AI intelligent customer service response method and system based on remote digital service

Abstract

The invention provides an AI intelligent customer service response method and system based on a remote digital service, and the method comprises the steps: firstly receiving a session request data flow inputted by a user through a remote session terminal, and extracting a dynamic semantic feature sequence containing a grammatical structure, an emotional tendency and a context dependence feature; inputting the data into an intention recognition model trained based on a historical session data multi-dimensional semantic relationship, and outputting an intention category label and confidence; matching a target dialogue stream node from a preset dialogue process library according to the label and the confidence coefficient; dynamic response content is generated based on the node response logic rule, and a response data stream is returned; and finally, collecting user feedback behavior data, updating the association weight of the intention recognition model according to an interaction validity index, optimizing the semantic mapping precision of the intention recognition model to similar session requests, and realizing more accurate, intelligent and adaptive customer service response.

Description

Technical Field

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

[0002] In today's era of rapid development of digital services, the demand for remote customer service is growing, and AI intelligent customer service has emerged and is widely used. Traditional AI intelligent customer service response methods have many limitations. Early customer service systems mostly used a keyword matching method, which could only simply identify user questions based on preset keywords, and had extremely limited ability to understand complex natural language expressions and semantics. When the user's question expression changes slightly or contains ambiguous semantics, it is difficult to accurately identify the user's intention, resulting in inaccurate or even wrong responses.

[0003] Some rule-based customer service systems that were subsequently developed have improved their ability to handle structured questions to a certain extent, but the rules are often fixed and limited, and cannot flexibly adapt to diverse user needs and the ever-changing language environment. For user expressions with emotional connotations or questions that are highly context-dependent, these systems are often helpless and cannot provide personalized services.

[0004] Some machine learning-based intelligent customer service methods that emerged later, although they can learn from large amounts of data, mostly focus on semantic features of a single dimension, ignoring the relationship between grammatical structure, sentiment, and context. This makes the understanding of user input text not comprehensive and in-depth enough in practical applications, and the accuracy and reliability of intent recognition are low. Moreover, these methods lack an effective feedback mechanism and cannot dynamically adjust and optimize their own performance based on actual user feedback, 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 the present application, an embodiment of the present application provides an AI intelligent customer service response method based on remote digital service, the method comprising: 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; Collect user feedback behavior data on the response data stream, 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 similar session requests.

[0006] On the other hand, an embodiment of the present application also provides a remote digital service system, including a processor and a machine-readable storage medium, wherein 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.

[0007] 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 containing grammatical structure, emotional tendency and context-dependent features. Compared with the traditional method that only focuses on surface semantics, it greatly improves the ability to understand the user's true intention. The dynamic semantic feature sequence is input into the intention recognition model generated by multi-dimensional semantic relationship training based on historical session data, and the user's intention category label and the corresponding intention confidence can be accurately output. This model based on multi-dimensional semantic relationship training effectively avoids the intention misjudgment caused by single-dimensional training, and enhances the accuracy and reliability of intention recognition. According to the intention category label and intention confidence, the target dialogue flow node is matched from the preset dialogue flow library. The target dialogue flow node contains the response logic rules and the multi-round dialogue jump path, which makes the customer service response more logical and coherent, and can flexibly respond to different user questions and multi-round dialogue scenarios, providing a better service experience. Based on the response logic rules in the target dialogue flow node, the dynamic response content is generated and returned to the user, realizing personalized and intelligent responses to meet the diverse needs of users. Collect user feedback behavior data on the response data stream, update the association weights corresponding to the intent category labels in the intent recognition model based on the interaction effectiveness index, and continuously optimize the model's semantic mapping accuracy for similar conversation requests, so that AI intelligent customer service can continue to learn and evolve, adapt to the language habits and demand changes of different users, and continuously improve service quality and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1It is a schematic diagram of the execution flow of the AI ​​intelligent customer service response method based on remote digital service provided in an embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the hardware architecture of the remote digital service system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an AI intelligent customer service response method based on remote digital service provided by an embodiment of the present application. The AI ​​intelligent customer service response method based on remote digital service is introduced in detail below.

[0011] Step S110, receiving a session request data stream input by a user through a remote session terminal, and extracting a dynamic semantic feature sequence in the session request data stream, wherein the dynamic semantic feature sequence includes grammatical structure features, sentiment tendency features, and context dependency features of the text input by the user.

[0012] 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 dropped significantly. I want to know what solutions you can provide." For this input session request data stream, the system first extracts grammatical structure features. From a grammatical point of view, this is a complex sentence containing causal relationships. The main clause is "I want to know what solutions you can provide", and the clause "I bought a smart phone from you before, but now the battery life has dropped significantly" is used to describe the background of the event. In terms of emotional tendency features, since the user mentioned that there was a problem with the product and sought a solution, it was initially judged that there was a certain negative emotional tendency, but it was not extremely dissatisfied, but more of an attitude of seeking help. The context-dependent feature is reflected in the correlation with the previous purchase behavior, that is, the purchase of a smart phone and now facing battery life problems. The system comprehensively extracts these dynamic semantic feature sequences through technical means such as lexical and syntactic analysis of the entire sentence and semantic role labeling, preparing for subsequent intent recognition.

[0013] Step S120, input the dynamic semantic feature sequence into a pre-trained intent recognition model, and output the user intent category label and the corresponding intent confidence. 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.

[0014] In this embodiment, this intention recognition model is constructed based on a large amount of historical conversation data of the e-commerce platform. For example, the historical conversation data contains different types of conversation records of many users for product problem feedback, after-sales service consultation, product use questions, etc. In the model training process, by semantically slicing these historical conversation data, the model learns the multi-dimensional semantic relationship for semantic fragment units such as "seeking solutions to product problems". For the semantic features of the current user input about seeking solutions to the battery life problem of smartphones, the model is calculated by a complex internal neural network. The model may output the user intention category label as "after-sales service-product failure solution" and output the corresponding intention confidence, such as 0.8. This means that the model is 80% sure that the user's intention is to seek a solution to the product failure in after-sales service. This confidence is based on the calculation result of the association weight between the model's input semantic features such as grammatical structure, emotional tendency and context dependence and the preset "after-sales service-product failure solution" intention category. If the input semantic features are highly matched with the "product failure solution" semantic pattern in the historical data, the confidence will be high; otherwise it will be low.

[0015] Step S130, according to the intent category label and the intent confidence, matching the target dialog flow node from the preset dialog flow library, the target dialog flow node including the response logic rule and the multi-round dialog jump path corresponding to the intent category label.

