Online customer service intelligent quality inspection system and method based on artificial intelligence

Abstract

The invention relates to the technical field of customer service quality inspection, in particular to an online customer service intelligent quality inspection system and method based on artificial intelligence. And automatically extracting a user demand keyword, an emotion expression keyword and a potential violation term keyword, and generating a structured keyword sequence. Performing emotion analysis on each dialogue round through a Transform model, and accurately outputting a customer emotion classification and an intensity score; and semantic correlation of the context is carried out through a neural network model. And automatically identifying the content of each round of dialogue and counting illegal verbal skills. Furthermore, the customer emotion value, the context coherence score and the occurrence frequency of violation verbal skills are input into a quality inspection scoring model, a comprehensive quality inspection score is automatically calculated, and whether the service is qualified or not is judged according to the comprehensive quality inspection score, so that the dynamic evaluation of the service quality is realized, the quality inspection efficiency is improved, and the customer experience is truly reflected.

Description

Technical Field

[0001] The present invention relates to the technical field of customer service quality inspection, and in particular to an online customer service intelligent quality inspection system and method based on artificial intelligence. Background Art

[0002] Online customer service has become a crucial component for e-commerce platforms to enhance user experience and strengthen their market competitiveness. To ensure customer service quality, e-commerce companies have generally established customer service quality inspection mechanisms to assess the quality of customer service conversations. Traditional customer service quality inspection methods typically rely on manual spot checks, where inspectors randomly check selected customer service conversations based on pre-set specifications or standards, assessing customer service attitude, business compliance, and standardization of conversational techniques. However, manual quality inspection methods present the following prominent issues:

[0003] First, manual spot checks can hardly cover all service conversations, and the quality inspection rate is extremely low, resulting in a large number of potential problems not being discovered in a timely manner, making it difficult to meet the growing business scale and quality management needs. Secondly, manual judgment is highly subjective. Limited by the experience and judgment standards of quality inspectors, there are large differences between different personnel, resulting in insufficient consistency and objectivity in quality inspection results. Thirdly, manual quality inspection is inefficient and cannot achieve dynamic monitoring and rapid feedback on real-time conversations, which affects the timely rectification and closed-loop management of service issues. In addition, with the popularization of new service methods such as intelligent customer service and online text communication, the structure of service conversations has become more complex, and traditional manual quality inspection has become difficult to cope with quality inspection tasks in scenarios such as high frequency, long conversations, and multiple rounds of interaction. Existing technologies have introduced simple automation technologies such as keyword matching and sensitive word detection to improve quality inspection efficiency, but these rule-based and template-based methods still have significant limitations in practical applications. For example, keyword recognition cannot deeply understand the customer's true intentions, and it is difficult to discover implicit service defects or violation risks; sensitive word filtering is difficult to adapt to the changing service language and context, and is prone to misjudgment and omission; in addition, most existing automated quality inspection tools are unable to comprehensively consider emotional changes, contextual logical coherence and business knowledge associations during the service process, resulting in a single dimension of service quality evaluation, which is difficult to truly reflect the customer experience and hinders service quality improvement. Summary of the Invention

[0004] To solve the above problems, the present invention provides an online customer service intelligent quality inspection system and method based on artificial intelligence.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] An artificial intelligence-based online customer service intelligent quality inspection method includes the following steps:

[0007] Get the service conversation text;

[0008] Perform semantic analysis on the service conversation text to obtain keyword sequences;

[0009] Based on the keyword sequence, the Transformer model is used to determine the customer sentiment value in each conversation round, and the neural network model is used to calculate the contextual coherence score of the service conversation text. The number of occurrences of illegal language is calculated based on the keyword sequence.

[0010] The customer sentiment value, context coherence score, and number of instances of illegal language in each conversation round are input into a pre-built quality inspection scoring model to calculate the quality inspection score of the service conversation text.

[0011] The service qualification is determined based on the quality inspection score of the service dialogue text.

[0012] Furthermore, the service conversation text includes several rounds of conversation text data between the customer service staff and the customer, wherein the conversation text data includes the speaker identification, timestamp information and text content; the service conversation text also includes customer order information and customer satisfaction evaluation.

[0013] Furthermore, the semantic parsing of the service conversation text includes the following steps:

[0014] Based on the pre-built knowledge graph recognition model, entity recognition is performed on the service conversation text to obtain initial text data containing entity annotations of product name, order number, service category, and question type.

[0015] In the initial text data, the text sequence is feature-labeled and boundary-determined through a conditional random field model, user demand keywords, emotion expression keywords, and potential illegal terminology keywords are extracted to generate a structured keyword sequence.

