An e-commerce customer service response system and method based on a multi-model routing technology

Abstract

This invention discloses an e-commerce customer service response system and method based on multi-model routing technology, relating to the field of e-commerce technology. The system includes: collecting user input text and user behavior data; performing preliminary intent matching on the user input text request using a regular expression rule base in the e-commerce domain to obtain a regular expression matching result; inputting the regular expression matching result and the user request text into a BERT multimodal classification model to obtain a precise intent recognition result; calculating a routing decision score using a dynamic weight formula based on the regular expression matching result and the precise intent recognition result to obtain the target business module; and generating personalized response content by querying a product knowledge graph through the target business module and combining it with user behavior data. This invention calculates the routing decision score using a dynamic weight formula based on the regular expression matching result and the precise intent recognition result, and generates personalized response content by querying a product knowledge graph through the target business module and combining it with user behavior data.

Description

Technical Field

[0001] This invention relates to the field of e-commerce technology, and in particular to an e-commerce customer service response system and method based on multi-model routing technology. Background Technology

[0002] In the e-commerce sector, intelligent customer service systems have become a key technology for improving user service efficiency. Existing conventional methods typically employ a rule-based preliminary intent recognition combined with a precise classification scheme using a single deep learning model. For example, after matching high-frequency keywords using regular expressions, a pre-trained language model (such as BERT) is used for fine-grained intent judgment, and then routing to the appropriate business module based on static rules. Such solutions can effectively handle user inquiries in standard scenarios.

[0003] However, existing conventional methods mainly suffer from a lack of dynamic adaptability in the collaborative mechanism between rules and models. Fixed-weight fusion methods are difficult to cope with the ever-changing user expressions and complex contexts in e-commerce scenarios. Static routing strategies may ignore the impact of visual information on decision-making. Personalized response generation is separated from routing decisions. User behavior data is only used in the final response stage and does not participate in routing optimization, resulting in the inability to identify the specific needs of high-value users in early decision-making. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an e-commerce customer service response method based on multi-model routing technology to solve the problems of insufficient collaborative adaptability between rules and models and the separation between personalized response and routing decision in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an e-commerce customer service response method based on multi-model routing technology, which includes,

[0008] Collect user input text and user behavior data, and use a regular expression rule library in the e-commerce field to perform preliminary intent matching on user input text requests to obtain regular expression matching results;

[0009] The regular expression matching results and the user request text are input into the BERT multimodal classification model to obtain accurate intent recognition results;

[0010] The regular expression matching results and the precise intent recognition results are used to calculate the routing decision score using a dynamic weighting formula to obtain the target business module.

[0011] By querying the product knowledge graph through the target business module and combining it with user behavior data, personalized response content is generated.

[0012] Personalized response content is adapted to cross-platform formats to generate customer service responses.

[0013] As a preferred embodiment of the e-commerce customer service response method based on multi-model routing technology described in this invention, the method includes the following steps: collecting user input text and user behavior data.

[0014] Receive raw text requests from users, perform basic cleaning, and generate standardized text;

[0015] User behavior data is obtained by accessing the e-commerce platform's behavior log database using the user ID.

[0016] As a preferred embodiment of the e-commerce customer service response method based on multi-model routing technology described in this invention, the method involves: performing preliminary intent matching on the user's input text request using a regular expression rule library in the e-commerce domain to obtain the regular expression matching result, including the following steps.

[0017] The system compares user-input text requests with regular expression templates from a regular expression rule library for the e-commerce domain to match keyword combinations.

[0018] Based on the matched keyword combinations, calculate the confidence score of the current regular expression match to obtain the regular expression match result.

[0019] As a preferred embodiment of the e-commerce customer service response method based on multi-model routing technology described in this invention, the method involves inputting the regular expression matching result and the user request text into the BERT multimodal classification model to obtain accurate intent recognition results, including the following steps:

[0020] The regular expression matching results and the user request text are used to generate a multimodal input text sequence;

[0021] Perform BERT standard preprocessing on the multimodal input text sequence to generate text vectors;

[0022] By loading the BERT base model using PyTorch, adding a multimodal fusion layer, and outputting an improved BERT base model architecture;

[0023] The improved BERT basic model architecture is pre-trained using e-commerce domain corpus to output the initial BERT model. The regular matching results and user behavior data are encoded into feature vectors to output the multimodal training dataset.

