A large model-driven customer complaint attribution system

Abstract

The present application provides a large model driven customer complaint attribution system, which relates to the field of large model driving. After encoding each record in the knowledge base in the same way as the customer feedback text data, the customer feedback semantics is processed with each record semantics respectively by significantly guiding semantic features between features, so as to evaluate the matching degree of customer feedback semantics and each record semantics in the knowledge base, thereby returning several records with high matching degree with customer feedback semantics as matching results. In this way, records related to customer feedback semantics can be automatically matched from the knowledge base, and provided as context information to LLM to ensure the accuracy of customer complaint classification, thereby enhancing the flexibility and adaptability of the customer complaint attribution system, and providing a basis for merchants to handle customer feedback information in a timely manner and provide better services.

Description

Technical Field

[0001] The present application relates to the field of big model driving, and more specifically, in an embodiment of the present application, to a big model driven customer complaint attribution system. Background Art

[0002] With the development of e-commerce, intelligent customer service systems have become an important tool for merchants to handle customer inquiries. Especially for cross-border e-commerce, dealing with a large number of customer feedback from customers from different regions around the world and speaking multiple languages ​​is a huge challenge. Traditional processing methods usually rely on manual customer service, but due to language barriers and the large number of customer feedback, this method is inefficient and prone to errors.

[0003] At present, some existing technologies mainly operate by pre-training classification models when handling customer complaints (customer complaints). This means that before the model is deployed, a large amount of labeled data is required to train the model so that it can identify and classify different customer complaint situations. This approach limits the flexibility and accuracy of existing customer complaint attribution systems. In addition, it is not possible to flexibly convey specific customer complaints as exceptions to the model and tell the model to perform special processing on the exceptions.

[0004] In recent years, the development of large language models (LLMs) has provided a new way to solve the above technical problems. LLMs have the ability to process complex natural language descriptions across languages, understand customer feedback and perform attribution analysis, such as identifying problems such as "poor product quality" or "logistics delays". Compared with traditional methods, LLMs' powerful understanding eliminates the need for pre-training of specialized classification models, and can better handle non-standardized language expressions such as modal particles, dialect slang, etc. By applying large language models, the accuracy of customer complaint classification and the adaptability of the system can be improved.

[0005] In addition, in order to further improve the accuracy and adaptability of classification, retrieval enhancement generation (RAG) technology is also introduced. RAG uses an embedding model to convert natural language customer feedback text data into vector form and store it in the knowledge base. When a new query request is received, the system will also convert this request into vector form, and compare it with the vector in the knowledge base, find the most relevant record as context information and add it to the prompt word of LLM to obtain prompt information, so as to optimize the customer complaint classification results. This method is particularly suitable for complex cases that require special treatment, and can be used as a method to correct or intervene in the expected results of LLM. That is to say, in particular, in the technical solution of the present application, the knowledge base is mainly used to store some customer complaint cases that are particularly difficult or require special treatment. If these prompt information are not given to LLM in the prompt word, then only using the existing knowledge of LLM, the expected customer complaint classification results cannot be obtained. Therefore, by applying RAG technology, it can be used as a method to correct or intervene in the expected results of LLM. For some ordinary customer complaints, the results of which can be smoothly inferred according to common sense, there is no need to store their results in RAG.

[0006] Therefore, an optimized customer complaint attribution system is desired. Summary of the invention

[0007] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a large model-driven customer complaint attribution system, which uses the same encoding method as the customer feedback text data for each record in the knowledge base, and then performs feature-by-feature significant guided semantic feature interaction processing on the customer feedback semantics and the semantics of each record, so as to evaluate the matching degree of the customer feedback semantics and the semantics of each record in the knowledge base, thereby returning a number of records with a high matching degree with the customer feedback semantics as matching results. In this way, records related to customer feedback semantics can be automatically matched from the knowledge base, and provided as context information to the LLM to ensure the accuracy of customer complaint classification, thereby enhancing the flexibility and adaptability of the customer complaint attribution system, and providing a basis for merchants to handle customer feedback information in a timely manner and provide better services.

[0008] According to one aspect of the present application, a large model-driven customer complaint attribution system is provided, which includes: a customer feedback data collection module, used to obtain customer feedback text data; a customer feedback text semantic encoding module, used to perform semantic embedding encoding on the customer feedback text data to obtain a customer feedback semantic encoding vector; a record matching module, used to match a number of records from a knowledge base based on the customer feedback semantic encoding vector, the records including customer feedback record text data and classification labels; a large model prompt word filling module, used to fill the customer feedback text data and the number of records as context information into the prompt words of a large language model to obtain prompt information; a customer complaint attribution classification module, used to determine a customer complaint classification result of the customer feedback text data based on the prompt information.

[0009] In the above-mentioned large model-driven customer complaint attribution system, the record matching module includes: a knowledge base record semantic embedding coding unit, which is used to perform semantic embedding coding on each record in the knowledge base to obtain a set of record semantic embedding coding vectors; a semantic feature joint significant fusion unit, which is used to perform semantic feature joint significant fusion on each record semantic embedding coding vector in the set of the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain a customer feedback-record semantic joint significant fusion representation, including: a joint implicit feature analysis activation processing subunit, which is used to perform joint implicit feature analysis and activation processing on the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain a customer feedback-record semantic conditional feature; a feature interaction fusion subunit, which is used to perform feature interaction fusion on the customer feedback semantic coding vector and the record semantic embedding coding vector with the customer feedback-record semantic conditional feature as conditional information to obtain the customer feedback-record semantic joint significant fusion representation; a matching degree estimation value analysis unit, which is used to perform matching degree evaluation based on the customer feedback-record semantic joint significant fusion representation, determine a matching degree estimation value, and determine whether to return a certain record as a matching result.

