence level in the aforementioned embodiments, and will not be repeated here.

[0275] The corresponding dimensions may differ for different products, and the dimensions for each product can be set according to the actual situation. This application does not impose any restrictions.

[0276] As an example, if the product is a computer, its corresponding dimensions could include price, performance, screen size, weight, etc.; if the product is a refrigerator, its corresponding dimensions could include price, power consumption, capacity, etc.

[0277] The third confidence level is the target parameter of the target product indicated by the third user intent across multiple dimensions. This target parameter can be clarity. Accordingly, the third confidence level represents the clarity of the product indicated by the user intent across multiple dimensions, indicating whether the user has provided sufficiently clear information about their needs across those dimensions.

[0278] After determining the third user intent, a candidate product recommendation list is obtained by filtering in the preset knowledge graph. The candidate product recommendation list contains multiple candidate products corresponding to the third user intent, as well as the dimensional information of each candidate product in each dimension.

[0279] The process of filtering the candidate product recommendation list from the preset knowledge graph can be referred to the process of determining the first candidate recommendation information based on the first user intent in the aforementioned method embodiment, and will not be elaborated here.

[0280] It should be noted that this embodiment is an explanation for the product recommendation scenario. In other recommendation scenarios, such as travel guide recommendation scenarios, a candidate travel guide recommendation list can be determined first in a preset knowledge graph, and then the corresponding target recommendation information (recommended travel guide) can be determined in the travel guide recommendation list.

[0281] 503. In response to the third confidence level being greater than the preset confidence threshold, determine the target recommendation information corresponding to the target product from the candidate product recommendation list.

[0282] The preset confidence threshold is a confidence threshold set for the target product, used to judge the overall confidence of the third-party session information.

[0283] If the third confidence level is greater than the preset confidence threshold, it indicates that enough clear user demand information has been collected. Then, the target recommendation information can be determined from the candidate product recommendation list based on the known information. Since each candidate product in the candidate product recommendation list is close to the user's demand for the target product, the target recommendation information corresponding to the target product can be determined from the candidate product recommendation list with high recommendation accuracy.

[0284] In one possible implementation, if the third confidence level is not greater than the preset confidence threshold, the user can be prompted to provide fourth session information, which supplements the known requirement information in the first session information.

[0285] The fourth session information refers to the session information provided by the user again in the product recommendation scenario. The process of prompting the user to provide the fourth session information is explained in the previous embodiments, and will not be repeated here.

[0286] When determining target recommendation information, if it is impossible to satisfy all needs, then the aforementioned method of determining target recommendation information in the case of conflicting needs can be adopted. Moreover, cognitive guidance information can be included in the output to enhance the persuasiveness to users.

[0287] In this embodiment, in the product recommendation scenario, a knowledge graph is pre-constructed. The pre-constructed knowledge graph includes a set of candidate products in the product recommendation scenario, dimensional information associated with the candidate product set, and user evaluation information. The third user intent in the third conversation information within the scenario is identified, and based on this third user intent, a candidate product recommendation list in the pre-constructed knowledge graph and a third confidence level of the third user intent are determined. This third confidence level characterizes the target parameters of the target product indicated by the third user intent in at least one dimension. The candidate product recommendation list includes at least one candidate product in the product recommendation scenario and dimensional information associated with at least one candidate product. When the third confidence level is greater than a pre-set confidence threshold, the target recommendation information corresponding to the target product is determined from the candidate product recommendation list, thus achieving the goal of accurate recommendation for the product recommendation scenario.

[0288] Figure 6 This is a schematic diagram of an information recommendation method provided in an embodiment of this application, including the following steps:

[0289] 601. Receive user input;

[0290] Receive session information input by the user, which can be input information for any object.

[0291] 602. Identify user intent;

[0292] The intent is obtained by identifying the user's intent through the intent recognition model on the input session information.

[0293] This user intent represents the user's needs in this instance.

[0294] 603. Preset knowledge graph filtering;

[0295] Based on the user's intent, candidate recommendation information corresponding to the multiple dimensions of needs contained in the user's intent is selected from the preset knowledge graph.

[0296] This preset knowledge graph can update the information of candidate objects within it according to a set period. This update can include the candidate objects themselves, or their attribute information. It also needs to update the user evaluation information of the candidate objects according to a set period. Both of these updates can involve adjusting the parameters of each candidate object in the preset knowledge graph to reflect the updated information.