[0024] The multimodal training dataset is input into the initial BERT model, and fine-tuned using a joint loss function to output the BERT multimodal classification model.

[0025] By inputting the text vector into the BERT multimodal classification model, accurate intent recognition results are obtained.

[0026] As a preferred embodiment of the e-commerce customer service response method based on multi-model routing technology described in this invention, the method involves: calculating a routing decision score by applying a dynamic weighting formula to the regular expression matching result and the precise intent recognition result to obtain the target business module, including the following steps.

[0027] Extract the intent type and confidence from the regular expression matching results, and the BERT intent score and contextual coherence score from the precise intent recognition results to form structured routing input data;

[0028] Analyze user behavior data to obtain session status, and dynamically calculate weight coefficients based on session status to obtain dynamic weights;

[0029] The confidence score and the BERT intent score are dynamically weighted together to calculate the routing decision score.

[0030] By extracting the accuracy-response latency balance point for different intent categories from user behavior data, setting a threshold for consultation intents, and comparing the routing decision score with the consultation intent threshold, the target business module is obtained.

[0031] As a preferred embodiment of the e-commerce customer service response method based on multi-model routing technology described in this invention, the method generates personalized response content by querying the product knowledge graph through the target business module and combining it with user behavior data, including the following steps:

[0032] Based on the target business module, the knowledge graph query interface is used to obtain the initial product attribute data;

[0033] Extract recent behavior data from user behavior data and output user profile feature vectors;

[0034] Input product attribute data and user profile feature vectors into the strategy engine, and generate response strategies using rule templates;

[0035] Based on the response logic, appropriate content is retrieved from the multimedia resource library to form the original response material. The original response material is then sorted and cropped in a personalized manner to generate personalized response content.

[0036] As a preferred embodiment of the e-commerce customer service response method based on multi-model routing technology described in this invention, the method involves: cross-platform format adaptation of personalized response content to form a customer service response, including the following steps.

[0037] The text in the personalized response content is parsed, and the UI adaptation template is matched according to the terminal type to obtain the rendering rules;

[0038] The personalized response content is transformed according to rendering rules to obtain an optimized content package, which is then embedded with session tracking identifiers to form a customer service response.

[0039] Secondly, the present invention provides an e-commerce customer service response system based on multi-model routing technology, including a regular expression matching result module, which collects user input text and user behavior data, performs preliminary intent matching on user input text requests through a regular expression rule library in the e-commerce field, and obtains regular expression matching results.

[0040] The precise intent recognition result module inputs the regular expression matching result and the user request text into the BERT multimodal classification model to obtain the precise intent recognition result;

[0041] The target business module is obtained by calculating the routing decision score using a dynamic weighting formula based on the regular expression matching results and the precise intent recognition results.

[0042] The personalized response content module queries the product knowledge graph by the target business module and combines it with user behavior data to generate personalized response content.

[0043] The customer service response module adapts personalized response content to cross-platform formats to generate customer service responses.

[0044] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the e-commerce customer service response method based on multi-model routing technology as described in the first aspect of the present invention.

[0045] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the e-commerce customer service response method based on multi-model routing technology as described in the first aspect of the present invention.

[0046] The beneficial effects of this invention are as follows: By using a dynamic weighting formula to calculate the routing decision score through regular expression matching results and accurate intent recognition results, and by querying the product knowledge graph through the target business module and combining it with user behavior data to generate personalized response content, the routing decision score is calculated through a dynamic weighting formula. This achieves dynamic fusion of rule matching and deep learning results, and adaptively adjusts the weight allocation according to the real-time session state. This ensures that the routing decision retains the efficiency of rule matching while also possessing the semantic understanding depth of model recognition, achieving the effect of accurately matching business scenario requirements. By generating personalized response content through the collaborative processing of knowledge graph and user behavior data, a deep correlation analysis between product attributes and user profiles is achieved. Based on a strategy engine, multimodal response materials are dynamically combined, ensuring that the generated customer service responses are both professional and accurate while meeting users' personalized preferences, significantly improving the service quality and user experience of e-commerce customer service. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of an e-commerce customer service response method based on multi-model routing technology.