[0016] In this embodiment, according to the intent category label of "after-sales service-product failure solution" obtained in the previous step and the intent confidence of 0.8, the target dialogue flow node is matched from the preset e-commerce service dialogue process library. The dialogue process library contains many dialogue flow nodes for different intent categories. For example, for the dialogue flow node of "after-sales service-product failure solution", there may be multiple response logic branches, such as failure solutions for different product types (smartphones, computers, etc.), processing procedures for different failure causes (hardware problems, software problems, etc.). First, the system filters out the candidate dialogue flow node set related to "after-sales service-product failure solution" according to the intent category label. Then these candidate nodes are prioritized. Assume that one of the nodes is an after-sales service process specifically for smartphone battery problems, and its historical response success rate is high, because many users have obtained satisfactory solutions to similar battery life problems through this process before, and the user satisfaction index is also high. And the consistency with the current session context (purchasing a smartphone and battery life problems) is very strong. The priority score of this node will be very high. At the same time, considering the intent confidence of 0.8, this node may be selected as the target dialogue flow node after weighted sum calculation. If this target dialogue flow node contains multiple rounds of dialogue jump paths, for example, after initially asking the user about battery usage habits, if the user's answer indicates that the battery life may be reduced due to software problems, then the trigger condition threshold of the subsequent node (such as the node that guides the user to update the software) will be dynamically adjusted to make it easier to trigger, so as to further solve the problem. If in the subsequent dialogue, the user proposes a new semantic feature, such as "I think this battery problem may have occurred because I used an accessory you recommended", this new feature does not match the current target dialogue flow node, then the dialogue flow backtracking mechanism will be triggered, and the new target dialogue flow node will be matched again according to the new intent category label and confidence.

[0017] Step S140, generating dynamic response content based on the response logic rule in the target dialog flow node, and returning a response data stream containing the dynamic response content to the user through the remote session terminal.

[0018] For example, for the selected target dialog flow node, the response logic rule it contains specifies how to generate the response content. For example, the response logic rule of this node specifies a structured template, in which the fixed text field may be "Dear user, regarding the battery life problem of the smartphone you reported, we have the following solutions:", and the variable parameter slot may include information such as battery model, purchase time, whether it is within the warranty period, etc. Filling data is extracted from the context-dependent features of the current session, such as obtaining the battery model "XX model" and the purchase time "XX year XX month XX day" from the user's previous purchase record, and it is found that it is still within the warranty period. These filling data are injected into the structured template to generate the initial response text "Dear user, regarding the battery life problem of the smartphone you reported, we have the following solutions: The XX model smartphone you purchased was purchased on XX year XX month XX day and is still within the warranty period. We recommend that you send the phone back to our after-sales service center for free inspection and repair." Then, the fluency of this initial response text is optimized through the natural language generation model, such as adjusting the sentence structure, adding some connecting words, etc., to generate more fluent and natural dynamic response content. After that, the dynamic response content is processed by multimodal conversion, such as converting text content into voice information, and displaying some visual elements on the e-commerce customer service interface, such as the address map of the after-sales service center, pictures of the shipping process, etc., 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, the sandbox environment is used to simulate the user's potential feedback path for this dynamic response content. For example, the simulated user may further ask for the contact information of the after-sales service center, or ask how long it will take for the repair. According to the simulation results, if it is found that the user is more likely to ask for the contact information of the after-sales service center, the parameter slot priority in the response logic rule is adjusted to advance the provision of the after-sales service center contact information to a more important position.

[0019] Step S150, collecting user feedback behavior data on the response data stream, 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.

[0020] In this embodiment, after the user receives the response data stream, feedback behavior data is generated. For example, the user quickly replies, "OK, I will send the phone back according to your process, thank you." The interaction effectiveness index is extracted from this feedback behavior data. The shorter the user response delay, the more satisfied the user is with the reply; the more satisfied the subsequent session request (here, to perform the operation according to the reply) is. Figure 1The consistency is high, no new intent appears; and there are no correction labels for manual review. The intent judgment error value of the intent recognition model in the current session is calculated based on these interaction effectiveness indicators. Due to good user feedback, this error value is low. The incremental learning algorithm is used to feed back this low error value to the parameter adjustment layer of the intent recognition model, and dynamically update the association weight corresponding to the intent category label "after-sales service-product failure solution". For example, the semantic mapping relationship between "battery life problem" and "after-sales service-product failure solution" may be strengthened in the model. If the cumulative error value of an intent category label (such as "after-sales service-product failure solution") exceeds the preset error value in multiple similar after-sales service scenarios, this may indicate that the model's performance on this intent category is not accurate enough. At this time, the 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 about product failure after-sales service. For example, the relationship between semantic features such as new failure types and new user feedback methods and intent categories may be re-analyzed. The updated intent recognition model is deployed to the shadow system. By comparing the intent recognition accuracy of the new and old intent recognition models on the test data set (containing more product failure after-sales service session samples), if the new model has a higher accuracy, it is decided to switch the updated intent recognition model to the corresponding e-commerce business production environment, thereby improving the semantic mapping accuracy of the model when processing similar session requests and providing better after-sales service answers to users.

[0021] Based on the above steps, 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 containing grammatical structure, emotional tendency and context-dependent features. Compared with the traditional method that only focuses on surface semantics, it greatly improves the ability to understand the user's true intention. The dynamic semantic feature sequence is input into the intention recognition model generated by multi-dimensional semantic relationship training based on historical session data, and the user's intention category label and the corresponding intention confidence can be accurately output. This model based on multi-dimensional semantic relationship training effectively avoids the intention misjudgment caused by single-dimensional training and enhances the accuracy and reliability of intention recognition. According to the intention category label and intention confidence, the target dialogue flow node is matched from the preset dialogue flow library. The target dialogue flow node contains the response logic rules and the multi-round dialogue jump path, which makes the customer service response more logical and coherent, and can flexibly respond to different user questions and multi-round dialogue scenarios, providing a better service experience. Based on the response logic rules in the target dialogue flow node, the dynamic response content is generated and returned to the user, realizing personalized and intelligent responses to meet the diverse needs of users. Collect user feedback behavior data on the response data stream, update the association weights corresponding to the intent category labels in the intent recognition model based on the interaction effectiveness index, and continuously optimize the model's semantic mapping accuracy for similar conversation requests, so that AI intelligent customer service can continue to learn and evolve, adapt to the language habits and demand changes of different users, and continuously improve service quality and efficiency.

[0022] In a possible implementation, the training step of the intent recognition model specifically includes: Step S210, obtaining a historical conversation data set, performing semantic slicing processing on each conversation record in the historical conversation data set, and generating a plurality of semantic segment units, each semantic segment unit containing at least one expression content of a complete semantic intent.

[0023] In this embodiment, the e-commerce platform has accumulated a large amount of historical session data, which covers a variety of user interaction situations. For example, user sessions about product consultation, such as "I want to know the processor performance of your laptop and whether it can run large games"; after-sales service-related sessions, such as "The clothes I bought have a thread breakage, what is your return and exchange policy"; and inquiries about promotional activities, such as "Do you have any promotional activities for mobile phones recently?" etc. Each session record in these historical session data sets is semantically sliced ​​to decompose complex session records into multiple semantic segment units. Take the conversation record "I bought one of your headphones before, the sound quality is very good, but now the sound in my 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 one of your headphones before, the sound quality is very good", "Now the sound in my right ear has become smaller", "I want to know what's going on", "Can you repair or replace it for me?" Each semantic fragment unit contains at least one complete semantic intent expression content. For example, the semantic fragment unit "the sound in my right ear has become smaller" expresses the intention that there is a problem with the product, and "can you repair or replace it for me" expresses the intention to seek after-sales service.

[0024] Step S220 , 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.