[0297] 604. Obtain the set of candidate recommendation information;

[0298] The candidate recommendation information obtained from the preset knowledge graph will be recorded as a candidate recommendation information set.

[0299] 605. Determine the confidence level;

[0300] A confidence level detection model is used to determine the confidence level of each dimension of the user's intent.

[0301] 606. Track the dialogue status;

[0302] This step uses a dialogue state tracking module to track the session state, combining historical sessions with the currently received session, and determining the confidence level for each dimension based on the session context.

[0303] Then, by combining the confidence levels of each dimension, the overall confidence level of the user's intent is determined.

[0304] 607. Determine if the confidence level is greater than 0.8;

[0305] The 0.8 is a preset confidence threshold. The overall confidence of the user's intent is compared with the preset confidence threshold. If it is greater than the threshold, step 608 is executed; otherwise, step 610 is executed.

[0306] This threshold can be selected as a larger or smaller value as needed. The 0.8 is only used as an example and does not limit the specific value of the preset reliability threshold.

[0307] 608. Generate target recommendation information;

[0308] If the confidence level is greater than 0.8, the target recommendation information is selected from the aforementioned set of candidate recommendation information.

[0309] The target recommendation information can be output as recommended content for users.

[0310] 609. Receive user feedback;

[0311] After recommending the target information to the user, you can receive feedback from the user regarding the target information.

[0312] The feedback can include whether the target recommended product was purchased, or whether the target recommended product was viewed.

[0313] The user feedback is provided to the tracking dialogue state model so that the reinforcement learning approach used by the tracking dialogue state model can optimize the strategy of the prompt questions it generates.

[0314] 610. Generate a prompt question;

[0315] During the process of tracking the dialogue status, the confidence level of each dimension in the conversation is known. If a certain dimension fails to collect any requirements or fails to collect sufficiently clear requirements, then a prompt message is generated for that dimension to prompt the user to provide the requirements for that dimension or clearer requirements in the subsequent conversation information.

[0316] The prompt question uses reinforcement learning to select the question that best matches the current conversation state from a set of preset question templates. The user can then input response information to the question, thus guiding the conversation toward the desired outcome.

[0317] Related technologies include rule-based dialogue systems (which employ static knowledge graphs). This application compares these related technologies with the information recommendation method provided in this application.

[0318] Table 1 below shows the accuracy comparison of the information recommendation method in this solution provided by this application with that of the recommendation method in related technologies.

[0319] Table 1

[0320]

[0321] As can be seen from Table 1 above, the information recommendation method in this solution has a higher accuracy rate when dealing with clear needs. When dealing with ambiguous or contradictory needs, the accuracy of the information recommendation method in this solution is significantly improved compared to traditional recommendation methods.

[0322] Table 2 below shows the information recommendation method provided in this application, and compares it with recommendation methods in related technologies in several aspects.

[0323] Table 2

[0324]

[0325] In Table 2 above, text refers to user reviews, metadata is the attribute information of candidate objects, and behavior refers to the conversation process with users.

[0326] As shown in Table 2 above, compared with traditional recommendation methods, the information recommendation method in this scheme requires fewer average dialogue rounds, has higher cold-start accuracy, and higher multimodal data utilization. AUC (Area Under Curve) is a binary classification evaluation metric; the higher the value, the higher the classification accuracy.

[0327] The above describes an information recommendation method provided by an embodiment of this application. The following describes the apparatus for performing the above information recommendation method.

[0328] Please see Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device applying an information recommendation method, as provided in an embodiment of this application. Figure 7 As shown, the electronic device 700 includes:

[0329] At least one processor 701, on which a target application runs, the target application being capable of independently executing or invoking at least one artificial intelligence model to perform the following operations:

[0330] Used to identify a first user intent in the first session information; based on the first user intent, determine candidate recommendation information and the confidence level of the first user intent; the confidence level is used to characterize the target parameter of the target object indicated by the first user intent in at least one dimension; in response to the confidence level being greater than a preset confidence threshold, determine the target recommendation information corresponding to the target object from the candidate recommendation information;

[0331] The processor can be a data processing component in an electronic device, such as a CPU (Central Processing Unit) or an NPU (Neural-network Processing Unit). This application does not limit the specific component of the processor.