[0049] Figure 2 This is a schematic diagram of an e-commerce customer service response system based on multi-model routing technology.

[0050] Figure 3 This is a flowchart of the regular expression matching results.

[0051] Figure 4 This is a flowchart of the target business module. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an e-commerce customer service response method based on multi-model routing technology, including the following steps:

[0056] S1. Collect user input text and user behavior data.

[0057] S1.1 Receive the user's input of raw text, perform basic cleaning, and generate standardized text.

[0058] Furthermore, invisible control characters are matched and removed using regular expressions. Bilingual stop word filtering is performed using the English stop word list from the NLTK library and the Chinese stop word list from Harbin Institute of Technology. The Unicode normalized form C (NFC) is used to standardize the text encoding conversion to ensure character encoding consistency during subsequent processing and generate standardized text.

[0059] S1.2. Use the user ID to access the e-commerce platform's behavior log database to obtain user behavior data.

[0060] Furthermore, an SSL encrypted channel is established to connect to the e-commerce platform's database. Then, parameterized SQL queries are executed to prevent SQL injection. The query results are encapsulated in JSON-LD standard format, with fields including core dimensions such as timestamp, behavior type, and associated product ID. The data acquisition process uses the OAuth2.0 protocol for authentication to ensure compliance with GDPR data privacy regulations and generate user behavior data.

[0061] S2. Perform preliminary intent matching on the user's input text request using a regular expression rule library in the e-commerce field to obtain the regular expression matching result.

[0062] S2.1 Compare the user's input text request with the regular expression templates in the e-commerce domain regular expression rule library to match keyword combinations.

[0063] Furthermore, the user-input text request is compared line by line with the regular expression templates in the e-commerce domain regular expression rule library. The match function of the re library is used to perform pattern matching to identify text fragments containing specific keyword combinations. The e-commerce domain regular expression rule library includes product consultation templates and after-sales service templates. During the matching process, the position and length of the matched keywords are recorded. When both the main keyword (e.g., "return") and the auxiliary keyword (e.g., "policy") are matched, it is determined to be a strong match. The comparison result is output as a structured matching record, with fields including the ID of the matched regular expression template, the list of matched keywords, and the matching position, thus obtaining the matched keyword combination.

[0064] S2.2 Calculate the confidence score of the current regular expression match based on the matched keyword combination to obtain the regular expression match result.

[0065] Furthermore, the system counts the number of core keywords hit (e.g., 1 point each for "return" and "policy"), calculates the position weight based on the keyword position (e.g., a core keyword appearing at the beginning of a sentence increases the weight by 0.2), and then combines this with the matching length coefficient (e.g., the coefficient is 1.2 when the matched text accounts for more than 50% of the whole sentence) to perform a weighted calculation, calculates the confidence of the current regular expression match, and obtains the regular expression matching result.

[0066] Specifically, the expression is,

[0067] ;

[0068] in, For confidence level, Based on, For positional weights, To match the length coefficient, This is the normalization factor.

[0069] S3. Input the regular expression matching results and the user request text into the BERT multimodal classification model to obtain accurate intent recognition results.

[0070] S3.1. Combine the regular expression matching results and the user request text to generate a multimodal input text sequence.

[0071] Furthermore, the intent type field (e.g., return policy consultation) in the regular expression matching result is concatenated with the user request text to form an annotated enhanced text sequence in the format of [intent tag] original request text (e.g., [return policy consultation] What is the return policy). The confidence value in the regular expression matching result is converted into a 32-bit floating-point feature vector and aligned with the BERT word vector of the user request text in terms of dimension to generate a multimodal input text sequence.

