AI Intelligent Customer Service Response Method and System Based on Remote Digital Services

Abstract

The present application provides an AI intelligent customer service response method and system based on remote digital services. First, it receives the session request data stream input by the user through the remote session terminal, and extracts the dynamic semantic feature sequence containing syntactic structure, sentiment tendency and context-dependent features. Then it inputs the sequence into the intent recognition model trained based on the multi-dimensional semantic relationships of historical session data, and outputs the intent category label and confidence level. Next, it matches the target dialogue flow node from the preset dialogue flow library according to the label and confidence level. Then it generates dynamic response content based on the node response logic rules and returns the response data stream. Finally, it collects the user feedback behavior data, updates the associated weights of the intent recognition model according to the interaction effectiveness index, and optimizes the semantic mapping accuracy of the model for the same type of session requests, so as to achieve more accurate, intelligent and adaptive customer service responses.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology. Specifically, it relates to an AI intelligent customer service response method and system based on remote digital services. Background Art

[0002] In the current era of the rapid development of digital services, the demand for remote customer service is increasing day by day, and AI intelligent customer service emerges and is widely used. Traditional AI intelligent customer service response methods have many limitations. Early customer service systems mostly adopted keyword matching methods. This method can only simply identify user problems based on preset keywords, and its ability to understand complex natural language expressions and semantics is extremely limited. When the expression of the user's question changes slightly or contains fuzzy semantics, it is difficult to accurately identify the user's intention, resulting in inaccurate or even incorrect responses.

[0003] Subsequently, some rule-based customer service systems developed, although they improved the processing ability of structured problems to a certain extent, the formulation of rules is often fixed and limited, and it is impossible to flexibly adapt to diverse user needs and changing language environments. For expressions with emotional colors or problems with strong context dependence from users, these systems are often helpless and cannot provide personalized services.

[0004] Some intelligent customer service methods based on machine learning that emerged later, although they can learn from a large amount of data, mostly only focus on single-dimensional semantic features, ignoring the grammatical structure, emotional tendency, and the association between contexts. This makes the understanding of the user input text in actual applications not comprehensive and in-depth enough, and the accuracy and reliability of intention recognition are relatively low. Moreover, these methods lack an effective feedback mechanism and cannot dynamically adjust and optimize their own performance according to the actual feedback of users, making it difficult to continuously improve service quality and user experience. Summary of the Invention

[0005] In view of the above-mentioned problems, in combination with the first aspect of this application, the embodiments of this application provide an AI intelligent customer service response method based on remote digital services. The method includes:

[0006] Receiving the session request data stream input by the user through the remote session terminal, and extracting the dynamic semantic feature sequence in the session request data stream. The dynamic semantic feature sequence includes the grammatical structure feature, emotional tendency feature, and context dependence feature of the user input text;

[0007] Inputting the dynamic semantic feature sequence into a pre-trained intention recognition model, and outputting the user intention category label and the corresponding intention confidence. The intention recognition model is trained and generated based on the multi-dimensional semantic relationships in the historical session data and is used to map the association weights between semantic features and preset intention categories;

[0008] Match a target dialogue flow node from a preset dialogue flow library according to the intention category label and intention confidence level, where the target dialogue flow node includes a response logic rule corresponding to the intention category label and a multi-round dialogue jump path;

[0009] Generate dynamic response content based on the response logic rule in the target dialogue flow node, and return a response data stream including the dynamic response content to the user through the remote session terminal;

[0010] Collect the feedback behavior data of the user on the response data stream, and update the association weight corresponding to the intention category label in the intention recognition model according to the interaction effectiveness index in the feedback behavior data, and optimize the semantic mapping accuracy of the intention recognition model for the same type of session requests.

[0011] On the other hand, an embodiment of the present application further provides a remote digital service system, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0012] Based on the above aspects, the embodiment of the present application can capture the deep semantic information of the user input text more comprehensively and accurately by receiving the user session request data stream and extracting the dynamic semantic feature sequence including the syntactic structure, sentiment tendency and context-dependent features. Compared with the traditional method that only focuses on the surface semantics, the ability to understand the true intention of the user is greatly improved. Inputting the dynamic semantic feature sequence into the intention recognition model trained based on the multi-dimensional semantic relationship of historical session data can accurately output the user intention category label and the corresponding intention confidence level. This model trained based on the multi-dimensional semantic relationship effectively avoids the intention misjudgment caused by single-dimensional training and enhances the accuracy and reliability of intention recognition. Matching the target dialogue flow node from the preset dialogue flow library according to the intention category label and intention confidence level, the target dialogue flow node includes the response logic rule and the multi-round dialogue jump path, making the customer service response more logical and coherent, and can flexibly handle different questions and multi-round dialogue scenarios of the user, providing a better service experience. Generating dynamic response content based on the response logic rule in the target dialogue flow node and returning it to the user realizes personalized and intelligent response, meeting the diverse needs of users. Collecting the feedback behavior data of the user on the response data stream, updating the association weight corresponding to the intention category label in the intention recognition model according to the interaction effectiveness index, and continuously optimizing the semantic mapping accuracy of the model for the same type of session requests, enabling the AI intelligent customer service to continuously learn and evolve, adapt to the language habits and demand changes of different users, and continuously improve the service quality and efficiency. Brief Description of the Drawings

[0013] Figure 1 It is a schematic execution flowchart of the AI intelligent customer service response method based on remote digital services provided by an embodiment of the present application.

[0014] Figure 2 It is a schematic hardware architecture diagram of the remote digital service system provided by an embodiment of the present application. Detailed Embodiments

[0015] The present application will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flowchart of the AI intelligent customer service response method based on remote digital services provided by an embodiment of the present application. The AI intelligent customer service response method based on remote digital services will be introduced in detail below.

[0016] Step S110: Receive the session request data stream input by the user through the remote session terminal, and extract the dynamic semantic feature sequence in the session request data stream. The dynamic semantic feature sequence includes the syntactic structure feature, emotional tendency feature, and context dependence feature of the user input text.

[0017] In this embodiment, in the e-commerce service scenario, the user sends a session request data stream through the customer service chat interface (i.e., the remote session terminal) of the e-commerce platform. For example, the user inputs "I bought a smart phone from you before, but now the battery life has decreased significantly. I want to know what solutions you can provide." For this input session request data stream, the system first extracts the syntactic structure feature. Syntactically, this is a complex sentence containing a causal relationship. The main clause is "I want to know what solutions you can provide", and the subordinate clause "I bought a smart phone from you before, but now the battery life has decreased significantly" is used to describe the event background. In terms of the emotional tendency feature, since the user mentions that the product has problems and seeks solutions, it is initially judged that there is a certain negative emotional tendency, but not extreme dissatisfaction. More importantly, it is an attitude of seeking help. The context dependence feature is reflected in the relevance to the previous purchase behavior, that is, the purchase of a smart phone and the current battery life problem. The system comprehensively extracts these dynamic semantic feature sequences through technical means such as lexical and syntactic analysis and semantic role annotation of the entire sentence, so as to prepare for subsequent intent recognition.

[0018] Step S120: Input the dynamic semantic feature sequence into the pre-trained intent recognition model, and output the user intent category label and the corresponding intent confidence. The intent recognition model is trained based on the multi-dimensional semantic relationships in the historical session data and is used to map the association weights between semantic features and preset intent categories.

[0019] In this embodiment, this intent recognition model is constructed based on a large amount of historical conversation data on the e-commerce platform. For example, the historical conversation data contains various types of conversation records such as numerous users' feedback on product problems, after-sales service consultations, and product usage questions. During the model training process, by performing semantic slicing on these historical conversation data, for semantic fragment units similar to "seeking solutions for product problems", the model learned the multi-dimensional semantic relationships therein. For the semantic features of the current user input regarding the battery life problem of the smartphone and seeking solutions, the model undergoes complex internal neural network calculations. The model may output the user intent category label as "After-sales Service - Product Fault Solution", and at the same time output the corresponding intent confidence, such as 0.8. This means that the model is 80% certain that the user's intent is to seek a product fault solution in after-sales service. This confidence is based on the calculation result of the association weight between the semantic features such as the syntactic structure, sentiment tendency, and context dependence of the input and the preset "After-sales Service - Product Fault Solution" intent category. If the semantic features of the input highly match the "Product Fault Solution" semantic pattern in the historical data, the confidence will be higher; otherwise, it will be lower.

[0020] Step S130, according to the intent category label and intent confidence, match the target dialogue flow node from the preset dialogue flow library, and the target dialogue flow node includes the response logic rule corresponding to the intent category label and the multi-round dialogue jump path.

[0021] In this embodiment, according to the "After-sales Service - Product Fault Solution" intent category label obtained in the previous step and the intent confidence of 0.8, the target conversation flow node is matched from the preset e-commerce service conversation flow library. The conversation flow library contains numerous conversation flow nodes for different intent categories. For example, for the conversation flow node of "After-sales Service - Product Fault Solution", there may be multiple response logic branches, such as the fault solutions for different product types (smartphones, computers, etc.), the processing flows for different fault reasons (hardware problems, software problems, etc.). First, the system filters out the candidate conversation flow node set related to "After-sales Service - Product Fault Solution" according to the intent category label. Then, priority scores are given to these candidate nodes. Suppose one of the nodes is a after-sales service process specifically for smartphone battery problems, and its historical response success rate is relatively high because many users have obtained satisfactory solutions through this process for similar battery life problems before, and the user satisfaction index is also very high. And the degree of consistency with the current conversation context (purchased a smartphone and there is a problem with battery life) is very strong. The priority score of this node will be very high. At the same time, considering the intent confidence of 0.8, after weighted summation calculation, this node may be selected as the target conversation flow node. If this target conversation flow node contains a multi-round conversation jump path, for example, after initially asking the user about their battery usage habits, if the user's answer indicates that it may be a software problem causing the battery life to decline, then the trigger condition threshold of the subsequent node (such as the node guiding the user to perform a software update) will be dynamically adjusted to make it easier to be triggered in order to further solve the problem deeply. If in the subsequent conversation, the user presents a new semantic feature, such as "I think this battery problem may have occurred after I used a certain accessory recommended by you", and this new feature does not match the current target conversation flow node, at this time, the conversation flow backtracking mechanism will be triggered, and a new target conversation flow node will be re-matched according to the new intent category label and confidence level.