[0025] For example, for the semantic segment unit "The sound in the right ear has become smaller now", it is input into the bidirectional recurrent neural network layer, and the forward semantic propagation vector and backward semantic propagation vector of the semantic segment unit in the historical session data are captured based on the hidden state output of the bidirectional recurrent neural network layer. For example, the forward semantic propagation vector may contain semantic information related to the usage of the earphones mentioned above, and the backward semantic propagation vector may be related to the subsequent solution-seeking. The forward semantic propagation vector and the backward semantic propagation vector are concatenated to generate the initial context association vector of the semantic segment unit. Then, the initial context association vector is subjected to multi-head self-attention weighting processing, and the long-distance dependency weights between the vocabulary units within the semantic segment unit are calculated to generate an enhanced context association vector with an attention mask. For example, in this semantic segment, there is a semantic association between "right ear" and "the sound has become smaller", and this weighted processing can more accurately reflect this relationship. 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, and a grammatical dependency tree structure with the core predicate verb as the root node is generated. In the sentence “the sound in my right ear has become smaller”, “sound” is the subject and “become smaller” is the predicate. The grammatical dependency tree structure obtained by analysis can clearly present the relationship between the components. After 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, and generating a set of grammatical modifiers, each modifier in the set of grammatical modifiers 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 modifier. In this example, if there is a modifier such as “obvious”, its sentiment intensity coefficient is determined by matching with the sentiment dictionary. According to the distribution density of the sentiment intensity coefficient in the set of grammatical modifiers, the sentiment polarity value of the semantic fragment unit is calculated. Here, since it is just describing the problem, the sentiment polarity value may be neutral. Expand the enhanced context association vector into a two-dimensional tensor in the order of vocabulary units, and embed the dependency type code corresponding to the vocabulary unit in the grammatical dependency tree structure in the row and column index position of the two-dimensional tensor. The normalized scalar of the sentiment polarity value is spliced ​​at each word unit position in the two-dimensional tensor to generate a three-dimensional semantic feature tensor containing a context-related vector, a grammatical dependency tree structure, and a sentiment polarity value. The three-dimensional semantic feature tensor is compressed in the channel dimension and pooled in the spatial dimension to eliminate redundant noise data in the three-dimensional semantic feature tensor and generate a dense semantic feature matrix after dimensionality reduction. The dense semantic feature matrix is ​​then spliced ​​with the feature matrix of the adjacent semantic fragment unit in the historical conversation data by time step alignment, and a position encoding vector is injected to mark the temporal position of the semantic fragment unit in the complete conversation. The spliced ​​feature matrix is ​​normalized by sliding window based on the temporal position to balance the contribution weights of the 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, the feature fusion coefficients of the context association vector, grammatical dependency tree structure and sentiment polarity value are dynamically allocated to finally generate a multidimensional semantic feature matrix.

[0026] Step S230, input the multidimensional semantic feature matrix into the initial neural network model, and adjust the model parameters of the initial neural network model through the back propagation algorithm so that the loss function value between the predicted intent category output by the initial neural network model and the manually labeled true intent label is minimized.

[0027] 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 headphone problem semantic fragment as an example, the model may predict its intent category as "after-sales service-product failure feedback". This predicted intent category is then compared with the manually annotated true intent label (the manually annotated true intent label here is also "after-sales service-product failure feedback"), and the model parameters of the initial neural network model are adjusted through the back propagation algorithm to minimize the loss function value between the predicted intent category output by the model and the manually annotated true intent label. This process is repeated continuously, a large number of semantic fragment units are trained, the model parameters are gradually optimized, and the model's ability to learn the relationship between different semantic features and intent categories is improved.

[0028] Step S240, introducing an attention weight allocation mechanism into the output layer of the initial neural network model, dynamically adjusting the contribution of different semantic fragment units to the final intent category determination, and generating the intent recognition model.

[0029] Continuing with the example of the semantic fragment of the headphone problem, different semantic fragment units may have different importance when the model determines the final intent category. For example, the semantic fragment unit "the sound in the right ear has become smaller" may be more critical for determining the intent category of "after-sales service-product failure feedback", while the semantic fragment unit "the sound quality of the headphone was very good before" is relatively less important. Through the attention weight allocation mechanism, the contribution of different semantic fragment units to the final intent category determination is dynamically adjusted. In this way, when determining the intent category, the model can more reasonably use the information of each semantic fragment unit, thereby improving the accuracy of intent recognition and generating the final intent recognition model.

[0030] Step S250, injecting noise semantic fragment units into the intent recognition model through adversarial training, so as to improve the robustness of the intent recognition model to fuzzy semantics and ambiguous expressions.

[0031] For example, in e-commerce scenarios, 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 "that thing" is a relatively vague expression. By injecting noisy semantic fragment units containing such fuzzy semantics or ambiguous expressions, the intent recognition model can learn how to deal with such situations during the training process. If the intent recognition model can correctly identify the intent category when faced with these noisy semantic fragment units, 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 a fuzzy expression such as "that thing", it can still accurately determine that the user is referring to a certain product purchased before and is reporting a product problem through the previously learned semantic relations and features, thereby improving the ability to cope with various complex semantic situations in actual e-commerce service scenarios.

[0032] In a possible implementation, step S220 includes: Step S221, input the semantic segment unit into a bidirectional recurrent neural network layer, and capture the forward semantic propagation vector and backward semantic propagation vector of the semantic segment unit in the historical session data based on the hidden state output of the bidirectional recurrent neural network layer.

[0033] In this embodiment, consider a semantic segment unit in an e-commerce conversation, "I am very disappointed with the camera function of this mobile phone, and the photos are always blurry." First, this semantic segment unit is input into the bidirectional recurrent neural network layer. The bidirectional recurrent neural network layer processes the words in this semantic segment unit in sequence, and captures the forward semantic propagation vector and backward semantic propagation vector of the semantic segment unit in the historical session data based on its hidden state output. 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 "camera function of this mobile phone", the forward semantic propagation vector contains descriptive information about a specific mobile phone and its camera function, which is a semantic continuation of the discussion or product introduction about the mobile phone that may have existed before. The backward semantic propagation vector builds a semantic relationship from the end of the sentence to the beginning. For the "photos are always blurry", the backward semantic propagation vector contains semantic information related to the photo quality problem, which may be related to the semantics of seeking solutions or expressing dissatisfaction in the future.

[0034] Step S222, concatenate the forward semantic propagation vector and the backward semantic propagation vector to generate an initial context association vector of the semantic fragment unit, and perform multi-head self-attention weighted processing on the initial context association vector, calculate the long-distance dependency weights between the vocabulary units within the semantic fragment unit, and generate an enhanced context association vector with an attention mask.

[0035] This initial context association vector integrates the semantic information captured from two directions. Then, this initial context association vector is subjected to multi-head self-attention weighted processing, and the long-distance dependency weights between the vocabulary units within the semantic segment unit are calculated to generate an enhanced context association vector with an attention mask. In this semantic segment, there is a semantic association between "photographing function" and "blurry". Through multi-head self-attention weighted processing, this long-distance dependency weight can be accurately calculated. For example, the vocabulary unit "photographing function" may have a higher dependency weight with the vocabulary unit "blurry", indicating that the two are closely related in semantic understanding. The calculation result of this dependency weight is reflected in the enhanced context association vector, and the relationship between important vocabulary units is highlighted through the attention mask.

[0036] Step S223, based on the lexical dependency weights in the enhanced context association vector, the hierarchical relationship of 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.

