Interaction strategy recommendation method and device, electronic equipment and storage medium

Abstract

The invention provides an interaction strategy recommendation method which can be applied to the technical field of artificial intelligence and financial science and technology. The method comprises the following steps: acquiring user portrait data and real-time interaction data; generating a multi-dimensional user static feature vector based on the user portrait data; determining a current dynamic dialogue context state based on the real-time interaction data; inputting the current dynamic dialogue context state and a preset service target parameter into a gating weight network, and generating a group of dynamic fusion weights through the gating weight network; performing weighted fusion on the multi-dimensional user static feature vector according to the dynamic fusion weight to obtain a target user dynamic feature; and based on the dynamic characteristics of the target user, matching and generating a recommended interaction strategy from a preset interaction strategy library. The invention further provides an interaction strategy recommendation device, electronic equipment and a storage medium.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, specifically to the field of financial technology, and more specifically to an interactive strategy recommendation method, apparatus, electronic device, and storage medium. Background Technology

[0002] In current intelligent outbound calling and online customer service systems, achieving accurate personalized interaction strategy recommendations is crucial for improving service quality and efficiency. However, related technologies have significant limitations: rule-based recommendation engines rely on manually pre-defined strategies, whose rules and weights are fixed and cannot dynamically adjust priorities based on real-time dialogue progress (such as changes in user emotions and intentions), resulting in high recommendation bias rates and poor adaptability; while static label-based matching models incorporate some customer profiles, they can only utilize structured labels and cannot deeply integrate multimodal heterogeneous features such as voice emotion, interactive text, and behavioral sequences, and lack the ability to respond to real-time dialogue context, leading to rigid recommendation strategies and low feature coverage. Therefore, an intelligent recommendation method for interaction strategy recommendation is urgently needed. Summary of the Invention

[0003] In view of the above problems, this application provides an interactive strategy recommendation method, apparatus, electronic device, and storage medium.

[0004] According to a first aspect of this application, an interaction strategy recommendation method is provided, comprising: acquiring user profile data and real-time interaction data with user authorization; generating a multi-dimensional user static feature vector based on the user profile data, wherein the multi-dimensional user static feature vector is a set of feature vectors representing the user's historical stability characteristics, generated based on multimodal historical information contained in the user profile data; determining the current dynamic dialogue context state based on the real-time interaction data; inputting the current dynamic dialogue context state and preset business target parameters into a gating weight network, generating a set of dynamic fusion weights through the gating weight network, wherein the number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension; performing weighted fusion of the multi-dimensional user static feature vector according to the dynamic fusion weights to obtain target user dynamic features; and matching and generating a recommended interaction strategy from a preset interaction strategy library based on the target user dynamic features.

[0005] According to an embodiment of this application, generating a multi-dimensional user static feature vector based on the user profile data includes: generating a basic attribute vector based on the basic attribute data in the user profile data; generating a behavior pattern vector based on historical behavior log data; generating a text semantic vector based on historical interactive text data through a semantic encoding model; and generating a speech emotion feature vector based on historical speech data through an acoustic analysis model.

[0006] According to an embodiment of this application, determining the current dynamic dialogue context state based on the real-time interaction data includes: performing real-time emotion recognition on the current round of speech in the real-time interaction data to obtain a real-time emotion tag; performing real-time intent recognition on the current round of dialogue text in the real-time interaction data to obtain the user's real-time intent; obtaining the current dialogue round information; and constructing the current dynamic dialogue context state based on the real-time emotion tag, the user's real-time intent, and the dialogue round information.

[0007] According to an embodiment of this application, the gated weight network is a trained multilayer perceptron. The step of generating a set of dynamic fusion weights through the gated weight network includes: concatenating the current dynamic dialogue context state and the preset business target parameters into an input vector; inputting the input vector into the multilayer perceptron for nonlinear transformation to obtain the original weight values; normalizing the original weight values ​​and outputting the dynamic fusion weights.

[0008] According to an embodiment of this application, the method further includes: in response to determining that the multi-dimensional user static feature vector satisfies a preset sparsity condition, inputting the multi-dimensional user static feature vector and a pre-stored set of historical customer feature vectors into a prediction model trained by a meta-learning framework; calculating the multi-dimensional user static feature vector and the set of historical customer feature vectors through the prediction model, and outputting a similar prototype feature vector; and using the similar prototype feature vector as the target user dynamic feature to execute a matching and generation recommendation interaction strategy.

[0009] According to an embodiment of this application, the step of matching and generating a recommended interaction strategy from a preset interaction strategy library based on the target user's dynamic features includes: calculating the attention similarity between the target user's dynamic features and multiple strategy vectors in the preset interaction strategy library; determining at least one candidate strategy based on the attention similarity; inputting the candidate strategy and the current dynamic dialogue context state into a controllable text generation model, and outputting the recommended interaction strategy.

[0010] According to an embodiment of this application, the step of weighting and fusing the multi-dimensional user static feature vectors according to the dynamic fusion weights to obtain the target user dynamic features includes: inputting the real-time interaction data into a recurrent neural network unit and outputting weight adjustment information; correcting the dynamic fusion weights based on the weight adjustment information; and using the corrected dynamic fusion weights to weight and fuse the multi-dimensional user static feature vectors to obtain the target user dynamic features.

[0011] A second aspect of this application provides an interaction strategy recommendation device, comprising: an acquisition module, configured to acquire user profile data and real-time interaction data upon obtaining user authorization; a first generation module, configured to generate a multi-dimensional user static feature vector based on the user profile data, wherein the multi-dimensional user static feature vector is a set of feature vectors representing the historical stability characteristics of the user, generated based on multimodal historical information contained in the user profile data; a second generation module, configured to determine the current dynamic dialogue context state based on the real-time interaction data; and a dynamic fusion weight generation module, configured to input the current dynamic dialogue context state and preset business target parameters into a gated weight network, and generate a set of dynamic fusion weights through the gated weight network, wherein the number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension.

