NLP-based customer service dialogue quality detection method and system

Abstract

The invention provides an NLP-based customer service dialogue quality detection method and system, and relates to the technical field of natural language processing, and the method comprises the steps: obtaining an e-commerce platform customer service dialogue text stream in real time; semantic understanding is conducted on the dialogue text flow, a semantic understanding result is output, and the semantic understanding result comprises a user consultation intention, key question elements and customer service inquiry information entity integrity; based on the semantic understanding result, performing sentiment analysis on the dialogue text stream, and outputting a structured sentiment analysis result; setting an initial state point and a termination state point in the dialogue processing path based on the semantic understanding result and the sentiment analysis result; the initial state point is a semantic vector fusing core problem elements fed back by the user for the first time and a current user emotional state value. According to the invention, through full-process intelligent processing, dynamic monitoring and accurate optimization of customer service quality are realized, and user satisfaction, customer service efficiency and business normalization are improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for detecting the quality of customer service dialogues based on NLP. Background Technology

[0002] With the explosive growth of e-commerce and internet finance, enterprise customer service systems have become the core hub for user interaction. However, current e-commerce platforms handle tens of millions of customer service conversations daily, and three major development trends are emerging:

[0003] The exponential growth in data volume: The massive accumulation of unstructured dialogue text has resulted in the coverage rate of traditional manual sampling methods being less than 0.1% in some cases, which may be insufficient to fully reflect service quality.

[0004] The increasing complexity and personalization of user needs: the proliferation of colloquial expressions, dialect vocabulary, and multi-turn interaction scenarios may increase the difficulty of understanding user needs;

[0005] Service quality requirements are being upgraded: users have increasingly higher demands for response efficiency and emotional experience, and traditional keyword-based evaluation methods may no longer be able to meet the needs of refined management.

[0006] Traditional methods rely on keyword matching, which may fail to handle semantic variations and result in low accuracy in intent recognition. Some lack domain knowledge support, making it difficult to analyze complex business logic and potentially leading to the omission of key issue elements. Others assess emotions through the frequency of sentiment words, which may overlook transitional logic, resulting in high error rates in emotion judgment. Some methods fail to correlate customer service reassurance behaviors with changes in user emotions, making it difficult to quantify the effectiveness of reassurance. Some methods use business process compliance as a single assessment dimension, which may ignore information completeness and the effectiveness of emotional reassurance, leading to a disconnect between assessment results and the actual user experience. Static assessment standards may not be suitable for dynamic scenarios and lack real-time correction mechanisms. Some dialogue data is unstructured, making it difficult to support multi-dimensional analysis and thus lacking real-time monitoring and early warning capabilities, potentially causing problem discovery to lag behind user complaints. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a customer service dialogue quality detection method and system based on NLP. Through intelligent processing of the entire process, it can realize dynamic monitoring and precise optimization of customer service quality, thereby improving user satisfaction, customer service efficiency and business standardization.

[0008] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0009] Firstly, an NLP-based method for assessing the quality of customer service conversations, the method comprising:

[0010] Step S1: Obtain the real-time text stream of customer service conversations on the e-commerce platform;

[0011] Step S2: Perform semantic understanding on the dialogue text stream and output the semantic understanding results, which include: user consultation intent, key question elements, and completeness of customer service inquiry information entities;

[0012] Step S3: Based on the semantic understanding results, perform sentiment analysis on the dialogue text stream and output the structured sentiment analysis results;

[0013] Step S4: Based on the semantic understanding results and sentiment analysis results, set a starting state point and an ending state point in the dialogue processing path; the starting state point is the semantic vector of the core question elements of the user's first feedback and the current user's emotional state value; the ending state point is the semantic vector of the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold; connect the starting state point and the ending state point to form a reference straight line vector as the expected processing path; map the real-time matched customer service solution to the current trajectory vector; calculate the directional deviation angle and magnitude difference value of the current trajectory vector relative to the reference straight line vector, and generate a dynamic correction coefficient.

[0014] Step S5: Based on the dynamic correction coefficient, perform weighted calibration on the business specification compliance score to generate quality assessment indicators;

[0015] Step S6: Based on the quality assessment indicators, perform service quality monitoring and assessment data storage.

[0016] Further, step S1: Real-time acquisition of the e-commerce platform customer service dialogue text stream, including:

[0017] Real-time access to the dialogue data interface of the e-commerce platform's customer service system to obtain a mixed text stream containing user inquiry messages and customer service reply messages;

[0018] The acquired mixed text stream is segmented into sentences and labeled with roles in real time, and a structured dialogue sequence is output; each sentence in the structured dialogue sequence is labeled with the speaking role and timestamp;

[0019] The system removes automatically generated advertising text from the structured dialogue sequence, standardizes colloquial abbreviations and dialect words, filters non-textual symbols, and outputs a dialogue text stream.