[0010] In the above-mentioned large model-driven customer complaint attribution system, the joint implicit feature analysis activation processing subunit includes: a joint implicit feature capture convolutional coding secondary subunit, which is used to input the customer feedback semantic coding vector and the record semantic embedding coding vector into the joint implicit feature capture network to obtain a customer feedback-record semantic joint implicit feature vector; a customer feedback-record semantic joint implicit feature activation secondary subunit, which is used to perform feature activation based on the Sigmoid function on the customer feedback-record semantic joint implicit feature vector to obtain a customer feedback-record semantic conditional feature vector as the customer feedback-record semantic conditional feature.

[0011] In the above-mentioned large model-driven customer complaint attribution system, the joint implicit feature capture convolutional coding secondary sub-unit is used to: after adding the customer feedback semantic coding vector and the record semantic embedding coding vector by position point, multiply the obtained customer feedback-record semantic sum vector with the weight matrix and then add it by position with the bias vector to obtain the customer feedback-record semantic joint interaction vector; use the tanh function to process the customer feedback-record semantic joint interaction vector to obtain the customer feedback-record semantic joint implicit feature vector.

[0012] In the above-mentioned large model-driven customer complaint attribution system, the feature interaction fusion subunit includes: a customer feedback semantic contribution calculation secondary subunit, which is used to calculate the customer feedback semantic contribution of the customer feedback semantic encoding vector relative to the customer feedback-record semantic conditional feature vector; a record semantic contribution calculation secondary subunit, which is used to calculate the record semantic contribution of the record semantic embedding encoding vector relative to the customer feedback-record semantic conditional feature vector; a weighted modulation secondary subunit, which is used to normalize the customer feedback semantic contribution and the record semantic contribution, and use the normalized customer feedback semantic contribution and the normalized record semantic contribution to normalize the customer feedback semantic encoding vector and the record semantic contribution. The semantic embedding coding vector is weighted modulated to obtain a modulated customer feedback semantic coding vector and a modulated record semantic embedding coding vector; a secondary subunit of significant guided interaction between features is used to use the modulated customer feedback semantic coding vector as a query vector, the modulated record semantic embedding coding vector as a key vector and the customer feedback-record semantic conditional feature vector as a value vector, and input the modulated customer feedback semantic coding vector, the modulated record semantic embedding coding vector and the customer feedback-record semantic conditional feature vector into a significant guided interaction module between features based on a converter structure to obtain a customer feedback-record semantic joint significant fusion representation vector as the customer feedback-record semantic joint significant fusion representation.

[0013] In the above-mentioned large model-driven customer complaint attribution system, the customer feedback semantic contribution calculation secondary subunit is used to: calculate the positional division of the customer feedback semantic encoding vector and the customer feedback-record semantic conditional feature vector to obtain a customer feedback semantic contribution vector; calculate the logarithmic function value with base two of the absolute value of each eigenvalue of the customer feedback semantic contribution vector to obtain a customer feedback semantic contribution logarithmic vector; calculate the positional dot product of the customer feedback semantic encoding vector and the customer feedback semantic contribution logarithmic vector, and perform positional multiplication on the obtained dot product vector to obtain a customer feedback semantic contribution value; calculate an exponential function with the natural constant e as the base and the customer feedback semantic contribution value as the exponent to obtain the customer feedback semantic contribution.

[0014] In the above-mentioned large model-driven customer complaint attribution system, the weighted modulation secondary sub-unit is used to: calculate the sum of the customer feedback semantic contribution and the record semantic contribution to obtain the total sum of customer feedback-record semantic contributions; respectively divide the customer feedback semantic contribution and the record semantic contribution by the total sum of customer feedback-record semantic contributions to obtain the normalized customer feedback semantic contribution and the normalized record semantic contribution; perform position point multiplication on the customer feedback semantic coding vector and the normalized customer feedback semantic contribution to obtain the modulated customer feedback semantic coding vector; perform position point multiplication on the record semantic embedding coding vector and the normalized record semantic contribution to obtain the modulated record semantic embedding coding vector.

[0015] In the above-mentioned large model-driven customer complaint attribution system, the feature-to-feature significant guided interaction secondary sub-unit is used to: vector-multiply the modulated customer feedback semantic coding vector with the transposed vector of the modulated record semantic embedding coding vector, divide the obtained modulated customer feedback-record semantic association matrix by the square root of the length of the modulated record semantic embedding coding vector by position to obtain a modulated customer feedback-record semantic association scaling matrix; use a softmax function to activate the modulated customer feedback-record semantic association scaling matrix, and multiply the obtained modulated customer feedback-record semantic association scaling activation matrix with the customer feedback-record semantic conditional feature vector to obtain the customer feedback-record semantic joint significant fusion representation vector.

[0016] In the above-mentioned large model-driven customer complaint attribution system, the matching estimate analysis unit is used to: input the customer feedback-record semantic joint significant fusion representation vector into the decoder-based matching evaluation module to obtain a matching estimate; based on the comparison between the matching estimate and a preset threshold, determine whether to return a certain record as a matching result.

[0017] Compared with the prior art, the present application provides a large model driven customer complaint attribution system, which uses the same encoding method as the customer feedback text data for each record in the knowledge base, and then performs feature-significant guided semantic feature interaction processing on the customer feedback semantics and each record semantics, so as to evaluate the matching degree of the customer feedback semantics and the semantics of each record in the knowledge base, thereby returning a number of records with a high matching degree with the customer feedback semantics as matching results. In this way, records related to customer feedback semantics can be automatically matched from the knowledge base, and provided as context information to the LLM to ensure the accuracy of customer complaint classification, thereby enhancing the flexibility and adaptability of the customer complaint attribution system, and providing a basis for merchants to handle customer feedback information in a timely manner and provide better services. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 The system block diagram of the customer complaint attribution system driven by a large model according to an embodiment of the present application.