[0016] Furthermore, determining the customer sentiment value in each conversation round by using the Transformer model includes the following steps:

[0017] Segmenting the structured keyword sequence according to the dialogue turns to obtain a text sequence grouped by turns;

[0018] The text sequence of each round of dialogue is concatenated with the text of adjacent historical rounds as the input of the Transformer model. The Transformer model, which has been fine-tuned with sentiment-annotated corpus, extracts sentiment features from the input data and outputs the corresponding customer sentiment classification results and sentiment intensity scores.

[0019] Furthermore, the calculating of the context coherence score of the service dialogue text by the neural network model comprises the following steps:

[0020] Based on the service conversation text processed by entity recognition and keyword extraction, each round of customer service response is paired with the customer question in the previous round to form a conversation turn pair dataset;

[0021] The conversation turn pair dataset is input into a semantic matching neural network built based on a bidirectional encoder and a coreference resolution mechanism, and features are extracted and quantified for the semantic correlation between the customer service response and the customer question in each conversation turn, and a contextual coherence score is output.

[0022] Furthermore, the semantic matching neural network based on the bidirectional encoder and the coreference resolution mechanism is constructed by the following steps:

[0023] Based on historical service conversation data, a training sample set containing reference relationship annotations and contextual relevance labels is constructed, and the customer service response and the corresponding customer question in each conversation round are used as input sample pairs;

[0024] Generate context embedding representations for the input sample pairs using a pre-trained language model, perform entity restoration and semantic completion on pronouns and omitted subjects in the text, and obtain a resolved context embedding vector;

[0025] The resolved context embedding vector is input into a bidirectional encoder structure. The encoder performs feature fusion and high-dimensional semantic representation on the input sample pair, and outputs a matching score representing the semantic relevance between the customer service response and the customer question.

[0026] Based on the context-dependent labels of the training samples, the network parameters are optimized through supervised learning.

[0027] Furthermore, the step of calculating the number of occurrences of illegal speech words based on the keyword sequence includes the following steps:

[0028] Based on structured keyword sequences, the service conversation text is input into a pre-built text classification model. The model identifies illegal speech in each round of conversation and outputs the corresponding illegal labels.

[0029] The number of occurrences of illegal words in the service conversation text is counted according to the illegal tags.

[0030] Furthermore, the quality inspection scoring model is constructed by the following steps:

[0031] Based on historical service conversation data, we collect corresponding customer sentiment values, context coherence scores, and the number of times illegal language usage occurs. We then combine this with the customer review data corresponding to each service conversation to construct a training sample set with customer review labels.

[0032] Inputting the training sample set into a scoring regression model, modeling the mapping relationship between customer sentiment value, context coherence score, number of occurrences of illegal speech words and customer evaluation, and training to obtain a quality inspection scoring model;

[0033] Cross-validation is used to evaluate the performance of the scoring model and dynamically adjust the model parameters.

[0034] Furthermore, determining the service eligibility based on the quality inspection score of the service dialogue text includes:

[0035] The quality inspection score of each service dialogue text is compared with the preset service qualification score threshold. If the quality inspection score is higher than or equal to the score threshold, the service dialogue text is judged as a qualified service. If the quality inspection score is lower than the score threshold, the service dialogue text is judged as an unqualified service.

[0036] An AI-based online customer service intelligent quality inspection system, applied to any of the aforementioned AI-based online customer service intelligent quality inspection methods, comprising:

[0037] Text collection module, used to obtain service conversation text;

[0038] Semantic parsing module, used to perform semantic parsing on service conversation text to obtain keyword sequences;

[0039] The multi-dimensional standard calculation module is used to determine the customer sentiment value in each conversation round based on the keyword sequence using the Transformer model, and calculate the contextual coherence score of the service conversation text using the neural network model. The module also calculates the number of occurrences of illegal speech based on the keyword sequence.

[0040] The quality inspection and scoring module is used to input the customer sentiment value, context coherence score, and number of illegal speech patterns in each conversation round into a pre-built quality inspection and scoring model to calculate the quality inspection score of the service conversation text;

[0041] The qualification determination module is used to determine the service qualification based on the quality inspection score of the service dialogue text.