[0020] Further, step S2: Perform semantic understanding on the dialogue text stream and output the semantic understanding result, which includes: user inquiry intent, key question elements, and the completeness of customer service inquiry information entities, including:

[0021] Extract user statements from the dialogue text stream, parse user intent based on domain knowledge graph, output user consultation intent category, and activate the corresponding structured extraction template according to intent category;

[0022] Based on structured extraction templates, nodes matching the template are located in the dependency tree, and key issue element values ​​are extracted.

[0023] Based on the key issue element values, the rule base is called to generate a list of required inquiry entities. That is, each key issue element value triggers a set of supplementary entities that customer service must confirm. Named entity recognition is performed on the customer service reply text to extract the customer service entity set. The extracted customer service entity set is compared with the list of required inquiry entities, and a boolean flag for entity integrity is output.

[0024] Integrate user consultation intent, key question element values, and entity integrity Boolean flags to output structured semantic understanding results.

[0025] Further, step S3: Based on the semantic understanding results, perform sentiment analysis on the dialogue text stream and output structured sentiment analysis results, including:

[0026] Based on the user's consultation intent category and key question element values ​​from the semantic understanding results, the user's statements are extracted from the dialogue text stream. Then, based on the timestamp of the dialogue text stream, the complete statement of the user's last feedback question is obtained. An e-commerce domain sentiment dictionary is loaded to match sentiment words in the complete statement of the user's last feedback question and obtain basic sentiment scores. Then, these sentiment word scores are calculated by weighting. The transition conjunctions in the complete statement of the user's last feedback question are detected and the statement is segmented into sub-units. The calculated sentiment word scores are decayed or reversed according to the order of the sub-units to generate the user's emotional state value.

[0027] Based on the user's emotional state value, emotion adjustment keywords are extracted from the customer service reply text. When at least one keyword is detected, the customer service representative is marked as having an intention to soothe the user. After the user's reply is marked as having an intention to soothe the user, the emotional state value is calculated again for the next reply to generate a new emotional state value. Based on the new emotional state value and the previous user emotional state value, an emotional change value is calculated. It is determined whether the emotional change value exceeds the threshold. If it does, and the customer service representative has an intention to soothe the user, the soothing is marked as effective, and a soothing effect mark is obtained.

[0028] Integrate user emotional state values ​​and soothing effect markers to output structured sentiment analysis results.

[0029] Further, step S4: Based on the semantic understanding results and sentiment analysis results, set a starting state point and an ending state point in the dialogue processing path; the starting state point is the semantic vector of the core question elements fused with the user's initial feedback and the current user's emotional state value; the ending state point is the semantic vector of the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold; connect the starting state point and the ending state point to form a reference straight line vector as the expected processing path; map the real-time matched customer service solution to the current trajectory vector; calculate the directional deviation angle and magnitude difference value of the current trajectory vector relative to the reference straight line vector, and generate dynamic correction coefficients, including:

[0030] Based on semantic understanding results, the core problem elements of the user's initial feedback are extracted. These core problem elements are processed and transformed into semantic vector representations to obtain the problem semantic vector. Sentiment analysis results are input to extract the current user's emotional state value. The problem semantic vector and the current user emotional state value are fused to generate an initial state vector. The initial state vector is input into the platform's business rule base to obtain a standardized solution matching the current problem type. The standardized solution is processed and transformed into a semantic vector representation to obtain a solution semantic vector. Sentiment analysis association rules are input to extract the target user's emotional threshold. The solution semantic vector and the target user's emotional threshold are fused to generate a termination state vector.

[0031] Connect the initial state vector and the final state vector to generate the reference line vector;

[0032] Based on the resolution semantic vector in the termination state vector, a semantic transformation benchmark is used; the real-time customer service solution is mapped to the same semantic space, and the current trajectory vector is output.

[0033] Based on the reference line vector and the current trajectory vector, calculate the directional deviation angle between the two vectors: calculate the angle deviation value using the cosine theorem of the vector angle; calculate the difference in magnitude between the two vectors: take the absolute difference between the magnitude of the reference line vector and the magnitude of the current trajectory vector; generate a dynamic correction coefficient by combining the directional deviation angle and the difference in magnitude.