[0022] Step S140, generate dynamic response content based on the response logic rules in the target conversation flow node, and return a response data stream containing the dynamic response content to the user through the remote session terminal.

[0023] For example, for the selected target dialogue flow node, the response logic rules it contains specify how to generate the response content. For example, the response logic rules of this node specify a structured template, where the fixed text field may be "Dear user, regarding the smartphone battery life issue you reported, we have the following solutions:", and the variable parameter slots may include information such as the battery model, purchase time, and whether it is within the warranty period. Extract the filling data from the context-dependent features of the current session. For example, obtain from the user's previous purchase record that the battery model is "Model XX", the purchase time is "XX / XX / XX", and it is found that it is still within the warranty period. Inject this filling data into the structured template to generate the initial response text "Dear user, regarding the smartphone battery life issue you reported, we have the following solutions: The smartphone of Model XX you purchased was bought on XX / XX / XX and is still within the warranty period. We suggest that you send the phone back to our after-sales service center for free inspection and repair." Then, optimize the fluency of this initial response text through a natural language generation model, such as adjusting the sentence structure and adding some connecting words, to generate a more smooth and natural dynamic response content. After that, perform multi-modal conversion processing on the dynamic response content, such as converting the text content into voice information, and at the same time display some visual elements on the e-commerce customer service interface, such as the address map of the after-sales service center and pictures of the shipping process, to generate a composite response data stream containing text, voice, and visual elements, and return it to the user through the remote session terminal (e-commerce customer service chat interface). Before sending this response data stream, simulate the potential feedback paths of the user to this dynamic response content through a sandbox environment. For example, simulate that the user may further ask for the contact information of the after-sales service center, or ask how long the repair will take, etc. According to the simulation results, if it is found that the possibility of the user asking for the contact information of the after-sales service center is relatively high, adjust the parameter slot priority in the response logic rules to bring the provision of the contact information of the after-sales service center to a more important position.

[0024] Step S150, collect the feedback behavior data of the user on the response data stream, and update the association weight corresponding to the intent category label in the intent recognition model according to the interaction effectiveness index in the feedback behavior data, and optimize the semantic mapping accuracy of the intent recognition model for the same type of session requests.

[0025] In this embodiment, after the user receives the response data stream, feedback behavior data will be generated. For example, the user quickly replies "Okay, I will send the phone back according to your process, thank you." Extract the interaction effectiveness index from this feedback behavior data. The shorter the user response delay duration indicates that the user is more satisfied with the reply; the intent of the subsequent session request (here it is to operate according to the reply) Figure 1The consistency is relatively high, and no new intentions appear; moreover, there are no corrective labels for manual review and annotation. Calculate the intention determination error value of the intention recognition model in the current session based on these interactive effectiveness indicators. Since the user feedback is good, this error value is low. Use the incremental learning algorithm to feedback this low error value to the parameter adjustment layer of the intention recognition model, and dynamically update the association weight corresponding to the intention category label of "After-sales service - Product fault solution". For example, in the model, the semantic mapping relationship between "Battery life problem" and "After-sales service - Product fault solution" may be strengthened. If in multiple similar after-sales service scenarios, the cumulative error value of a certain intention category label (such as "After-sales service - Product fault solution") exceeds the preset error value, this may indicate that the model's performance in this intention category is not accurate enough. At this time, trigger the local retraining process of the intention recognition model, and use the newly added session data on product fault after-sales service to re-optimize the mapping relationship of the multi-dimensional semantic feature matrix. For example, it may re-analyze the relationship between new semantic features such as new fault types and new user feedback methods and the intention category. Deploy the updated intention recognition model to the shadow system. By comparing the intention recognition accuracy rates of the new and old intention recognition models on the test dataset (including more session samples of product fault after-sales service), if the accuracy rate of the new model is higher, determine to switch the updated intention recognition model to the corresponding e-commerce business production environment, so as to improve the semantic mapping accuracy of the model when processing similar session requests and better provide after-sales service answers for users.

[0026] Based on the above steps, in the embodiment of the present application, by receiving the user session request data stream and extracting the dynamic semantic feature sequence including syntactic structure, sentiment tendency and context-dependent features, the deep semantic information of the user input text can be captured more comprehensively and accurately. Compared with the traditional method that only focuses on the surface semantics, the ability to understand the true intention of the user is greatly improved. Inputting the dynamic semantic feature sequence into the intention recognition model trained based on the multi-dimensional semantic relationship of historical session data can accurately output the user intention category label and the corresponding intention confidence. This model trained based on the multi-dimensional semantic relationship effectively avoids the intention misjudgment caused by single-dimensional training and enhances the accuracy and reliability of intention recognition. Matching the target dialogue flow node from the preset dialogue flow library according to the intention category label and intention confidence, the target dialogue flow node includes response logic rules and multi-turn dialogue jump paths, making the customer service response more logical and coherent, being able to flexibly handle different questions and multi-turn dialogue scenarios of users, and providing a better service experience. Generating dynamic response content based on the response logic rules in the target dialogue flow node and returning it to the user realizes personalized and intelligent response, meeting the diverse needs of users. Collecting the feedback behavior data of the user on the response data stream, updating the association weight corresponding to the intention category label in the intention recognition model according to the interaction effectiveness index, and continuously optimizing the semantic mapping accuracy of the model for the same type of session requests, enabling the AI intelligent customer service to continuously learn and evolve, adapt to the language habits and demand changes of different users, and continuously improve the service quality and efficiency.

[0027] In a possible implementation manner, the training steps of the intention recognition model specifically include:

[0028] Step S210, obtaining a historical session data set, performing semantic slicing processing on each session record in the historical session data set to generate a plurality of semantic segment units, and each semantic segment unit includes the expression content of at least one complete semantic intention.

[0029] In this embodiment, the e-commerce platform has accumulated a vast amount of historical conversation data, which covers various user interaction scenarios. For example, conversations about product inquiries, such as "I want to know the processor performance of your laptop. Can it run large games?"; conversations related to after-sales services, like "There is a loose thread on the clothes I bought. What's your return and exchange policy?"; and inquiries about promotional activities, such as "Do you have any promotional activities for mobile phones recently?" and so on. Each conversation record in these historical conversation data sets is subjected to semantic slicing processing to decompose the complex conversation record into multiple semantic fragment units. Taking the conversation record "I bought a pair of your headphones before. The sound quality is very good, but now the sound in the right ear has become smaller. I want to know what's going on. Can you repair or replace it for me" as an example, it can be cut into semantic fragment units such as "I bought a pair of your headphones before. The sound quality is very good", "Now the sound in the right ear has become smaller", "I want to know what's going on", "Can you repair or replace it for me", etc. Each semantic fragment unit contains the expression content of at least one complete semantic intention. For example, the semantic fragment unit "the sound in the right ear has become smaller" expresses the intention of a problem with the product, and "Can you repair or replace it for me" expresses the intention of seeking after-sales service.

[0030] Step S220: Perform feature enhancement processing on each semantic fragment unit, extract the context association vector, syntactic dependency tree structure, and sentiment polarity value of the semantic fragment unit, and generate a multi-dimensional semantic feature matrix.

[0031] For example, for the semantic fragment unit "The sound in the right ear has become quieter now", it is input into the bidirectional recurrent neural network layer, and based on the hidden state output of the bidirectional recurrent neural network layer, the forward semantic propagation vector and the backward semantic propagation vector of this semantic fragment unit in the historical conversation data are captured. For instance, the forward semantic propagation vector may contain semantic information related to the previously mentioned headphone usage, and the backward semantic propagation vector may be related to the subsequent search for solutions. The forward semantic propagation vector and the backward semantic propagation vector are concatenated to generate the initial context association vector of this semantic fragment unit. Then, the initial context association vector is processed with multi-head self-attention weighting to calculate the long-distance dependence weights between the lexical units within the semantic fragment unit, generating an enhanced context association vector with an attention mask. For example, in this semantic fragment, there is a semantic association between "right ear" and "the sound has become quieter", and this relationship can be more accurately reflected through this weighting process. Based on the lexical dependence weights in the enhanced context association vector, the hierarchical relationship of the subject-predicate-object components in this semantic fragment unit is parsed to generate a syntactic dependency tree structure with the core predicate verb as the root node. In "The sound in the right ear has become quieter", "sound" is the subject and "has become quieter" is the predicate. By analyzing and obtaining the syntactic dependency tree structure, the relationship between the components can be clearly presented. Each leaf node in the syntactic dependency tree structure is traversed to extract the adjectival phrases and adverbial phrases that have a direct modifying relationship with the core predicate verb. After generating the set of syntactic modifying components, each modifying component in the set of syntactic modifying components is pattern-matched with a predefined sentiment dictionary to identify the part-of-speech tag and sentiment intensity coefficient of the sentiment polarity keyword in the modifying component. In this example, if there is a modifier such as "obviously", its sentiment intensity coefficient is determined through matching with the sentiment dictionary. According to the distribution density of the sentiment intensity coefficients in the set of syntactic modifying components, the sentiment polarity value of this semantic fragment unit is calculated. Here, since it is just a description of the problem, the sentiment polarity value may be neutral. The enhanced context association vector is unfolded into a two-dimensional tensor in the order of lexical units, and the dependency relationship type encoding of the corresponding lexical unit in the syntactic dependency tree structure is embedded at the row and column index positions of the two-dimensional tensor. The normalized scalar of the sentiment polarity value is concatenated at each lexical unit position in the two-dimensional tensor to generate a three-dimensional semantic feature tensor containing the context association vector, the syntactic dependency tree structure, and the sentiment polarity value. Channel dimension compression and spatial dimension pooling operations are performed on the three-dimensional semantic feature tensor to eliminate the redundant noise data in the three-dimensional semantic feature tensor, generating a reduced-dimensional dense semantic feature matrix. Then, the dense semantic feature matrix is time-step aligned and concatenated with the feature matrices of adjacent semantic fragment units in the historical conversation data, and a position encoding vector is injected to mark the temporal position of this semantic fragment unit in the complete conversation. Based on the temporal position, a sliding window normalization process is performed on the concatenated feature matrix to balance the contribution degree weights of semantic fragment units at different time steps to the multi-dimensional semantic feature matrix.According to the variance distribution of each channel dimension in the normalized feature matrix, dynamically allocate the feature fusion coefficients of the context correlation vector, the syntactic dependency tree structure, and the sentiment polarity value, and finally generate a multi-dimensional semantic feature matrix.