[0012] The fusion module is used to perform weighted fusion of the multi-dimensional user static feature vector according to the dynamic fusion weight to obtain the target user dynamic features; and the matching and recommendation module is used to match and generate a recommended interaction strategy from a preset interaction strategy library based on the target user dynamic features.

[0013] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0014] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0015] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0016] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0017] Figure 1 This diagram illustrates an application scenario of the interactive strategy recommendation method according to an embodiment of this application.

[0018] Figure 2 A flowchart illustrating an interactive strategy recommendation method according to an embodiment of this application is shown schematically.

[0019] Figure 3 A flowchart illustrating a cold-start recommendation method based on meta-learning according to an embodiment of this application is shown schematically.

[0020] Figure 4 This schematically illustrates a structural block diagram of an interactive strategy recommendation device according to an embodiment of the present application; and

[0021] Figure 5 A block diagram schematically illustrates an electronic device suitable for implementing an interactive strategy recommendation method according to an embodiment of this application. Detailed Implementation

[0022] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0024] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0025] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0026] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0027] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0028] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0029] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to activities that automatically analyze and evaluate an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then make decisions. Here, "expert decision-making" refers to activities where individuals specializing in a particular field, possessing specialized experience, knowledge, and skills, and reaching a certain level of professional expertise make decisions.

[0030] This application provides an interaction strategy recommendation method, comprising: acquiring user profile data and real-time interaction data with user authorization; generating a multi-dimensional user static feature vector based on the user profile data, wherein the multi-dimensional user static feature vector is a set of feature vectors representing the user's historical stability characteristics, generated based on multimodal historical information contained in the user profile data; determining the current dynamic dialogue context state based on the real-time interaction data; inputting the current dynamic dialogue context state and preset business target parameters into a gating weight network, generating a set of dynamic fusion weights through the gating weight network, wherein the number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension; weighting and fusing the multi-dimensional user static feature vector according to the dynamic fusion weights to obtain target user dynamic features; and matching and generating a recommended interaction strategy from a preset interaction strategy library based on the target user dynamic features.

[0031] Figure 1 The diagram illustrates an application scenario of the interactive strategy recommendation method according to an embodiment of this application.

[0032] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0033] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as web browser applications, instant messaging tools, email clients, social platform software, etc. (for example only).

[0034] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0035] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0036] It should be noted that the interaction strategy recommendation method provided in this embodiment can generally be executed by server 105. Correspondingly, the interaction strategy recommendation device provided in this embodiment can generally be located in server 105. The interaction strategy recommendation method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the interaction strategy recommendation device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0037] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0038] The following will be based on Figure 1 The described scene, through Figures 2-5 The interaction strategy recommendation method according to the embodiments of this application will be described in detail.

[0039] Figure 2 A flowchart illustrating an interactive strategy recommendation method according to an embodiment of this application is shown.

[0040] like Figure 2 As shown, the interaction strategy recommendation method in this embodiment includes operations S210 to S260, which can be executed by a server or a terminal device.

[0041] When operating S210, with user authorization, user profile data and real-time interaction data are obtained.

[0042] In one example, before initiating the interaction process, the system must ensure that it has obtained explicit authorization from the user for data processing, and that this authorization complies with relevant laws and regulations. The acquired data mainly includes two categories: User profile data: This refers to a collection of historical data related to the target user extracted from persistent storage such as Customer Relationship Management (CRM) systems and data warehouses. It covers multimodal historical information, including but not limited to: basic customer attributes (such as age, occupation, and region), historical product holdings and transaction records, historical customer service work order texts, voice recordings of historical calls, and behavioral logs on the bank's app or website. Real-time interaction data: This refers to the currently ongoing voice dialogue stream and its derived real-time text stream. Specifically, the system acquires the user's current voice data through a real-time audio interface and converts it into text data in real time using an Automatic Speech Recognition (ASR) engine, forming a dialogue text stream. This data stream carries the instantaneous state information of the dialogue.

[0043] In operation S220, a multi-dimensional static feature vector of the user is generated based on the user profile data.

[0044] According to an embodiment of this application, the multi-dimensional user static feature vector is a set of feature vectors generated based on the multimodal historical information contained in the user profile data, used to characterize the historical stability features of the user.

[0045] In one example, the system extracts and vectorizes features from the acquired user profile data to construct a comprehensive static feature representation. This multi-dimensional user static feature vector is generated based on the multimodal historical information contained in the user profile data, and is a set of feature vectors used to characterize the historical stability features of the user. In a specific embodiment, this set is constructed by processing multiple data modalities in parallel, including:

[0046] Basic attribute vector: Structured basic attributes (such as occupational codes, regional codes, and asset grade segments) are encoded using one-hot encoding or embedding encoding to form a numerical vector.

[0047] Behavioral pattern vector: Construct a time series graph or sequence from the user's historical transaction sequence, APP click stream and other behavioral logs, input it into a graph neural network (GNN) or a long short-term memory network (LSTM), extract its high-frequency behavioral patterns and preferences, and output it as a behavioral embedding vector.

[0048] Text semantic vectors: These are generated by collecting users' historical work orders and online consultation texts, and using a pre-trained semantic model (such as BERT) fine-tuned with financial domain corpora to encode the text and generate text semantic vectors that represent their historical focus and semantic tendencies.

[0049] Voice emotion feature vector: Acoustic features such as Mel-spectrogram are extracted from the user's historical call recordings. Convolutional neural networks (such as ResNet) are used to classify emotions (such as calm, anger, and joy), and output a probability distribution vector representing the stable emotional tendencies in the user's historical calls.

[0050] In operation S230, the current dynamic dialogue context state is determined based on the real-time interaction data.

