Instruction processing method and device based on social intention recognition, storage medium and equipment

Abstract

The invention discloses an instruction processing method and device based on social intent recognition, a storage medium and equipment, relates to the technical field of the Internet, and mainly aims to solve the problem of low service efficiency caused by invalid instruction execution of an existing social application. Comprising the steps of determining a social intent based on an application addition count value and user session content; if the social intention is a non-absolute rejection intention, determining a multi-modal feature and a meta-learning weight based on the user session content, and determining an intention score based on the multi-modal feature and the meta-learning weight, the meta-learning weight being determined based on a context feature of the user session content; and when the intention score is greater than a preset intention threshold value, executing a social application adding event, and determining a target social execution instruction based on an adding response and the updated application adding count value, so as to execute a corresponding operation based on the target social execution instruction.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to an instruction processing method, apparatus, storage medium, and device based on social intent recognition. Background Technology

[0002] With the widespread adoption of social media across various industries, private domain operations have become a crucial means for businesses to acquire and retain customers. In some online marketing scenarios, marketers can increase the success rate of product marketing by establishing communication connections with customers through social applications. Accurately identifying customers' communication intentions on social media and initiating add requests at the appropriate time is a key issue in improving customer conversion rates.

[0003] Currently, existing methods for identifying social intent typically rely on simple keyword matching or large language models (LLMs). However, these methods are ineffective when keyword coverage is incomplete or keyword matching only matches single sentences, failing to consider the context of the dialogue and thus failing to identify implicit intents. Furthermore, the high frequency of LLM calls leads to high operational costs and the inability to integrate speech acoustic features and behavioral characteristics results in poor accuracy in complex scenarios, diminishing the effectiveness of triggering relevant commands in social applications and significantly increasing useless operations. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, storage medium, and device for processing instructions based on social intent recognition, with the main purpose of solving the problem of low business efficiency caused by invalid instruction execution in existing social applications.

[0005] According to one aspect of this application, a method for processing instructions based on social intent recognition is provided, comprising: Determine social intent based on application add counts and user conversation content; If the social intent is a non-absolute rejection intent, then multimodal features and meta-learning weights are determined based on the user session content, and an intent score is determined based on the multimodal features and the meta-learning weights, wherein the meta-learning weights are determined based on the contextual features of the user session content. When the intent score is greater than the preset intent threshold, a social application add event is executed, and a target social execution instruction is determined based on the add response and the updated application add count value, so as to execute the corresponding operation based on the target social execution instruction.

[0006] Further, the step of determining multimodal features and meta-learning weights based on the user session content, and determining the intent score based on the multimodal features and the meta-learning weights, includes: Based on the user session content, speech acoustic features, text semantic features, and temporal behavioral features are extracted respectively, and contextual features in the user session content are also extracted; The context features are predicted based on the meta-learning prediction model that has been trained, and meta-learning weights are obtained. The meta-learning prediction model includes a lightweight first hidden layer, a second hidden layer, an input layer, and an output layer. The intention score is obtained by weighted summation of the speech acoustic features, the text semantic features, the temporal behavioral features, and the meta-learning weights.

[0007] Furthermore, the method also includes: If the social intent is to agree to add, then the social application add event is executed, and the add response from the social application is obtained to update the application add count value.

[0008] Furthermore, determining the target social execution instruction based on the added response and the updated application add count value includes: If the response to adding is "add," then the target social execution instruction is determined to be a script broadcast instruction. If the add response is "not added" and the application add count matches the first preset number of times condition, then the target social execution instruction is determined to be "send application add request".

[0009] Furthermore, the application addition count value includes a first count value and a second count value, and the step of determining the target social execution instruction based on the addition response and the updated application addition count value further includes: If the add response is "not added", and the first count value matches the second preset number of times condition, and the second count value matches the third preset number of times condition, then it is determined that the target social execution instruction is to send a non-application add request; The first count value is used to count the add response of the social application, and the second count value is used to count the add response of the non-social application.

[0010] Furthermore, the method also includes: When the intent score is less than or equal to a preset intent threshold, the social intent is determined to be an absolute rejection intent.

[0011] Furthermore, determining social intent based on application-added count values ​​and user session content includes: If the application adds a count value less than a preset threshold, then the number of rejection keywords in each statement of the user's conversation content is extracted. If the number of rejection keywords is less than a preset rejection threshold, then the social intent is determined to be a non-absolute rejection intent. If the number of rejection keywords is greater than or equal to a preset rejection threshold, then the social intent is determined to be an absolute rejection intent.

[0012] According to another aspect of this application, a command processing apparatus based on social intent recognition is provided, comprising: The first determination module is used to determine social intent based on application-added count values ​​and user session content. The second determining module is used to determine multimodal features and meta-learning weights based on the user session content if the social intent is a non-absolute rejection intent, and to determine an intent score based on the multimodal features and the meta-learning weights, wherein the meta-learning weights are determined based on the contextual features of the user session content. The execution module is used to execute a social application add event when the intent score is greater than a preset intent threshold, and to determine a target social execution instruction based on the add response and the updated application add count value, so as to perform the corresponding operation based on the target social execution instruction.