[0032] Step S230: Input the multi-dimensional semantic feature matrix into the initial neural network model, and adjust the model parameters of the initial neural network model through the backpropagation algorithm to minimize the loss function value between the predicted intent category output by the initial neural network model and the true intent label manually annotated.

[0033] For example, the initial neural network model is a multi-layer neural network structure, including an input layer, multiple hidden layers, and an output layer. During the training process, the model will calculate based on the input multi-dimensional semantic feature matrix and output the predicted intent category. Taking the previous semantic segment of the headphone problem as an example, the model may predict its intent category as "After-sales service - Product fault feedback". Then compare this predicted intent category with the true intent label manually annotated (here the true intent label manually annotated is also "After-sales service - Product fault feedback"), and adjust the model parameters of the initial neural network model through the backpropagation algorithm to minimize the loss function value between the predicted intent category output by the model and the true intent label manually annotated. This process is repeated continuously, training a large number of semantic segment units, gradually optimizing the model parameters, and improving the model's learning ability for the relationship between different semantic features and intent categories.

[0034] Step S240: Introduce an attention weight allocation mechanism in the output layer of the initial neural network model to dynamically adjust the contribution degree of different semantic segment units to the final intent category determination, and generate the intent recognition model.

[0035] Continuing with the semantic segment of the headphone problem as an example, when the model determines the final intent category, different semantic segment units may have different importance. For example, the semantic segment unit "The sound in the right ear has become smaller" may be more critical for determining the intent category of "After-sales service - Product fault feedback", while the semantic segment unit "The sound quality of the previous headphones was very good" is relatively less important. Through the attention weight allocation mechanism, dynamically adjust the contribution degree of different semantic segment units to the final intent category determination. In this way, when determining the intent category, the model can more reasonably utilize the information of each semantic segment unit, thereby improving the accuracy of intent recognition and generating the final intent recognition model.

[0036] Step S250: Inject noisy semantic segment units into the intent recognition model through an adversarial training method to enhance the robustness of the intent recognition model to fuzzy semantics and ambiguous expressions.

[0037] For example, in the e-commerce scenario, there are many cases of fuzzy semantics and ambiguous expressions. For example, a user may say "There is something wrong with the thing I bought", where "the thing" is a relatively fuzzy expression. By injecting noise semantic fragment units containing such fuzzy semantics or ambiguous expressions, the intent recognition model can learn how to handle such situations during training. If the intent recognition model can correctly identify the intent category when facing these noise semantic fragment units, then the robustness of the intent recognition model to fuzzy semantics and ambiguous expressions is improved. For example, when the model processes a semantic fragment unit containing the fuzzy expression "the thing", through the previously learned semantic relationships and features, it can still accurately determine that the user is referring to a certain product purchased before and is reporting a product problem, thus improving the ability to handle various complex semantic situations in the actual e-commerce service scenario.

[0038] In a possible implementation manner, step S220 includes:

[0039] Step S221, inputting the semantic fragment unit into a bidirectional recurrent neural network layer, and outputting a forward semantic propagation vector and a backward semantic propagation vector that capture the semantic fragment unit in the historical session data based on the hidden state of the bidirectional recurrent neural network layer.

[0040] In this embodiment, consider a semantic fragment unit in an e-commerce session: "The camera function of this mobile phone disappoints me. The photos are always blurry." First, input this semantic fragment unit into the bidirectional recurrent neural network layer. The bidirectional recurrent neural network layer processes the words in this semantic fragment unit in sequence and outputs a forward semantic propagation vector and a backward semantic propagation vector that capture the semantic fragment unit in the historical session data based on its hidden state. The forward semantic propagation vector reflects the semantic information accumulated from the beginning of the sentence to the current word. In this example, for the part "The camera function of this mobile phone", the forward semantic propagation vector contains descriptive information about a specific mobile phone and its camera function, and this information is a semantic continuation related to the previous possible discussion about the mobile phone or product introduction. The backward semantic propagation vector constructs semantic relationships from the end of the sentence to the beginning. For the part "The photos are always blurry", the backward semantic propagation vector contains semantic information related to the photo quality problem, and this information may be related to the subsequent search for solutions or expression of dissatisfaction.

[0041] Step S222, performing a concatenation operation on the forward semantic propagation vector and the backward semantic propagation vector to generate an initial context association vector of the semantic fragment unit, and performing a multi-head self-attention weighting process on the initial context association vector to calculate the long-distance dependence weights between the lexical units within the semantic fragment unit, and generating an enhanced context association vector with an attention mask.

[0042] This initial context correlation vector synthesizes semantic information captured from two directions. Then, multi-head self-attention weighting is performed on this initial context correlation vector to calculate the long-distance dependence weights between each lexical unit within the semantic segment unit, generating an enhanced context correlation vector with an attention mask. In this semantic segment, there is a semantic association between the "photo-taking function" and "blurry". Through multi-head self-attention weighting, this long-distance dependence weight can be accurately calculated. For example, the lexical unit "photo-taking function" may have a relatively high dependence weight with the lexical unit "blurry", indicating a close relationship between the two in semantic understanding. The calculation result of this dependence weight is reflected in the enhanced context correlation vector, and the relationship between important lexical units is highlighted through the attention mask.

[0043] Step S223: Based on the lexical dependence weights in the enhanced context correlation vector, analyze the hierarchical relationship of the subject, verb, and object components in the semantic segment unit, and generate a syntactic dependency tree structure with the core predicate verb as the root node.

[0044] For example, in the sentence "The photo-taking function of this mobile phone disappoints me very much, and the photos are always blurry", "photo-taking function" is the subject, "disappoints" is the core predicate verb, "me" is the indirect object, "very much" is the direct object, "photos" is another subject, and "blurry" is the predicate. Through such analysis, a syntactic dependency tree structure is constructed, clearly presenting the syntactic relationships between the components in the sentence. This syntactic dependency tree structure helps to deeply understand the semantic structure of the sentence and provides a basis for subsequent sentiment analysis and semantic feature extraction.

[0045] Step S224: Traverse each leaf node in the syntactic dependency tree structure, extract the adjectival phrases and adverbial phrases that have a direct modification relationship with the core predicate verb, generate a set of syntactic modification components, and then match each modification component in the set of syntactic modification components with a predefined sentiment dictionary to identify the part-of-speech tag and sentiment intensity coefficient of the sentiment polarity keyword in the modification component.

[0046] For example, in this case, the adjectival phrase "very disappointed" has a direct modification relationship with the core predicate verb "disappoints". Match each modification component in the set of syntactic modification components with a predefined sentiment dictionary to identify the part-of-speech tag and sentiment intensity coefficient of the sentiment polarity keyword in the modification component. In the phrase "very disappointed", "disappointed" is the sentiment polarity keyword, with the part of speech being an adjective, and the sentiment intensity coefficient can be determined as a relatively high negative intensity, such as -0.8 (assuming -1 represents extremely negative, 0 represents neutral, and 1 represents extremely positive) according to the predefined sentiment dictionary.

[0047] Step S225: Calculate the sentiment polarity value of the semantic fragment unit according to the distribution density of the sentiment intensity coefficient in the set of grammatical modification components. The sentiment polarity value is a quantitative score for positive sentiment, negative sentiment, or neutral sentiment.

[0048] In this example, since there is only one sentiment polarity keyword and its intensity is high, the sentiment polarity value of this semantic fragment unit is negative sentiment, and the quantitative score is -0.8, indicating that the user has an obvious negative attitude towards the camera function of this mobile phone.

[0049] Step S226: Unfold the enhanced context association vector into a two-dimensional tensor in the order of lexical units, and embed the dependency relationship type encoding of the corresponding lexical unit in the grammatical dependency tree structure at the row and column index positions of the two-dimensional tensor.

[0050] For example, if the dependency relationship between "camera function" and "makes" is a subject-predicate relationship, embed the encoding representing the subject-predicate relationship at the corresponding row and column index positions. Concatenate the normalized scalar of the sentiment polarity value at each lexical unit position of the two-dimensional tensor to generate a three-dimensional semantic feature tensor containing the context association vector, the grammatical dependency tree structure, and the sentiment polarity value. In this example, in addition to the original context association vector information and grammatical dependency relationship encoding at each lexical unit position, the normalized information of the sentiment polarity value is added, such as the value after normalizing -0.8.

[0051] Step S227: Concatenate the normalized scalar of the sentiment polarity value at each lexical unit position of the two-dimensional tensor to generate a three-dimensional semantic feature tensor containing the context association vector, the grammatical dependency tree structure, and the sentiment polarity value. Perform channel dimension compression and spatial dimension pooling operations on the three-dimensional semantic feature tensor to eliminate redundant noise data in the three-dimensional semantic feature tensor and generate a reduced-dimensional dense semantic feature matrix.

[0052] In this process, the data in the three-dimensional semantic feature tensor is processed by a specific algorithm to remove some information that contributes less or is repetitive to semantic expression, and a more compact and effective dense semantic feature matrix is obtained.