[0332] Should Figure 7 The image shows a processor, but the number of processors is not limited thereto. Different numbers of processors can be set according to actual needs. This application does not limit the number of processors.

[0333] This artificial intelligence model can be used to implement the functions of the pre-set reliability detection model and the preset intent recognition model in the aforementioned method embodiments.

[0334] Display device 702 is used to display target recommendation information.

[0335] The display device can be a display screen, an audio playback device, etc., and can output the specific target recommendation information.

[0336] If the display device is a display screen, the target recommendation information can be played and output in the form of text, images, or videos.

[0337] It should be noted that the specific process and detailed explanation of this processor function can be found in the explanations in the foregoing method embodiments, and will not be repeated here.

[0338] In this embodiment, a confidence level parameter is set, which characterizes the target parameters of the target object indicated by the user's intent in at least one set dimension. These target parameters can represent the clarity of the user's intent in the corresponding dimension. The processor first identifies the first user intent in the first session information; based on the first user intent, it determines the first candidate recommendation information and the confidence level of the first user intent. This confidence level can represent the target parameters of the target object indicated by the user intent in the first session information in multiple dimensions; in response to the confidence level being greater than a preset confidence threshold, the target recommendation information corresponding to the target object is determined from the first candidate recommendation information; the output device outputs the target recommendation information. By quantifying the user's intent in the user's session information with confidence, the first candidate recommendation information is initially determined based on this user intent, and the target recommendation information is determined from the first candidate recommendation information when the confidence level is greater than the preset confidence threshold. Through this two-stage selection process, the range of recommendation information can be gradually narrowed down, and by using a clearer user intent to determine the target recommendation information, accurate recommendations can be achieved.

[0339] In one possible implementation, this first session information belongs to the session information in the product recommendation scenario;

[0340] The target application is used to determine a candidate product recommendation list in a preset knowledge graph and a first confidence level of the first user intent based on the first user intent; the first confidence level is used to characterize the target product indicated by the first user intent in at least one dimension; the preset knowledge graph includes a set of candidate products in the product recommendation scenario and dimensional information and user evaluation information associated with the set of candidate products; the candidate product recommendation list includes at least one candidate product in the product recommendation scenario and dimensional information associated with at least one candidate product.

[0341] In response to a first confidence level greater than a preset confidence threshold, target recommendation information corresponding to the target product is determined from the candidate product recommendation list.

[0342] The target application can be an intelligent agent or an artificial intelligence assistant that can interact with the user, receive conversation information input by the user, and provide feedback on the conversation information. This feedback may include prompting the user to provide second conversation information and recommendation information.

[0343] It should be noted that for the specific processing procedure in this product recommendation scenario, please refer to the explanation of the corresponding embodiments of steps 501 to 503 in the aforementioned method embodiments, which will not be repeated here.

[0344] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the information recommendation methods provided in this application.

[0345] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the information recommendation methods provided in this application.

[0346] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0347] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0348] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0349] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0350] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0351] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. An information recommendation method, comprising: Identify the first user intent in the first session information; Based on the first user intent, determine the first candidate recommendation information and the first confidence level of the first user intent; The first confidence level is used to characterize the target parameters of the target object indicated by the first user intent in at least one dimension; In response to the confidence level being greater than a preset confidence threshold, target recommendation information corresponding to the target object is determined from the first candidate recommendation information.

2. The information recommendation method according to claim 1 further includes: In response to the first confidence level being less than or equal to the preset confidence threshold, the user is prompted to provide second session information; Identify the second user intent in the second session information; Based on the second user intent, a second confidence level of the second user intent is determined; the second confidence level is used to characterize the target object indicated by the second user intent in at least one dimension as a target parameter. Based on the second confidence level, target recommendation information corresponding to the target object is determined.