[0020] Figure 2 The block diagram is a record matching module in the customer complaint attribution system driven by a large model according to an embodiment of the present application.

[0021] Figure 3 The figure is a schematic diagram of the data flow of the record matching module in the customer complaint attribution system driven by a large model according to an embodiment of the present application.

[0022] Figure 4 It is a block diagram of a semantic feature joint significant fusion unit in a customer complaint attribution system driven by a large model according to an embodiment of the present application.

[0023] Figure 5 The block diagram is a joint implicit feature analysis activation processing subunit in the customer complaint attribution system driven by a large model according to an embodiment of the present application.

[0024] Figure 6 It is a block diagram of a feature interaction fusion subunit in a customer complaint attribution system driven by a large model according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Various exemplary embodiments, features and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0026] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0027] In addition, in order to better illustrate the present application, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present application can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present application.

[0028] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0029] In response to the above technical problems, this application proposes a large model-driven customer complaint attribution system. Figure 1 FIG. 1 is a system block diagram of a customer complaint attribution system driven by a large model according to an embodiment of the present application. Figure 1 As shown, according to the large model-driven customer complaint attribution system 100 of the embodiment of the present application, it includes: a customer feedback data collection module 110, which is used to obtain customer feedback text data; a customer feedback text semantic encoding module 120, which is used to perform semantic embedding encoding on the customer feedback text data to obtain a customer feedback semantic encoding vector; a record matching module 130, which is used to match a number of records from a knowledge base based on the customer feedback semantic encoding vector, wherein the records include customer feedback record text data and classification labels; a large model prompt word filling module 140, which is used to fill the customer feedback text data and the several records as context information into the prompt words of the large language model to obtain prompt information; a customer complaint attribution classification module 150, which is used to determine the customer complaint classification result of the customer feedback text data based on the prompt information.

[0030] In the above-mentioned large model-driven customer complaint attribution system 100, the customer feedback data collection module 110 is used to obtain customer feedback text data. It should be understood that customer feedback is an important basis for enterprises to understand user needs, identify problems and improvement directions. By analyzing customer feedback, enterprises can promptly discover deficiencies in products or services, so as to make targeted improvements, meet user expectations, and enhance customer loyalty. Specifically, in the technical solution of the present application, customer feedback is obtained through a customer service system.

[0031] In the above-mentioned large model-driven customer complaint attribution system 100, the customer feedback text semantic encoding module 120 is used to perform semantic embedding encoding on the customer feedback text data to obtain a customer feedback semantic encoding vector. It should be understood that semantic embedding encoding converts the original text data into a high-dimensional vector representation by mapping the text to a continuous vector space, so that texts with similar meanings in the vector space also have similar representations in the vector space to better capture the semantic information of the text. Specifically, first, the customer feedback text data is preprocessed, including removing noise (such as punctuation marks, special characters), word segmentation, and part-of-speech tagging, so as to normalize the customer feedback text and reduce unnecessary interference. Then, the preprocessed customer feedback text data is input into the BERT semantic embedding model. Among them, the input data needs to be converted into a format that the model can process, including segmenting the text into tokens, adding special tags (such as [CLS] and [SEP]), and constructing corresponding attention masks. After that, the processed input data is passed to the model, and the model will output the contextual embedding representation of each tag through multiple levels of transformation. Finally, by taking out the vector corresponding to the [CLS] tag output by the model, or averaging or weighted summing the vectors of all tags in the text, a fixed-dimensional semantic encoding vector is obtained. This vector can be regarded as a semantic embedding representation of customer feedback, that is, a customer feedback semantic encoding vector, to capture the semantic information therein and improve the usability of text data.

[0032] In the above-mentioned large model-driven customer complaint attribution system 100, the record matching module 130 is used to match a number of records from the knowledge base based on the customer feedback semantic encoding vector, and the records include customer feedback record text data and classification labels. It should be understood that matching a number of records from the knowledge base can more effectively associate customer feedback with existing data, and can achieve fast and accurate response and decision-making. Specifically, in a specific embodiment of the present application, the knowledge base records some customer complaints and their classifications in the form of Q&A question-answer pairs, for example: Q: The desk I received is already broken.

[0033] A: Classification, quality, material, breakage.

[0034] Q: Thank you.

[0035] A: Classification, service evaluation, thanks.

[0036] This structured recording method allows quick retrieval and acquisition of answers and handling suggestions related to specific customer complaints, helping customer service staff to respond to customer issues more accurately. At the same time, the clarity of classification allows similar customer complaints to be quickly summarized and analyzed, providing companies with data-driven insights to optimize products and services.

[0037] In particular, in the above-mentioned large model-driven customer complaint attribution system, the step of matching several records from the knowledge base based on the customer feedback semantic encoding vector is crucial. This is because by converting the customer feedback text into a semantic encoding vector and comparing it with the records in the knowledge base, the system can find the historical cases that are most similar to the current customer feedback. These historical cases and their classification labels are added to the prompt words of the large language model (LLM) as context information, thereby enhancing the LLM's understanding of the current feedback. This enhancement helps the LLM to classify customer complaints more accurately when processing new feedback. It is worth mentioning that the knowledge base stores customer complaint cases that are particularly difficult or require special processing after processing and classification. These cases represent problems encountered in the past and their solutions to some extent. By matching the current customer feedback semantics with the records in the knowledge base, the system can draw on past successful experiences to guide the handling of current problems and the generation of customer complaint classification results, avoid repeated errors, and ensure the consistency of customer complaint classification results.