[0034] Further, step S5: Based on the dynamic correction coefficient, the business specification compliance score is weighted and calibrated to generate quality assessment indicators, including:

[0035] Input a dynamic correction coefficient, which includes the path deviation value, to obtain an initial score for real-time matching of business specification compliance.

[0036] Based on the path deviation value in the dynamic correction coefficient, when the path deviation is ≤5%, the initial score of business specification compliance remains unchanged. When the path deviation is >5%, the score deduction ratio is calculated. The compensation factor is obtained by subtracting the deduction ratio from 1. The initial score of business specification compliance is multiplied by the compensation factor to obtain the calibrated business specification compliance score. If the calibrated business specification compliance score is lower than 60 points, it is forcibly corrected to 60 points.

[0037] Input the calibration business specification compliance score, the information inquiry integrity score converted from the entity integrity Boolean mark, and the emotional reassurance effectiveness score converted from the reassurance effect mark. Calculate the weighted values ​​of the calibration business specification compliance score multiplied by 50%, the information inquiry integrity score multiplied by 30%, and the emotional reassurance effectiveness score multiplied by 20%, respectively, according to the preset weights of 50% for business specification compliance, 30% for information integrity, and 20% for emotional reassurance. Then, add the three together and output the total dialogue quality score.

[0038] Quality assessment indicators are generated based on the calibration scores for business compliance, information inquiry completeness, emotional reassurance effectiveness, and total dialogue quality.

[0039] Further, step S6: Based on quality assessment indicators, perform service quality monitoring and assessment data storage, including:

[0040] The total score of the conversation quality is extracted from the quality assessment indicators. When the total score is ≥80, it is marked as excellent service and the green light access command is activated; when the total score is ≤60 and <80, it is marked as qualified service and the yellow light monitoring command is activated; when the total score is <60, it is marked as poor service and the red light alarm command is activated. The service marking results are obtained. For conversations marked as poor service, a conversation freeze command is sent to the customer service system.

[0041] Based on service tagging results and quality assessment indicators, standardized storage records are generated, including: a unique dialogue identifier, a calibrated business specification compliance score, an information inquiry completeness score, an emotional reassurance effectiveness score, a total dialogue quality score, and service tagging results; the storage records are then partitioned by time and written into a quality inspection result table in a distributed database.

[0042] Based on the quality inspection result table data, the quality inspection result table is queried in groups by customer service employee number to calculate the occurrence rate of poor service for each customer service representative. The quality inspection result table is also queried in groups by problem type to count the number of cases with an emotional reassurance effectiveness score of <60. The data on the occurrence rate of poor service and the failure of reassurance cases are integrated to generate a management analysis report containing trend charts, and the report is pushed to the customer service management platform.

[0043] Secondly, an NLP-based customer service dialogue quality inspection system includes:

[0044] The acquisition module is used to acquire the real-time text stream of customer service conversations on e-commerce platforms.

[0045] The semantic understanding module is used to perform semantic understanding on the dialogue text stream and output the semantic understanding results, which include: user consultation intent, key question elements, and the completeness of customer service inquiry information entities.

[0046] The sentiment analysis module is used to perform sentiment analysis on the dialogue text stream based on the semantic understanding results and output structured sentiment analysis results.

[0047] The dynamic correction module is used to set a starting state point and an ending state point in the dialogue processing path based on semantic understanding results and sentiment analysis results. The starting state point is a semantic vector that integrates the core question elements of the user's initial feedback and the current user's emotional state value. The ending state point is a semantic vector that integrates the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold. The starting state point and the ending state point are connected to form a reference straight line vector, which serves as the expected processing path. The real-time matched customer service solution is mapped to the current trajectory vector. The deviation angle and magnitude difference of the current trajectory vector relative to the reference straight line vector are calculated to generate dynamic correction coefficients.

[0048] The calibration module is used to perform weighted calibration of the business specification compliance score based on dynamic correction coefficients, and generate quality assessment indicators.

[0049] The storage module is used to perform service quality monitoring and evaluation data storage based on quality assessment indicators.