[0051] In one example, the system performs real-time analysis of interactive data, specifically including: Real-time emotion recognition: Inputting the current user's speech fragment into a lightweight real-time acoustic emotion analysis model, outputting a real-time emotion label (e.g., "neutral," "confused," "impatient"). Real-time intent recognition: Inputting the current user's dialogue text obtained from ASR (Automatic Sentence Recognition) into a real-time intent classification model to identify the user's current immediate intent (e.g., "inquiring about rates," "rejecting," "requesting explanation"). Context information extraction: Recording and acquiring metadata such as the current dialogue turn and the duration of the dialogue. The system integrates the above real-time emotion label, user real-time intent, and dialogue turn information to construct a structured current dynamic dialogue context state vector, which serves as a digital summary of the instantaneous dialogue situation.

[0052] In operation S240, the current dynamic dialogue context state and preset business target parameters are input into the gating weight network, and a set of dynamic fusion weights are generated through the gating weight network.

[0053] According to an embodiment of this application, the number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension.

[0054] In one example, the network in this embodiment consists of a multilayer perceptron (MLP) and a softmax output layer. Its workflow is as follows: Input concatenation: The current dynamic dialogue context state vector, representing the current situation, is concatenated with a preset business objective parameter vector representing the core task of this interaction (e.g., [activation: 0, conversion: 1, risk warning: 0]) to form a unified input vector. Nonlinear mapping and weight generation: This input vector is mapped to an array of original weight scores through multilayer nonlinear transformations of the MLP. Weight normalization: The original weight score array is normalized by the softmax layer to ensure that the sum of all weights is 1, and the final output is a set of dynamic fusion weights. The number of dynamic fusion weights strictly matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension. For example, if the static feature vector contains 4 dimensions (attribute, behavior, text, speech), the gating network outputs an array containing 4 weight values, which respectively determine the importance of these 4 vectors in subsequent fusion.

[0055] In operation S250, the multi-dimensional user static feature vector is weighted and fused according to the dynamic fusion weight to obtain the target user dynamic features.

[0056] In one example, the system performs a weighted summation of the dynamic fusion weights generated by operation S240 with the corresponding dimension of the user static feature vectors generated by operation S220. Specifically, assume the dynamic fusion weights are [w1, w2, w3, w4] and the static feature vectors are [V attr V behav V text V audio ], then the target user's dynamic features V dynamic =w1*V attr +w2*V behav +w3*V text +w4* V audio The resulting V dynamic It is no longer a static profile, but a comprehensive feature representation that is dynamically generated based on the current dialogue context and business objectives, reflecting "how to understand the user in the current scenario".

[0057] In operation S260, based on the target user's dynamic characteristics, a recommended interaction strategy is matched and generated from a preset interaction strategy library.

[0058] In one example, the attention similarity between the target user's dynamic features and the feature vectors of all strategy templates in a pre-defined interaction strategy library is calculated. This library stores a large number of standard strategy templates labeled with feature vectors. Through an attention mechanism, the system can focus on the strategy dimension most relevant to the current dynamic features, calculate a matching score, and select one or more candidate strategy templates with the highest matching score. The selected candidate strategy templates and their context, along with the current dynamic dialogue context state, are input into a controllable text generation model (e.g., a model based on a T5 architecture). This model receives instructions to personalize the candidate templates, including but not limited to: entity replacement based on user profiles (e.g., filling in product names or customer surnames), adjusting tone and wording based on real-time sentiment tags, and strengthening key information based on business objectives. Finally, the model outputs a fully personalized, fluent recommendation interaction strategy text that fits the current dialogue, for use by agents or read aloud by a speech synthesis system.

[0059] The interactive strategy recommendation method provided in this application first constructs a multi-dimensional static user feature vector based on multimodal historical information to form a stable user profile. Then, it combines the current dynamic dialogue context state determined by real-time interaction data to capture the instantaneous situation. Using a gated weight network, it generates dynamic fusion weights that match the static feature dimensions based on the dynamic state and business objectives. This achieves dynamic and contextualized fusion of user long-term and short-term features, generating dynamic features of the target user that accurately represent the current needs. Finally, strategy matching and generation are performed based on these dynamic features. This ensures that the recommended interactive strategy is not only rooted in the user's historical profile but also responds in real time to the progress of the dialogue and the user's immediate feedback. This significantly improves the accuracy, adaptability, and personalization of strategy recommendation, effectively overcoming the problems of rigid strategies, inability to dynamically adjust, and poor cold start effect in existing technologies. This effectively improves business processing efficiency and customer satisfaction.

[0060] According to embodiments of this application, a multi-dimensional user static feature vector is generated based on the user profile data, including: generating a basic attribute vector based on the basic attribute data in the user profile data; generating a behavior pattern vector based on historical behavior log data; generating a text semantic vector based on historical interactive text data through a semantic encoding model; and generating a speech emotion feature vector based on historical speech data through an acoustic analysis model.

[0061] In one example, the module first processes the structured basic attribute fields in the user profile data. These fields typically include, but are not limited to, discrete or categorical data such as customer age segments, occupational category codes, registered regions, and asset level classifications. This module uses one-hot encoding or embedding layer techniques to map each attribute value to a numerical vector, and then concatenates or pools all attribute vectors to generate a comprehensive basic attribute vector, representing the user's demographics and basic social attributes in a machine-readable form. Secondly, it processes the user's historical behavior log data, such as transaction records, product browsing history, app function click sequences, and page dwell time. This module typically uses time-series models or graph neural networks to capture sequence dependencies and complex relationships in user behavior. In one specific implementation, the user behavior sequence is constructed as a timestamp-based behavior graph, which is then input into a graph neural network for node and edge embedding learning; for example, a long short-term memory network can also be used directly to model the behavior sequence. The module outputs a behavior pattern vector, which encapsulates deep-seated behavioral pattern features such as the user's transaction habits, risk preferences, and activity cycles.