[0038] Specifically, in the process of matching several records from the knowledge base based on the customer feedback semantic encoding vector, the technical concept of the present application is to use the same encoding method as the customer feedback text data for each record in the knowledge base, and then perform significant guided semantic feature interaction processing between the customer feedback semantics and the semantics of each record, so as to evaluate the matching degree of the customer feedback semantics and the semantics of each record in the knowledge base, and thus return several records with a high matching degree with the customer feedback semantics as matching results. In this way, records related to customer feedback semantics can be automatically matched from the knowledge base, and provided as context information to LLM to ensure the accuracy of customer complaint classification. In this way, the defects caused by the traditional method can be avoided, the flexibility and adaptability of the customer complaint attribution system are enhanced, and a basis is provided for merchants to handle customer feedback information in a timely manner and provide better services.

[0039] Figure 2 The block diagram is a record matching module in the customer complaint attribution system driven by a large model according to an embodiment of the present application. Figure 3 FIG. 1 is a schematic diagram of data flow of a record matching module in a customer complaint attribution system driven by a large model according to an embodiment of the present application. Figure 2 and Figure 3As shown, the record matching module 130 includes: a knowledge base record semantic embedding coding unit 131, which is used to perform semantic embedding coding on each record in the knowledge base to obtain a set of record semantic embedding coding vectors; a semantic feature joint significant fusion unit 132, which is used to perform semantic feature joint significant fusion on each record semantic embedding coding vector in the set of the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain a customer feedback-record semantic joint significant fusion representation; a matching degree estimation value analysis unit 133, which is used to perform matching degree evaluation based on the customer feedback-record semantic joint significant fusion representation, determine a matching degree estimation value, and determine whether to return a certain record as a matching result.

[0040] Specifically, in the knowledge base record semantic embedding coding unit 131, semantic embedding coding is performed on each record in the knowledge base to obtain a set of record semantic embedding coding vectors. It should be understood that traditional text matching methods often rely on keyword matching and are difficult to capture the deep semantic relationship between texts. Through semantic embedding coding, text data can be converted into high-dimensional vectors, thereby retaining the basic information of the text, reflecting the semantic characteristics of the text, and achieving more accurate similarity comparison and retrieval in the vector space. Specifically, by using the Word2Vec model, the word vector is trained by predicting the context vocabulary using the concept of the context window, and a vector representation corresponding to each vocabulary is obtained to reflect its position in the semantic space. Then, in order to obtain the semantic embedding coding vector of each record, it is necessary to perform an aggregation operation on the vocabulary vectors contained in the text. Among them, the vectors of all words in each record are extracted, and then the arithmetic mean of these vectors is calculated to perform an aggregation operation. Finally, the semantic embedding coding vectors of all records are collected together to form a vector set, that is, a set of record semantic embedding coding vectors. This set can be used for subsequent similarity calculations and information retrieval.

[0041] Specifically, in the semantic feature joint significant fusion unit 132, the semantic feature joint significant fusion is performed on each record semantic embedding coding vector in the set of the customer feedback semantic encoding vector and the record semantic embedding coding vector to obtain the customer feedback-record semantic joint significant fusion representation. It should be understood that the customer feedback semantic encoding vector contains the semantic embedding coding features related to the customer feedback text data, and each record semantic embedding coding vector in the set of the record semantic embedding coding vector contains the semantic embedding coding features of several records matched from the knowledge base. In order to be able to perform significant feature interactive fusion for the semantic information of the two, so as to generate an enhanced and more discriminative feature representation, and provide a basis for subsequent matching degree estimation and record return, in the technical solution of the present application, the customer feedback semantic encoding vector and each record semantic embedding coding vector in the set of the record semantic embedding coding vector are further subjected to semantic feature joint significant fusion to obtain the customer feedback-record semantic joint significant fusion representation. Among them, the customer feedback-record semantic joint significant fusion representation can better reflect the essential relationship and significant interactive fusion features between customer feedback and knowledge base records. This vector not only contains the original semantic information, but also integrates the potential complex associations between the two, strengthens the importance of features, and reduces the interference of irrelevant information, thus providing a more accurate basis for subsequent matching evaluation and classification. This is crucial to improving the accuracy of customer complaint classification because it ensures that the system can make decisions based on the most relevant information.

[0042] Figure 4 FIG. 1 is a block diagram of a semantic feature joint significant fusion unit in a customer complaint attribution system driven by a large model according to an embodiment of the present application. Figure 4 As shown, the semantic feature joint significant fusion unit 132 includes: a joint implicit feature analysis activation processing subunit 1321, which is used to perform joint implicit feature analysis and activation processing on the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain a customer feedback-record semantic conditional feature; a feature interaction fusion subunit 1322, which is used to perform feature interaction fusion on the customer feedback semantic coding vector and the record semantic embedding coding vector with the customer feedback-record semantic conditional feature as conditional information to obtain the customer feedback-record semantic joint significant fusion representation.

[0043] Figure 5 FIG. 1 is a block diagram of a joint implicit feature analysis activation processing subunit in a customer complaint attribution system driven by a large model according to an embodiment of the present application. Figure 5As shown, the joint implicit feature analysis activation processing subunit 1321 includes: a joint implicit feature capture convolutional coding secondary subunit 1321-1, which is used to input the customer feedback semantic coding vector and the record semantic embedding coding vector into a joint implicit feature capture network to obtain a customer feedback-record semantic joint implicit feature vector; a customer feedback-record semantic joint implicit feature activation secondary subunit 1321-2, which is used to perform feature activation based on a Sigmoid function on the customer feedback-record semantic joint implicit feature vector to obtain a customer feedback-record semantic conditional feature vector as the customer feedback-record semantic conditional feature.