[0053] Step S228: Align and concatenate the dense semantic feature matrix with the feature matrices of adjacent semantic fragment units in the historical conversation data in time steps, inject position encoding vectors to mark the temporal positions of the semantic fragment units in the complete conversation, and perform sliding window normalization processing on the concatenated feature matrix based on the temporal positions to balance the contribution degree weights of semantic fragment units at different time steps to the multi-dimensional semantic feature matrix.

[0054] For example, if there was a semantic fragment unit previously about the appearance design of this mobile phone, align and splice the feature matrices of these two semantic fragment units in the order of their appearance in the conversation. Inject a position encoding vector during the splicing process to mark the temporal position of the semantic fragment unit in the complete conversation, and perform a sliding window normalization process on the spliced feature matrix to balance the contribution weight of the semantic fragment units at different time steps to the multi-dimensional semantic feature matrix. In this example, if the semantic fragment unit about the camera function is relatively late in the conversation, through the sliding window normalization process, adjust its weight in the entire multi-dimensional semantic feature matrix according to its position to ensure that its contribution to the overall semantic understanding is balanced with the previous semantic fragment unit about the appearance design.

[0055] Step S229: Dynamically allocate the feature fusion coefficients of the context association vector, the syntactic dependency tree structure, and the sentiment polarity value according to the variance distribution of each channel dimension in the normalized feature matrix, and generate the multi-dimensional semantic feature matrix.

[0056] For example, if in the normalized feature matrix, the variance of the channel dimension related to the sentiment polarity value is large, it indicates that the discrimination degree of the sentiment polarity in this semantic fragment unit is high, then a higher feature fusion coefficient will be allocated to the sentiment polarity value; if the variance of the channel dimension related to the syntactic dependency tree structure is small, it indicates that the importance of the syntactic structure in this semantic fragment unit is relatively low, and a lower feature fusion coefficient will be allocated to the syntactic dependency tree structure. Through such dynamic allocation, the generated multi-dimensional semantic feature matrix can more effectively reflect the comprehensive semantic features of the semantic fragment unit and provide more accurate input data for the subsequent intent recognition model.

[0057] Consider another semantic fragment unit in an e-commerce conversation: "I'm very interested in the new promotion you launched. It seems very cost-effective." Similarly, input this semantic fragment unit into the bidirectional recurrent neural network layer to obtain the forward and backward semantic propagation vectors. The forward semantic propagation vector contains information related to the user's previous browsing or consumption behavior on the e-commerce platform, while the backward semantic propagation vector is related to possible subsequent operations (such as participating in the promotion). Concatenate these two vectors to obtain the initial context correlation vector, and then perform multi-head self-attention weighting to obtain the enhanced context correlation vector. In this semantic fragment, there is a relatively high dependence weight between "on the promotion" and "interested". When constructing the syntactic dependency tree structure, "I" is the subject, "interested" is the predicate, and "promotion" is the object. Extract the modifier "very" related to the core predicate verb "interested". By matching with the sentiment dictionary, "very" is a sentiment intensity modifier, the sentiment polarity keyword "interested" has a positive sentiment, and the sentiment intensity coefficient is 0.6. The sentiment polarity value of the entire semantic fragment unit is positive sentiment, and the quantitative score is 0.6. Convert it into a three-dimensional semantic feature tensor containing the context correlation vector, syntactic dependency tree structure, and sentiment polarity value according to the above process. After compression and pooling operations, obtain the dense semantic feature matrix. Concatenate it with the feature matrix of the adjacent semantic fragment unit, inject the position encoding vector, and perform sliding window normalization. Finally, dynamically allocate the feature fusion coefficient according to the variance distribution to generate the multi-dimensional semantic feature matrix. In this way, whether it is a semantic fragment unit with positive or negative sentiment, it can accurately extract semantic features through such a process and integrate them into the multi-dimensional semantic feature matrix, providing comprehensive and accurate semantic information input for the intent recognition model.

[0058] In the e-commerce service scenario, there are various semantic fragment units, such as user evaluations of products, feedback on after-sales services, and inquiries about logistics. By performing such detailed feature enhancement processing on each semantic fragment unit, semantic information can be comprehensively and deeply mined and integrated into the multi-dimensional semantic feature matrix, thereby improving the intent recognition model's understanding and processing capabilities for different semantic situations and better supporting the identification and response to user needs in e-commerce services.

[0059] In one possible implementation manner, step S130 specifically includes:

[0060] Step S131, according to the intent category label, screen a set of candidate dialogue flow nodes from the dialogue flow library, and each node in the set of candidate dialogue flow nodes includes at least one response logic branch associated with the intent category label.

[0061] In this embodiment, it is assumed that the intent category label obtained after the intent recognition of the user input session request is "Product After-sales - Return and Exchange Consultation", and the intent confidence is 0.75. First, according to this intent category label, a set of candidate dialogue flow nodes is screened from the dialogue flow library. The dialogue flow library contains numerous dialogue flow nodes for different e-commerce service scenarios, and these nodes construct various possible dialogue flows. For the intent category label of "Product After-sales - Return and Exchange Consultation", the nodes in the set of candidate dialogue flow nodes are all related to returns and exchanges. One of the nodes may be the answer to the return and exchange policy for non-damaged products, and the response logic branches included are different return and exchange time limits, return and exchange methods (such as by mail, in-store, etc.); another node may be the return and exchange process for damaged products, and the response logic branches involve damage appraisal processes, liability division, etc. Each node has at least one response logic branch associated with the intent category label of "Product After-sales - Return and Exchange Consultation".

[0062] Step S132, perform a priority score on each node in the set of candidate dialogue flow nodes. The basis for the priority score includes the response success rate of the node in the historical conversation, the user satisfaction index, and the degree of consistency with the current conversation context.

[0063] Taking the node of answering the return and exchange policy for non-damaged products as an example, check its response success rate in the historical conversation. If in a large number of similar conversations in the past, this node can successfully answer the user's questions about the return and exchange policy for non-damaged products, for example, the success rate reaches 80%, this provides a relatively high basic score for its priority score. Then look at the user satisfaction index. If the user shows a high degree of satisfaction through subsequent feedback (such as evaluation, asking again, etc.) after receiving the response of this node, for example, the satisfaction score is 0.8 (assuming a score between 0 - 1, with 1 being very satisfied), this will also increase its priority score. At the same time, consider the degree of consistency with the current conversation context. If the user mentions in the conversation "The clothes I bought have no problems, but the size is not suitable. I want to know if I can exchange them", this node is highly consistent with the current conversation context because it is an answer to the return and exchange policy for non-damaged products, and this high degree of consistency will further increase its priority score.

[0064] For another node of the return and exchange process for damaged products, assume its historical response success rate is 70% and the user satisfaction index is 0.7. Since the user clearly states that the product is not damaged, the degree of consistency with this node's context is relatively low, so the priority score will be lower than the previous node. In this way, a comprehensive priority score is performed on each node in the set of candidate dialogue flow nodes.

[0065] Step S133: Select the node with the highest score from the set of candidate dialogue flow nodes as the target dialogue flow node according to the weighted sum result of the intention confidence and the priority score.

[0066] Suppose the priority score for the node answering the return and exchange policy for non-damaged goods is 0.8 (after comprehensive calculation), the intention confidence is 0.75, and the weighted sum result is 0.75 * 0.8 = 0.6; the priority score for the node of the return and exchange process for damaged goods is 0.6, and the weighted sum result is 0.75 * 0.6 = 0.45. Since 0.6 is greater than 0.45, the node answering the return and exchange policy for non-damaged goods is selected as the target dialogue flow node.

[0067] Step S134: If the target dialogue flow node contains a multi-turn dialogue jump path, dynamically adjust the trigger condition threshold of the subsequent nodes in the jump path according to the context dependence characteristics of the current session.

[0068] If this target dialogue flow node contains a multi-turn dialogue jump path, for example, after answering the time limit for return and exchange in the return and exchange policy, the next possible jump node may be about the issue of who bears the return and exchange fees. At this time, dynamically adjust the trigger condition threshold of the subsequent nodes in the jump path according to the context dependence characteristics of the current session. Suppose in the user's conversation, it is mentioned that "I am your premium member. Is there a special return and exchange policy?" This newly added semantic feature indicates that the user's membership status may affect the return and exchange policy. Then, the trigger condition threshold of the nodes related to the special return and exchange policy for premium members can be lowered to make it easier to be triggered, so as to provide timely return and exchange information related to the user's identity in the subsequent conversation.

[0069] Step S135: When it is detected that the newly added semantic feature in the user session request does not match the current target dialogue flow node, trigger the dialogue flow backtracking mechanism, and re-execute the step of matching the target dialogue flow node from the preset dialogue flow library according to the intention category label and intention confidence to match a new target dialogue flow node.

[0070] For example, during the process of answering the return and exchange policy for undamaged goods, the user suddenly says, "Actually, I bought this product during your promotional event, and there are some special coupon usage situations. Will this affect the return and exchange?" This newly added semantic feature does not match the current node for answering the return and exchange policy of undamaged goods because the previous nodes did not consider the impact of coupon usage during the promotional event on return and exchange. At this time, the dialogue flow backtracking mechanism is triggered, and based on the new intention category label (which may become "After-sales of Goods - Consultation on Return and Exchange during Promotional Events") and intention confidence level (recalculated or adjusted based on previous results), a new set of candidate dialogue flow nodes is filtered from the dialogue flow library. It may include nodes specifically for the relationship between return and exchange during promotional events and coupon usage. Then, following the above steps of priority scoring, weighted summation, etc., a new target dialogue flow node is rematched to ensure that the dialogue can accurately respond to the user's needs.