[0013] Further, the second determining module is specifically used to extract speech acoustic features, text semantic features, and temporal behavior features based on the user session content, and to extract context features from the user session content; to predict the context features based on the meta-learning prediction model that has completed model training, to obtain meta-learning weights, wherein the meta-learning prediction model includes a lightweight first hidden layer, a second hidden layer, an input layer, and an output layer; and to perform a weighted summation of the speech acoustic features, the text semantic features, the temporal behavior features, and the meta-learning weights to obtain an intent score.

[0014] Furthermore, the device also includes: If the social intent is to agree to add, then the social application add event is executed, and the add response from the social application is obtained to update the application add count value.

[0015] Furthermore, the execution module is specifically used to determine the target social execution instruction as a speech broadcast instruction if the add response is "add successfully"; and to determine the target social execution instruction as "send application add request" if the add response is "add unsuccessfully" and the application add count value matches the first preset number of times condition.

[0016] Furthermore, the application addition count value includes a first count value and a second count value. The execution module is specifically used to determine that the target social execution instruction is to send a non-application addition request if the addition response is a failure to add, and the first count value matches a second preset number of times condition, and the second count value matches a third preset number of times condition; wherein, the first count value is used to count the addition response of the social application, and the second count value is used to count the addition response of the non-social application.

[0017] Furthermore, the device also includes: The third determining module is used to determine that the social intent is an absolute rejection intent when the intent score is less than or equal to a preset intent threshold.

[0018] Further, the first determining module is specifically used to extract the number of rejection keywords in each statement in the user session content when the application addition count value is less than a preset number threshold; if the number of rejection keywords is less than a preset rejection threshold, the social intent is determined to be a non-absolute rejection intent; if the number of rejection keywords is greater than or equal to the preset rejection threshold, the social intent is determined to be an absolute rejection intent.

[0019] According to another aspect of this application, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform an operation corresponding to the above-described instruction processing method based on social intent recognition.

[0020] According to another aspect of this application, a computer device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described instruction processing method based on social intent recognition.

[0021] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a method, apparatus, storage medium, and device for instruction processing based on social intent recognition. Compared with the prior art, the embodiments of this application determine social intent based on application addition count and user session content. If the social intent is not an absolute rejection intent, multimodal features and meta-learning weights are determined based on the user session content, and an intent score is determined based on the multimodal features and the meta-learning weights. The meta-learning weights are determined based on the contextual features of the user session content. When the intent score is greater than a preset intent threshold, a social application addition event is executed, and a target social execution instruction is determined based on the addition response and the updated application addition count. The corresponding operation is executed based on the target social execution instruction. Through a multimodal feature fusion mechanism, multidimensional features of voice, text, and behavior are combined with meta-learning weights to achieve the accurate triggering of intent recognition-related instructions, significantly improving the accuracy of intent recognition, reducing the operating cost of calling the LLM model, avoiding unnecessary LLM model calls, and controlling the number of additions through counter management to prevent frequent triggering of addition instructions and reduce resource waste of social application addition instructions.

[0022] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an instruction processing method based on social intent recognition provided in an embodiment of this application is shown; Figure 2 This application illustrates an instruction processing system architecture diagram provided by an embodiment of the present application; Figure 3 This illustration shows a schematic diagram of the entire instruction processing flow provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of a social intent scoring process provided in an embodiment of this application; Figure 5 This illustration shows a block diagram of an instruction processing device based on social intent recognition, according to an embodiment of this application. Figure 6 A schematic diagram of the structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] It should be noted that the terms "first," "second," etc., 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 data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] The embodiments of this invention can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0027] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0028] Based on this, in one embodiment, the present invention provides a method for processing medical web pages. Taking the application of this method to computer devices such as servers as an example, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, such as intelligent medical systems and digital medical platforms.

[0029] This application provides an instruction processing method based on social intent recognition, such as... Figure 1 As shown, the method includes: 101. Determine social intent based on application-added counter values ​​and user session content.

[0030] In this embodiment, a social application refers to an application that provides social functions to users, such as WeChat or QQ. Through these social functions, different marketing personnel can add users as friends in the social application to push marketing services. Before this, marketing personnel or marketing robots can communicate by phone through the current execution terminal to obtain user conversation content. The user conversation content can include conversations between marketing personnel or marketing robots and different customers. It can be a pre-recorded and played conversation or a random conversation recording; this embodiment does not impose specific limitations. Additionally, the application add count value refers to the number of times the user responds to the response after the current execution terminal triggers an add request. For example, it can be the number of rejected add responses or the number of successful add responses, which can be counted using a counter; this embodiment does not impose specific limitations.

[0031] It should be noted that, in this embodiment, social intent refers to the user's intention to establish a social conversation with marketers or marketing bots. In different social applications, this can be understood as the intention to add friends. In determining social intent, a threshold can be set for the application's add count value, and the determination can be made in conjunction with the identified content of the user's conversation. This embodiment does not impose specific limitations.

[0032] For example, if the number of rejection responses is 2 and the keyword "really not needed" is identified in the user's conversation content, then the social intent is determined to be an absolute rejection intent.

[0033] 102. If the social intent is a non-absolute rejection intent, then multimodal features and meta-learning weights are determined based on the user's conversation content, and an intent score is determined based on the multimodal features and the meta-learning weights.