[0044] Specifically, in the joint implicit feature capture convolutional coding secondary subunit 1321-1, the customer feedback semantic coding vector and the record semantic embedding coding vector are input into the joint implicit feature capture network to obtain a customer feedback-record semantic joint implicit feature vector. It should be understood that by inputting the semantic embedding coding vectors of customer feedback and knowledge base records into the joint implicit feature capture network, a customer feedback-record semantic joint implicit feature vector can be generated. This new vector not only contains the information of the original features, but also captures the intrinsic connection between customer feedback and knowledge base records, thereby generating a richer and more discriminative customer feedback-record semantic joint implicit feature representation.

[0045] In an embodiment of the present application, the joint implicit feature capture convolutional coding secondary subunit 1321-1 includes: adding the customer feedback semantic coding vector and the record semantic embedding coding vector point by point, multiplying the obtained customer feedback-record semantic sum vector with the weight matrix, and then adding it with the bias vector by position to obtain the customer feedback-record semantic joint interaction vector; using the tanh function to process the customer feedback-record semantic joint interaction vector to obtain the customer feedback-record semantic joint implicit feature vector.

[0046] Specifically, in the customer feedback-record semantic joint implicit feature activation secondary subunit 1321-2, the customer feedback-record semantic joint implicit feature vector is subjected to feature activation based on the Sigmoid function to obtain a customer feedback-record semantic condition feature vector as the customer feedback-record semantic condition feature. It should be understood that the activation process of the customer feedback-record semantic joint implicit feature vector using the Sigmoid function can amplify the most critical parts for describing the customer feedback-record semantic comprehensive conditions, while suppressing irrelevant noise information. This ensures that the model focuses on features that are more important for subsequent matching degree estimation and customer complaint classification.

[0047] Figure 6FIG. 1 is a block diagram of a feature interaction fusion subunit in a customer complaint attribution system driven by a large model according to an embodiment of the present application. Figure 6 As shown, the feature interaction fusion subunit 1322 includes: a customer feedback semantic contribution calculation secondary subunit 1322-1, which is used to calculate the customer feedback semantic contribution of the customer feedback semantic encoding vector relative to the customer feedback-record semantic conditional feature vector; a record semantic contribution calculation secondary subunit 1322-2, which is used to calculate the record semantic contribution of the record semantic embedding encoding vector relative to the customer feedback-record semantic conditional feature vector; a weighted modulation secondary subunit 1322-3, which is used to normalize the customer feedback semantic contribution and the record semantic contribution, and use the normalized customer feedback semantic contribution and the normalized record semantic contribution to normalize the customer feedback semantic encoding vector and the record semantic conditional feature vector. The record semantic embedding coding vector is weighted modulated to obtain a modulated customer feedback semantic coding vector and a modulated record semantic embedding coding vector; a feature-to-feature significant guided interaction secondary subunit 1322-4 is used to use the modulated customer feedback semantic coding vector as a query vector, the modulated record semantic embedding coding vector as a key vector and the customer feedback-record semantic conditional feature vector as a value vector, and input the modulated customer feedback semantic coding vector, the modulated record semantic embedding coding vector and the customer feedback-record semantic conditional feature vector into a feature-to-feature significant guided interaction module based on a converter structure to obtain a customer feedback-record semantic joint significant fusion representation vector as the customer feedback-record semantic joint significant fusion representation.

[0048] Specifically, in the customer feedback semantic contribution calculation secondary subunit 1322-1 and the record semantic contribution calculation secondary subunit 1322-2, the contribution of the customer feedback and record semantic encoding vectors to the customer feedback-record semantic conditional feature vector is calculated, and the importance of each semantic feature to the final representation can be determined to obtain the customer feedback semantic contribution and the record semantic contribution. Through such an evaluation, the system can identify which features are more important for the final feature representation and for subsequent matching estimation and customer complaint classification tasks.

[0049] In an embodiment of the present application, the customer feedback semantic contribution calculation secondary subunit 1322-1 includes: calculating the positional division of the customer feedback semantic encoding vector and the customer feedback-record semantic conditional feature vector to obtain a customer feedback semantic contribution vector; calculating the logarithmic function value with base two of the absolute value of each eigenvalue of the customer feedback semantic contribution vector to obtain a customer feedback semantic contribution logarithmic vector; calculating the positional dot product of the customer feedback semantic encoding vector and the customer feedback semantic contribution logarithmic vector, and performing positional multiplication on the obtained dot product vector to obtain a customer feedback semantic contribution value; calculating an exponential function with the natural constant e as the base and the customer feedback semantic contribution value as the exponent to obtain the customer feedback semantic contribution.

[0050] Specifically, in the weighted modulation secondary subunit 1322-3, the customer feedback semantic contribution and the record semantic contribution are normalized, and the normalized customer feedback semantic contribution and the normalized record semantic contribution are used to weighted modulate the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain the modulated customer feedback semantic coding vector and the modulated record semantic embedding coding vector. It should be understood that by normalizing the customer feedback semantic contribution and the record semantic contribution, and adjusting the original customer feedback semantic coding vector and the record semantic embedding coding vector accordingly, it can be ensured that the more important customer feedback semantic features and record semantic features occupy a greater weight in the final feature representation. This helps the model better understand and process customer feedback information, and interact and match the semantic features of the two.

[0051] In an embodiment of the present application, the weighted modulation secondary subunit 1322-3 includes: calculating the sum of the customer feedback semantic contribution and the record semantic contribution to obtain the total sum of customer feedback-record semantic contributions; respectively dividing the customer feedback semantic contribution and the record semantic contribution by the total sum of customer feedback-record semantic contributions to obtain the normalized customer feedback semantic contribution and the normalized record semantic contribution; performing position point multiplication on the customer feedback semantic coding vector and the normalized customer feedback semantic contribution to obtain the modulated customer feedback semantic coding vector; performing position point multiplication on the record semantic embedding coding vector and the normalized record semantic contribution to obtain the modulated record semantic embedding coding vector.