[0037] For example, in the sentence "I am very disappointed with the camera function of this phone, the photos are always blurry", "camera function" is the subject, "let" is the core predicate verb, "I" is the indirect object, "disappointed" is the direct object, "photo" is another subject, and "blurry" is the predicate. Through such analysis, a grammatical dependency tree structure is constructed, which clearly presents the grammatical relationship between the components in the sentence. This grammatical 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.

[0038] Step S224, traverse each leaf node in the grammatical dependency tree structure, extract adjective phrases and adverbial phrases that have a direct modification relationship with the core predicate verb, generate a set of grammatical modifiers, and then perform pattern matching on each modifier in the set of grammatical modifiers with a predefined sentiment dictionary to identify the part-of-speech tags and sentiment intensity coefficients of the sentiment polarity keywords in the modifiers.

[0039] For example, in this example, the adjective phrase "very disappointed" has a direct modification relationship with the core predicate verb "let". Each modifying component in this set of grammatical modifiers is pattern matched with the predefined sentiment dictionary to identify the part-of-speech tags and sentiment intensity coefficients of the sentiment polarity keywords in the modifying components. In the phrase "very disappointed", "disappointed" is a sentiment polarity keyword with an adjective part of speech. The sentiment intensity coefficient can be determined as a high negative intensity, such as -0.8, according to the predefined sentiment dictionary (assuming -1 represents extremely negative, 0 represents neutral, and 1 represents extremely positive).

[0040] Step S225, 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 neutrality.

[0041] In this example, since there is only one sentiment polarity keyword with a high intensity, the sentiment polarity value of this semantic fragment unit is negative sentiment, and the quantitative score is -0.8, indicating that users have a clear negative attitude towards the camera function of this mobile phone.

[0042] Step S226, 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.

[0043] For example, if the dependency relationship between "photographing function" and "let" is a subject-predicate relationship, the code representing the subject-predicate relationship is embedded in the corresponding row and column index positions. 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 a context-related vector, a grammatical dependency tree structure, and a sentiment polarity value. In this example, in addition to the original context-related vector information and grammatical dependency relationship encoding, each lexical unit position also adds normalized information of the sentiment polarity value, such as the normalized value of -0.8.

[0044] Step S227, splicing the normalized scalar of the sentiment polarity value at each vocabulary unit position of the two-dimensional tensor, generating 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, eliminating redundant noise data in the three-dimensional semantic feature tensor, and generating a dense semantic feature matrix after dimensionality reduction.

[0045] In this process, the data in the three-dimensional semantic feature tensor is processed through a specific algorithm to remove some information that contributes little to the semantic expression or is repeated, and a more compact and effective dense semantic feature matrix is ​​obtained.

[0046] Step S228, align and splice the dense semantic feature matrix with the feature matrix of adjacent semantic fragment units in the historical conversation data in time steps, inject a position encoding vector to mark the temporal position of the semantic fragment unit in the complete conversation, and perform sliding window normalization processing on the spliced ​​feature matrix based on the temporal position to balance the contribution weights of the semantic fragment units at different time steps to the multidimensional semantic feature matrix.

[0047] For example, if there is a semantic fragment unit about the appearance design of this mobile phone before, the feature matrices of these two semantic fragment units are aligned and spliced ​​according to their order in the session. In the splicing process, the position encoding vector is injected to mark the temporal position of the semantic fragment unit in the complete session, and the spliced ​​feature matrix is ​​normalized by sliding window based on the temporal position to balance the contribution weights of semantic fragment units at different time steps to the multidimensional semantic feature matrix. In this example, if the semantic fragment unit about the camera function is at a later position in the session, its weight in the entire multidimensional semantic feature matrix is ​​adjusted according to its position through sliding window normalization to ensure that it is balanced with the previous semantic fragment unit about the appearance design in terms of contribution to the overall semantic understanding.

[0048] Step S229, dynamically allocating feature fusion coefficients of the context association vector, grammatical dependency tree structure, and sentiment polarity value according to the variance distribution of each channel dimension in the normalized feature matrix, to generate the multi-dimensional semantic feature matrix.

[0049] For example, if the variance of the channel dimension related to the sentiment polarity value is large in the normalized feature matrix, it means that the sentiment polarity has a high degree of discrimination in this semantic segment unit, so a higher feature fusion coefficient will be assigned to the sentiment polarity value; if the variance of the channel dimension related to the grammatical dependency tree structure is small, it means that the importance of the grammatical structure in this semantic segment unit is relatively low, so a lower feature fusion coefficient will be assigned to the grammatical dependency tree structure. Through such dynamic allocation, the generated multidimensional semantic feature matrix can more effectively reflect the comprehensive semantic features of the semantic segment unit and provide more accurate input data for the subsequent intent recognition model.

[0050] Consider another semantic segment unit in an e-commerce conversation: "I am very interested in your new promotion, and I think it is a good deal." Similarly, this semantic segment unit is input 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 promotions). The two vectors are concatenated to obtain the initial context association vector, and the enhanced context association vector is obtained after multi-head self-attention weighted processing. In this semantic segment, there is a high dependency weight between "interested in promotions" and "interested in promotions." When constructing the grammatical dependency tree structure, "I" is the subject, "interested" is the predicate, and "promotions" is the object. Extract the modifier "very" related to the core predicate verb "interested". By matching with the sentiment dictionary, "very" is used as the sentiment intensity modifier. The sentiment polarity keyword "interested" is positive sentiment, and the sentiment intensity coefficient is 0.6. The sentiment polarity value of the entire semantic segment unit is positive sentiment, and the quantitative score is 0.6. According to the above process, it is converted into a three-dimensional semantic feature tensor containing context association vectors, grammatical dependency tree structures and sentiment polarity values. After compression and pooling operations, a dense semantic feature matrix is ​​obtained, which is spliced ​​with the feature matrix of adjacent semantic fragment units, injected with position encoding vectors, and normalized by sliding windows. Finally, feature fusion coefficients are dynamically allocated according to variance distribution to generate a multidimensional semantic feature matrix. In this way, whether it is a semantic fragment unit with positive or negative emotions, the semantic features can be accurately extracted and integrated into the multidimensional semantic feature matrix through such a set of processes, providing comprehensive and accurate semantic information input for the intention recognition model.

[0051] In e-commerce service scenarios, there are various semantic fragment units, such as user evaluation of products, feedback on after-sales service, inquiries about logistics, etc. By performing such detailed feature enhancement processing on each semantic fragment unit, it is possible to fully and deeply mine semantic information and integrate it into a multi-dimensional semantic feature matrix, thereby improving the intention recognition model's ability to understand and process different semantic situations, and better provide support for user demand identification and response in e-commerce services.

[0052] In a possible implementation, step S130 specifically includes: Step S131, based on the intention category label, filter a set of candidate dialog flow nodes from the dialog flow library, each node in the set of candidate dialog flow nodes includes at least one response logic branch associated with the intention category label.