3. The information recommendation method according to claim 2, wherein determining the target recommendation information corresponding to the target object based on the second confidence level includes: If the dimensions included in the second confidence level are of the same type as the dimensions included in the first confidence level, then based on the second user intent, the second candidate recommendation information is obtained by filtering from the first candidate recommendation information. In response to the second confidence level being greater than the preset confidence threshold, target recommendation information corresponding to the target object is determined from the second candidate recommendation information; or If the dimensions included in the second confidence level are of a different type than the dimensions included in the first confidence level, a third candidate recommendation information is obtained by filtering from the preset knowledge graph based on the first user intent and the second user intent. In response to the second confidence level being greater than a preset confidence threshold, target recommendation information corresponding to the target object is determined from the third candidate recommendation information.

4. The information recommendation method according to claim 1, wherein determining the first candidate recommendation information based on the first user intent includes: Extract the corresponding demand keywords from the first user intent; Based on the required keywords, the first candidate recommendation information is obtained by filtering in a preset knowledge graph. The preset knowledge graph includes a set of candidate objects in the target scenario and dimensional information and user evaluation information associated with the set of candidate objects. The first candidate recommendation information includes at least one candidate object in the target scenario and dimensional information associated with the at least one candidate object.

5. The information recommendation method according to claim 1, wherein determining a first confidence level of the first user intent based on the first user intent includes: Extract the corresponding demand keywords from the first user intent; Using a confidence level detection model, determine the confidence level of the required keywords in at least one dimension; The first confidence level is calculated based on the confidence level corresponding to each demand keyword.

6. The information recommendation method according to claim 5, wherein calculating the first confidence level based on the confidence level corresponding to each demand keyword includes at least one of the following: Determine the frequency of keywords corresponding to at least one dimension in the first conversation information; based on the frequency of keywords corresponding to the at least one dimension, determine the weight of the required keyword under the at least one dimension; The first confidence level is calculated based on the confidence level and weight corresponding to each demand keyword; the frequency is positively correlated with the weight; the keyword is the word corresponding to the demand keyword. Determine the target parameter and session round corresponding to at least one dimension in the first session information; based on the target parameter and session round, determine the weight of the demand keyword in at least one dimension; The first confidence level is calculated based on the confidence level and weight corresponding to each of the aforementioned demand keywords; The target parameter is negatively correlated with the weight, and the session rounds are negatively correlated with the weight; Based on the type of dimension in the first user intent, a preset weight value is assigned, and the preset weight value is determined as the weight of the demand keyword in at least one dimension; The first confidence level is calculated based on the confidence level and weight corresponding to each of the aforementioned demand keywords.

7. The information recommendation method according to claim 2, wherein prompting the user to provide second session information includes at least one of the following: If the first dimension of the preset dimension type is missing from the first confidence level, the user is prompted to provide the requirement information corresponding to the first dimension. If the confidence level of the second dimension in the first confidence level is less than the target confidence level threshold, the user is prompted to provide the required information corresponding to the second dimension.

8. The information recommendation method according to claim 1, further comprising: Identify the current recommendation scenario; Based on the current recommendation scenario, determine the target confidence threshold for at least one dimension; or In response to the absence of a target object in the preset knowledge graph that corresponds to the first user intent, a third user intent is generated, and the recommended object and / or related information corresponding to the third user intent are output; the third user intent is generated after adjusting the variable values ​​of the target object in at least one dimension.

9. An electronic device comprising at least one processor and a target application running on the processor, the target application being capable of independently executing or invoking at least one artificial intelligence model to perform the following operation: identifying a first user intent in first session information; Based on the first user intent, candidate recommendation information and the confidence level of the first user intent are determined; the confidence level is used to characterize the target parameters of the target object indicated by the first user intent in at least one dimension. In response to the confidence level being greater than a preset confidence threshold, target recommendation information corresponding to the target object is determined from the candidate recommendation information; A display device is used to display the target recommendation information.

10. The electronic device according to claim 9, wherein the first session information belongs to session information in a product recommendation scenario; The target application is used to determine a candidate product recommendation list in a preset knowledge graph and a first confidence level of the first user intent based on the first user intent. The first confidence level is used to characterize the target parameter of the target product indicated by the first user intent in at least one dimension; The preset knowledge graph includes a set of candidate products in the product recommendation scenario and dimensional information and user evaluation information associated with the set of candidate products. The candidate product recommendation list includes at least one candidate product in the product recommendation scenario and dimensional information associated with the at least one candidate product. In response to the first confidence level being greater than the preset confidence threshold, target recommendation information corresponding to the target product is determined from the candidate product recommendation list.

Images