[0062] In one example, a pre-trained semantic encoding model processes a user's historical interaction text data, such as historical customer service chat logs, email inquiries, and ticket descriptions. The pre-trained semantic encoding model, such as the BERT model based on the Transformer architecture, is used. Before implementation, the model needs to be fine-tuned for domain adaptation using corpora from the financial or customer service domains. During processing, the cleaned historical text is input into the model, and the output of the [CLS] flag or the mean pooling of the last hidden state is typically used to generate a fixed-dimensional text semantic vector. This vector deeply encodes the core issues, expression habits, and potential needs and intentions of the user in historical communications. Finally, the user's historical voice interaction data (such as past call recordings) is analyzed. The processing flow begins with speech signal preprocessing and feature extraction, such as calculating Mel frequency cepstral coefficients or Mel spectrograms. These acoustic features are fed into a trained acoustic analysis model, typically based on a convolutional neural network architecture, whose training objective is to classify speech segments by emotion. The model does not directly output discrete emotion labels, but rather an emotion probability distribution vector (such as the predicted probability for categories like "calm," "satisfied," "anxious," and "angry"). By weighted averaging or clustering the analysis results of a sufficient number of historical speech segments from the user, a speech emotion feature vector representing their stable emotional tendency is generated. Through processing via the four parallel channels mentioned above, the system ultimately combines the generated basic attribute vector, behavioral pattern vector, text semantic vector, and speech emotion feature vector to form the aforementioned "multi-dimensional user static feature vector" set, providing comprehensive and in-depth static feature input for subsequent dynamic fusion and decision-making.

[0063] According to an embodiment of this application, determining the current dynamic dialogue context state based on the real-time interaction data includes: performing real-time emotion recognition on the current round of speech in the real-time interaction data to obtain a real-time emotion tag; performing real-time intent recognition on the current round of dialogue text in the real-time interaction data to obtain the user's real-time intent; obtaining the current dialogue round information; and constructing the current dynamic dialogue context state based on the real-time emotion tag, the user's real-time intent, and the dialogue round information.

[0064] In one example, the system captures the user's speech data for the current conversation turn via a real-time audio streaming interface. This speech segment is fed into a lightweight real-time acoustic sentiment analysis model. This model typically uses Mel spectrograms as input features and employs efficient architectures such as convolutional neural networks. Trained on a large amount of sentiment-labeled speech data, it can quickly output the sentiment classification result for the current speech segment, i.e., a real-time sentiment label (e.g., "calm," "pleasant," "confused," "impatient," or "angry"). This process is completed within hundreds of milliseconds, ensuring immediate feedback. Next, real-time intent recognition is performed. Parallel to sentiment recognition, the system uses an automatic speech recognition engine to convert the user's speech from the same turn into text in real time. This text is then input into a real-time intent classification model. This model is fine-tuned based on a pre-trained language model (such as a lightweight version of BERT), and its output is the user's real-time intent, i.e., a judgment of the core purpose of the user's current statement (e.g., "asking for product information," "expressing refusal intention," "requesting human assistance," "confirming the processing procedure"). This step transforms unstructured dialogue text into machine-understandable semantic actions. The system acquires and maintains the meta-information of the dialogue. This primarily includes the current dialogue turn information, i.e., the count of rounds the system has interacted with the user since the start of the dialogue. This information is a crucial indicator for judging the stage, pace, and depth of the dialogue. Finally, a dynamic dialogue context state vector is constructed. The system characterizes and vectorizes the aforementioned real-time sentiment tags, user real-time intent, and dialogue turn information. For example, sentiment tags and intent tags are converted into dense vectors using an embedding table, and turn information is standardized or segmented for encoding. Subsequently, these vectors are concatenated and fused and dimensionality-reduced through a lightweight fully connected network, ultimately forming a fixed-dimensional, structured current dynamic dialogue context state vector. This state vector integrates the user's "current emotional state," "immediate speaking purpose," and "stage of the dialogue," constituting a comprehensive digital snapshot of the current interaction moment, providing direct contextual input for the dynamic decision-making of the gating weight network.

[0065] According to an embodiment of this application, the gated weight network is a trained multilayer perceptron. The step of generating a set of dynamic fusion weights through the gated weight network includes: concatenating the current dynamic dialogue context state and the preset business target parameters into an input vector; inputting the input vector into the multilayer perceptron for nonlinear transformation to obtain the original weight values; and normalizing the original weight values ​​to output the dynamic fusion weights.

[0066] In one example, the current dynamic dialogue context state vector, representing the instantaneous situation, is concatenated with a preset business objective parameter vector defining the core task of this interaction to form a comprehensive input vector. The business objective parameters are set during the initialization of the interaction task and are typically encoded in vector form (e.g., for a "credit card installment marketing" task, the parameter vector might emphasize the "conversion" dimension; for a "repayment reminder" task, it might emphasize the "compliance completion" and "customer experience" dimensions). This concatenation operation integrates two types of information: "how the user is now" and "what we need to do at this moment," serving as a joint basis for network decision-making. A deep nonlinear transformation is then performed. The input vector obtained from the concatenation is fed into a multilayer perceptron (MLP). This MLP is typically composed of multiple fully connected layers stacked together, with nonlinear activation functions (such as the ReLU function) between layers. Each layer performs an abstract mathematical transformation on the input data. Through this stacked nonlinear transformation, the network can learn and model an extremely complex, nonlinear mapping relationship from the "comprehensive situation" to the "preliminary judgment of feature importance." After the final hidden layer calculation, the network outputs a raw, unverified array of weight scores, i.e., the raw weight values. The dimensions of this array are fixed during design to ensure its length strictly matches the total number of dimensions of the multi-dimensional user static feature vector to be weighted. The resulting raw weight value array is then normalized. In a typical embodiment, the Softmax function is used as the normalizer. The Softmax function converts each raw weight value into a probability value between 0 and 1, ensuring that the sum of all transformed weight values ​​is 1. This step makes the weights on different dimensions comparable, directly representing the relative importance or attention allocation ratio of that dimension feature in the current specific context; its probabilistic characteristics make the output stable and easy for subsequent calculations. The normalized array is the final output dynamic fusion weight. For example, if the static feature vector contains four dimensions: attribute, behavior, text, and speech, the network outputs a four-dimensional weight vector [w1, w2, w3, w4], where w1+w2+w3+w4=1, and w1 represents the fusion weight that the basic attribute vector should occupy in the current dialogue round.