[0052] Specifically, in the feature-to-feature significant guided interaction secondary subunit 1322-4, the modulated customer feedback semantic coding vector is used as a query vector, the modulated record semantic embedding coding vector is used as a key vector, and the customer feedback-record semantic conditional feature vector is used as a value vector. The modulated customer feedback semantic coding vector, the modulated record semantic embedding coding vector, and the customer feedback-record semantic conditional feature vector are input into the feature-to-feature significant guided interaction module based on the converter structure to obtain the customer feedback-record semantic joint significant fusion representation vector as the customer feedback-record semantic joint significant fusion representation. It should be understood that the modulated feature vectors are interactively fused through the feature-to-feature significant guided interaction module based on the converter structure. This adaptive feature fusion method enables the model to simultaneously examine the interaction between customer feedback semantic features and record semantic features from multiple perspectives, and discover the most valuable information flow path therefrom.

[0053] In an embodiment of the present application, the feature-to-feature significant guided interaction secondary subunit 1322-4 includes: vector multiplying the modulated customer feedback semantic coding vector with the transposed vector of the modulated record semantic embedding coding vector, dividing the obtained modulated customer feedback-record semantic association matrix by the square root of the length of the modulated record semantic embedding coding vector by position to obtain a modulated customer feedback-record semantic association scaling matrix; using a softmax function to activate the modulated customer feedback-record semantic association scaling matrix, and multiplying the obtained modulated customer feedback-record semantic association scaling activation matrix with the customer feedback-record semantic conditional feature vector to obtain the customer feedback-record semantic joint significant fusion representation vector.

[0054] Specifically, in the feature interaction fusion subunit 1322, the customer feedback-record semantic conditional feature is used as conditional information to perform feature interaction fusion on the customer feedback semantic encoding vector and the record semantic embedding encoding vector to obtain the customer feedback-record semantic joint significant fusion representation. It should be understood that by performing semantic feature joint significant fusion processing on the customer feedback semantic encoding vector and the record semantic embedding encoding vector, the joint significant key features and potential exercises between the customer feedback semantics and the record semantics can be fused, the expression ability of the semantic association features with higher relevance to the subsequent tasks is enhanced, and the interference of irrelevant information is reduced, thereby providing a basis for the subsequent matching evaluation and customer complaint classification, which is crucial to improving the accuracy of customer complaint classification.

[0055] More specifically, the semantic feature joint significant fusion is performed on the customer feedback semantic encoding vector and each record semantic embedding encoding vector in the set of the record semantic embedding encoding vector using the following semantic feature joint formula to obtain the customer feedback-record semantic joint significant fusion representation; wherein the semantic feature joint formula is: .

[0056] in, and are the customer feedback semantic encoding vector and the record semantic embedding encoding vector respectively, To add by location, and are the weight matrix and bias vector respectively, for function, The implicit feature vector for the semantic joint of the customer feedback-record, for function, represents the semantic condition feature vector of the customer feedback record, is the feature value of each position in the customer feedback semantic encoding vector, is the feature value of each position in the record semantic embedding encoding vector, For the customer feedback-record the feature values ​​of each position in the semantic condition feature vector, is the number of feature values ​​in the customer feedback semantic encoding vector and the record semantic embedding encoding vector, represents the logarithmic function value with base 2, represents an exponential function with the natural constant e as the base, Provide semantic contribution feedback to customers, To record semantic contribution, and are the normalized customer feedback semantic contribution and the normalized record semantic contribution, and They are the modulated customer feedback semantic encoding vector and the modulated record semantic embedding encoding vector, yes The transposed vector of yes Length, represents vector multiplication, yes function, is the customer feedback-record semantic joint significant fusion representation vector.

[0057] In an embodiment of the present application, the match estimation value analysis unit 133 includes: inputting the customer feedback-record semantic joint significant fusion representation vector into a match estimation module based on a decoder to obtain a match estimation value; based on the comparison between the match estimation value and a preset threshold, determining whether to return a certain record as a match result. It should be understood that the decoding regression is performed using the joint significant fusion feature between the customer feedback semantics and each record semantics, so as to perform a match estimation on the customer feedback semantics and each record semantics in the knowledge base. Further, based on the comparison between the match estimation value and the preset threshold, it is determined whether to return a certain record as a match result. In particular, in response to the match estimation value being greater than the preset threshold, it is determined to return a certain record as a match result. In this way, records related to customer feedback semantics can be automatically matched from the knowledge base, and provided as context information to the LLM to ensure the accuracy of customer complaint classification. In this way, the defects caused by the traditional method can be avoided, the flexibility and adaptability of the customer complaint attribution system are enhanced, and a basis is provided for merchants to handle customer feedback information in a timely manner and provide better services.

[0058] In the above-mentioned large model driven customer complaint attribution system 100, in the large model prompt word filling module 140, the customer feedback text data and the plurality of records are used as context information to fill in the prompt words of the large language model to obtain prompt information, so that more relevant and targeted prompt information can be generated. This method can help the model understand specific user needs, situations and backgrounds, so as to provide more accurate and effective responses or suggestions.

[0059] It should be understood that the Large Language Model (LLM) is an artificial intelligence technology based on deep learning, which is designed to process and generate natural language text. In the technical solution of the present application, since the initial knowledge base is empty, the classification result is mainly determined by the reasoning ability of the LLM itself. Professionals are required to make some corrections to the classification of the system, and store the corrected classification results of the original classification errors or no similar records in the knowledge base into the knowledge base, and the classification results of the LLM can be enhanced through RAG, that is, added to the Prompt through RAG retrieval. In this way, according to the prompt information, LLM can improve the classification effect and reduce the error rate. Furthermore, with the continuous accumulation of data, the number of useful records in the knowledge base gradually increases, which provides more effective reference information for the model. Since customer feedback often has similarities in content, after a period of manual intervention and the establishment of a relatively complete knowledge base, the need for manual participation will be greatly reduced. At this time, the accuracy of LLM's answers has been significantly improved, and more efficient automated responses can be achieved. It is worth mentioning that the Prompt template is a template used to organize prompt information, context, and customer complaint information when inputting into the large model. One of the templates is described below.