[0053] In this embodiment, it is assumed that the intent category label obtained after the conversation request input by the user is "after-sales service of goods - return and exchange consultation" after intent recognition, and the intent confidence is 0.75. First, according to this intent category label, the candidate dialogue flow node set is screened from the dialogue process library. The dialogue process library contains many dialogue flow nodes for different e-commerce service scenarios, and these nodes construct various possible dialogue processes. For the intent category label of "after-sales service of goods - return and exchange consultation", the nodes in the candidate dialogue flow node set are all related to return and exchange. One of the nodes may be a return and exchange policy answer for the case where the goods are not damaged, and the answer logic branches included have different return and exchange time limits, return and exchange methods (such as mail, store, etc.); another node may be a return and exchange process for the case where the goods are damaged, and the answer logic branch involves damage identification process, responsibility division and other contents. Each node has at least one answer logic branch associated with the intent category label of "after-sales service of goods - return and exchange consultation".

[0054] Step S132, performing a priority score on each node in the candidate dialog flow node set, wherein the basis for the priority score includes the response success rate of the node in the historical conversation, the user satisfaction index and the consistency degree of the current conversation context.

[0055] Take the node of answering the return and exchange policy for undamaged goods as an example, and check the success rate of its answer in the historical session. If in a large number of similar sessions in the past, this node can successfully answer the user's questions about the return and exchange policy for undamaged goods, for example, the success rate reaches 80%, which provides a high basic score for its priority score. Looking at the user satisfaction index, if the user shows high satisfaction through subsequent feedback (such as evaluation, re-asking, etc.) after accepting the answer to this node, for example, the satisfaction score is 0.8 (assuming a score between 0-1, 1 is very satisfied), this will also increase its priority score. At the same time, consider the consistency of the current session context. If the user mentions in the session, "There is nothing wrong with the clothes I bought, but the size is not suitable. I want to know if I can exchange it", this node is highly consistent with the current session context because it is an answer to the return and exchange policy for undamaged goods. This high consistency will further improve its priority score.

[0056] For another node in the return and exchange process for damaged goods, assuming that its historical response success rate is 70% and the user satisfaction index is 0.7, since the user clearly stated that the goods were not damaged, the degree of context consistency with this node is low, so the priority score will be lower than that of the previous node. In this way, each node in the candidate dialogue flow node set is comprehensively prioritized.

[0057] Step S133: Select a node with the highest score from the set of candidate dialog flow nodes as the target dialog flow node according to the weighted sum of the intent confidence and the priority score.

[0058] Assume that the priority score of the node for the policy answer for non-damaged goods is 0.8 (after comprehensive calculation), the intent confidence is 0.75, and the weighted sum is 0.75 * 0.8 = 0.6; the priority score of the node for the process of returning and exchanging damaged goods is 0.6, and the weighted sum is 0.75 * 0.6 = 0.45. Since 0.6 is greater than 0.45, the node for the policy answer for non-damaged goods is selected as the target dialogue flow node.

[0059] Step S134: If the target dialog flow node includes a multi-round dialog jump path, the trigger condition threshold of the subsequent node in the jump path is dynamically adjusted according to the context-dependent characteristics of the current conversation.

[0060] If the target dialogue flow node contains multiple dialogue jump paths, for example, after answering the return and exchange time limit in the return and exchange policy, the next possible jump node is about the cost of return and exchange. At this time, the trigger condition threshold of the subsequent nodes in the jump path is dynamically adjusted according to the context-dependent features of the current session. Assuming that the user's conversation mentions "I am your premium member, is there a special return and exchange policy?", this newly added semantic feature indicates that the user's membership may affect the return and exchange policy, then the trigger condition threshold of the node related to the premium member special return and exchange policy can be lowered to make it easier to be triggered, so that the return and exchange information related to the user's identity can be provided in a timely manner in subsequent conversations.

[0061] Step S135, 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, and the step of matching the target dialog flow node from the preset dialog flow library according to the intent category label and intent confidence is re-executed to match the new target dialog flow node.

[0062] For example, when answering the question about the return and exchange policy for undamaged goods, the user suddenly said, "Actually, I bought this product during your promotion. 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 for undamaged goods, because the previous node did not consider the impact of coupon usage during promotions on returns and exchanges. At this point, the dialogue flow backtracking mechanism is triggered, and a new set of candidate dialogue flow nodes is screened from the dialogue flow library based on the new intent category label (which may become "Product After-Sales - Return and Exchange Consultation during Promotion") and intent confidence (recalculated or adjusted based on previous results). This may include nodes specifically for the relationship between returns and exchanges and coupon usage during promotions, and then the new target dialogue flow nodes are re-matched according to the above-mentioned priority scoring, weighted summation and other steps to ensure that the dialogue can accurately respond to user needs.

[0063] For another example, the user's initial intent category label is identified as "product consultation - product function understanding", and the intent confidence is 0.8. The candidate dialogue flow node set selected from the dialogue process library may include nodes for introducing different types of product functions, such as electronic product function introduction nodes, clothing product function introduction nodes, etc. For the electronic product function introduction node, its answer success rate in historical conversations is 75%, and the user satisfaction index is 0.7. If the user asks about the product function of a mobile phone and a specific model of the mobile phone is mentioned in the current conversation, this node has a high degree of consistency with the current conversation context, and its priority score is calculated comprehensively. For the clothing product function introduction node, the corresponding answer success rate and user satisfaction index are calculated based on its historical data, and the degree of consistency with the current conversation context is considered to perform priority scoring. According to the weighted sum of the intent confidence and priority score, the node with the highest score is selected as the target dialogue flow node. If the target dialogue flow node contains multiple 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 conversation, such as the mention of the interest in the camera function, the trigger condition threshold of the node related to the mobile phone camera function can be adjusted to make it easier to trigger. If the user mentions "I want to know if this phone can be used abroad" during the conversation, this newly added semantic feature does not match the current target dialogue flow node, triggering the dialogue flow backtracking mechanism to re-match the target dialogue flow node to adapt to the new user needs.

[0064] In e-commerce service scenarios, this method of matching target dialogue flow nodes from a preset dialogue process library based on intent category labels and intent confidence can effectively select appropriate dialogue flow nodes based on the user's intent and the context of the conversation, and can dynamically adjust according to the progress of the conversation to ensure the continuity and accuracy of the conversation, thereby improving user experience and service efficiency.

[0065] In a possible implementation, step S140 specifically includes: Step S141, parsing the response logic rules in the target dialog flow node, and determining a structured template of the response content, wherein the structured template includes a fixed text field and a variable parameter slot.

[0066] In this embodiment, after the target dialog flow node is determined, for example, in the aforementioned "product after-sales - return and exchange consultation" scenario, the target dialog flow node is a node for answering the policy for return and exchange of non-damaged products.

[0067] The answer logic rule of this target dialogue flow node contains a structured template. The fixed text field may be "Dear customer, regarding the return and exchange policy for non-damaged goods you inquired about, the following is the policy:". This part of the text is relatively fixed and is used to give users a clear response guide at the beginning of the answer. The variable parameter slot is related to the specific return and exchange policy, such as the time limit for return and exchange, the method of return and exchange (mail, store, etc.), etc.

[0068] Step S142, 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.

[0069] In terms of user historical behavior preferences, if it is found from the user's historical purchase records that the user often chooses to receive goods by mail, this may affect the recommendation of return and exchange methods. Real-time session environment parameters, such as the current logistics and distribution situation, if the logistics and distribution are tight, it may affect the time limit for returns and exchanges. External database query results, query the return and exchange policy database of the e-commerce platform, and obtain the specific return and exchange time for non-damaged goods within 7 days from the date of purchase. The return and exchange methods support two methods: mail and store.