[0034] In this embodiment, the social intent can be an absolute rejection intent, meaning the user absolutely does not want to add friends, or a non-absolute rejection intent, meaning the user may want to add friends. When the social intent is a non-absolute rejection intent, it indicates a possible intention to add friends further. To avoid frequent triggering of invalid commands and algorithm calls, the current execution end first extracts multimodal features from the user's conversation content and determines the meta-learning weights. These multimodal features may include speech acoustic features, text semantic features, and temporal behavioral features. The meta-learning weights are determined based on the contextual features of the user's conversation content. Then, the intent score is determined based on the multimodal features and the meta-learning weights. In this embodiment, the intent score can be expressed as a score for the intention to add friends, such as by weighting the multimodal features and the meta-learning weights, without specific limitations.

[0035] 103. When the intent score is greater than the preset intent threshold, execute the social application add event, and determine the target social execution instruction based on the add response and the updated application add count value.

[0036] In this embodiment, the obtained intent score is compared with a preset intent threshold to determine whether the social application add event can be further executed. The social application add event can be a trigger event for adding friends from different social applications, including but not limited to social application add requests and SMS add requests; this embodiment does not impose specific limitations. Furthermore, after executing the social application add event, the received add response can be either "agree to add" or "disagree to add," updating the application add count. After updating the application add count, the target social execution instruction to be executed by the current execution terminal is determined, and the corresponding operation is performed based on the target social execution instruction.

[0037] It should be noted that the social execution commands set in the current execution terminal can include commands such as script broadcasting commands, sending application add requests, sending non-application add requests, and calling the Language Processing Model (LLM). These commands are determined based on different add responses and updated application add count values, so as to use social intent as the trigger condition for the commands, which greatly reduces the system resources of invalid commands and model calls.

[0038] In a scenario where a Language Processing Model (LLM) is invoked, the current turn can be obtained from `kwargs`, with a default value of 0. Executing Layer 1, i.e., user session detection, involves calling the `has_user_spoken` method to check if the user actually spoke. For example, it searches the last user message in the conversation history backwards, checking if the message is marked as silent (e.g., "Customer didn't speak," "ASR405Fuse"), and whether the message begins with "ASR" or is empty. If the user didn't speak, the result and verification flag are written to Redis and returned. Furthermore, after obtaining the intent score, the confidence level of the intent scoring using multimodal features is also obtained. If the confidence level is greater than 0.8 (high confidence), a social application add request instruction is directly generated without invoking the LLM model for intent recognition. If the confidence level is less than or equal to 0.8 (low confidence), a model invocation instruction is generated to execute subsequent steps for calling the LLM model's intent recognition method, such as... Figure 2 As shown, the embodiments of this application do not specifically limit the intent recognition based on the LLM model.

[0039] In a scenario where a user agrees to add a social application, the user conversation includes: a sales conversation: "I'm from Pacific Insurance. We can send you a quote for this BYD car under your name right away. I'm calling to add you on WeChat to send you the quote so you can compare," the user's response is "Okay, add me," the sales conversation is "Okay, I'll add you later, please agree," and the user's response is "Okay." After the current execution end detects that the user has spoken and does not detect any rejection keywords, it further performs multimodal feature fusion judgment. First, three-dimensional features are extracted: voice features include a fast speech rate (e.g., 5.2 words / second), few pauses (e.g., 1 pause), and a speaking percentage of 45%; text features include short but positive replies with 100% positive vocabulary; and behavioral features include a fast response time (e.g., 1.5 seconds), no interruptions, and 2 dialogue turns. Then, feature scores are calculated: voice_score = 0.75, text_score = 0.85, and behavior_score = 0.70. Meta-learning weight prediction predicts weights based on the dialogue context: weights=[0.25, 0.55, 0.20]. This indicates that short dialogue text has high weights. The intention score for a certain message is represented as: 0.25×0.75 + 0.55×0.85 + 0.20×0.70 = 0.79. Finally, the score of 0.79 is determined to be higher than the threshold of 0.6, and the model prediction confidence is 0.79, which is relatively high. Therefore, the social application add event is executed, indicating that it is determined that the social application can add friends, i.e., a friend add request is sent to the message application. At this time, LLM is not called for further intent recognition. Compared with calling LLM for intent judgment, it takes about 2-3 seconds, which greatly saves costs.

[0040] In this embodiment of the application, for the complete instruction processing process, such as Figure 3As shown, multiple layers of filtering can be configured, including a user session detection layer and a rejection intent detection layer. The user session detection layer defines a list of silence markers, including ASR error tags such as "Customer didn't speak," "ASR405 Fuse," and "ASR406 Overtime User NoSpeaking." It searches for the last user message in the user session from the end to the beginning, traversing the conversation history messages from the end to the beginning to find messages with the role 'user'. For each found message, the system performs the following checks in sequence, including but not limited to checking if the message content completely matches any item in the silence marker list, such as checking if the message content starts with "ASR," and checking if the message content is empty after removing spaces. Finally, it returns the result: if any of the above checks pass, it returns False if no user message is found, meaning the user did not speak; if the user message contains actual content, it returns True, meaning the user has spoken. The rejection intent detection layer defines a list of rejection keywords, such as "No need," "Forget it," "Let it go," "Don't need," etc. Examine the last 6 user messages in the user's conversation content. Iterate through the last 6 messages in the conversation history and filter out messages with the role 'user'. Match for rejection keywords. For each user message, check if its content contains any item in the rejection keyword list. If it does, increment the rejection count by 1 and break out of the inner loop. Each message only needs to be counted once.