[0072] S3.2 Perform BERT standard preprocessing on the multimodal input text sequence to generate text vectors.

[0073] Furthermore, the enhanced text sequence is converted into a sub-word ID sequence using BERTTokenizer, positional encoding is added to special tokens, and the regular matching confidence feature vector and user behavior statistical features are respectively subjected to Min-Max normalization to ensure that the numerical range is consistent with the distribution of BERT word vectors, generating a 768-dimensional text vector representation, where the first 512 dimensions are BERT word vectors, the middle 128 dimensions are normalized confidence features, and the last 128 dimensions are user behavior features, thus generating a text vector.

[0074] S3.3 Load the BERT base model using PyTorch, add a multimodal fusion layer, and output an improved BERT base model architecture.

[0075] Furthermore, the weights of the bert-base-uncased pre-trained model are loaded using the from_pretrained method of the PyTorch framework. A multimodal fusion layer is inserted after the [CLS] marker in the BERT base model. This fusion layer consists of two fully connected layers. The first layer concatenates the BERT hidden layer output (768-dimensional) with regularized matching confidence features (128-dimensional) and user behavior features (128-dimensional) to form a 1024-dimensional vector. The second layer reduces the dimensionality to 768-dimensional using the ReLU activation function. The improved BERT base model architecture maintains the original 12-layer Transformer structure, only adding a multimodal feature fusion path in the sequence output part. The final output dimension is consistent with the original BERT model.

[0076] S3.4. Use e-commerce domain corpus to perform domain adaptation pre-training on the improved BERT basic model architecture, output the initial BERT model, encode the regular matching results and user behavior data into feature vectors, and output the multimodal training dataset.

[0077] Furthermore, the improved BERT basic model architecture was pre-trained using e-commerce domain corpus (including product description text, customer service dialogue records, and user review data) for domain adaptation. Joint optimization was performed using masked language modeling and next sentence prediction tasks. During training, the parameters of the original BERT model's embedding layer were frozen, and only the parameters of the multimodal fusion layer and the last three Transformer layers were updated. The output was a domain-adapted initial BERT model. The intent type field in the regularization matching results was converted into a 32-dimensional sparse vector using One-Hot encoding, and the numerical features in the user behavior data (such as the number of returns in the last 30 days) were normalized into a 32-dimensional dense vector using Z-score. The two types of feature vectors were concatenated to form a 64-dimensional multimodal feature vector, which together with the pre-processed e-commerce domain text corpus constituted the multimodal training dataset.

[0078] S3.5 Input the multimodal training dataset into the BERT initial model, fine-tune it using the joint loss function, and output the BERT multimodal classification model.

[0079] Furthermore, the multimodal training dataset is input into the initial BERT model for fine-tuning training. A joint loss function consisting of cross-entropy loss and KL divergence loss is used. The cross-entropy loss function calculates the difference between the predicted intent and the true label, while the KL divergence loss function constrains the consistency of the distribution of multimodal feature vectors and BERT text representations. During training, the AdamW optimizer is used with a learning rate of 5e-5 and a batch size of 32. The training is iterated for 10 epochs on the training set. After fine-tuning, the optimal checkpoint is retained, and the BERT multimodal classification model is output.

[0080] S3.6 Input the text vector into the BERT multimodal classification model to obtain accurate intent recognition results.

[0081] Furthermore, the text vector is input into the BERT multimodal classification model to perform forward inference. The text vector first obtains contextual representation through the 12-layer Transformer encoder of the BERT model, and then interacts with the multimodal feature vector in the fusion layer. The fused representation is then processed by the classification head (fully connected layer + Softmax) to calculate the probability distribution of 28 types of intents. The intent with the highest probability value is selected as the main intent graph output, and the second highest probability intent is retained as an auxiliary result to generate a structured and accurate intent recognition result.

[0082] S4. Calculate the routing decision score by combining the regular expression matching results and the precise intent recognition results using a dynamic weighting formula to obtain the target business module.

[0083] S4.1 Extract the intent type and confidence level from the regular expression matching results, as well as the BERT intent score and contextual coherence score from the precise intent recognition results, to form structured routing input data.