[0067] It should be noted that all parameters (weights and biases of each layer) in this gating weight network are trained using a large amount of historical dialogue data and its success labels (such as whether a conversion or complaint occurred) through supervised learning or reinforcement learning. The training objective is to enable the network to learn to output fusion weights that achieve the optimal business results for the final strategy recommendation in a given context, thereby realizing the automation and intelligence of weight allocation and replacing the fixed rules preset by human experience in traditional systems.

[0068] Figure 3 A flowchart illustrating a cold-start recommendation method based on meta-learning according to an embodiment of this application is shown. Figure 3 As shown, operations S310 to S330 are included. Operations S310 to S330 describe an intelligent processing branch for users with sparse features, namely a cold-start recommendation path based on meta-learning. This path is activated when the main process determines that the user's static features are insufficient. It aims to generate a reasonable and usable feature representation for new users or long-tail users by transferring knowledge from group experience, thereby ensuring the coverage and robustness of the recommendation system.

[0069] In operation S310, in response to determining that the multi-dimensional user static feature vector satisfies the preset sparsity condition, the multi-dimensional user static feature vector and the pre-stored set of historical customer feature vectors are input into the prediction model trained by the meta-learning framework.

[0070] In one example, a predefined decision module evaluates whether the multi-dimensional user static feature vector meets a preset sparsity condition. This sparsity condition is typically set based on the effective information density of the vector; for example, it is considered sparsity when the vector's magnitude is below a threshold, the proportion of non-zero elements is too small, or the uncertainty score based on confidence is too high. Once the condition is met, the system triggers a cold start process. This process calls a prediction model trained by a meta-learning framework and inputs the current sparse static feature vector along with a pre-stored set of historical customer feature vectors containing a massive number of registered users' complete static feature vectors. Meta-learning frameworks, such as Model-Agnostic Meta-Learning (MAML), are designed to enable models to "learn to adapt quickly." During the training phase, the model does not learn to match a specific user but learns a general, fast similarity matching and prototype inference algorithm by repeatedly training on massive "simulated cold start tasks" (i.e., deliberately masking some features from rich data to construct sparse samples).

[0071] In operation S320, the prediction model is used to calculate the multi-dimensional user static feature vector and the historical customer feature vector set, and output a similar prototype feature vector.

[0072] In one example, the predictive model, based on its acquired meta-capabilities, rapidly computes the input sparse vector and the historical vector set. Its computational goal is not to find a single most similar user, but rather to locate a "group" of multiple similar users within the high-dimensional feature space for the sparse vector. The model quickly identifies the potential group best matching the new user by measuring the distance between the sparse vector and the cluster centers or representative vectors in the historical set. Subsequently, the model outputs the center vector of this group or the representative vector obtained through aggregation computation, called the similarity prototype feature vector. This prototype vector is not any real individual, but rather a stable vector distilled from the similar group data, representing the common characteristics of this type of user. It effectively compensates for the lack of information about a single new user.

[0073] In operation S330, the similar prototype feature vector is used as the target user's dynamic feature to perform the matching and generation of a recommendation interaction strategy.

[0074] In one example, the system directly assigns the similar prototype feature vector obtained in operation S320 to the sparse feature user, and uses it as the target user dynamic feature for the current user in this round of dialogue. Afterwards, the process jumps to the matching generation step of the main process, i.e., operation S260, to match and generate strategies using this prototype feature. Through this mechanism, this application enables the system to provide relatively reasonable and personalized strategy recommendations based on collective intelligence when facing users with scarce data, thereby significantly improving the system's applicability and user experience in cold start scenarios.

[0075] According to an embodiment of this application, matching and generating recommended interaction strategies from a preset interaction strategy library based on the target user's dynamic features includes: calculating the attention similarity between the target user's dynamic features and multiple strategy vectors in the preset interaction strategy library; and determining at least one candidate strategy based on the attention similarity.

[0076] The candidate strategy and the current dynamic dialogue context state are input into the controllable text generation model to output the recommended interaction strategy.

[0077] In one example, an attention mechanism is used to calculate the strategy matching degree. The system pre-builds and maintains a preset interaction strategy library. Each standard interaction strategy template in the library, such as "product introduction opening remarks," "rate Q&A," and "closing a transaction invitation," is encoded into a fixed-dimensional strategy vector through a feature extraction model. During matching, the currently generated target user dynamic feature vector is used as the query vector, and all strategy vectors in the strategy library are used as key and value vectors, respectively, and then substituted into the attention mechanism for calculation. Specifically, by calculating the dot product between the target feature vector and each strategy vector, scaling it, and applying the Softmax function for normalization, a set of attention similarity scores is obtained. This score is not a simple cosine similarity, but dynamically measures the relative importance of different dimensions of each strategy template in the strategy library under the current specific user characteristics, thereby achieving focused and context-aware matching. Candidate strategies are determined based on the matching scores. The system sorts the calculated attention similarity scores and selects one or more strategy templates with the highest scores as candidate strategies. In one embodiment, the system selects the top three (TOP-3) as the candidate set to provide some diversity and selection space for subsequent generation stages. The final strategy is output using a controllable generative model. The candidate strategies determined in the previous step (usually in their original text template form) and the current dynamic dialogue context state vector representing the real-time situation are input into a controllable text generation model (e.g., a model based on the T5 architecture). This model operates in a "text-to-text" paradigm, and its core task is to conditionally rewrite the input text based on control signals. In this scenario, the "current dynamic dialogue context state" serves as the core control signal. Based on the real-time sentiment, intent, and other information contained in this state vector, the model performs a series of controllable rewriting operations on the candidate strategy template, mainly including: 1) Entity replacement: Replacing generic placeholders in the template with specific user and product information based on user profiles; 2) Sentiment and tone adaptation: Adjusting the strength of sentences and the emotional tone of words based on real-time sentiment tags (e.g., changing neutral statements to have a reassuring or positive tone); 3) Information simplification or expansion: Summarizing or refining the template content based on the dialogue rounds and complexity. Finally, the model outputs a fully personalized natural language text that seamlessly integrates with the current dialogue flow, thus completing the end-to-end generation from features to an executable strategy.