[0042] The beneficial effects of the present invention are:

[0043] This invention captures the full text of service conversations between customer service representatives and customers, performs semantic parsing on the captured service conversations, and further extracts keywords related to user needs, emotional expressions, and potentially offensive terms to generate structured keyword sequences. This processing approach not only improves the ability to understand complex conversation scenarios but also overcomes the bottleneck of traditional rule-based matching, which is poorly adaptable to diverse service contexts. Based on the keyword sequence, a Transformer model performs sentiment analysis on each conversation turn, accurately identifying the customer's true emotional state during the conversation and outputting sentiment classification results and sentiment intensity scores. Simultaneously, a neural network model assesses the contextual coherence of the service conversation texts, quantifying the semantic relevance between customer service responses and customer questions, effectively avoiding the subjective bias and lack of consistency inherent in manual judgment. Furthermore, based on the structured keyword sequence, a text classification model is used to automatically identify and count offensive terms in each conversation turn, promptly identifying and warning of potential service risks. This overcomes the problems of missed and misclassified sensitive word filtering methods in diverse contexts. Finally, multi-dimensional features such as customer sentiment, contextual coherence, and the number of instances of non-compliant language in each conversational round are fed into a pre-built quality inspection scoring model. This automatically calculates a comprehensive quality inspection score for the service conversation text and uses this score to determine whether the service is qualified. This creates an objective, standardized, and dynamic closed-loop evaluation and management of service quality. This overall solution not only improves quality inspection efficiency, reduces manual intervention, but also truly reflects the customer experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a step-by-step flowchart of an intelligent quality inspection method for online customer service based on artificial intelligence.

[0045] Figure 2 It is a flow chart of the construction steps of the semantic matching neural network constructed based on the bidirectional encoder and reference resolution mechanism in the present invention. DETAILED DESCRIPTION

[0046] See also Figure 1-Figure 2 As shown, the present invention relates to an online customer service intelligent quality inspection method based on artificial intelligence, comprising the following steps:

[0047] Get the service conversation text;

[0048] Perform semantic analysis on the service conversation text to obtain keyword sequences;

[0049] Based on the keyword sequence, the Transformer model is used to determine the customer sentiment value in each conversation round, and the neural network model is used to calculate the contextual coherence score of the service conversation text. The number of occurrences of illegal language is calculated based on the keyword sequence.

[0050] The customer sentiment value, context coherence score, and number of instances of illegal language in each conversation round are input into a pre-built quality inspection scoring model to calculate the quality inspection score of the service conversation text.

[0051] The service qualification is determined based on the quality inspection score of the service dialogue text.

[0052] In some embodiments, using an e-commerce platform's customer service center as an example, the system automatically collects and archives all daily online text conversations between customer service representatives and customers, ensuring complete coverage of customer service data throughout their lifecycle. Subsequently, entity recognition and semantic parsing are performed on the collected service conversation texts. Specifically, a named entity recognition model based on the e-commerce knowledge graph is employed to accurately annotate key information such as product names, order numbers, service categories, and question types. The system further uses a conditional random field (CRF) model to perform feature annotation and boundary analysis on the text sequences, automatically extracting keywords related to user needs, emotional expressions, and potentially offensive terms to form structured keyword sequences. This allows for highly adaptable extraction of complex multi-turn conversations and industry-specific terminology, significantly outperforming traditional rule-based or template-based keyword recognition methods. After obtaining the keyword sequences, the system uses a Transformer model to extract sentiment features from each conversation text and its historical context. Supervised fine-tuning is performed on the sentiment-annotated corpus, enabling the model to not only output fine-grained sentiment categories (such as satisfaction, doubt, anger, and anxiety), but also quantify the intensity scores of customer sentiments in each turn, comprehensively reflecting the evolving trajectory of the customer service experience. Compared with the existing automatic quality inspection tools that only judge the positive and negative emotional tendencies, this embodiment has achieved significant innovations in the resolution, domain adaptability and context modeling of customer emotion recognition. At the same time, the system conducts in-depth semantic correlation analysis on customer service replies and the previous round of customer questions by constructing a semantic matching neural network based on a bidirectional encoder and a reference resolution mechanism. The specific process is that the system first restores and completes the pronouns and omitted subjects in each dialogue round, and then inputs the resolved text into the bidirectional encoder structure to obtain a high-dimensional embedding representation, and further calculates the contextual coherence score between the customer service reply and the customer question. This method can effectively identify problems such as context inconsistency, irrelevant answers, and information omissions, significantly improving the quality inspection's ability to understand and judge complex dialogue scenarios, and breaking through the existing problems of inaccuracy and lack of consistency caused by relying on manual subjective judgment or simple sentence vector comparison. To ensure compliance with service regulations, the system uses a text classification model based on extracted structured keyword sequences to detect and count illegal language in each round of conversation text. This system can automatically identify inappropriate language, sensitive word variations, and industry taboos. During the training phase, the model integrates multi-source illegal samples with a dynamically expanded library of banned words, thus addressing the technical difficulties of traditional sensitive word matching, which has poor contextual adaptability and is prone to missed detections. Ultimately, the system inputs the customer sentiment value, contextual coherence score, and number of occurrences of illegal language in each round of service conversation as multi-dimensional features into a pre-trained quality inspection scoring model. This model uses historical customer evaluation data as labels and, through supervised learning, comprehensively evaluates the impact of various features on the customer experience to generate a comprehensive quality inspection score for each service conversation.If the quality inspection score is lower than the threshold set by the platform, the system will automatically mark the service as unqualified and output rectification suggestions, realizing an intelligent closed loop of quality inspection, feedback and rectification.