[0071] Another example is that the user's initial intention category label is recognized as "Product Consultation - Understanding Product Functions", and the intention confidence level is 0.8. The set of candidate dialogue flow nodes filtered from the dialogue flow library may include nodes for introducing the functions of different types of products, such as nodes for introducing the functions of electronic products and nodes for introducing the functions of clothing products. For the node for introducing the functions of electronic products, the response success rate in the historical conversation is 75%, and the user satisfaction index is 0.7. If the user asks about the functions of a mobile phone and the specific model of the mobile phone is mentioned in the current conversation, the consistency of this node with the current conversation context is relatively high, and its priority score is calculated comprehensively. For the node for introducing the functions of clothing products, the corresponding response success rate and user satisfaction index are calculated based on its historical data, and the priority score is determined considering the consistency with the current conversation context. According to the weighted summation result of the intention confidence level and the priority score, the node with the highest score is selected as the target dialogue flow node. If this target dialogue flow node contains multi-round dialogue jump paths, for example, after introducing the basic functions of the mobile phone, the next jump node may be about the advanced functions or accessory functions of the mobile phone. Based on other semantic features in the user's conversation, such as mentioning the concern about the camera function, the trigger condition threshold for the node related to the camera function of the mobile phone can be adjusted to make it easier to be triggered. If during the conversation the user says, "I want to know if this mobile phone can be used abroad", this newly added semantic feature does not match the current target dialogue flow node, triggering the dialogue flow backtracking mechanism to rematch the target dialogue flow node to adapt to the new user needs.

[0072] In the e-commerce service scenario, this method of matching target dialogue flow nodes from a preset dialogue flow library based on intent category tags and intent confidence can effectively select appropriate dialogue flow nodes according to the user's intent and the context of the conversation, and can dynamically adjust according to the progress of the conversation to ensure the coherence and accuracy of the conversation, improving the user experience and service efficiency.

[0073] In a possible implementation manner, step S140 specifically includes:

[0074] Step S141, parsing the response logic rules in the target dialogue flow node to determine the structured template of the response content, where the structured template includes fixed text fields and variable parameter slots.

[0075] In this embodiment, after determining the target dialogue flow node, for example, in the previously mentioned "Product After-sales - Return and Exchange Consultation" scenario, the target dialogue flow node is the node for answering the return and exchange policy for undamaged products.

[0076] The response logic rules of this target dialogue flow node include a structured template. The fixed text field may be "Dear customer, the return and exchange policy for undamaged products you inquired about is as follows:", and this part of the text is relatively fixed and is used to give the user a clear response guidance at the beginning of the response. The variable parameter slots are the content related to the specific return and exchange policy, such as the time limit for return and exchange, the methods of return and exchange (by mail, in-store, etc.).

[0077] Step S142, extracting the filling data corresponding to the variable parameter slots from the context-dependent features of the current session, where the filling data includes the user's historical behavior preferences, real-time session environment parameters, and external database query results.

[0078] In terms of the user's historical behavior preferences, if it is found from the user's historical purchase records that the user often chooses to receive products by mail, this may affect the recommendation of the return and exchange method. Real-time session environment parameters, such as the current logistics distribution situation, if the logistics distribution is relatively tense, may affect the time limit for return and exchange. External database query results, querying the return and exchange policy database of the e-commerce platform, it is obtained that the specific time limit for return and exchange of undamaged products is within 7 days from the purchase date, and the return and exchange methods support both mail and in-store.

[0079] Step S143, injecting the filling data into the structured template to generate an initial response text, and optimizing the fluency of the initial response text through a natural language generation model to generate the dynamic response content.

[0080] Then, inject these filled data into the structured template to generate an initial response text, such as "Dear customer, the return and exchange policy for undamaged goods you inquired about is as follows: You can return or exchange the goods within 7 days from the date of purchase. The return and exchange methods support both mailing and in-store. According to your historical purchase habits, mailing may be a more familiar way for you." Then optimize the fluency of this initial response text through a natural language generation model. This model may adjust the sentence structure. For example, it may optimize "According to your historical purchase habits, mailing may be a more familiar way for you." to "Your historical purchase habits show that you may be more familiar with the mailing method." to generate a more smooth and natural dynamic response content.

[0081] Step S144, perform multi-modal conversion processing on the dynamic response content to generate a composite response data stream including text, voice, and visual elements.

[0082] For example, perform multi-modal conversion processing on this dynamic response content. Convert the text content into voice information, and the voice can be broadcast in a kind and professional voice style. At the same time, add visual elements to the e-commerce customer service interface, such as showing an animation demonstration of the return and exchange process, or marking the locations of the stores where returns and exchanges can be made in-store on a map, etc., to generate a composite response data stream including text, voice, and visual elements.

[0083] Step S145, before sending the response data stream, simulate the potential feedback paths of users to the dynamic response content through a sandbox environment, and adjust the parameter slot priorities in the response logic rules according to the simulation results.

[0084] For example, simulate that users may further ask who will bear the mailing cost, or whether an appointment is required for in-store returns and exchanges. If the simulation results show that the possibility of users asking about the mailing cost is relatively high, adjust the parameter slot priorities in the response logic rules, and move the description of the mailing cost to a more important position in advance, so as to be able to respond to users' needs more timely and accurately in subsequent possible conversations.

[0085] In a possible implementation manner, step S150 specifically includes:

[0086] Step S151, extract interaction effectiveness indicators from the feedback behavior data, and the interaction effectiveness indicators include the user response delay duration, the Figure 1 consistency of subsequent session requests, and the correction labels manually reviewed and marked.

[0087] For example, in terms of the user response delay duration, if the user replies within 1 minute after receiving the reply, this is a relatively short response delay duration, which usually indicates that the user is more concerned about the reply and may be relatively satisfied. The intention of subsequent session requestsFigure 1 In terms of consistency, if the user replies "Okay, then I choose to return or exchange the goods by mail. How do I operate specifically?", this subsequent conversation request is consistent with the previous intention of returning or exchanging goods consultation, and no new intention appears. Regarding the correction labels marked by manual review, assuming that no content that needs to be corrected is found by manual review, that is, there are no correction labels.

[0088] Step S152, calculate the intention determination error value of the intention recognition model in the current conversation according to the interaction effectiveness index, and the intention determination error value is positively correlated with the deviation degree between the actual needs of the user and the intention output by the model.

[0089] Due to the short length of the user response delay, the subsequent conversation requests are Figure 1 consistent and there are no correction labels, this error value is relatively low. This intention determination error value is positively correlated with the deviation degree between the actual needs of the user and the intention output by the model. In this case, the intention of "Product after-sales - Return / Exchange Consultation" output by the model highly matches the actual needs of the user, so the error value is close to 0.

[0090] Step S153, use the incremental learning algorithm to feedback the error value to the parameter adjustment layer of the intention recognition model, and dynamically update the associated weight corresponding to the intention category label.

[0091] For example, use the incremental learning algorithm to feedback this error value to the parameter adjustment layer of the intention recognition model, and dynamically update the associated weight corresponding to the intention category label of "Product after-sales - Return / Exchange Consultation". For example, in the model, the associated weight between semantic features related to the return and exchange policy (such as "undamaged goods", "within 7 days", "mailing and in-store methods", etc.) and the intention of "Product after-sales - Return / Exchange Consultation" may be fine-tuned to further strengthen the connection between them.

[0092] Step S154, when it is detected that the cumulative error value of the same intention category label exceeds the preset error value, trigger the local re-training process of the intention recognition model, and use the new session data to re-optimize the mapping relationship of the multi-dimensional semantic feature matrix.

[0093] Suppose in multiple conversations of return and exchange consultations, in one case, due to a misunderstanding of the return and exchange policy for special goods (such as customized goods), the intention determination error value is relatively large. When the cumulative error values related to "Product after-sales - Return / Exchange Consultation" exceed the preset error value, the local re-training process will be triggered. Use the new session data to re-optimize the mapping relationship of the multi-dimensional semantic feature matrix. The new session data may include more conversation records about the return and exchange of special goods, re-analyze the relationship between the semantic features and intentions in these conversations, and adjust the feature fusion coefficients in the multi-dimensional semantic feature matrix, etc.

[0094] Step S155: Deploy the updated intent recognition model to the shadow system. By comparing the intent recognition accuracies of the new and old intent recognition models on the test data set, determine whether to switch the updated intent recognition model to the corresponding business production environment.

[0095] For example, in the shadow system, compare the intent recognition accuracies of the new and old intent recognition models on the test data set. The test data set contains a large number of e-commerce service session samples, including various types of sessions such as product consultations and after-sales services. If the intent recognition accuracy of the new model on the test data set is higher than that of the old model, for example, the accuracy of the old model is 80% and the accuracy of the new model is 85%, it is determined to switch the updated intent recognition model to the corresponding business production environment, so as to improve the semantic mapping accuracy of the model when processing similar session requests and better support the identification of user needs in e-commerce services.

[0096] In the e-commerce service scenario, this process of generating dynamic response content based on target dialogue flow nodes and updating the intent recognition model according to feedback behavior data can continuously optimize the accuracy of the service and the user experience, and improve the effectiveness and adaptability of the e-commerce service system when processing user sessions.

[0097] In a possible implementation manner, the method further includes:

[0098] Step S310: Real-time monitor the abnormal semantic patterns in the user session request, where the abnormal semantic patterns include high-frequency repeated queries, semantic logic conflicts, and sensitive keyword combinations.

[0099] In this embodiment, during the session interaction process of the e-commerce service, the abnormal semantic patterns in the user session request can be real-time monitored. For example, in terms of high-frequency repeated queries, if the user asks "What is the warranty policy for the mobile phone I bought" multiple times and repeats it more than 3 times within a short period (such as within 5 minutes), this constitutes an abnormal semantic pattern of high-frequency repeated queries. In terms of semantic logic conflicts, the user says "I want to buy this product, but I don't want to spend money and I don't want to use coupons", and this expression of buying a product without wanting to spend money or use coupons has a semantic logic conflict. In terms of sensitive keyword combinations, if sensitive keyword combinations such as "fraud" and "illegal" that are contrary to the normal business of the e-commerce service and may involve risks appear in the user's session, such as "Is your promotion activity a fraud".