[0041] Finally, determine whether to reject. If the rejection count is greater than or equal to 2, return True, meaning the user has rejected; otherwise, return False, meaning the user has not rejected.

[0042] In another embodiment of this application, for further understanding and explanation, the steps of determining multimodal features and meta-learning weights based on the user session content, and determining the intent score based on the multimodal features and the meta-learning weights, include: Based on the user session content, speech acoustic features, text semantic features, and temporal behavioral features are extracted respectively, and contextual features in the user session content are also extracted; The context features are predicted based on the meta-learning prediction model that has been trained, and the meta-learning weights are obtained. The intention score is obtained by weighted summation of the speech acoustic features, the text semantic features, the temporal behavioral features, and the meta-learning weights.

[0043] To assess social intent, the current execution end first extracts speech acoustic features, text semantic features, and temporal behavioral features based on the user's conversation content when determining the intent score. Specifically, speech acoustic features (voice_features) can be extracted from the ASR recognition results corresponding to the user's conversation content. These features may include speech rate features, number of pauses, volume variation, and speaking duration percentage. Speech rate refers to the number of words per second, calculated by dividing the number of characters in the speech text by the speech duration; a fast speech rate indicates urgency, while a slow speech rate indicates hesitation. The number of pauses is calculated from the ASR results; more pauses indicate hesitation. Volume variation is calculated by determining the standard deviation of the volume; large volume variations indicate emotional fluctuations. Speaking duration percentage is calculated by dividing the user's speaking time by the total conversation time; a high percentage indicates active participation. Text semantic features (text_features) can be extracted from the user's conversation content and may include response length, percentage of active words, number of interrogative sentences, and personalized expressions. Response length refers to the average number of words per response, reflecting the user's expressive engagement. The percentage of positive vocabulary refers to the frequency of positive words such as "good," "okay," and "great," reflecting the user's level of interest. The number of interrogative sentences refers to the frequency of interrogative markers such as "how much," "how," "how," and "how," reflecting the user's willingness to inquire. Personalized expression refers to the frequency of first-person pronouns such as "I think," "I consider," "I feel," and "I believe," reflecting the user's initiative. Temporal behavioral features (behavior_features) can be extracted from user conversation content and may include average response time, number of interruptions, number of dialogue rounds, and changes in engagement. Average response time refers to the average time interval between the salesperson finishing speaking and the user starting to speak, reflecting the user's responsiveness. Number of interruptions refers to the number of times the user interrupts the salesperson through overlapping speech detection, reflecting the user's engagement. Number of dialogue rounds refers to the number of rounds the user speaks, reflecting the depth of the conversation. Changes in engagement refer to the difference in the amount of speech between the first three rounds and the last three rounds, reflecting the dynamic changes in the user's attitude.

[0044] In this embodiment, when the current execution end extracts contextual features from the user's conversation content, the extracted features may include, but are not limited to, basic features, quality features, user profiles, and behavioral features. Basic features include dialogue turns (e.g., the number of times a user speaks), dialogue duration, and the proportion of user turns (which can be based on the number of user speeches divided by the total number of messages). Quality features include the mean ASR confidence score (reflecting speech recognition quality) and the standard deviation of text length (reflecting text diversity). User profiles (which can be obtained from a CRM) include user age groups (e.g., 20-30, 30-40, 40-50, 50+), and purchase history (e.g., high-value, medium-value, low-value, new customer). Behavioral features include average response time and the number of interruptions. The above contextual features can be further normalized to ensure the comparability of features with different dimensions.

[0045] It should be noted that after obtaining the context features, the current execution end predicts the context features based on the meta-learning prediction model that has completed model training, and obtains the meta-learning weights. In this process, the meta-learning prediction model adopts a lightweight neural network, including a lightweight first hidden layer, a second hidden layer, an input layer, and an output layer. In some embodiments, the first hidden layer contains 32 neurons and uses the ReLU activation function as the first hidden layer feature vector h1 output by the first hidden layer, calculated as h1 = ReLU(W1×context_normalized + b1), where W1 is a 32×input_dim weight matrix and b1 is a 1×32 bias vector. The second hidden layer contains 16 neurons and uses the ReLU activation function as the second hidden layer feature vector h2 output by the second hidden layer, calculated as h2 = ReLU(W2×h1 + b2), where W2 is a 16×32 weight matrix and b2 is a 1×16 bias vector. The output layer contains three neurons, corresponding to the logits of the three-dimensional features. The calculation formula is logits = W3 × h2 + b3, where W3 is a 3×16 weight matrix and b3 is a 1×3 bias vector. Furthermore, the logits of the output layer can be softmax normalized to obtain the weights of the three-dimensional features [w_voice, w_text, w_behavior]. Softmax normalization ensures that the sum of the weights is 1 (w_voice + w_text + w_behavior = 1), and that each weight value is between 0 and 1. This embodiment does not impose specific limitations on this.

[0046] Finally, the current execution end performs a weighted summation of the speech acoustic features, the text semantic features, the temporal behavioral features, and the meta-learning weights to obtain the intent score.