[0053] Furthermore, the service conversation text includes several rounds of conversation text data between the customer service staff and the customer, wherein the conversation text data includes the speaker identification, timestamp information and text content; the service conversation text also includes customer order information and customer satisfaction evaluation.

[0054] Specifically, the platform integrates text data from all historical and real-time online conversations between customer service personnel and customers. Each service conversation text contains not only the actual speech content, but also the speaker identifier (e.g., "customer," "customer service"), timestamp information (message sent to the second), and the complete text of the conversation. This accurately restores the chronological relationship of the conversation and the details of the interaction between the two parties, facilitating subsequent multi-dimensional analysis such as tracking sentiment changes, analyzing service response timeliness, and identifying responsible parties. Furthermore, the platform structurally associates each service conversation with the corresponding customer order information (e.g., order number, product name, order time, order status, etc.), enabling the system to automatically retrieve and extract business data directly related to the current conversation. For example, when a customer inquires about order logistics progress, after-sales service procedures, or product details, the system can integrate the specific order context to perform intelligent contextual analysis and business process consistency verification. After each service round, the platform automatically collects customer satisfaction ratings, typically including a rating (e.g., a five-star scale), a brief review text, or tags. This subjective customer feedback data is archived along with the service conversation data to serve as training labels for subsequent AI quality inspection models and as a benchmark for calibrating quality inspection results.

[0055] Furthermore, the semantic parsing of the service conversation text includes the following steps:

[0056] Based on the pre-built knowledge graph recognition model, entity recognition is performed on the service conversation text to obtain initial text data containing entity annotations of product name, order number, service category, and question type.

[0057] In the initial text data, the text sequence is feature-labeled and boundary-determined through a conditional random field model, user demand keywords, emotion expression keywords, and potential illegal terminology keywords are extracted to generate a structured keyword sequence.

[0058] In some embodiments, the system first obtains multiple rounds of conversation text between customer service and customers. The system pre-trains a knowledge graph recognition model, built with e-commerce industry characteristics, for entity recognition tasks. This model can automatically detect and annotate core entities strongly related to the business from the conversation text, such as product names, order numbers, service categories, and specific question types. Specifically, the model utilizes standard product and service names, common problem descriptions, and multi-level attribute information stored in the knowledge graph to perform entity recognition and semantic normalization on the input text sequence, outputting structured entity annotation data. This addresses the entity extraction challenges caused by the diverse expressions of the original text and complex industry terminology variants. After completing entity annotation, the system further processes the initial entity annotation text, using a conditional random field (CRF) model to perform sequence annotation and boundary determination on the text sequence. The model inputs the contextual features, part-of-speech tags, previously identified entity categories, and supplementary attributes related to the knowledge graph for each word or subword in the text. Based on the globally optimal annotation path, it outputs a label for the functional role of each word in the sentence (such as "user needs," "emotional expression," "illegal language," etc.). By jointly modeling word sequences using a CRF model, we can effectively capture contextual dependencies and inter-label transition patterns, thereby accurately distinguishing keyword boundaries and categories. For example, if a user expresses, "Why hasn't my order 123456 been shipped yet?", the system can label "order 123456" as the order number entity, "not shipped" as a logistics issue under the service category, and "why still" as a keyword expressing emotion, ultimately outputting a result represented by a structured keyword sequence.

[0059] Furthermore, determining the customer sentiment value in each conversation round by using the Transformer model includes the following steps:

[0060] Segmenting the structured keyword sequence according to the dialogue turns to obtain a text sequence grouped by turns;

[0061] The text sequence of each round of dialogue is concatenated with the text of adjacent historical rounds as the input of the Transformer model. The Transformer model, which has been fine-tuned with sentiment-annotated corpus, extracts sentiment features from the input data and outputs the corresponding customer sentiment classification results and sentiment intensity scores.