[0070] 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.

[0071] Then, these filled data are injected into the structured template to generate the 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 goods within 7 days from the date of purchase. The return and exchange methods support two methods: mail and store. According to your historical purchasing habits, mail may be a method you are more familiar with." Then, the fluency of this initial response text is optimized through the natural language generation model. This model may adjust the structure of the sentence, for example, optimizing "According to your historical purchasing habits, mail may be a method you are more familiar with." to "Your historical purchasing habits show that you may be more familiar with mailing.", thereby generating more fluent and natural dynamic response content.

[0072] Step S144, performing multimodal conversion processing on the dynamic response content to generate a composite response data stream including text, voice and visualization elements.

[0073] For example, the dynamic response content can be converted into multimodal processing. The text content can be converted into voice information, and the voice can be broadcast in a friendly and professional voice style. At the same time, visual elements can be added to the e-commerce customer service interface, such as displaying an animated demonstration of the return and exchange process, or marking the store locations where returns and exchanges can be made on the map, etc., to generate a composite response data stream containing text, voice and visual elements.

[0074] Step S145, before sending the response data stream, simulate the potential feedback path of the user to the dynamic response content through the sandbox environment, and adjust the parameter slot priority in the response logic rule according to the simulation result.

[0075] For example, the simulated user may further ask who will bear the postage, or whether an appointment is required in advance for in-store returns and exchanges. If the simulation results show that the user is more likely to ask about postage, the parameter slot priority in the response logic rule is adjusted to move the explanation of postage to a more important position, so that the user's needs can be responded to more promptly and accurately in subsequent possible conversations.

[0076] In a possible implementation, step S150 specifically includes: Step S151, extracting interaction effectiveness indicators from the feedback behavior data, wherein the interaction effectiveness indicators include user response delay time, subsequent session request intention, Figure 1 Consistency and manual review of the corrected labels.

[0077] For example, in terms of user response delay, if the user replies within 1 minute after receiving the reply, this is a short response delay, which usually means that the user is more concerned about the reply and may be satisfied. Figure 1In terms of consistency, if the user replies "OK, then I choose to return or exchange the product by mail, how do I do it specifically?", this subsequent conversation request is consistent with the previous return or exchange consultation intention, and no new intention has emerged. In terms of the correction label marked by manual review, assuming that no manual review has found content that needs to be corrected, there is no correction label.

[0078] Step S152, 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 needs and the model output intention.

[0079] Due to the short user response delay, subsequent session requests Figure 1 The error value is low if the intention judgment error value is consistent and there is no correction label. The intention judgment error value is positively correlated with the degree of deviation between the user's actual needs and the model output intention. In this case, the "product after-sales - return and exchange consultation" intention output by the model is highly matched with the user's actual needs, so the error value is close to 0.

[0080] Step S153: Use an incremental learning algorithm to feed back the error value to the parameter adjustment layer of the intent recognition model, and dynamically update the associated weight corresponding to the intent category label.

[0081] For example, the incremental learning algorithm is used to feed back this error value to the parameter adjustment layer of the intent recognition model, and dynamically update the association weight corresponding to the intent category label "after-sales service - return and exchange consultation". For example, in the model, the association weight between semantic features related to the return and exchange policy (such as "no damaged goods", "within 7 days", "mailing and store delivery methods", etc.) and the "after-sales service - return and exchange consultation" intent may be fine-tuned to further strengthen the connection between them.

[0082] Step S154, when it is detected that the cumulative error value of the same intent category label exceeds the preset error value, the 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.

[0083] Assume that in multiple return and exchange consultation sessions, a certain error in intent determination is large due to a misunderstanding of the return and exchange policy for special products (such as customized products). When the cumulative error value related to "product after-sales - return and exchange consultation" exceeds the preset error value, the local retraining process will be triggered. Use the newly added session data to re-optimize the mapping relationship of the multidimensional semantic feature matrix. The newly added session data may include more session records about the return and exchange of special products, re-analyze the relationship between the semantic features and intent in these sessions, and adjust the feature fusion coefficient in the multidimensional semantic feature matrix.

[0084] Step S155, 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.

[0085] For example, in the shadow system, the intent recognition accuracy of the new and old intent recognition models on the test data set is compared. The test data set contains a large number of e-commerce service conversation samples, including various types of conversations such as product consultation and after-sales service. If the new model's intent recognition accuracy on the test data set is higher than that of the old model, for example, the old model's accuracy is 80% and the new model's accuracy is 85%, it is determined that the updated intent recognition model will be switched to the corresponding business production environment, thereby improving the model's semantic mapping accuracy when processing similar conversation requests, and better supporting user demand identification in e-commerce services.

[0086] In the e-commerce service scenario, this process of generating dynamic response content based on the target dialogue flow node and updating the intent recognition model based on feedback behavior data can continuously optimize the accuracy of the service and user experience, and improve the effectiveness and adaptability of the e-commerce service system in processing user sessions.

[0087] In a possible implementation, the method further includes: Step S310 , monitoring abnormal semantic patterns in user session requests in real time, wherein the abnormal semantic patterns include high-frequency repeated queries, semantic logic conflicts, and sensitive keyword combinations.

[0088] In this embodiment, during the conversation interaction of the e-commerce service, the abnormal semantic patterns in the user's conversation request can be monitored in real time. For example, in terms of high-frequency repeated queries, if the user repeatedly asks "What is the warranty policy of the mobile phone I bought?" and repeats the question more than 3 times in a short period of time (such as within 5 minutes), this constitutes an abnormal semantic pattern of high-frequency repeated queries. In terms of semantic logic conflicts, the user said "I want to buy this product, but I don't want to spend money or use coupons." This statement of buying a product without spending money or using 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 e-commerce services and may involve risks appear in the user's conversation, such as "Is your promotion a fraud?"

[0089] Step S320, when the abnormal semantic pattern is identified, a risk handling plan is triggered, and the risk handling plan includes switching to manual customer service intervention, injecting a verification question and answer process, and limiting the conversation response frequency.

[0090] For example, for high-frequency repeated queries, if the severity level is judged to be low (because the user may not understand or pay attention to the previous answer), the handling strategy may be to inject a verification question and answer process. For more serious anomalies such as semantic logic conflicts (which may imply malicious or unreasonable user requests), it may trigger a switch to manual customer service intervention. For more serious situations such as sensitive keyword combinations, in addition to switching to manual customer service intervention, the frequency of conversation responses may also be limited, such as limiting users to only one message within 10 minutes, to prevent possible malicious behavior or further expansion of risks.

[0091] Step S330, recording the processing process of the abnormal semantic pattern, generating a risk handling case library, which is used to optimize the recognition boundary of the intention recognition model for abnormal semantics and the risk response nodes in the dialogue process library.

[0092] Wherein, step S320 specifically includes: Step S321 , selecting a matching processing action combination from a preset processing strategy set according to the type and severity level of the abnormal semantic pattern.