[0060] “You are an after-sales expert who is good at categorizing customer feedback to help merchants improve the quality of their products and services. Here is a customer service feedback, {feedback}. Some similar examples are: {sample}. Your previous classification is: {category}. Please refer to these examples, but they do not necessarily have to be completely consistent with them.

[0061] Now, please select the most appropriate customer complaint category from [{categories}] and output it. Do not output any extra characters. "Among them, {feedback} is filled with the customer complaint information, {sample} is filled with some categories that are closest to the customer complaint information found in the knowledge base, {category} is filled with the parent category that was previously filtered out, and {categories} is filled with the category selection at the current level.

[0062] Specifically, in the technical solution of this application, a large language model (LLM) is introduced into the system to avoid pre-training classification models. By using large language models (ChatGPT4, Qwen) that support multiple languages ​​to understand and respond to instructions in different languages, a flexible interactive experience is provided for users. Among them, these large language models have excellent instruction compliance and reasoning capabilities, and can handle situations where the input and output languages ​​are inconsistent.

[0063] In this way, by filling the customer feedback text data and the multiple records as context information into the prompt words of the large language model, the prompt information obtained may include a suggested solution, a detailed explanation of the customer's problem, or improvement measures for customer feedback. In this way, the model can respond to user needs more effectively and improve customer satisfaction.

[0064] In the above-mentioned large model-driven customer complaint attribution system 100, the customer complaint attribution classification module 150 is used to determine the customer complaint classification results of the customer feedback text data based on the prompt information. It should be understood that customer feedback texts often contain a variety of emotions and opinions, which may involve multiple aspects such as product quality, service attitude, functional requirements, etc. In order to accurately classify these feedbacks, it is necessary to rely on contextual information to assist the model in understanding the specific intentions of customers. For example, if the customer feedback mentions "slow response of after-sales service", combined with the service process, historical response time and other information in the relevant records, it can help the model better judge the specific classification of this feedback. By combining customer feedback with relevant records, the generated prompt information provides a rich context for the model, which not only helps to understand the specific needs of customers, but also improves the accuracy and efficiency of classification.

[0065] It is worth mentioning that in the technical solution of the present application, a method that supports multi-level classification is designed. Specifically, the input of the system is customer complaint information, and the output is multiple classifications from top to bottom. In each level of customer complaint classification, after a classification is selected, the sub-classification in the json file will be automatically read as a candidate for the next level of classification, until there is no sub-classification in the next level, and all results are output. Among them, the customer complaint table is organized into json, and the classification category of each level is automatically determined by the program, and a loop is used to go deeper layer by layer. In this way, when the customer complaint classification table is modified, there is no need to change the program or retrain the model. You only need to add or subtract fields in the json file to complete the system update. Here, the customer complaint classification table in json format is the merchant's specific classification of customer complaints. It is organized into json format to facilitate program reading and adaptive filling into the prompt.

[0066] In summary, the customer complaint attribution system 100 driven by the large model of the embodiment of the present application is explained. After encoding each record in the knowledge base in the same way as the customer feedback text data, the customer feedback semantics is processed with each record semantics respectively by significantly guiding semantic features between features, so as to evaluate the matching degree of the customer feedback semantics and each record semantics in the knowledge base, thereby returning a number of records with a high matching degree with the customer feedback semantics as matching results. In this way, records related to customer feedback semantics can be automatically matched from the knowledge base, and provided as context information to the LLM to ensure the accuracy of customer complaint classification, thereby enhancing the flexibility and adaptability of the customer complaint attribution system, and providing a basis for merchants to handle customer feedback information in a timely manner and provide better services.

[0067] As described above, the customer complaint attribution system 100 driven by the big model according to the embodiment of the present application can be implemented in various terminal devices. In one example, the customer complaint attribution system 100 driven by the big model can be integrated into the terminal device as a software module and / or a hardware module. For example, the customer complaint attribution system 100 driven by the big model can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the customer complaint attribution system 100 driven by the big model can also be one of the many hardware modules of the terminal device.

[0068] Alternatively, in another example, the big model-driven customer complaint attribution system 100 and the terminal device may also be separate devices, and the big model-driven customer complaint attribution system 100 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0069] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0070] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0071] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0072] It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0073] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any attached figure mark in the claims should not be regarded as limiting the claims involved.

[0074] In addition, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the system claim can also be implemented by one unit or device through software or hardware. The second and other words are used to indicate names, but not to indicate any particular order.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit of the technical solution of the present invention.