[0062] In some embodiments, all keyword sequences are first segmented according to the conversation turns between customer service representatives and customers. Using word embedding methods (such as Word2Vec and BERT embedding), each turn's keyword sequence is converted into a high-dimensional vector, resulting in text sequence vectors grouped by turn. To capture the dynamic changes in customer emotions and the influence of cross-turn context, the algorithm concatenates the text vector of the current conversation turn with the text vectors of a certain number of previous turns to form a composite input representation that contains temporal information and contextual dependencies. This composite vector input passes through a multi-layer Transformer structure. The model uses a self-attention mechanism to calculate correlation weights across all input sequences, automatically learning the global dependencies between information from different turns for emotion recognition. The model has been fine-tuned using a large dataset of emotion-annotated e-commerce customer service conversations, enabling it to accurately identify emotion categories (such as anxiety, satisfaction, anger, and neutrality) and their intensity scores in complex conversation scenarios. The training objective utilizes a joint optimization of a multi-label classification loss and an emotion intensity regression loss to ensure the model outputs fine-grained emotion labels and quantitative intensity estimates. During the actual inference phase, the model automatically extracts emotion-related features from the input vector for each round of conversation and outputs the customer's emotion classification and corresponding emotion intensity score for that round. For example, when a customer expresses the sentiment of "My order still hasn't shipped, and I'm worried the courier is lost" in several consecutive rounds, the model can integrate the content of the current and previous rounds to correctly classify the emotion as "anxiety" and assign a high intensity score, providing direct basis for subsequent risk warnings and service interventions. This method effectively improves the accuracy and industry adaptability of emotion recognition in multi-round conversation scenarios by fusing deep semantics with temporal features.

[0063] Furthermore, the calculating of the context coherence score of the service dialogue text by the neural network model comprises the following steps:

[0064] Based on the service conversation text processed by entity recognition and keyword extraction, each round of customer service response is paired with the customer question in the previous round to form a conversation turn pair dataset;

[0065] The conversation turn pair dataset is input into a semantic matching neural network built based on a bidirectional encoder and a coreference resolution mechanism, and features are extracted and quantified for the semantic correlation between the customer service response and the customer question in each conversation turn, and a contextual coherence score is output.

[0066] In some embodiments, entity recognition and keyword extraction are first performed on all service conversation texts to generate structured text data. The algorithm module pairs each customer service response with the previous customer question, automatically generating a training and inference dataset containing a large number of "question-response" rounds. This operation effectively restores the logical order and semantic dependencies of multiple rounds of conversation, providing a data foundation for subsequent contextual coherence analysis. At the model level, a semantic matching neural network incorporating a coreference resolution mechanism is used as the core architecture. Specifically, for each "customer question-customer response" text pair, the system first automatically identifies and restores pronouns, demonstratives, or omitted subjects based on the industry knowledge graph, lexical rules, and context window, completing implicit entities at the input layer and enhancing semantic alignment. After entity completion, each text pair is fed into a pre-trained bidirectional encoder (such as a BiLSTM or bidirectional Transformer) to generate a contextual embedding vector, capturing the deep semantic features of the sentence in context. The embedding vectors are fused through concatenation or a mutual attention mechanism and further fed into a fully connected layer or semantic similarity calculation module, outputting a continuous-valued contextual coherence score. This score characterizes the responsiveness and logical consistency of the customer service reply to the semantics of the previous customer's question. During the training phase, the contextual consistency labels manually scored or judged by experts in historical annotated samples are used as supervisory signals, and the mean square error loss or correlation measurement is used to optimize the model parameters. In actual reasoning and service quality inspection, the model can identify problems such as "irrelevant answers", "content discontinuity", "information omissions" and so on for complex question-and-answer scenarios, cross-turn references or multiple supplementary expressions, quantify them into coherence scores, and provide high-credibility feature inputs for subsequent comprehensive quality inspection scores. Compared with traditional approaches based on keyword co-occurrence or simple sentence vector distances, this method has significant advantages in context perception, logical reasoning and semantic fine-grained modeling.

[0067] Furthermore, the semantic matching neural network based on the bidirectional encoder and the coreference resolution mechanism is constructed by the following steps:

[0068] Based on historical service conversation data, a training sample set containing reference relationship annotations and contextual relevance labels is constructed, and the customer service response and the corresponding customer question in each conversation round are used as input sample pairs;

[0069] Generate context embedding representations for the input sample pairs using a pre-trained language model, perform entity restoration and semantic completion on pronouns and omitted subjects in the text, and obtain a resolved context embedding vector;

[0070] The resolved context embedding vector is input into a bidirectional encoder structure. The encoder performs feature fusion and high-dimensional semantic representation on the input sample pair, and outputs a matching score representing the semantic relevance between the customer service response and the customer question.

[0071] Based on the context-dependent labels of the training samples, the network parameters are optimized through supervised learning.