[0078] According to an embodiment of this application, the step of weighting and fusing the multi-dimensional user static feature vectors according to the dynamic fusion weights to obtain the target user dynamic features includes: inputting the real-time interaction data into a recurrent neural network unit and outputting weight adjustment information; correcting the dynamic fusion weights based on the weight adjustment information; and using the corrected dynamic fusion weights to weight and fuse the multi-dimensional user static feature vectors to obtain the target user dynamic features.

[0079] Embodiments of this application also include an online adaptive and optimization closed loop, designed to enable the system to dynamically optimize its core decision-making logic based on the user's immediate reactions. This closed loop drives a recurrent neural network through real-time feedback, enabling iterative fine-tuning of the gating weight network's future decisions, thus allowing the system's strategy recommendation capabilities to continuously evolve throughout the dialogue process.

[0080] As an example, the real-time interaction data acquired in this operation could be, for instance, real-time user feedback data after the execution of a recommended interaction strategy, such as through agent announcements or intelligent voice announcements. The system immediately initiates multi-dimensional monitoring and analysis of the user's next round of responses to obtain real-time user feedback data based on the recommended interaction strategy. Real-time feedback data includes at least a voice emotion change vector. By comparing the user's voice segments before and after hearing the recommended strategy, the difference in emotional state or the new emotional probability distribution calculated by a real-time acoustic emotion model is used to quantify the immediate impact of the strategy on the user's emotions (e.g., from "confused" to "calm," or from "neutral" to "impatient"). Real-time feedback data also includes textual intent and content feedback. Through ASR and intent recognition models, the system analyzes the user's response text immediately following the recommended strategy to obtain new real-time user intent and key entity words, used to determine whether the strategy has been accepted, rejected, or raised new questions. Real-time feedback data also includes dialogue behavior signals, such as whether the user interrupts before the strategy is finished, the delay in response, and changes in non-textual features of tone.

[0081] In one example, the fused real-time feedback data is encoded into a vector form and input into a recurrent neural network unit. In a preferred embodiment, a Gated Recurrent Unit (GRU) is used. The GRU maintains a hidden state that implicitly encodes the dialogue history and adjustment memories up to the present. The GRU calculates the weight adjustment increment signal by combining the newly arrived feedback data vector with the current hidden state, i.e., the weight adjustment information. This signal does not directly modify the already output strategy, but is used to iteratively adjust the dynamically fused weights generated by the gated weight network in the next iteration.

[0082] In one example, the adjustment signal can serve as an additional input or bias condition for the gated weight network when calculating the dynamic fusion weights for the next round, thereby subtly altering its weight allocation strategy. In one feasible implementation, the weight adjustment information is used as an enhancement input. In this implementation, the system establishes a cache unit to temporarily store the weight adjustment information calculated by the recurrent neural network unit (GRU) based on the feedback from the previous round. When the next round (round N+1) of recommendation is needed, the system performs the following operations: obtains the new current dynamic dialogue context state (denoted as S). {n+1} The input vector is formed by combining the preset business target parameters (denoted as G) with the cached weight adjustment information Δ and concatenating them together to form an enhanced input vector, i.e., [S]. {n+1} The enhanced input vector is then fed into the gated weight network. This means that the decision-making basis of the gated weight network in this round includes not only the latest situation (S, G, Δ). {n+1} The algorithm, along with the objective (G), explicitly includes an empirical summary (Δ) based on the effects of the previous round of interactions. The gated weighted network performs forward computation based on this enhanced information, which incorporates historical experience. Its final output, dynamically fused weights, inherently and autonomously references and integrates feedback from the previous round, thereby adjusting the weight allocation. This approach enables the network to intelligently weigh historical experience against the current new context.

[0083] In another feasible embodiment, the weight adjustment information is used as a dynamic bias. In this implementation, the weight adjustment information Δ is directly used to fine-tune the mathematical calculation process within the gated weight network. Specifically, the Δ vector is used as a dynamic bias term. After the gated weight network completes its multilayer perceptron nonlinear transformation and obtains the original weight scores, before being sent to the Softmax layer for normalization, the dynamic bias term Δ is added element-wise to the original weight score vector. If the original score is Z, the corrected input is Z + Δ, which is then normalized by the Softmax function to output the final weights. This operation directly and explicitly increases or decreases the proportion of a specific feature dimension in the final weights. For example, the Δ vector is converted into a scaling factor vector S, S = 1 + αΔ, where α is the scaling coefficient. During the forward propagation of the network, when calculating a specific hidden layer, the activation value output of that layer is multiplied element-wise with the scaling factor vector S before being passed to the next layer. This operation temporarily and dynamically amplifies or reduces the signal strength of specific feature paths flowing to subsequent parts of the network, thereby indirectly but effectively affecting the weight distribution of the final output. By using the modified dynamic fusion weights to perform weighted fusion of multi-dimensional user static feature vectors, dynamic features of the target user are obtained.