[0100] Step S320: When the abnormal semantic pattern is recognized, trigger a risk handling plan, where the risk handling plan includes switching to manual customer service intervention, injecting a verification question-and-answer process, and restricting the session response frequency.

[0101] For example, for high-frequency repeated queries, if it is determined that their severity level is relatively low (because it may just be that the user did not understand or notice the previous response), the disposal strategy may be to inject a verification Q&A process. For relatively serious anomalies such as semantic logic conflicts (which may imply malicious or unreasonable requests from the user), it may trigger a switch to manual customer service intervention. For cases with a relatively high severity level that contain sensitive keyword combinations, in addition to switching to manual customer service intervention, it may also limit the session response frequency. For example, the user may be restricted to sending only one message within 10 minutes to prevent the possible spread of malicious behavior or risks.

[0102] Step S330: Record the processing process of the abnormal semantic pattern to generate a risk disposal case library, which is used to optimize the recognition boundary of the abnormal semantics by the intent recognition model and the risk response nodes in the dialogue flow library.

[0103] Among them, step S320 specifically includes:

[0104] Step S321: Select a matching combination of disposal actions from a preset set of disposal strategies according to the type and severity level of the abnormal semantic pattern.

[0105] Taking high-frequency repeated queries as an example, according to their type (query repetition) and severity level (low), select a matching combination of disposal actions from a preset set of disposal strategies. In this case, the disposal strategy set may stipulate that for high-frequency repeated queries with a low severity level, the disposal action of injecting a verification Q&A process is adopted. This verification Q&A process will dynamically generate verification questions based on information related to the user's historical behavior. Suppose the user has previously purchased a mobile phone and asked about mobile phone accessories. Then the dynamically generated verification question may be "You previously asked us about mobile phone accessories. Do you remember which accessory it was?" Adjust the session trust score according to the accuracy of the user's answer. If the user answers correctly, increase the session trust score. For example, increase it from the initial 0.5 to 0.6, indicating that the user may be a normal but somewhat confused user. If the user answers incorrectly, decrease the session trust score, for example, decrease it to 0.4.

[0106] Step S322: Before performing the switch to manual customer service, re-confirm the intent of the current session through the intent recognition model. If the intent category label after re-confirmation is inconsistent with the initial category label, re-match the target dialogue flow node.

[0107] Before performing the disposal action of switching to a human customer service, for example, in the case of semantic logic conflicts, the intent recognition model is used to re-confirm the intent of the current conversation. Suppose the initial intent category label of the user is recognized as "Product Purchase - Special Purchase Requirements" (here, the special purchase requirements refer to unreasonable requirements such as not wanting to spend money or use coupons to purchase products). By analyzing the conversation content again through the intent recognition model, it may be found that the user's real intent is to want to know if there are other preferential methods or if there are free gifts, etc. The intent category label after re-confirmation becomes "Product Purchase - Consultation on Preferential Methods". Since the intent category label after re-confirmation is inconsistent with the initial category label, the target dialogue flow node is re-matched. The re-matched target dialogue flow node may be a node for answering product preferential methods, and this node contains the response logic rules for various possible preferential methods (such as full reduction, free gifts, etc.).

[0108] Step S323, when injecting the verification Q&A process, dynamically generate verification questions related to the user's historical behavior, and adjust the session trust score according to the accuracy of the user's answer.

[0109] Step S324, when the session trust score is lower than the set trust score, enable an enhanced review scheme for subsequent session requests. The enhanced review scheme includes real-time semantic feature desensitization processing and manual review delay of the response content.

[0110] When injecting the verification Q&A process, after generating verification questions based on the user's historical behavior as mentioned above, adjust the session trust score according to the accuracy of the user's answer. Taking another example, if the user has purchased clothes historically and asked about the material of the clothes, the dynamically generated verification question is "What is the material of the clothes you purchased before?" If the user answers correctly, the session trust score is increased from 0.5 to 0.6. When the session trust score is lower than the set trust score (assuming the set trust score is 0.5), enable an enhanced review scheme for subsequent session requests. For example, if the user's session trust score drops to 0.4, for subsequent user session requests such as "I want to see other products", enable real-time semantic feature desensitization processing, that is, process some semantic features that may involve privacy or sensitivity (such as some privacy-related information in the user's browsing history) to avoid overusing this information during the processing. At the same time, perform a manual review delay on the response content, that is, do not immediately send the response content automatically generated by the customer service to the user, but wait for manual review to ensure the accuracy and security of the response content.

[0111] Step S325, after the execution of the risk disposal protocol is completed, return a disposal result notification to the user, and collect the user's satisfaction feedback on the disposal process to optimize the priority configuration of the disposal strategy set.

[0112] After the execution of the risk disposal agreement is completed, a disposal result notification is returned to the user. For example, if it is an injection verification Q&A process and the user passes the verification, the user is sent a notification: "Dear user, you have passed our verification process, and we will continue to answer your questions related to the product." If it is to switch to manual customer service intervention, the user is notified: "Dear user, since your problem is relatively complex, we have transferred you to the manual customer service. Please wait patiently." And the satisfaction feedback of the user on the disposal process is collected to optimize the priority configuration of the disposal strategy set. If the user replies "I am very satisfied with this verification process and think it is reasonable", then in the disposal strategy set, the priority of the injection verification Q&A process for similar situations may be increased; if the user expresses dissatisfaction, such as "I think it takes too long for you to transfer me to the manual customer service", then the relevant disposal strategies for manual customer service transfer will be optimized, such as adjusting the transfer process or resource allocation, to improve user satisfaction. At the same time, record the processing process of abnormal semantic patterns and generate a risk disposal case library. In this risk disposal case library, it contains the processing details of various abnormal semantic patterns (high-frequency repeated queries, semantic logic conflicts, sensitive keyword combinations), such as the specific content of the abnormal semantic pattern, the triggered disposal plan, the user's feedback, etc. This risk disposal case library is used to optimize the recognition boundary of the abnormal semantics of the intent recognition model. For example, by analyzing a large number of high-frequency repeated query cases, adjust the judgment threshold of the query frequency of the intent recognition model to make it more accurately identify abnormal situations. It is also used to optimize the risk response nodes in the dialogue flow library. For example, according to the user's satisfaction feedback on the transfer to the manual customer service, adjust the response logic rules of the manual customer service transfer node in the dialogue flow library to improve the ability of the entire e-commerce service system to handle abnormal situations.

[0113] In a possible implementation manner, the method further includes:

[0114] Step S410, perform interpretability annotation on the dynamic response content to generate a response decision link report, where the response decision link report includes intent recognition basis, dialogue flow node selection logic, and filling data source.

[0115] Step S420, when the user initiates an appeal request, extract the relevant evidence chain from the response decision link report to generate an automated appeal response content.

[0116] Step S430, perform a matching degree analysis between the appeal response content and the user's actual appeal semantics, and dynamically adjust the detail level and presentation method of the interpretability annotation.

[0117] Among them, the step S410 specifically includes:

[0118] Step S411, record the attention weight distribution of each semantic fragment unit when the intent recognition model processes the current session request.

[0119] Step S412: Extract the calculation details of the priority score for the selected target dialogue flow node, including the weight assignment of candidate nodes and the context consistency verification result.

[0120] Step S413: Trace the full-link log of the filled data from querying the external database to injecting it into the response template, and mark the authenticity and validity of the data source.

[0121] Step S414: Integrate the attention weight distribution, priority score calculation details, and data link log into a structured report, and display the key decision path through a visualization chart.

[0122] Step S415: Conduct a privacy compliance check on the structured report and filter out sensitive content regarding the user's personal identity information.

[0123] In this embodiment, in the e-commerce service, after generating the dynamic response content, for example, the response content in the previously mentioned "Product After-sales - Return and Exchange Consultation" scenario. Perform interpretability annotation on this dynamic response content to generate a response decision link report. First, record the attention weight distribution of each semantic segment unit when the intent recognition model processes the current session request. For the user's session request "There is nothing wrong with this piece of clothing I bought, it's just the wrong size, and I want to know if I can exchange it", when the intent recognition model processes this request, semantic segment units such as "clothing", "wrong size", and "can I exchange" will be assigned different attention weights. Suppose the semantic segment unit "wrong size" is assigned a relatively high attention weight because it is a key piece of information closely related to the return and exchange policy, while the attention weight of the semantic segment unit "There is nothing wrong" is relatively low.

[0124] Next, extract the calculation details of the priority score for the selected target dialogue flow node. Among many candidate dialogue flow nodes, for example, the node for answering the return and exchange policy for non-damaged goods is selected. The calculation details of its priority score include the weight assignment of candidate nodes and the context consistency verification result. Regarding the weight assignment of candidate nodes, different weights may be assigned based on factors such as the response success rate and user satisfaction index in historical conversations. For example, this node for answering the return and exchange policy for non-damaged goods is assigned a relatively high weight because its historical response success rate is relatively high (assumed to be 80%). Regarding the context consistency verification result, since the user clearly mentions that the clothes are not damaged but just the wrong size, which is highly consistent with the return and exchange policy of this node for non-damaged goods, this factor also plays an important role in the priority score calculation.

[0125] Then, trace the full-link log of the filled data from querying the external database to injecting it into the response template. In this example, the filled data includes information such as the time limit for returns and exchanges, and the methods of returns and exchanges. After querying this information from the external database, inject it into the response template. The full-link log records the specific operations of querying from the database, such as the SQL statements queried, the query time, and the accuracy verification of the query results. At the same time, mark the authenticity and validity of the data source. For example, confirm that the return and exchange policy data in the database is formulated and regularly updated by the official of the e-commerce platform, so the data source is true and valid.

[0126] Integrate these attention weight distributions, priority score calculation details, and data link logs into a structured report. This structured report presents all relevant information in a clear and organized manner. And display the key decision paths through visual charts. For example, use a bar chart to show the attention weights of different semantic segment units, use a line chart to show the change of the priority scores of candidate nodes with different factors (such as historical response success rate, user satisfaction), and use a flow chart to show the process of the filled data from querying to injecting into the response template. Finally, conduct a privacy compliance check on this structured report to filter out sensitive content related to user personal identity information. In this process, check whether the report contains sensitive information such as the user's name, ID number, bank card number, etc. If so, delete or anonymize it.