[0047] In some embodiments, when extracting contextual features, the current execution end can call the `extract_multimodal_features` method to extract three-dimensional features from the corresponding dialogue history and speech text in the user session content through ASR recognition: speech acoustic features, text semantic features, and temporal behavioral features. Furthermore, a feature scoring system for each dimension can be used to calculate a score for each feature, serving as a numerical representation of the contextual features. For example, as... Figure 4 As shown, the speech score is calculated as `voice_score = sigmoid(w_voice_base · voice_features + b_voice)`, mapping speech features to the range of 0-1. The text score is calculated as `text_score = sigmoid(w_text_base · text_features + b_text)`, mapping text features to the range of 0-1. The behavior score is calculated as `behavior_score = sigmoid(w_behavior_base · behavior_features + b_behavior)`, mapping behavior features to the range of 0-1. The current execution end calls the `meta_learning_weight_prediction` method through the meta-learning weight prediction system to dynamically predict the weight combination of the three-dimensional features [w_voice, w_text, w_behavior] based on the dialogue context features. Finally, a dynamic weighted summation system can be used to sum the scores of the three dimensions using the predicted weights to obtain the final addition intention score, expressed as: final_score = w_voice × voice_score + w_text × text_score + w_behavior × behavior_score. Here, a preset intent threshold of 0.6 is preferred; if final_score is greater than the threshold, it is determined that there is an intention to add the item.

[0048] In some embodiments, before determining the meta-learning weights, the meta-learning prediction model is pre-trained. Specifically, this includes: loading a training dataset, where each sample contains dialogue context features and three-dimensional features. The dialogue context features (context_features) include dialogue turns, duration, ASR confidence, user profile, etc., while the three-dimensional features (voice_features, text_features, behavior_features) include speech, text, and behavioral features—labels (where 0 indicates no social applications are added, and 1 indicates social applications are added). During model initialization, the neural network parameters are initialized. The first layer includes weights W1 and biases b1, where W1 is a [32, input_dim] matrix and b1 is a 1×32 vector. The second layer includes weights W2 and biases b2, where W2 is a 16×32 matrix and b2 is a 1×16 vector. The output layer includes weights W3 and biases b3, where W3 is a 31×16 matrix and b3 is a 1×3 vector. During multi-round training, the number of iterations can be selected from 100 to 500 (num_epochs = 100-500). Forward propagation involves inputting contextual features into the network for each batch (batch_size = 32), passing them through two ReLU activation layers and Softmax normalization to obtain the predicted weights. Simultaneously, the binary cross-entropy loss function is used to calculate the difference between the predicted probabilities and the true labels. During backpropagation, the gradient of the loss with respect to the parameters of each layer is calculated, including the gradients with respect to the softmax layer, hidden layers, and input layer. Additionally, the Adam optimizer (learning_rate = 0.001) can be used to update all weights and bias parameters. After training is complete, a validation set is evaluated every 10 epochs to assess model performance, outputting the validation set accuracy. All parameters after model training are saved as a "meta_learning_weight_predictor.pkl" file for online inference.

[0049] In some embodiments, after determining the meta-learning weights, the weights can be adaptively adjusted after each user session to improve the accuracy of intent recognition. For example, in a short dialogue scenario, the sales conversation is "I'm from Pacific Insurance, add me on WeChat to get a quote," and the user conversation is "Okay, add me." In this case, the dialogue rounds consist of one round of introduction, and the predicted meta-learning weights are weights=[0.20, 0.70, 0.10], indicating that the dialogue text weights are significantly increased. That is, in the short dialogue, the user's explicit expression of "Okay, add me" is the most reliable, hence the increased weight. As another example, in a long dialogue scenario, the sales conversation is "...(multiple rounds of introduction)...", the user conversation is "Hmm...(hesitating)...how about it?", the sales conversation is "...(continue introducing)...", the user conversation is "Oh, I see...", the sales conversation is "...(further introducing)...", and the user conversation is "Okay, then add me." The final number of dialogue rounds is 7. The engagement changes as the first 3 rounds are short replies, and the last 3 rounds are long replies. The meta-learning weight prediction is weights=[0.30, 0.40, 0.30], indicating an increase in behavioral weights. This means that in long conversations, changes in engagement better reflect genuine intent. For example, in a high-quality voice dialogue scenario, the sales conversation might be "...(detailed introduction)...", while the user conversation might be "Okay (fast speech, few pauses, stable volume)". In this case, the average confidence score obtained from ASR recognition is 0.95, which is relatively high. The speech rate is 5.5 words / second, indicating a fast pace and urgency. Therefore, the meta-learning weight prediction is weights=[0.45, 0.35, 0.20], indicating an increase in speech weights. This means that in high-quality voice dialogues, speech rate and pauses reflect the user's emotions.

[0050] In another embodiment of this application, for further illustrative purposes, the steps further include: If the social intent is to agree to add, then the social application add event is executed, and the add response from the social application is obtained to update the application add count value.

[0051] To ensure the effectiveness of intent recognition and trigger corresponding add events based on valid intents, thereby reducing the need for social application commands, the current execution terminal can directly trigger the social application add event when it determines that the social intent is to agree to add a friend, such as sending a friend request to the social application. This embodiment of the application does not impose specific limitations on this. Simultaneously, after executing the social application add event, the current execution terminal obtains feedback from the social application in real time, i.e., receives the add response from the social application. If the user chooses to add a friend, the add response is "Add approved"; if the user chooses not to add a friend, the add response is "Add rejected," thereby updating the application's add count.