Claims

1. A customer complaint attribution system driven by a large model, characterized in that: include: Customer feedback data collection module, used to obtain customer feedback text data; A customer feedback text semantic encoding module, used for performing semantic embedding encoding on the customer feedback text data to obtain a customer feedback semantic encoding vector; A record matching module, used to match a number of records from a knowledge base based on the customer feedback semantic encoding vector, wherein the records include customer feedback record text data and classification labels; A large model prompt word filling module, used to fill the customer feedback text data and the plurality of records as context information into the prompt words of the large language model to obtain prompt information; A customer complaint attribution classification module, used to determine a customer complaint classification result of the customer feedback text data based on the prompt information; Wherein, the record matching module includes: A knowledge base record semantic embedding coding unit, used for performing semantic embedding coding on each record in the knowledge base to obtain a set of record semantic embedding coding vectors; A semantic feature joint significant fusion unit, used for performing semantic feature joint significant fusion on each record semantic embedding coding vector in the set of the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain a customer feedback-record semantic joint significant fusion representation, including: a joint implicit feature analysis activation processing subunit, used for performing joint implicit feature analysis and activation processing on the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain a customer feedback-record semantic conditional feature; a feature interactive fusion subunit, used for performing feature interactive fusion on the customer feedback semantic coding vector and the record semantic embedding coding vector with the customer feedback-record semantic conditional feature as conditional information to obtain the customer feedback-record semantic joint significant fusion representation; A matching degree estimation value analysis unit, configured to perform matching degree evaluation based on the customer feedback-record semantic joint significant fusion representation, determine a matching degree estimation value, and determine whether to return a certain record as a matching result; Wherein, the feature interaction fusion subunit includes: A customer feedback semantic contribution calculation secondary subunit, used to calculate the customer feedback semantic contribution of the customer feedback semantic encoding vector relative to the customer feedback-record semantic condition feature vector; A record semantic contribution calculation secondary subunit, used to calculate the record semantic contribution of the record semantic embedding coding vector relative to the customer feedback-record semantic conditional feature vector; The weighted modulation secondary subunit is used to normalize the customer feedback semantic contribution and the record semantic contribution, and use the normalized customer feedback semantic contribution and the normalized record semantic contribution to weighted modulate the customer feedback semantic coding vector and the record semantic embedding coding vector to obtain the modulated customer feedback semantic coding vector and the modulated record semantic embedding coding vector.

2. The large model-driven customer complaint attribution system according to claim 1 is characterized in that: The joint implicit feature analysis activation processing subunit includes: A joint implicit feature capture convolutional coding secondary subunit, used for inputting the customer feedback semantic coding vector and the record semantic embedding coding vector into a joint implicit feature capture network to obtain a customer feedback-record semantic joint implicit feature vector; The customer feedback-record semantic joint implicit feature activation secondary subunit is used to perform feature activation based on the Sigmoid function on the customer feedback-record semantic joint implicit feature vector to obtain a customer feedback-record semantic conditional feature vector as the customer feedback-record semantic conditional feature.

3. The large model-driven customer complaint attribution system according to claim 2 is characterized in that: The joint implicit feature capture convolutional coding secondary sub-unit is used to: After adding the customer feedback semantic encoding vector and the record semantic embedding encoding vector by position, the obtained customer feedback-record semantic sum vector is multiplied by the weight matrix and then added by position with the bias vector to obtain the customer feedback-record semantic joint interaction vector; The customer feedback-record semantic joint interaction vector is processed using a tanh function to obtain the customer feedback-record semantic joint implicit feature vector.

4. The large model-driven customer complaint attribution system according to claim 3 is characterized in that: The feature interaction fusion subunit further includes: The feature-to-feature significant guided interaction secondary subunit is used to use the modulated customer feedback semantic coding vector as a query vector, the modulated record semantic embedding coding vector as a key vector and the customer feedback-record semantic conditional feature vector as a value vector, and input the modulated customer feedback semantic coding vector, the modulated record semantic embedding coding vector and the customer feedback-record semantic conditional feature vector into a transformer-structured feature-to-feature significant guided interaction module to obtain a customer feedback-record semantic joint significant fusion representation vector as the customer feedback-record semantic joint significant fusion representation.

5. The large model-driven customer complaint attribution system according to claim 4 is characterized in that: The customer feedback semantic contribution calculation secondary subunit is used to: Calculating the positional division of the customer feedback semantic encoding vector and the customer feedback-record semantic conditional feature vector to obtain a customer feedback semantic contribution vector; Calculating a logarithmic function value with base 2 of the absolute value of each eigenvalue of the customer feedback semantic contribution vector to obtain a customer feedback semantic contribution logarithmic vector; Calculate the positional dot product of the customer feedback semantic encoding vector and the customer feedback semantic contribution logarithm vector, and perform positional multiplication on the obtained dot product vector to obtain the customer feedback semantic contribution value; An exponential function with the natural constant e as the base and the customer feedback semantic contribution value as the exponent is calculated to obtain the customer feedback semantic contribution degree.

6. The large model-driven customer complaint attribution system according to claim 5, characterized in that: The weighted modulation secondary subunit is used for: Calculating the sum of the customer feedback semantic contribution and the record semantic contribution to obtain a total sum of customer feedback-record semantic contributions; Respectively dividing the customer feedback semantic contribution and the record semantic contribution by the total sum of the customer feedback-record semantic contributions to obtain the normalized customer feedback semantic contribution and the normalized record semantic contribution; Multiplying the customer feedback semantic coding vector by the normalized customer feedback semantic contribution by position point to obtain the modulated customer feedback semantic coding vector; The record semantic embedding coding vector is multiplied by the normalized record semantic contribution degree at a position point to obtain the modulated record semantic embedding coding vector.

7. The large model-driven customer complaint attribution system according to claim 6, characterized in that: The inter-feature significant guided interaction secondary subunit is used for: Performing vector multiplication of the modulated customer feedback semantic coding vector and the transposed vector of the modulated record semantic embedding coding vector, and dividing the obtained modulated customer feedback-record semantic association matrix by the square root of the length of the modulated record semantic embedding coding vector by position to obtain a modulated customer feedback-record semantic association scaling matrix; The modulated customer feedback-record semantic association scaling matrix is ​​activated using a softmax function, and the obtained modulated customer feedback-record semantic association scaling activation matrix is ​​multiplied by the customer feedback-record semantic conditional feature vector to obtain the customer feedback-record semantic joint significant fusion representation vector.

8. The large model-driven customer complaint attribution system according to claim 7, characterized in that: The matching degree estimation value analysis unit is used for: Inputting the customer feedback-record semantic joint significant fusion representation vector into a decoder-based matching evaluation module to obtain a matching estimation value; Based on the comparison between the match degree estimate and a preset threshold, it is determined whether to return a certain record as a match result.

Images