[0072] Specifically, the system first uses large-scale service conversation history data to construct a training sample set with reference relationship annotations and context relevance labels. Each training sample consists of a set of "customer question-customer service response" pairs, and domain experts annotate their contextual coherence scores and reference resolution relationships. For example, it clarifies which pronouns (such as "it," "this," "that") or omitted subjects in the conversation refer to which products, orders, or services. In the feature construction phase, the system inputs the customer question and customer service response texts into a language model pre-trained with large-scale general and e-commerce domain-specific corpora to generate a high-dimensional semantic embedding vector containing contextual information. For pronouns, demonstrative words, or omitted subjects in the text, by combining the knowledge graph entity relationships and the context window, reference resolution and entity restoration operations are automatically performed, mapping words such as "it" or "this" back to specific products, services, or question entities, and obtaining the resolved context embedding representation. This step ensures that the model can capture the true semantic links in multi-round conversations without being affected by text omissions or ambiguous expressions. The resolved embedding vectors of the customer question and customer service response are then concatenated pairwise and fed into a bidirectional encoder, a BiLSTM structure. This encoder leverages forward and backward information to simultaneously capture complex semantic connections between sentences, deeply fuses sequence features, and outputs a high-dimensional feature representation that incorporates global information about the conversation. The system then generates a quantitative contextual coherence score for each pair of samples through a matching layer or similarity calculation layer, accurately reflecting the semantic and logical consistency between the customer service response and the customer question. During model training, expert-annotated contextual relevance labels serve as supervisory signals, and network parameters are backpropagated and optimized using a mean squared error loss function or a relevance maximization objective. The resulting model is capable of automatically distinguishing the coherence of responses across diverse real-world service conversations, particularly those with complex interrogation and long-range dependencies. Compared to traditional one-way, shallow sentence vector matching or hand-crafted rules, this algorithm automatically models the entire process of context, interrogation, and conversation flow, effectively enhancing the depth of contextual understanding.

[0073] Furthermore, the step of calculating the number of occurrences of illegal speech words based on the keyword sequence includes the following steps:

[0074] Based on structured keyword sequences, the service conversation text is input into a pre-built text classification model. The model identifies illegal speech in each round of conversation and outputs the corresponding illegal labels.

[0075] The number of occurrences of illegal words in the service conversation text is counted according to the illegal tags.

[0076] Specifically, based on the structured keyword sequences generated in the previous steps, each service conversation text is fed as an independent sample into a pre-built text classification model. This classification model typically utilizes a deep neural network architecture, such as BERT or BiLSTM, and is pre-trained using real historical e-commerce conversations and manually annotated illegal language data. It is capable of learning service violations, sensitive terms, and potential compliance risk expressions from complex language expressions. The model inputs the complete text of each conversation and its structured keyword features. The embedding layer extracts contextual representations, and a multi-layer neural network performs deep modeling of semantic features, business context, and keyword positions. During the inference phase, the model automatically performs multi-class inference on each input text, outputting a label indicating whether illegal language is present and the specific violation type. For example, when expressions such as "Go check the delivery yourself" or "Don't bother me anymore" appear in a conversation, the model can not only identify violations such as "inappropriate service attitude" or "customer shirking responsibility," but also accurately distinguish the compliance of language in different scenarios, such as general inquiries and after-sales complaints. The system aggregates violation labels from each conversation, automatically counting the occurrences of illegal language throughout the entire service conversation, and uses this information to generate quality inspection risk signatures. Compared to traditional, simple matching methods based on static sensitive word or keyword lists, this algorithmic process can adapt to complex scenarios such as expression variations, contextual associations, and semantic transformations, significantly improving the accuracy and practicality of illegal language identification.

[0077] Furthermore, the quality inspection scoring model is constructed by the following steps:

[0078] Based on historical service conversation data, we collect corresponding customer sentiment values, context coherence scores, and the number of times illegal language usage occurs. We then combine this with the customer review data corresponding to each service conversation to construct a training sample set with customer review labels.

[0079] Inputting the training sample set into a scoring regression model, modeling the mapping relationship between customer sentiment value, context coherence score, number of occurrences of illegal speech words and customer evaluation, and training to obtain a quality inspection scoring model;

[0080] Cross-validation is used to evaluate the performance of the scoring model and dynamically adjust the model parameters.