[0052] In another embodiment of this application, for further understanding and explanation, the step of determining the target social execution instruction based on the added response and the updated application add count value includes: If the response to adding is "add," then the target social execution instruction is determined to be a script broadcast instruction. If the add response is "not added" and the application add count matches the first preset number of times condition, then the target social execution instruction is determined to be "send application add request".

[0053] To improve the effectiveness of execution instructions and reduce the waste of resources associated with adding friends in social applications, the current execution terminal pre-configures different processing instructions for different add responses when determining the target social execution instruction. In some embodiments, if the add response is "Added successfully," the target social execution instruction is determined to be a scripted announcement instruction to notify the user to complete adding a friend in the social application and to notify the current execution terminal to complete the purpose of the current session. For example, an announcement such as "I've added you on WeChat [redacted], and we will communicate with you on WeChat [redacted] later" can complete or end the current session.

[0054] In addition, since the application add count can represent the number of times the social application has failed to add friends or the number of times it has succeeded, if the add response is "not added" and the application add count matches the first preset number condition, it means that the current user has not directly added friends in the social application, and the number of rejections has not reached the threshold or the number of successes is 0. Therefore, the current execution terminal determines that the target social execution instruction is to send an application add request to request the user to add friends in the social application again.

[0055] In another embodiment of this application, for further understanding and explanation, the step of determining the target social execution instruction based on the added response and the updated application add count value further includes: If the add response is "not added", and the first count value matches the second preset count condition, and the second count value matches the third preset count condition, then the target social execution instruction is determined to be sending a non-application add request.

[0056] To meet the intent recognition requirements in different addition scenarios and avoid invalid triggering of intent commands, the application addition count value in the current execution terminal includes a first count value and a second count value. The first count value is used to count the addition response from the social application, and the second count value is used to count the addition response from non-social applications. The second preset count condition can be a count condition set for the success or failure of adding a friend after the current execution terminal directly sends a request to the social application, such as a threshold for the number of unsuccessful attempts or a threshold for the number of successful attempts. The third preset count condition can be a count condition set for the success or failure of adding a friend after the current execution terminal indirectly sends a request to the social application (adding a friend through a non-social application), such as adding a friend via SMS. This embodiment of the application does not specifically limit the specific method used.

[0057] In a specific scenario where a user agrees but fails to add a certain WeChat account, the user session content includes: The sales conversation is: "I'm from xxx car insurance. We can send you a quote for this BYD car under your name right away. I'm calling to add you on WeChat." The user conversation is: "Okay, send it to me." The sales conversation is: "Okay, I'll add you later." The current execution end enters the multimodal feature fusion judgment step, which extracts features separately: speech features are moderate speech rate (4.5 words / second) and speaking proportion (40%); text features are short replies and positive vocabulary proportion (50%), such as "okay" as a positive vocabulary; behavioral features are fast response (2 seconds) and 1 dialogue turn. At this point, through meta-learning weight prediction, weights=[0.30, 0.50, 0.20] are obtained. The willingness to add WeChat is 0.65, which is greater than the preset threshold of 0.6, meaning WeChat needs to be added. Therefore, the add command is issued, which can be marked as assignAddWechat=1. At this point, the client reports an application failure, with the count set to 1. This can be marked as `assign_can_add_wechat=2`, indicating one failure count. The system checks all application add counts, including the first and second counts. If count 1 matches the second preset count condition (one rejection is met), count 2 matches the third preset count condition (if the SMS sending limit has not been reached). Therefore, the system can switch to the SMS process, allowing the addition of an application friend's information via SMS.

[0058] In some embodiments, a counter can be used to update the count value. In this case, the count value for any round can be stored in Redis for retrieval during counting. Specifically, the counter value system can retrieve the counter values ​​of the first and second count values ​​and their corresponding threshold conditions from Redis. This application embodiment does not provide specific limitations.

[0059] In another embodiment of this application, for further illustrative purposes, the steps further include: When the intent score is less than or equal to a preset intent threshold, the social intent is determined to be an absolute rejection intent.

[0060] To ensure the accuracy of intent recognition, the current execution end compares the obtained intent score with a preset intent threshold. This preset intent threshold can be set based on the social addition request, and this embodiment does not impose specific limitations. When the intent score is less than or equal to the preset intent threshold, it indicates that the user does not want to add friends based on the user's conversation content. Therefore, the social intent is determined to be an absolute rejection intent, so that other intent analysis models are not invoked or other instructions are triggered subsequently, thereby improving instruction processing efficiency and reducing system resource consumption.

[0061] In another embodiment of this application, for further understanding and explanation, the step of determining social intent based on application-added counter values ​​and user session content includes: If the application adds a count value less than a preset threshold, then the number of rejection keywords in each statement of the user's conversation content is extracted. If the number of rejection keywords is less than a preset rejection threshold, then the social intent is determined to be a non-absolute rejection intent. If the number of rejection keywords is greater than or equal to a preset rejection threshold, then the social intent is determined to be an absolute rejection intent.