[0127] When the user initiates an appeal request, for example, the user appeals that "the return and exchange response you gave me does not meet my expectations. I think there are other situations not considered". Extract the relevant evidence chain from the previously generated response decision link report to generate the content of the automated appeal response. From the intention recognition basis part in the report, check how the model understands the user's initial request and whether there are misunderstandings. From the dialogue flow node selection logic part, determine whether the correct dialogue flow node is selected to answer the user's question. From the filled data source part, check whether the filled data such as the return and exchange policy provided is accurate and complete. Generate the content of the automated appeal response based on this information, such as "Dear user, according to our records, when processing your return and exchange consultation, based on the information you mentioned such as 'the clothes are undamaged and the size is inappropriate', we selected the node to answer the return and exchange policy for undamaged goods. We queried the official database and obtained that the time limit for returns and exchanges is within 7 days from the purchase date, and the methods of returns and exchanges support both mailing and in-store. All this information is accurate and valid. If you have other special situations, please specify in detail."

[0128] Perform a matching degree analysis between the appeal response content and the actual appeal semantics of the user, and dynamically adjust the detail level and presentation method of the interpretability annotation. If the matching degree between the appeal response content and the actual appeal semantics of the user is relatively high, it indicates that the current response decision link report can already better explain the processing process, and the detail level of the interpretability annotation may not need to be adjusted, and the presentation method can also remain unchanged. However, if the matching degree is relatively low, for example, the user replies "I am a premium member, and you have not considered the impact of my membership rights on the return and exchange policy", this indicates that the current response decision link report does not contain relevant information about the impact of membership rights on the return and exchange policy. At this time, it is necessary to increase the detail level of the interpretability annotation, such as adding a check section on the relationship between membership rights and the return and exchange policy in the report, and adjusting the presentation method, such as adding a display section on the membership rights factor in the visualization chart, so as to more comprehensively display the decision-making process in future appeal processing or user inquiries.

[0129] In a possible implementation manner, the method further includes:

[0130] Step S510, count the trigger frequency and response success rate of each intent category label within a preset time window, and generate a service efficiency heat map.

[0131] Step S520, identify high-frequency and low-efficiency intent categories according to the service efficiency heat map, and initiate a targeted optimization task, where the targeted optimization task includes expanding the corresponding nodes in the dialogue process library, increasing the diversity of semantic training data, and adjusting the loss function weights of the intent recognition model.

[0132] Step S530, when a new domain intent category requirement is detected, generate a dialogue flow node prototype and a training data collection plan, and trigger a semi-supervised learning process to expand the coverage of the intent recognition model.

[0133] Step S540, gradually deploy the optimized dialogue process and intent recognition model to different user groups through a gray release mechanism, and determine the full-scale online strategy according to the test results.

[0134] In the e-commerce service scenario, within a preset time window (e.g., within a month), count the trigger frequencies and response success rates of each intent category label. For example, for the intent category label "Product Inquiry - Product Function Understanding", it is counted that it has been triggered 1000 times within this month, and the response success rate is 80% (i.e., there are 800 times when the user's questions about product functions have been successfully answered). For the intent category label "After-sales Service - Return and Exchange Inquiry", it has been triggered 500 times, and the response success rate is 70%. For the intent category label "Promotion Activity Inquiry", it has been triggered 800 times, and the response success rate is 85%, etc. Generate a service efficiency heat map based on these statistical data. In this heat map, the intent category labels are used as the horizontal axis, and the trigger frequency and response success rate are used as the vertical axis, and different colors are used to represent different value ranges. For example, the area with a high trigger frequency and a high response success rate is represented by green, indicating an efficient intent category; the area with a high trigger frequency but a low response success rate is represented by red, indicating a high-frequency and low-efficiency intent category; the area with a low trigger frequency and a high response success rate is represented by blue; the area with a low trigger frequency and a low response success rate is represented by yellow.

[0135] Identify high-frequency and low-efficiency intent categories according to the service efficiency heat map. For example, the intent category label "After-sales Service - Return and Exchange Inquiry" is identified as a high-frequency and low-efficiency intent category because its trigger frequency is relatively high (500 times) and the response success rate is relatively low (70%). Start a targeted optimization task to expand the corresponding nodes in the dialogue flow library. For example, add more return and exchange solution nodes for special situations (such as the impact of membership rights and return and exchange during promotional activities). Increase the diversity of semantic training data, collect more conversation data on different expressions of return and exchange inquiries, such as the ways of saying return and exchange by users in different regions and the special requirements for return and exchange of different types of goods. Adjust the loss function weights of the intent recognition model. For semantic features related to "After-sales Service - Return and Exchange Inquiry", increase their weights in the calculation of the loss function, so that the model pays more attention to these semantic features during the training process and improves the recognition accuracy of return and exchange inquiry intents.

[0136] When new domain intent category requirements are detected, for example, when an e-commerce platform starts to get involved in the financial service field, users may have inquiries about financial services such as installment payment and wealth management provided by the e-commerce platform. This is the new domain intent category requirement. Generate the prototype of the dialogue flow node and the training data collection plan. Regarding the prototype of the dialogue flow node, for the installment payment consultation node in financial services, its response logic rules may include installment payment policies for different goods, interest rate calculations for installment payments, application processes, and other content. The training data collection plan includes collecting professional terms in the financial field, possible ways for users to make inquiries (such as "I want to know how much the interest is for buying this mobile phone in 12 installments"), etc. Trigger the semi-supervised learning process to expand the coverage of the intent recognition model. During the semi-supervised learning process, use a small amount of labeled data (such as some typical conversations about financial service consultations manually labeled) and a large amount of unlabeled data (such as articles and user comments related to financial services collected from the Internet) to train the intent recognition model so that the model can recognize the intent categories in the new domain.

[0137] Gradually deploy the optimized dialogue process and intent recognition model to different user groups through the gray release mechanism, and determine the full-scale online strategy based on the test results. For example, first select 10% of the active users as the test group and deploy the optimized dialogue process and intent recognition model related to financial service consultations to this part of users. During this process, closely monitor the interaction between this part of users and the customer service, and count test results such as the accuracy of intent recognition and user satisfaction. If the test results are good, such as the intent recognition accuracy rate reaches over 85% and the user satisfaction rate reaches over 80%, then gradually expand the deployment scope, such as increasing the deployment ratio to 50% of the user group and continue the test. If good results are still maintained during this process, finally determine the full-scale online strategy, and comprehensively apply the optimized dialogue process and intent recognition model to the user service of the entire e-commerce platform to improve the service ability for new domain intent categories and the overall service efficiency.

[0138] Figure 2 Shows the hardware structure intent of the remote digital service system 100 provided by the embodiment of the present application, as Figure 2 shown, the remote digital service system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0139] In one possible design, the remote digital service system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the remote digital service system 100 can be a distributed system). In some embodiments, the remote digital service system 100 can be local or remote. For example, the remote digital service system 100 can access information and / or data stored in the machine-readable storage medium 120 via a network. As another example, the remote digital service system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the remote digital service system 100 can be implemented on a server. By way of example only, the server can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0140] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions that the remote digital service system 100 uses to execute or complete the exemplary methods described in this application.

[0141] In a specific implementation process, one or more processors 110 execute the computer-executable instructions stored in the machine-readable storage medium 120, so that the processors 110 can execute the AI intelligent customer service response method based on remote digital services in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected by a bus 130. The processors 110 can be used to control the sending and receiving actions of the communication unit 140.

[0142] For the specific implementation process of the processors 110, reference can be made to the various method embodiments executed by the above remote digital service system 100. Their implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0143] In addition, an embodiment of the present application also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above AI intelligent customer service response method based on remote digital services is implemented.

[0144] It should be noted that, in order to simplify the description of the present application disclosure and thus help the understanding of one or more invention embodiments, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. An AI intelligent customer service response method based on remote digital service, characterized in that: The method comprises: Receiving a session request data stream input by a user through a remote session terminal, extracting a dynamic semantic feature sequence in the session request data stream, wherein the dynamic semantic feature sequence includes a grammatical structure feature, a sentiment tendency feature, and a context-dependent feature of a text input by the user; Inputting the dynamic semantic feature sequence into a pre-trained intent recognition model, and outputting a user intent category label and a corresponding intent confidence, wherein the intent recognition model is generated based on multi-dimensional semantic relationship training in historical session data and is used to map the association weights between semantic features and preset intent categories; According to the intent category label and the intent confidence, a target dialog flow node is matched from a preset dialog flow library, wherein the target dialog flow node includes a response logic rule and a multi-round dialog jump path corresponding to the intent category label; Generate dynamic response content based on the response logic rule in the target dialog flow node, and return a response data stream containing the dynamic response content to the user through the remote session terminal; Collecting user feedback behavior data on the response data stream, and updating the association weight corresponding to the intent category label in the intent recognition model according to the interaction effectiveness index in the feedback behavior data, so as to optimize the semantic mapping accuracy of the intent recognition model for the same type of session requests; The training steps of the intent recognition model specifically include: Acquire a historical conversation data set, perform semantic slicing processing on each conversation record in the historical conversation data set, and generate a plurality of semantic segment units, each of which contains at least one expression content of a complete semantic intent; Performing feature enhancement processing on each semantic segment unit, extracting the context association vector, grammatical dependency tree structure and sentiment polarity value of the semantic segment unit, and generating a multi-dimensional semantic feature matrix; Inputting the multidimensional semantic feature matrix into an initial neural network model, and adjusting the model parameters of the initial neural network model through a back propagation algorithm so as to minimize the loss function value between the predicted intent category output by the initial neural network model and the manually annotated true intent label; Introducing an attention weight allocation mechanism into the output layer of the initial neural network model to dynamically adjust the contribution of different semantic segment units to the final intent category determination, thereby generating the intent recognition model; Injecting noise semantic fragment units into the intent recognition model through adversarial training to improve the robustness of the intent recognition model to fuzzy semantics and ambiguous expressions; The step of matching the target dialog flow node from a preset dialog flow library according to the intent category label and the intent confidence level specifically includes: According to the intention category label, a candidate dialog flow node set is screened from the dialog flow library, each node in the candidate dialog flow node set includes at least one answer logic branch associated with the intention category label; Prioritize each node in the candidate dialog flow node set, wherein the priority rating is based on the answer success rate of the node in the historical conversation, the user satisfaction index, and the consistency of the current conversation context; According to a weighted sum of the intent confidence and the priority score, selecting a node with the highest score from the set of candidate dialog flow nodes as the target dialog flow node; If the target dialog flow node includes a multi-round dialog jump path, dynamically adjusting the trigger condition threshold of the subsequent node in the jump path according to the context-dependent characteristics of the current conversation; When it is detected that the newly added semantic features in the user session request do not match the current target dialog flow node, the dialog flow backtracking mechanism is triggered to re-execute the step of matching the target dialog flow node from the preset dialog flow library according to the intent category label and intent confidence to match the new target dialog flow node.