[0081] In some embodiments, historical service conversation data is first compiled. For each conversation, multi-dimensional quality features, such as customer sentiment, contextual coherence scores, and the number of instances of non-compliant language, are calculated. Post-service customer satisfaction ratings, scores, and opinion tags are then structurally associated with these features to form a high-quality training sample set. This training set covers a wide range of customer service scenarios and rating distributions, helping to improve the model's generalization capabilities. During the quality inspection and scoring model construction phase, the system uses a gradient boosting tree as the core modeling tool. The input features of the training sample set are customer sentiment, contextual coherence scores, and the number of instances of non-compliant language, while the target output is a customer evaluation label or score. The model is trained by minimizing a loss function, such as the mean squared error (MSE), between predicted scores and actual customer evaluations. It automatically learns the weights and nonlinear relationships between each quality feature and the customer's subjective evaluation, achieving a multi-dimensional quantitative mapping of service quality. During training, algorithmic optimization techniques such as feature normalization, regularization constraints, and feature importance analysis can be employed to improve model stability and interpretability. To ensure the robust generalization capabilities of the scoring model, the system employs K-fold cross-validation in the modeling process. This method divides the training set into multiple subsets, looping through training and validation to evaluate the model's fit and error distribution under varying data distributions. The optimal model parameters are selected through cross-validation, and regular dynamic fine-tuning is conducted based on actual business feedback. This ensures that the quality inspection scoring model remains highly sensitive and adaptable to customer experience and service risks as the business evolves. Ultimately, the trained scoring model automatically outputs an objective, traceable comprehensive service quality score for any new service conversation, providing a scientific basis for service qualification assessment and refined management.

[0082] Furthermore, determining the service eligibility based on the quality inspection score of the service dialogue text includes:

[0083] The quality inspection score of each service dialogue text is compared with the preset service qualification score threshold. If the quality inspection score is higher than or equal to the score threshold, the service dialogue text is judged as a qualified service. If the quality inspection score is lower than the score threshold, the service dialogue text is judged as an unqualified service.

[0084] Specifically, the system has pre-set service qualification score thresholds by quality inspectors based on historical quality inspection data and business needs. Whenever a new service conversation quality inspection score is generated, the system will automatically and strictly compare the score with the preset threshold. If the score is higher than or equal to the threshold, the system automatically marks the conversation as a qualified service and enters the normal archiving and operational data analysis process. If the score is lower than the threshold, the system will mark the conversation as an unqualified service, and can simultaneously trigger an early warning, record low-scoring feature items in detail, output corresponding rectification suggestions, or push them to dedicated quality inspectors for review. The judgment process is fully automated to ensure the efficiency and standardization of service quality screening, realize real-time service quality control of batch and full conversations, and provide high-reliability basic data support for subsequent refined operations and service improvements. It significantly reduces manual intervention links and improves the overall service compliance and customer experience assurance capabilities of the e-commerce platform.

[0085] The present invention also includes an artificial intelligence-based online customer service intelligent quality inspection system, which is applied to any of the aforementioned artificial intelligence-based online customer service intelligent quality inspection methods, including:

[0086] Text collection module, used to obtain service conversation text;

[0087] Semantic parsing module, used to perform semantic parsing on service conversation text to obtain keyword sequences;

[0088] The multi-dimensional standard calculation module is used to determine the customer sentiment value in each conversation round based on the keyword sequence using the Transformer model, and calculate the contextual coherence score of the service conversation text using the neural network model. The module also calculates the number of occurrences of illegal speech based on the keyword sequence.

[0089] The quality inspection and scoring module is used to input the customer sentiment value, context coherence score, and number of illegal speech patterns in each conversation round into a pre-built quality inspection and scoring model to calculate the quality inspection score of the service conversation text;

[0090] The qualification determination module is used to determine the service qualification based on the quality inspection score of the service dialogue text.

[0091] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. An artificial intelligence-based online customer service intelligent quality inspection method, characterized in that: The following steps are involved: Get the service conversation text; Perform semantic analysis on the service conversation text to obtain keyword sequences; Based on the keyword sequence, the Transformer model is used to determine the customer sentiment value in each conversation round, and the contextual coherence score of the service conversation text is calculated using a semantic matching neural network. The number of occurrences of illegal language is also calculated based on the keyword sequence. The customer sentiment value, contextual coherence score, and number of instances of illegal language in each conversation round are input into a pre-built quality inspection scoring model to calculate the quality inspection score of the service conversation text. The customer sentiment value, contextual coherence score, and number of instances of illegal language all include corresponding labels. The quality inspection scoring model uses a scoring regression model to model the relationship between the quality inspection score and the labels. The service qualification is determined based on the quality inspection score of the service dialogue text.

2. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 1, characterized in that: The service conversation text includes several rounds of conversation text data between the customer service staff and the customer, wherein the conversation text data includes the speaker identification, timestamp information and text content; the service conversation text also includes customer order information and customer satisfaction evaluation.

3. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 1, characterized in that: The semantic parsing of the service conversation text comprises the following steps: Based on the pre-built knowledge graph recognition model, entity recognition is performed on the service conversation text to obtain initial text data containing entity annotations of product name, order number, service category, and question type. In the initial text data, the text sequence is feature-labeled and boundary-determined through a conditional random field model, user demand keywords, emotion expression keywords, and potential illegal terminology keywords are extracted to generate a structured keyword sequence.

4. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 3, characterized in that: Determining the customer sentiment value in each conversation round using the Transformer model includes the following steps: Segmenting and vectorizing the structured keyword sequence according to the dialogue turns to obtain text sequence vectors grouped by turns; The text sequence vector of each round of dialogue is concatenated with the text vectors of adjacent historical rounds as the input of the Transformer model. The Transformer model, which has been fine-tuned with sentiment-annotated corpus, extracts sentiment features from the input data and outputs the corresponding customer sentiment classification results and sentiment intensity scores.

5. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 1, characterized in that: Calculating the context coherence score of the service conversation text by using a semantic matching neural network includes the following steps: Based on the service conversation text processed by entity recognition and keyword extraction, each round of customer service response is paired with the customer question in the previous round to form a conversation turn pair dataset; The conversation turn pair dataset is input into a semantic matching neural network built based on a bidirectional encoder and a coreference resolution mechanism, and features are extracted and quantified for the semantic correlation between the customer service response and the customer question in each conversation turn, and a contextual coherence score is output.

6. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 5, characterized in that: The semantic matching neural network based on the bidirectional encoder and the coreference resolution mechanism is constructed by the following steps: Based on historical service conversation data, a training sample set containing reference relationship annotations and contextual relevance labels is constructed, and the customer service response and the corresponding customer question in each conversation round are used as input sample pairs; Generate context embedding representations for the input sample pairs using a pre-trained language model, perform entity restoration and semantic completion on pronouns and omitted subjects in the text, and obtain a resolved context embedding vector; The resolved context embedding vector is input into a bidirectional encoder structure. The encoder performs feature fusion and high-dimensional semantic representation on the input sample pair, and outputs a matching score representing the semantic relevance between the customer service response and the customer question. Based on the context-dependent labels of the training samples, the network parameters are optimized through supervised learning.

7. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 3, characterized in that: Calculating the number of occurrences of illegal words based on the keyword sequence includes the following steps: Based on structured keyword sequences, the service conversation text is input into a pre-built text classification model. The model identifies illegal speech in each round of conversation and outputs the corresponding illegal labels. The number of occurrences of illegal words in the service conversation text is counted according to the illegal tags.

8. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 1, characterized in that: The quality inspection scoring model is constructed by the following steps: Based on historical service conversation data, we collect corresponding customer sentiment values, contextual coherence scores, and the number of times illegal language usage occurs. We then combine this with the customer review data corresponding to each service conversation to construct a training sample set with customer review labels. Inputting the training sample set into a scoring regression model, modeling the mapping relationship between customer sentiment value, context coherence score, number of occurrences of illegal speech words and customer evaluation, and training to obtain a quality inspection scoring model; Cross-validation is used to evaluate the performance of the scoring model and dynamically adjust the model parameters.

9. The method for intelligent quality inspection of online customer service based on artificial intelligence according to claim 1, characterized in that: The determination of service eligibility based on the quality inspection score of the service dialogue text includes: The quality inspection score of each service dialogue text is compared with the preset service qualification score threshold. If the quality inspection score is higher than or equal to the score threshold, the service dialogue text is judged as a qualified service. If the quality inspection score is lower than the score threshold, the service dialogue text is judged as an unqualified service.

10. An artificial intelligence-based online customer service intelligent quality inspection system, applied to an artificial intelligence-based online customer service intelligent quality inspection method according to any one of claims 1 to 9, characterized in that: include: Text collection module, used to obtain service conversation text; Semantic parsing module, used to perform semantic parsing on service conversation text to obtain keyword sequences; The multi-dimensional standard calculation module is used to determine the customer sentiment value in each conversation round based on the keyword sequence using the Transformer model, and calculate the contextual coherence score of the service conversation text using the neural network model. The module also calculates the number of occurrences of illegal speech based on the keyword sequence. The quality inspection and scoring module is used to input the customer sentiment value, context coherence score, and number of illegal speech patterns in each conversation round into a pre-built quality inspection and scoring model to calculate the quality inspection score of the service conversation text; The qualification determination module is used to determine the service qualification based on the quality inspection score of the service dialogue text.

Images