[0084] Furthermore, the intent type field and normalized confidence score are extracted from the regular expression matching results, and the BERT intent score is obtained from the precise intent recognition results. The four types of data are then encapsulated according to a predefined field structure. The contextual coherence score is determined by analyzing the intent of the three most recent interactions in the dialogue history. Figure 1 Consistency is derived from the structured routing input data output.

[0085] S4.2 Analyze user behavior data to obtain session status, and dynamically calculate weight coefficients based on session status to obtain dynamic weights.

[0086] Furthermore, the system statistically analyzes the number of user queries, the average time interval between adjacent requests, and the distribution of operation types in the current session. It inputs the session state metrics into the dynamic weights to obtain -0.1 average time interval + 0.05 operation type entropy value, where the operation type entropy value reflects the dispersion of the operation distribution. This yields the dynamic weight coefficient (e.g., 0.45), and the weight value is limited to the range of [0.2, 0.8] to ensure routing stability. The system then outputs the normalized dynamic weights.

[0087] S4.3. Combine the confidence score with the BERT intent score using dynamic weights to calculate the routing decision score.

[0088] Furthermore, the confidence score of the regular expression matching result and the BERT intent score are weighted and fused using dynamic weight coefficients. The input parameters are range-checked to ensure that the confidence score and BERT intent score are in the range of [0,1], and the dynamic weight coefficient is in the range of [0.2,0.8]. The normalized routing decision score is then output.

[0089] Specifically, the expression is,

[0090] ;

[0091] in, For routing decision scores, For dynamic weighting coefficients, For regular expression matching confidence, Score the intent of BERT.

[0092] S4.4 Extract the accuracy-response latency balance point for different intent categories from user behavior data, set a threshold for consultation intent, compare the routing decision score with the consultation intent threshold, and obtain the target business module.

[0093] Furthermore, based on user behavior data analysis, the relationship between accuracy and response latency for each intent category is analyzed, an accuracy-latency curve is plotted, and the inflection point is selected as the equilibrium point. Based on the equilibrium point, a threshold for consultation intents is set, and the routing decision score is compared with the threshold of the corresponding intent category. When the routing decision score is greater than or equal to the threshold, the target business module is obtained.

[0094] S5. Generate personalized response content by querying the product knowledge graph through the target business module and combining it with user behavior data.

[0095] S5.1. Use the knowledge graph query interface based on the target business module to obtain the initial product attribute data.

[0096] Furthermore, based on the target business module, a SPARQL query statement is constructed, and the RDF endpoint interface of the e-commerce knowledge graph is accessed via the HTTP protocol. The query statement includes the product ID and the required attribute fields (such as material composition and washing method). The knowledge graph returns the initial product attribute data in JSON-LD format. The data fields strictly follow the Schema.org vocabulary specification. During the query process, the product ID is validated, and the returned data is checked for integrity to obtain the initial product attribute data.

[0097] S5.2 Extract recent behavior data from user behavior data and output user profile feature vector.

[0098] Furthermore, the system extracts the user behavior records for the past 30 days (including browsing product categories, purchase history, and customer service inquiry types) from the user behavior database. It then calculates the weights of behavioral keywords using the TF-IDF algorithm (e.g., 0.12 for pure cotton and 0.08 for returns). Combined with a time decay factor (daily decay coefficient of 0.95), it generates weighted behavioral features. Numerical features (e.g., 3 returns in the past 7 days) are Z-score standardized, and categorical features (e.g., preferred material type) are One-Hot encoded to output a user profile feature vector.

[0099] S5.3 Input product attribute data and user profile feature vectors into the strategy engine, and generate response strategies through rule templates.

[0100] Furthermore, product attribute data and user profile feature vectors are input into a strategy engine based on the Drools rule engine to match predefined business rule templates. During the execution of the rule engine, high-priority rules are matched first, and conflicting rules are handled using the nearest-active principle to generate response strategies.