[0093] Taking high-frequency repeated queries as an example, according to their type (repeated queries) and severity (low), a matching combination of treatment actions is selected from the preset treatment strategy set. In this case, the treatment strategy set may stipulate that for high-frequency repeated queries with low severity, the treatment action of injecting the verification question-and-answer process is adopted. This verification question-and-answer process dynamically generates verification questions based on the user's historical behavior-related information. Assuming that the user has purchased a mobile phone before and asked about the accessories of the mobile phone, the dynamically generated verification question may be "You asked about the accessories of our mobile phone before. Do you remember which accessory it was?" Adjust the session trust score based on the accuracy of the user's answer. If the user answers correctly, the session trust score is increased, for example, from the initial 0.5 to 0.6, indicating that the user may be a normal user who is just a little confused; if the user answers incorrectly, the session trust score is reduced, for example, to 0.4.

[0094] Step S322, before executing the manual customer service switch, the intent of the current conversation 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.

[0095] Before executing the action of switching to manual customer service, for example, in the case of semantic logic conflicts, the intent of the current session is reconfirmed through the intent recognition model. Assuming that the user's initial intent category label is identified as "commodity purchase-special purchase demand" (here, special purchase demand refers to unreasonable demands such as not wanting to spend money or use coupons to buy goods), the intent recognition model analyzes the conversation content again, and it may be found that the user's real intention is to find out whether there are other discount methods or gifts, etc., and the intent category label after reconfirmation becomes "commodity purchase-discount method consultation". Since the intent category label after reconfirmation 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 the product discount method, which contains the response logic rules for various possible discount methods (such as full discount, gifts, etc.).

[0096] Step S323, 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 according to the accuracy of the user's answer.

[0097] Step S324, when the session trust score is lower than the set trust score, an enhanced review scheme is enabled for subsequent session requests, and the enhanced review scheme includes real-time semantic feature desensitization processing and manual review delay of the response content.

[0098] When injecting the verification question-and-answer process, as mentioned above, after generating verification questions based on the user's historical behavior, adjust the session trust score based on the accuracy of the user's answer. For another example, if the user has purchased clothes in the past and asked about the material of the clothes, the dynamically generated verification question is "What material are the clothes you bought before made of?" 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), the enhanced review scheme is enabled 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", real-time semantic feature desensitization is enabled, that is, some semantic features that may involve privacy or are sensitive (such as some privacy-related information in the user's browsing history) are processed to avoid excessive use of this information during the processing process. At the same time, the manual review of the answer content is delayed, that is, the answer content automatically generated by the customer service is not sent to the user immediately, but waits for manual review to ensure the accuracy and security of the answer content.

[0099] Step S325, after the execution of the risk treatment protocol is completed, 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.

[0100] After the execution of the risk disposal agreement is completed, the disposal result notification is returned to the user. For example, if the injection verification question-and-answer process is completed and the user passes the verification, a notification is sent to the user, "Dear user, you have passed our verification process, and we will continue to answer your questions about the product." If the manual customer service is switched, the user is notified, "Dear user, because your question is more complicated, we have transferred you to the manual customer service, please wait patiently." And collect user satisfaction feedback on the disposal process to optimize the priority configuration of the disposal strategy set. If the user replies, "I am very satisfied with this verification process and feel it is very reasonable", then in the disposal strategy set, the priority of the injection verification question-and-answer process for similar situations may be increased; if the user expresses dissatisfaction, such as "I think you are too slow to transfer to the manual customer service", the relevant disposal strategies for the manual customer service transfer will be optimized, such as adjusting the transfer process or resource allocation to improve user satisfaction. At the same time, the processing process of abnormal semantic patterns is recorded to generate a risk disposal case library. This risk disposal case library 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, and user feedback. This risk handling case library is used to optimize the recognition boundary of the intent recognition model for abnormal semantics. For example, by analyzing a large number of high-frequency repeated query cases, the judgment threshold of the intent recognition model for query frequency is adjusted to enable it to more accurately identify abnormal situations. It is also used to optimize the risk response nodes in the dialogue process library. For example, based on user satisfaction feedback on manual customer service transfers, the response logic rules of the manual customer service transfer node in the dialogue process library are adjusted to improve the ability of the entire e-commerce service system to respond to abnormal situations.

[0101] In a possible implementation, the method further includes: Step S410, annotate the dynamic response content for interpretability and generate a response decision link report, wherein the response decision link report includes intent recognition basis, dialogue flow node selection logic and fill data source.

[0102] Step S420: When a user initiates a complaint request, relevant evidence chains are extracted from the response decision link report to generate automated complaint response content.

[0103] Step S430, dynamically adjusting the detail level and presentation method of the explainable annotation by performing a matching analysis between the complaint response content and the user's actual complaint semantics.

[0104] Wherein, the step S410 specifically includes: Step S411, recording the attention weight distribution of each semantic segment unit when the intent recognition model processes the current session request.

[0105] Step S412, 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.

[0106] Step S413, tracking the full-link log of the filling data from querying the external database to injecting the response template, marking the authenticity and validity of the data source.

[0107] Step S414, integrating the attention weight distribution, priority score calculation details and data link log into a structured report, and displaying the key decision path through a visual chart.

[0108] Step S415: Perform a privacy compliance check on the structured report to filter out sensitive content related to the user's personal identity information.

[0109] In this embodiment, in the e-commerce service, when dynamic response content is generated, for example, the response content in the previously mentioned "after-sales service of goods - return and exchange consultation" scenario. This dynamic response content is annotated with interpretability to generate a response decision link report. First, the attention weight distribution of each semantic segment unit when the intent recognition model processes the current session request is recorded. For the user's session request "There is no problem with the clothes I bought, but the size is not suitable. I want to know if I can change it", when the intent recognition model processes this request, the semantic segment units such as "clothes", "size is not suitable", and "can it be changed" will be assigned different attention weights. Assume that the semantic segment unit "size is not suitable" is assigned a higher attention weight because it is key information closely related to the return and exchange policy, while the attention weight of the semantic segment unit "no problem" is relatively low.

[0110] Next, the priority score calculation details of the selected target dialogue flow node are extracted. Among the many candidate dialogue flow nodes, the node for answering the policy on returns and exchanges for undamaged goods is selected. The priority score calculation details include the weight allocation of candidate nodes and the results of context consistency verification. In terms of the weight allocation of candidate nodes, different weights may be assigned based on factors such as the response success rate in historical conversations and user satisfaction indicators. For example, this node for answering the policy on returns and exchanges for undamaged goods is assigned a higher weight due to its high historical response success rate (assuming 80%). In terms of the context consistency verification results, because the user explicitly mentioned that the clothes were not damaged but the size was not suitable, it was highly consistent with the return and exchange policy of this node for undamaged goods. This factor also played an important role in the priority score calculation.

[0111] Then, track the full-link log of the populated data from the external database query to the injection into the response template. In this example, the populated data includes information such as the time limit for return and exchange, the method of return and exchange, etc. After querying this information from the external database, it is injected into the response template. The full-link log records the specific operations of the database query, such as the SQL statement of the query, the time of the query, the accuracy verification of the query results, etc. At the same time, mark the authenticity and validity of the data source, such as confirming that the return and exchange policy data in the database is officially formulated and regularly updated by the e-commerce platform, so the data source is authentic and valid.

[0112] These attention weight distributions, priority score calculation details, and data link logs are integrated into a structured report. This structured report presents all relevant information in a clear and organized manner. And the key decision paths are displayed through visual charts, such as using bar charts to show the attention weights of different semantic fragment units, using line charts to show the changes in the priority scores of candidate nodes with different factors (such as historical response success rate, user satisfaction), and using flow charts to show the process of filling data from query to injection into the response template. Finally, this structured report is checked for privacy compliance and sensitive content about the user's personal identity information is filtered. 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.