[0062] To improve the accuracy of social intent recognition and reduce system resource usage by large models, when determining social intent, specifically, when comparing the application add count with a preset threshold, if the application add count is less than the preset threshold, it means that the number of times the friend has been added has not reached the direct rejection limit. Therefore, the rejection keyword count of each statement in the user's conversation content can be further extracted. Rejection keywords can be "never mind", "no need", "don't need", etc., and this application embodiment does not make specific limitations.

[0063] In some embodiments, if the number of rejection keywords is less than a preset rejection threshold, it indicates that there is still an opportunity to add social friends; therefore, the social intention is determined to be a non-absolute rejection intention. If the number of rejection keywords is greater than or equal to the preset rejection threshold, it indicates that there is no opportunity to add social friends; therefore, the social intention is determined to be an absolute rejection intention. The preset rejection threshold can be set based on the needs of intention recognition, and this embodiment does not impose specific limitations.

[0064] It should be noted that when an absolute rejection intent is determined, the current execution end can directly refuse to issue any instructions in order to reduce the system resources required to trigger instructions in social applications.

[0065] In a user rejection scenario, the user conversation includes: the salesperson says, "I'm from xxxx car insurance, here's a quote for you," the user replies, "I don't need it," the salesperson says, "I understand your meaning perfectly, let me give you a discounted quote, listen to it first, comparing prices is always a good thing," and the user replies, "Never mind, I don't need it." At this point, the user has spoken, and rejection keywords, including "I don't need it" and "Never mind, I don't need it," are extracted and counted. `refusal_count=2`. The preset rejection threshold is 2; when `refusal_count>=2`, it's considered a rejection. The current execution end directly returns a result without adding social applications. Therefore, the LLM model can be avoided, multimodal feature fusion recognition can be omitted, and all instructions to add social applications can be blocked.

[0066] This application provides a method for instruction processing based on social intent recognition. This method determines social intent based on an application add count and user session content. If the social intent is not an absolute rejection intent, multimodal features and meta-learning weights are determined based on the user session content, and an intent score is determined based on the multimodal features and meta-learning weights. The meta-learning weights are determined based on the contextual features of the user session content. When the intent score is greater than a preset intent threshold, a social application add event is executed, and a target social execution instruction is determined based on the add response and the updated application add count. The corresponding operation is then executed based on the target social execution instruction. Through a multimodal feature fusion mechanism, multidimensional features of voice, text, and behavior are combined with meta-learning weights to achieve accurate triggering of intent-related instructions, significantly improving the accuracy of intent recognition, reducing the operational cost of calling the LLM model, avoiding unnecessary LLM calls, and controlling the number of add events through counter management to prevent frequent triggering of add instructions and reduce resource waste in social application add instructions.

[0067] Furthermore, as a response to the above Figure 1 To implement the method shown, this application provides an instruction processing device based on social intent recognition, such as... Figure 5 As shown, the device includes: The first determining module 21 is used to determine social intent based on the application-added count value and user session content; The second determining module 22 is used to determine multimodal features and meta-learning weights based on the user session content if the social intent is a non-absolute rejection intent, and to determine an intent score based on the multimodal features and the meta-learning weights, wherein the meta-learning weights are determined based on the contextual features of the user session content. The execution module 23 is used to execute a social application add event when the intent score is greater than a preset intent threshold, and to determine a target social execution instruction based on the add response and the updated application add count value, so as to perform the corresponding operation based on the target social execution instruction.

[0068] Further, the second determining module is specifically used to extract speech acoustic features, text semantic features, and temporal behavior features based on the user session content, and to extract context features from the user session content; to predict the context features based on the meta-learning prediction model that has completed model training, to obtain meta-learning weights, wherein the meta-learning prediction model includes a lightweight first hidden layer, a second hidden layer, an input layer, and an output layer; and to perform a weighted summation of the speech acoustic features, the text semantic features, the temporal behavior features, and the meta-learning weights to obtain an intent score.

[0069] Furthermore, the device also includes: If the social intent is to agree to add, then the social application add event is executed, and the add response from the social application is obtained to update the application add count value.

[0070] Furthermore, the execution module is specifically used to determine the target social execution instruction as a speech broadcast instruction if the add response is "add successfully"; and to determine the target social execution instruction as "send application add request" if the add response is "add unsuccessfully" and the application add count value matches the first preset number of times condition.

[0071] Furthermore, the application addition count value includes a first count value and a second count value. The execution module is specifically used to determine that the target social execution instruction is to send a non-application addition request if the addition response is a failure to add, and the first count value matches a second preset number of times condition, and the second count value matches a third preset number of times condition; wherein, the first count value is used to count the addition response of the social application, and the second count value is used to count the addition response of the non-social application.

[0072] Furthermore, the device also includes: The third determining module is used to determine that the social intent is an absolute rejection intent when the intent score is less than or equal to a preset intent threshold.

[0073] Further, the first determining module is specifically used to extract the number of rejection keywords in each statement in the user session content when the application addition count value is less than a preset number threshold; if the number of rejection keywords is less than a preset rejection threshold, the social intent is determined to be a non-absolute rejection intent; if the number of rejection keywords is greater than or equal to the preset rejection threshold, the social intent is determined to be an absolute rejection intent.