2. The AI ​​intelligent customer service response method based on remote digital service according to claim 1 is characterized in that: The step of performing feature enhancement processing on each semantic segment unit, extracting the context association vector, grammatical dependency tree structure and sentiment polarity value of the semantic segment unit, and generating a multi-dimensional semantic feature matrix includes: Inputting the semantic segment unit into a bidirectional recurrent neural network layer, and capturing a forward semantic propagation vector and a backward semantic propagation vector of the semantic segment unit in historical session data based on a hidden state output of the bidirectional recurrent neural network layer; The forward semantic propagation vector and the backward semantic propagation vector are concatenated to generate an initial context association vector of the semantic segment unit, and the initial context association vector is subjected to multi-head self-attention weighted processing to calculate the long-distance dependency weights between the vocabulary units within the semantic segment unit to generate an enhanced context association vector with an attention mask; Based on the lexical dependency weights in the enhanced context association vector, the hierarchical relationship between the subject, predicate and object components in the semantic segment unit is parsed to generate a grammatical dependency tree structure with the core predicate verb as the root node; Traversing each leaf node in the grammatical dependency tree structure, extracting adjective phrases and adverbial phrases that have a direct modification relationship with the core predicate verb, generating a set of grammatical modifiers, performing pattern matching on each modifier in the set of grammatical modifiers with a predefined sentiment dictionary, and identifying the part-of-speech tags and sentiment intensity coefficients of sentiment polarity keywords in the modifiers; Calculating the sentiment polarity value of the semantic segment unit according to the distribution density of the sentiment intensity coefficient in the grammatical modifier component set, wherein the sentiment polarity value is a quantitative score of positive sentiment, negative sentiment or neutral; Expanding the enhanced context association vector into a two-dimensional tensor according to the order of vocabulary units, and embedding the dependency type code corresponding to the vocabulary unit in the grammatical dependency tree structure at the row and column index positions of the two-dimensional tensor; splicing the normalized scalar of the sentiment polarity value at each vocabulary unit position of the two-dimensional tensor to generate a three-dimensional semantic feature tensor including a context association vector, a grammatical dependency tree structure and a sentiment polarity value, performing channel dimension compression and spatial dimension pooling operations on the three-dimensional semantic feature tensor to eliminate redundant noise data in the three-dimensional semantic feature tensor, and generating a dense semantic feature matrix after dimensionality reduction; The dense semantic feature matrix is ​​spliced ​​with the feature matrix of the adjacent semantic segment unit in the historical conversation data by time step alignment, a position encoding vector is injected to mark the temporal position of the semantic segment unit in the complete conversation, and a sliding window normalization process is performed on the spliced ​​feature matrix based on the temporal position to balance the contribution weights of the semantic segment units at different time steps to the multidimensional semantic feature matrix; According to the variance distribution of each channel dimension in the normalized feature matrix, the feature fusion coefficients of the context association vector, the grammatical dependency tree structure and the sentiment polarity value are dynamically allocated to generate the multi-dimensional semantic feature matrix.

3. The AI ​​intelligent customer service response method based on remote digital service according to claim 1 is characterized in that: The step of generating dynamic response content based on the response logic rule in the target dialog flow node specifically includes: Parsing the response logic rules in the target dialog flow node to determine a structured template of the response content, wherein the structured template includes a fixed text field and a variable parameter slot; Extracting filling data corresponding to the variable parameter slot from the context-dependent features of the current session, the filling data including user historical behavior preferences, real-time session environment parameters and external database query results; Injecting the filling data into the structured template to generate an initial response text, and optimizing the fluency of the initial response text through a natural language generation model to generate the dynamic response content; Performing multimodal conversion processing on the dynamic response content to generate a composite response data stream including text, voice and visualization elements; Before sending the response data stream, a potential feedback path of the user to the dynamic response content is simulated through a sandbox environment, and the parameter slot priority in the response logic rule is adjusted according to the simulation result.

4. The AI ​​intelligent customer service response method based on remote digital service according to claim 3 is characterized in that: The step of updating the association weight corresponding to the intent category label in the intent recognition model according to the interaction effectiveness index in the feedback behavior data, and optimizing the semantic mapping accuracy of the intent recognition model for similar session requests, specifically includes: Extracting interaction effectiveness indicators from the feedback behavior data, wherein the interaction effectiveness indicators include user response delay duration, intent consistency of subsequent session requests, and correction labels marked by manual review; Calculating the intention determination error value of the intention recognition model in the current session according to the interaction effectiveness index, wherein the intention determination error value is positively correlated with the degree of deviation between the user's actual demand and the model output intention; Using an incremental learning algorithm to feed back the error value to a parameter adjustment layer of the intent recognition model, and dynamically updating the associated weight corresponding to the intent category label; When it is detected that the accumulated error value of the same intent category label exceeds the preset error value, a local retraining process of the intent recognition model is triggered, and the mapping relationship of the multi-dimensional semantic feature matrix is ​​re-optimized using the newly added session data; Deploy the updated intent recognition model to the shadow system, and determine whether to switch the updated intent recognition model to the corresponding business production environment by comparing the intent recognition accuracy of the new and old intent recognition models on the test data set.

5. The AI ​​intelligent customer service response method based on remote digital service according to claim 1 is characterized in that: The method further comprises: Real-time monitoring of abnormal semantic patterns in user session requests, including high-frequency repeated queries, semantic logic conflicts, and sensitive keyword combinations; When the abnormal semantic pattern is identified, a risk handling plan is triggered, which includes switching to manual customer service intervention, injecting a verification question-and-answer process, and limiting the frequency of conversation responses; Record the processing process of the abnormal semantic pattern and generate a risk handling case library for optimizing the recognition boundary of the intention recognition model for abnormal semantics and the risk response nodes in the dialogue process library; The step of triggering a risk handling solution when the abnormal semantic pattern is identified specifically includes: According to the type and severity level of the abnormal semantic pattern, a matching treatment action combination is selected from the preset treatment strategy set; Before executing the manual customer service switch, the intent of the current session is reconfirmed through the intent recognition model. If the intent category label after reconfirmation is inconsistent with the initial category label, the target dialogue flow node is re-matched; When injecting the verification question-and-answer process, dynamically generate verification questions related to the user's historical behavior, and adjust the session trust score based on the accuracy of the user's answer; When the session trust score is lower than the set trust score, an enhanced review scheme is enabled for subsequent session requests, which includes real-time semantic feature desensitization and manual review delay of the response content; After the risk treatment protocol is executed, a treatment result notification is returned to the user, and user satisfaction feedback on the treatment process is collected to optimize the priority configuration of the treatment strategy set.

6. The AI ​​intelligent customer service response method based on remote digital service according to claim 1 is characterized in that: The method further comprises: Perform interpretable annotation on the dynamic response content to generate a response decision link report, wherein the response decision link report includes intent recognition basis, dialogue flow node selection logic and filling data source; When a user initiates a complaint request, relevant evidence chains are extracted from the response decision link report to generate automated complaint response content; By analyzing the matching degree between the complaint response content and the actual complaint semantics of the user, the detail level and presentation method of the explainable annotation are dynamically adjusted; The step of annotating the dynamic response content to be interpretable and generating a response decision link report specifically includes: Recording the attention weight distribution of each semantic segment unit when the intent recognition model processes the current session request; Extracting the priority score calculation details of the selected target dialog flow node, including the weight distribution of the candidate nodes and the context consistency verification results; Track the full-link log of the populated data from querying the external database to injecting the response template, and mark the authenticity and validity of the data source; Integrate the attention weight distribution, priority score calculation details and data link logs into a structured report, and display the key decision paths through visual charts; The structured report is subjected to a privacy compliance check to filter out sensitive content regarding the user's personally identifiable information.

7. The AI ​​intelligent customer service response method based on remote digital service according to claim 1 is characterized in that: The method further comprises: Count the trigger frequency and response success rate of each intent category label within the preset time window to generate a service performance heat map; Identify high-frequency and low-efficiency intent categories according to the service performance heat map, and initiate targeted optimization tasks, including expanding corresponding nodes in the dialogue process library, increasing the diversity of semantic training data, and adjusting the loss function weight of the intent recognition model; When a new domain intent category requirement is detected, a dialog flow node prototype and a training data collection plan are generated, and a semi-supervised learning process is triggered to expand the coverage of the intent recognition model; The optimized dialogue process and intent recognition model will be gradually deployed to different user groups through the grayscale release mechanism, and the full launch strategy will be determined based on the test results.

8. A remote digital service system, characterized in that: The remote digital service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI ​​intelligent customer service response method based on remote digital service as described in any one of claims 1 to 7.

Images