[0113] When a user initiates a complaint request, for example, the user complains that "the return and exchange reply you gave me does not meet my expectations, and I think there are other situations that have not been considered." Extract the relevant evidence chain from the previously generated response decision link report to generate an automated complaint response content. From the intention recognition basis part of the report, you can see how the model understands the user's initial request and whether there is any misunderstanding. 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 filling data source part, check whether the provided return and exchange policy and other filling data are accurate and complete. Generate automated complaint response content based on this information, such as "Dear user, according to our records, when processing your return and exchange consultation, we selected the undamaged product return and exchange policy answer node based on the information you mentioned, such as 'the clothes are not damaged, the size is not suitable'. We queried the official database and obtained the return and exchange time is within 7 days from the date of purchase. The return and exchange methods support both mail and store delivery. This information is accurate and valid. If you have other special circumstances, please explain in detail." Through the matching degree analysis between the complaint response content and the user's actual complaint semantics, the detail level and presentation method of the explainable annotation are dynamically adjusted. If the complaint response content matches the user's actual complaint semantics to a high degree, it means that the current response decision link report can already explain the processing process well, the detail level of the explainable annotation may not need to be adjusted, and the presentation method can remain unchanged. However, if the matching degree is low, for example, the user replies "I am a senior member, 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 explainable annotation, such as adding an inspection section on the relationship between member rights and return and exchange policies to the report, and adjust the presentation method, such as adding a display section on member rights factors in the visualization chart, so that the decision-making process can be more comprehensively displayed in future complaint processing or user queries.

[0114] In a possible implementation, the method further includes: Step S510, counting the triggering frequency and response success rate of each intent category tag within a preset time window to generate a service performance heat map.

[0115] Step S520, identifying high-frequency and low-efficiency intent categories based on the service performance heat map, and starting targeted optimization tasks, which include 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.

[0116] Step S530, 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.

[0117] Step S540: gradually deploy the optimized dialogue process and intent recognition model to different user groups through the grayscale release mechanism, and determine the full online strategy based on the test results.

[0118] In the e-commerce service scenario, the triggering frequency and response success rate of each intent category label are counted within a preset time window (for example, within one month). For example, for the intent category label "Product consultation - Product function understanding", it was triggered 1,000 times within this month, of which the response success rate was 80% (that is, 800 users' questions about product functions were successfully answered). For the intent category label "Product after-sales - Return and exchange consultation", it was triggered 500 times, and the response success rate was 70%. For the intent category label "Promotion consultation", it was triggered 800 times, and the response success rate was 85%, etc. A service efficiency heat map is generated based on these statistical data. In this heat map, the intent category label is used as the horizontal axis, the triggering frequency and the response success rate are used as the vertical axis, and different colors are used to represent different value ranges. For example, areas with high triggering frequency and high response success rate are represented by green, indicating efficient intent categories; areas with high triggering frequency but low response success rate are represented by red, indicating high-frequency and inefficient intent categories; areas with low triggering frequency and high response success rate are represented by blue; areas with low triggering frequency and low response success rate are represented by yellow.

[0119] Identify high-frequency and low-efficiency intent categories based on the service performance heat map, such as the "after-sales service - return and exchange consultation" intent category label, which is identified as a high-frequency and low-efficiency intent category due to its high trigger frequency (500 times) and relatively low response success rate (70%). Start targeted optimization tasks and expand the corresponding nodes in the dialogue process library, such as adding more return and exchange answer nodes for special situations (such as member rights impact, return and exchange during promotional activities, etc.). Increase the diversity of semantic training data and collect more conversation data on different expressions of return and exchange consultation, such as users in different regions say return and exchange, special requirements for return and exchange of different types of goods, etc. Adjust the loss function weight of the intent recognition model, and increase the weight of the semantic features related to "after-sales service - return and exchange consultation" in the loss function calculation, so that the model pays more attention to these semantic features during the training process and improves the recognition accuracy of return and exchange consultation intent.

[0120] When a new domain intent category demand is detected, for example, when an e-commerce platform begins to enter the financial services field, users may have inquiries about financial services such as installment payments and financial management provided by the e-commerce platform. This is a new domain intent category demand. Generate a dialog flow node prototype and a training data collection plan. In terms of dialog flow node prototypes, for the installment payment consultation node in financial services, its response logic rules may include installment payment policies for different products, installment payment interest rate calculation, application process, etc. The training data collection plan includes collecting professional terms in the financial field, possible user consultation methods (such as "I want to know how much interest I will pay for buying this mobile phone in 12 installments"), etc. Trigger a semi-supervised learning process to expand the coverage of the intent recognition model. In the semi-supervised learning process, a small amount of labeled data (such as some typical conversations about financial services consultation that are manually labeled) and a large amount of unlabeled data (such as financial services-related articles and user comments collected from the Internet) are used to train the intent recognition model so that the model can recognize intent categories in new fields.

[0121] The optimized dialogue process and intent recognition model are gradually deployed to different user groups through the grayscale release mechanism, and the full launch strategy is determined based on the test results. For example, 10% of active users are selected as the test group, and the optimized dialogue process and intent recognition model involving financial service consultation are deployed to this part of users. In this process, the interaction between these users and customer service is closely monitored, and the test results such as the accuracy of intent recognition and user satisfaction are counted. If the test results are good, such as the accuracy of intent recognition reaches more than 85% and the user satisfaction reaches more than 80%, the deployment scope is gradually expanded, such as increasing the deployment ratio to 50% of the user group, and continuing the test. If good results are still maintained in this process, the full launch strategy is finally determined, and the optimized dialogue process and intent recognition model are fully applied to the user services of the entire e-commerce platform to improve the service capabilities for new domain intent categories and the overall service efficiency.

[0122] Figure 2 The hardware structure of the remote digital service system 100 provided in the embodiment of the present application is shown as follows: Figure 2 As 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 .

[0123] 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 a machine-readable storage medium 120 via a network. For 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.

[0124] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the remote digital service system 100 to execute or use to complete the exemplary methods described in this application.

[0125] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the AI ​​intelligent customer service response method based on remote digital services in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0126] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned remote digital service system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0127] 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-mentioned AI intelligent customer service response method based on remote digital services is implemented.

[0128] It should be noted that in order to simplify the description disclosed in the present application and thereby facilitate the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, various features are sometimes grouped 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; Collect user feedback behavior data on the response data stream, 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 similar session requests.

2. The AI ​​intelligent customer service response method based on remote digital service according to claim 1 is characterized in that: 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; Noise semantic fragment units are injected into the intent recognition model through adversarial training to improve the robustness of the intent recognition model to fuzzy semantics and ambiguous expressions.

3. The AI ​​intelligent customer service response method based on remote digital service according to claim 2 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.

4. The AI ​​intelligent customer service response method based on remote digital service according to claim 2 is characterized in that: 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.

5. The AI ​​intelligent customer service response method based on remote digital service according to claim 4 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.

6. The AI ​​intelligent customer service response method based on remote digital service according to claim 5 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.

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: 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.

8. 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.

9. 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.

10. 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 9.

Images