[0074] This application provides an instruction processing device based on social intent recognition. This embodiment determines social intent based on an application addition count and user session content. If the social intent is not an absolute rejection intent, multimodal features and meta-learning weights are determined based on the user session content, and an intent score is determined based on the multimodal features and meta-learning weights. The meta-learning weights are determined based on the contextual features of the user session content. When the intent score is greater than a preset intent threshold, a social application addition event is executed, and a target social execution instruction is determined based on the addition response and the updated application addition count. The corresponding operation is then executed based on the target social execution instruction. Through a multimodal feature fusion mechanism, multidimensional features of voice, text, and behavior are combined with meta-learning weights to achieve accurate triggering of intent recognition-related instructions, significantly improving the accuracy of intent recognition, reducing the operational cost of calling the LLM model, avoiding unnecessary LLM calls, and controlling the number of additions through counter management to prevent frequent triggering of addition instructions and reduce resource waste in social application addition instructions.

[0075] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the instruction processing method based on social intent recognition in any of the above method embodiments.

[0076] Figure 6 The diagram shows a structural schematic of a computer device according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the computer device.

[0077] like Figure 6 As shown, the computer device may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0078] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0079] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0080] The processor 302 is used to execute program 310, specifically to execute the relevant steps in the above-described embodiment of the instruction processing method based on social intent recognition.

[0081] Specifically, program 310 may include program code that includes computer operation instructions.

[0082] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The computer device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0083] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0084] Specifically, program 310 can be used to cause processor 302 to perform the following operations: Determine social intent based on application add counts and user conversation content; If the social intent is a non-absolute rejection intent, then multimodal features and meta-learning weights are determined based on the user session content, and an intent score is determined based on the multimodal features and the meta-learning weights, wherein the meta-learning weights are determined based on the contextual features of the user session content. When the intent score is greater than the preset intent threshold, a social application add event is executed, and a target social execution instruction is determined based on the add response and the updated application add count value, so as to execute the corresponding operation based on the target social execution instruction.

[0085] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0086] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for processing instructions based on social intent recognition, characterized in that, include: Determine social intent based on application add counts and user conversation content; If the social intent is a non-absolute rejection intent, then multimodal features and meta-learning weights are determined based on the user session content, and an intent score is determined based on the multimodal features and the meta-learning weights, wherein the meta-learning weights are determined based on the contextual features of the user session content. When the intent score is greater than the preset intent threshold, a social application add event is executed, and a target social execution instruction is determined based on the add response and the updated application add count value, so as to execute the corresponding operation based on the target social execution instruction.

2. The method according to claim 1, characterized in that, The step of determining multimodal features and meta-learning weights based on the user session content, and determining the intent score based on the multimodal features and the meta-learning weights, includes: Based on the user session content, speech acoustic features, text semantic features, and temporal behavioral features are extracted respectively, and contextual features in the user session content are also extracted; The context features are predicted based on the meta-learning prediction model that has been trained, and meta-learning weights are obtained. The meta-learning prediction model includes a lightweight first hidden layer, a second hidden layer, an input layer, and an output layer. The intention score is obtained by weighted summation of the speech acoustic features, the text semantic features, the temporal behavioral features, and the meta-learning weights.

3. The method according to claim 1, characterized in that, The method further includes: If the social intent is to agree to add, then the social application add event is executed, and the add response from the social application is obtained to update the application add count value.

4. The method according to claim 3, characterized in that, The determination of the target social execution instruction based on the added response and the updated application add count value includes: If the response to adding is "add," then the target social execution instruction is determined to be a script broadcast instruction. If the add response is "not added" and the application add count matches the first preset number of times condition, then the target social execution instruction is determined to be "send application add request".

5. The method according to claim 4, characterized in that, The application addition count value includes a first count value and a second count value. The step of determining the target social execution instruction based on the addition response and the updated application addition count value further includes: If the add response is "not added", and the first count value matches the second preset number of times condition, and the second count value matches the third preset number of times condition, then it is determined that the target social execution instruction is to send a non-application add request; The first count value is used to count the add response of the social application, and the second count value is used to count the add response of the non-social application.

6. The method according to claim 1, characterized in that, The method further includes: When the intent score is less than or equal to a preset intent threshold, the social intent is determined to be an absolute rejection intent.

7. The method according to any one of claims 1-6, characterized in that, The method of determining social intent based on application-added counter values ​​and user session content includes: If the application adds a count value less than a preset threshold, then the number of rejection keywords in each statement of the user's conversation content is extracted. If the number of rejection keywords is less than a preset rejection threshold, then the social intent is determined to be a non-absolute rejection intent. If the number of rejection keywords is greater than or equal to a preset rejection threshold, then the social intent is determined to be an absolute rejection intent.

8. A command processing device based on social intent recognition, characterized in that, include: The first determination module is used to determine social intent based on application-added count values ​​and user session content. The second determining module is used to determine multimodal features and meta-learning weights based on the user session content if the social intent is a non-absolute rejection intent, and to determine an intent score based on the multimodal features and the meta-learning weights, wherein the meta-learning weights are determined based on the contextual features of the user session content. The execution module is used to execute a social application add event when the intent score is greater than a preset intent threshold, and to determine a target social execution instruction based on the add response and the updated application add count value, so as to perform the corresponding operation based on the target social execution instruction.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images