[0101] S5.4. Based on the response logic, retrieve the adapted content from the multimedia resource library to form the original response material. Perform personalized sorting and cropping on the original response material to generate personalized response content.

[0102] Furthermore, based on the basic response content and additional service information fields in the response strategy, matching image and video materials are retrieved from the multimedia resource library. Multiple conditions are combined using Elasticsearch's Boolean query to accurately match resources, outputting the original response material set. The materials are sorted based on the preference weights in the user profile feature vector, and the top N items (e.g., N=3 for mobile devices) are extracted to generate personalized response content.

[0103] S6. Adapt personalized response content to cross-platform formats to generate customer service responses.

[0104] S6.1. Parse the text in the personalized response content, match the UI adaptation template according to the terminal type, and obtain the rendering rules.

[0105] Furthermore, the text fields and multimedia resource identifiers in the personalized response content are parsed, the terminal type is determined through User-Agent detection, and UI adaptation templates that match the current terminal type and content type are queried from the UI template library. The UI adaptation template query conditions include the terminal resolution range, content presentation format, and interactive control requirements, and structured rendering rules are output.

[0106] S6.2. The personalized response content is converted according to the rendering rules to obtain an optimized content package, which is then embedded with a session tracking identifier to form a customer service response.

[0107] Furthermore, the text and multimedia resources in the personalized response content are formatted according to rendering rules. The text content is reformatted according to the specified font size and layout position, the multimedia resources are converted to an encoding format adapted to the terminal resolution, and platform-specific event bindings are added to the interactive controls. After the conversion is completed, an optimized content package containing text, media, and metadata is generated. Then, a session tracking identifier is embedded in the header of the content package to form a customer service response that conforms to the HTTP protocol.

[0108] This embodiment also provides an e-commerce customer service response system based on multi-model routing technology, including:

[0109] The regular expression matching result module collects user input text and user behavior data, and performs preliminary intent matching on the user input text request using a regular expression rule library in the e-commerce field to obtain the regular expression matching result.

[0110] The precise intent recognition result module inputs the regular expression matching result and the user request text into the BERT multimodal classification model to obtain the precise intent recognition result;

[0111] The target business module is obtained by calculating the routing decision score using a dynamic weighting formula based on the regular expression matching results and the precise intent recognition results.

[0112] The personalized response content module queries the product knowledge graph by the target business module and combines it with user behavior data to generate personalized response content.

[0113] The customer service response module adapts personalized response content to cross-platform formats to generate customer service responses.

[0114] This embodiment also provides a computer device applicable to the e-commerce customer service response method based on multi-model routing technology, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the e-commerce customer service response method based on multi-model routing technology proposed in the above embodiment.

[0115] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0116] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the e-commerce customer service response method based on multi-model routing technology as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0117] In summary, this invention calculates routing decision scores using a dynamic weighting formula based on regular expression matching results and precise intent recognition results. It also generates personalized response content by querying the product knowledge graph through the target business module and combining it with user behavior data. The dynamic weighting formula for calculating routing decision scores achieves a dynamic fusion of rule matching and deep learning results. The weight allocation is adaptively adjusted based on real-time session status, ensuring that routing decisions retain the efficiency of rule matching while also possessing the semantic understanding depth of model recognition. This achieves the effect of accurately matching business scenario requirements. The collaborative processing of knowledge graphs and user behavior data to generate personalized response content enables deep correlation analysis between product attributes and user profiles. Based on a strategy engine, multimodal response materials are dynamically combined, ensuring that the generated customer service responses are both professional and accurate while meeting users' personalized preferences, significantly improving the service quality and user experience of e-commerce customer service.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An e-commerce customer service response method based on multi-model routing technology, characterized in that: include, Collect user input text and user behavior data, and use a regular expression rule library in the e-commerce field to perform preliminary intent matching on user input text requests to obtain regular expression matching results; The regular expression matching results and the user request text are input into the BERT multimodal classification model to obtain accurate intent recognition results, including the following steps: The regular expression matching results and the user request text are used to generate a multimodal input text sequence; Perform BERT standard preprocessing on the multimodal input text sequence to generate text vectors; By loading the BERT base model using PyTorch, adding a multimodal fusion layer, and outputting an improved BERT base model architecture; The improved BERT basic model architecture is pre-trained using e-commerce domain corpus to output the initial BERT model. The regular matching results and user behavior data are encoded into feature vectors to output the multimodal training dataset. The multimodal training dataset is input into the initial BERT model, and fine-tuned using a joint loss function to output the BERT multimodal classification model. By inputting text vectors into the BERT multimodal classification model, accurate intent recognition results are obtained; The routing decision score is calculated using a dynamic weighting formula based on the regular expression matching results and the precise intent recognition results to obtain the target business module, including the following steps: Extract the intent type and confidence from the regular expression matching results, and the BERT intent score and contextual coherence score from the precise intent recognition results to form structured routing input data; Analyze user behavior data to obtain session status, and dynamically calculate weight coefficients based on session status to obtain dynamic weights; The confidence score and the BERT intent score are dynamically weighted together to calculate the routing decision score. The accuracy-response latency balance point for different intent categories is extracted from user behavior data. A threshold for consultation intent is set, and the routing decision score is compared with the consultation intent threshold to obtain the target business module. By querying the product knowledge graph through the target business module and combining it with user behavior data, personalized response content is generated. Personalized response content is adapted to cross-platform formats to generate customer service responses.