[0084] Based on the above-described interaction strategy recommendation method, this application also provides an interaction strategy recommendation device. The following will combine... Figure 4The device is described in detail.

[0085] Figure 4 A schematic diagram illustrating the structure of an interactive strategy recommendation apparatus according to an embodiment of this application is provided. Figure 4 As shown, the interactive strategy recommendation device 400 of this embodiment includes an acquisition module 410, a first generation module 420, a second generation module 430, a dynamic fusion weight generation module 440, a fusion module 450, and a matching recommendation module 460.

[0086] The acquisition module 410 is used to acquire user profile data and real-time interaction data when authorized by the user. In one embodiment, the acquisition module 410 can be used to perform the operation S210 described above, which will not be repeated here.

[0087] The first generation module 420 is used to generate a multi-dimensional user static feature vector based on the user profile data. The multi-dimensional user static feature vector is a set of feature vectors generated based on the multimodal historical information contained in the user profile data, used to characterize the historical stability features of the user. In one embodiment, the first generation module 420 can be used to perform the operation S220 described above, which will not be repeated here.

[0088] The second generation module 430 is used to determine the current dynamic dialogue context state based on the real-time interaction data. In one embodiment, the second generation module 430 can be used to perform the operation S230 described above, which will not be repeated here.

[0089] The dynamic fusion weight generation module 440 is used to input the current dynamic dialogue context state and preset business target parameters into the gated weight network, and generate a set of dynamic fusion weights through the gated weight network. The number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension. In one embodiment, the dynamic fusion weight generation module 440 can be used to perform the operation S240 described above, which will not be repeated here.

[0090] The processing suggestion generation module 440 is used to generate processing suggestions based on the key features and the influence weights of the key features. In one embodiment, the processing suggestion generation module 440 can be used to perform the operation S240 described above, which will not be repeated here.

[0091] The fusion module 450 is used to perform weighted fusion of the multi-dimensional user static feature vector according to the dynamic fusion weight to obtain the target user dynamic features. In one embodiment, the fusion module 450 can be used to perform the operation S250 described above, which will not be repeated here.

[0092] The matching and recommendation module 460 is used to match and generate recommended interaction strategies from a preset interaction strategy library based on the target user's dynamic characteristics. In one embodiment, the matching and recommendation module 460 can be used to perform the operation S260 described above, which will not be repeated here.

[0093] According to an embodiment of this application, the first generation module further includes: a first generation submodule, a second generation submodule, a third generation submodule, and a fourth generation submodule.

[0094] The first generation submodule generates a basic attribute vector based on the basic attribute data in the user profile data; the second generation submodule generates a behavior pattern vector based on historical behavior log data; the third generation submodule generates a text semantic vector based on historical interactive text data through a semantic encoding model; and the fourth generation submodule generates a speech emotion feature vector based on historical speech data through an acoustic analysis model.

[0095] According to an embodiment of this application, the second generation module further includes an emotion recognition submodule, an intent recognition submodule, an acquisition submodule, and a construction submodule.

[0096] The emotion recognition submodule performs real-time emotion recognition on the current round of speech in the real-time interaction data to obtain real-time emotion tags; the intent recognition submodule performs real-time intent recognition on the current round of dialogue text in the real-time interaction data to obtain the user's real-time intent; the acquisition submodule acquires the current dialogue round information; and the construction submodule constructs the current dynamic dialogue context state based on the real-time emotion tags, the user's real-time intent, and the dialogue round information.

[0097] According to an embodiment of this application, the dynamic fusion weight generation module is specifically used to concatenate the current dynamic dialogue context state and the preset business target parameters into an input vector; input the input vector into the multilayer perceptron for nonlinear transformation to obtain the original weight value; and normalize the original weight value to output the dynamic fusion weight.

[0098] According to an embodiment of this application, the device further includes a cold start module.

[0099] The cold start module is used to respond to the determination that the multi-dimensional user static feature vector satisfies the preset sparsity condition by inputting the multi-dimensional user static feature vector and the pre-stored set of historical customer feature vectors into the prediction model trained by the meta-learning framework; calculating the multi-dimensional user static feature vector and the set of historical customer feature vectors through the prediction model to output a similar prototype feature vector; and using the similar prototype feature vector as the target user dynamic feature to execute the steps of matching and generating a recommendation interaction strategy.

[0100] According to embodiments of this application, the matching recommendation module includes a calculation submodule, a determination submodule, and a recommendation interaction strategy output submodule.

[0101] The calculation submodule is used to calculate the attention similarity between the target user's dynamic features and multiple policy vectors in the preset interaction policy library; the determination submodule is used to determine at least one candidate policy based on the attention similarity; and the recommended interaction policy output submodule is used to input the candidate policy and the current dynamic dialogue context state into a controllable text generation model and output the recommended interaction policy.

[0102] According to an embodiment of this application, the fusion module is further configured to input the real-time interactive data into a recurrent neural network unit and output weight adjustment information; correct the dynamic fusion weights based on the weight adjustment information; and use the corrected dynamic fusion weights to perform weighted fusion of the multi-dimensional user static feature vectors to obtain the target user dynamic features.