2. The e-commerce customer service response method based on multi-model routing technology as described in claim 1, characterized in that: Collect user input text and user behavior data. Includes the following steps, Receive raw text requests from users, perform basic cleaning, and generate standardized text; User behavior data is obtained by accessing the e-commerce platform's behavior log database using the user ID.

3. The e-commerce customer service response method based on multi-model routing technology as described in claim 2, characterized in that: The user's input text request is initially matched using a regular expression rule library for e-commerce, and the regular expression matching results are obtained. Includes the following steps, The system compares user-input text requests with regular expression templates from a regular expression rule library for the e-commerce domain to match keyword combinations. Based on the matched keyword combinations, calculate the confidence score of the current regular expression match to obtain the regular expression match result.

4. The e-commerce customer service response method based on multi-model routing technology as described in claim 3, characterized in that: By querying the product knowledge graph through the target business module and combining it with user behavior data, personalized response content is generated. Includes the following steps, Based on the target business module, the knowledge graph query interface is used to obtain the initial product attribute data; Extract recent behavior data from user behavior data and output user profile feature vectors; Input product attribute data and user profile feature vectors into the strategy engine, and generate response strategies using rule templates; Based on the response logic, appropriate content is retrieved from the multimedia resource library to form the original response material. The original response material is then sorted and cropped in a personalized manner to generate personalized response content.

5. The e-commerce customer service response method based on multi-model routing technology as described in claim 4, characterized in that: To adapt personalized response content to cross-platform formats and generate customer service responses, the following steps are involved. The text in the personalized response content is parsed, and the UI adaptation template is matched according to the terminal type to obtain the rendering rules; The personalized response content is transformed according to rendering rules to obtain an optimized content package, which is then embedded with session tracking identifiers to form a customer service response.

6. An e-commerce customer service response system based on multi-model routing technology, based on the e-commerce customer service response method based on multi-model routing technology as described in any one of claims 1 to 5, characterized in that: include, The regular expression matching result module collects user input text and user behavior data, and performs preliminary intent matching on the user input text request using a regular expression rule library in the e-commerce field to obtain the regular expression matching result. The precise intent recognition result module inputs the regular expression matching result and the user request text into the BERT multimodal classification model to obtain the precise intent recognition result; The target business module is obtained by calculating the routing decision score using a dynamic weighting formula based on the regular expression matching results and the precise intent recognition results. The personalized response content module queries the product knowledge graph by the target business module and combines it with user behavior data to generate personalized response content. The customer service response module adapts personalized response content to cross-platform formats to generate customer service responses.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the e-commerce customer service response method based on multi-model routing technology as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the e-commerce customer service response method based on multi-model routing technology as described in any one of claims 1 to 5.

Images