[0103] According to embodiments of this application, any multiple modules among the acquisition module 410, the first generation module 420, the second generation module 430, the dynamic fusion weight generation module 440, the fusion module 450, and the matching recommendation module 460 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the acquisition module 410, the first generation module 420, the second generation module 430, the dynamic fusion weight generation module 440, the fusion module 450, and the matching recommendation module 460 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the acquisition module 410, the first generation module 420, the second generation module 430, the dynamic fusion weight generation module 440, the fusion module 450, and the matching recommendation module 460 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0104] Figure 5 A block diagram schematically illustrates an electronic device suitable for implementing an interactive strategy recommendation method according to an embodiment of this application. (See diagram below.) Figure 5As shown, an electronic device 500 according to an embodiment of this application includes a processor 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage portion 508 into a random access memory (RAM) 503. The processor 501 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 501 may also include onboard memory for caching purposes. The processor 501 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0105] RAM 503 stores various programs and data required for the operation of electronic device 500. Processor 501, ROM 502, and RAM 503 are interconnected via bus 504. Processor 501 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 502 and / or RAM 503. It should be noted that the programs may also be stored in one or more memories other than ROM 502 and RAM 503. Processor 501 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0106] According to embodiments of this application, the electronic device 500 may further include an input / output (I / O) interface 505, which is also connected to a bus 504. The electronic device 500 may also include one or more of the following components connected to the input / output (I / O) interface 505: an input section 506 including a keyboard, mouse, etc.; an output section 507 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN card, modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the input / output (I / O) interface 505 as needed. A removable medium 511, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 510 as needed so that computer programs read from it can be installed into the storage section 508 as needed.

[0107] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0108] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the interactive strategy recommendation method provided in the embodiments of this application.

[0109] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 509, and / or installed from a removable medium 511. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0110] In such an embodiment, the computer program can be downloaded and installed from a network via communication section 509, and / or installed from removable medium 511. When the computer program is executed by processor 501, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0111] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0112] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. An interactive strategy recommendation method, characterized in that, include: With user authorization, user profile data and real-time interaction data can be obtained. Based on the user profile data, a multi-dimensional user static feature vector is generated. The multi-dimensional user static feature vector is a set of feature vectors that are generated based on the multimodal historical information contained in the user profile data and are used to characterize the historical stability features of the user. The current dynamic dialogue context state is determined based on the real-time interactive data; The current dynamic dialogue context state and preset business target parameters are input into the gated weight network, and a set of dynamic fusion weights are generated through the gated weight network. The number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension. The multi-dimensional user static feature vector is weighted and fused according to the dynamic fusion weights to obtain the target user's dynamic features; and Based on the dynamic characteristics of the target user, a recommended interaction strategy is matched and generated from a preset interaction strategy library.

2. The method according to claim 1, characterized in that, The process of generating a multi-dimensional user static feature vector based on the user profile data includes: Generate a basic attribute vector based on the basic attribute data in the user profile data; Generate behavioral pattern vectors based on historical behavior log data; Based on historical interactive text data, text semantic vectors are generated through a semantic encoding model; and Based on historical speech data, speech emotion feature vectors are generated through acoustic analysis models.

3. The method according to claim 1, characterized in that, Determining the current dynamic dialogue context state based on the real-time interaction data includes: Real-time emotion recognition is performed on the current round of speech in real-time interactive data to obtain real-time emotion tags; Real-time intent recognition is performed on the current round of dialogue text in real-time interaction data to obtain the user's real-time intent; Get the current dialogue turn information; Based on the real-time emotion tags, the user's real-time intent, and the dialogue turn information, the current dynamic dialogue context state is constructed.

4. The method according to claim 1, characterized in that, The gated weight network is a trained multilayer perceptron, and the generation of a set of dynamic fusion weights through the gated weight network includes: The current dynamic dialogue context state and the preset business target parameters are concatenated into an input vector; The input vector is input into the multilayer perceptron and subjected to a nonlinear transformation to obtain the original weight values; and The original weight values ​​are normalized to output the dynamic fusion weights.

5. The method according to claim 1, characterized in that, The method further includes: In response to determining that the multi-dimensional user static feature vector satisfies the preset sparsity condition, the multi-dimensional user static feature vector and the pre-stored set of historical customer feature vectors are input into the prediction model trained by the meta-learning framework. The prediction model calculates the multi-dimensional user static feature vector and the historical customer feature vector set, and outputs a similar prototype feature vector; and The similar prototype feature vector is used as the target user's dynamic feature to perform the matching and generation of a recommendation interaction strategy.

6. The method according to claim 1, characterized in that, The step of matching and generating recommended interaction strategies from a preset interaction strategy library based on the target user's dynamic characteristics includes: Calculate the attention similarity between the target user's dynamic features and multiple policy vectors in the preset interaction policy library; At least one candidate strategy is determined based on the attention similarity. The candidate strategy and the current dynamic dialogue context state are input into the controllable text generation model to output the recommended interaction strategy.

7. The method according to claim 1, characterized in that, The step of weighting and fusing the multi-dimensional user static feature vectors according to the dynamic fusion weights to obtain the target user dynamic features includes: The real-time interactive data is input into a recurrent neural network unit, which outputs weight adjustment information. The dynamic fusion weights are corrected based on the weight adjustment information; and The multi-dimensional user static feature vector is weighted and fused using the corrected dynamic fusion weights to obtain the target user dynamic features.

8. An interactive strategy recommendation device, characterized in that, The device includes: The acquisition module is used to acquire user profile data and real-time interaction data with user authorization. The first generation module is used to generate a multi-dimensional user static feature vector based on the user profile data. The multi-dimensional user static feature vector is a set of feature vectors that are generated based on the multimodal historical information contained in the user profile data and are used to characterize the historical stability features of the user. The second generation module is used to determine the current dynamic dialogue context state based on the real-time interaction data; The dynamic fusion weight generation module is used to input the current dynamic dialogue context state and preset business target parameters into the gated weight network, and generate a set of dynamic fusion weights through the gated weight network. The number of dynamic fusion weights matches the number of dimensions of the multi-dimensional user static feature vector, and each weight corresponds to a static feature vector of one dimension. The fusion module is used to perform weighted fusion of the multi-dimensional user static feature vectors according to the dynamic fusion weights to obtain the target user dynamic features; and The matching and recommendation module is used to match and generate recommended interaction strategies from a preset interaction strategy library based on the dynamic characteristics of the target user.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images