[0050] Thirdly, a computing device, comprising:

[0051] one or more processors;

[0052] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0053] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0054] The above-described solution of the present invention has at least the following beneficial effects:

[0055] The system acquires real-time data to ensure data timeliness and purity, and enables structured dialogue; the semantic understanding component accurately deconstructs user needs and verifies information integrity; the sentiment analysis unit dynamically tracks user emotions and provides early warnings of service risks; the dynamic correction process uses vector deviation analysis to achieve intelligent service path planning; the calibration system uses dynamic weighting and lower limit protection to achieve multi-dimensional and accurate quality assessment; and the storage module supports real-time monitoring, automated intervention, and data-driven management decisions. Overall, this technology improves demand response efficiency and emotional service levels at the user experience level, and optimizes customer service efficiency, reduces costs, and strengthens risk control at the enterprise operation level. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a customer service dialogue quality detection method based on NLP, provided by an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram of a customer service dialogue quality detection system based on NLP, provided by an embodiment of the present invention. Detailed Implementation

[0058] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0059] like Figure 1 As shown, an embodiment of the present invention proposes a customer service dialogue quality detection method based on NLP, the method comprising the following steps:

[0060] Step S1: Obtain the real-time text stream of customer service conversations on the e-commerce platform;

[0061] Step S2: Perform semantic understanding on the dialogue text stream and output the semantic understanding results, which include: user consultation intent, key question elements, and completeness of customer service inquiry information entities;

[0062] Step S3: Based on the semantic understanding results, perform sentiment analysis on the dialogue text stream and output the structured sentiment analysis results;

[0063] Step S4: Based on the semantic understanding results and sentiment analysis results, set a starting state point and an ending state point in the dialogue processing path; the starting state point is the semantic vector of the core question elements of the user's first feedback and the current user's emotional state value; the ending state point is the semantic vector of the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold; connect the starting state point and the ending state point to form a reference straight line vector as the expected processing path; map the real-time matched customer service solution to the current trajectory vector; calculate the directional deviation angle and magnitude difference value of the current trajectory vector relative to the reference straight line vector, and generate a dynamic correction coefficient.

[0064] Step S5: Based on the dynamic correction coefficient, perform weighted calibration on the business specification compliance score to generate quality assessment indicators;

[0065] Step S6: Based on the quality assessment indicators, perform service quality monitoring and assessment data storage.

[0066] In this embodiment of the invention, dialogue data is captured in real time to ensure that subsequent analysis is based on the latest interaction information, avoiding service response deviations caused by lag. The complete recording of the dialogue text flow provides a comprehensive data foundation for multi-dimensional analysis. By identifying consultation intent and key issue elements, customer service can quickly grasp the core needs of users, reduce communication costs, and improve problem-solving efficiency. User emotions are quantified to assist customer service in adjusting communication strategies, promptly soothing negative emotions, and improving the service experience. Using a baseline straight line vector as a reference, and adapting to real-time dialogue scenarios through dynamic correction coefficients, the service ensures both compliance with business rules and flexibility. By calculating the difference between the direction deviation angle and the modulus, the execution deviation of customer service solutions is transformed into quantifiable indicators. Automated calibration of service paths is achieved through vector mapping and dynamic calculation, improving processing efficiency. Quality scores are weighted based on real-time dialogue scenarios, making the evaluation results more closely reflect actual service effects. Evaluation standards are adjusted differently for different problem types and user emotional states, improving the accuracy of quality monitoring. Through real-time monitoring and data storage, a management closed loop of "analysis-evaluation-optimization" is formed, providing data support for long-term service improvement.

[0067] In a preferred embodiment of the present invention, step S1: acquiring the e-commerce platform customer service dialogue text stream in real time may include:

[0068] Step S1.1: Access the dialogue data interface of the e-commerce platform's customer service system in real time to obtain a mixed text stream containing user inquiry messages and customer service reply messages;

[0069] Step S1.2: Perform real-time sentence segmentation and role labeling on the acquired mixed text stream, and output a structured dialogue sequence; each sentence in the structured dialogue sequence is labeled with the speaking role and timestamp;

[0070] Step S1.3: Remove the automatically generated advertising text from the structured dialogue sequence, standardize colloquial abbreviations and dialect words, filter non-textual content symbols, and output the dialogue text stream.

[0071] In this embodiment of the invention, the complete two-way dialogue content between the user and customer service is acquired, providing raw data support for subsequent semantic analysis and sentiment recognition; unstructured text is transformed into a three-dimensional structured sequence of role-statement-time, facilitating subsequent analysis by role dimension; timestamps and role annotations make the dialogue process traceable, supporting the reconstruction of the interaction process along the timeline, providing a temporal basis for service quality assessment; irrelevant information such as advertisements and symbols are removed to avoid non-business content interfering with the accuracy of semantic analysis; after standardizing spoken language and dialects, the dialogue text meets the input requirements of a unified semantic analysis model, reducing model recognition errors; the clean text stream reduces the amount of subsequent processing computation and improves the overall process speed.

[0072] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0073] In step S1.1 above, a long-lived data interface connection is established with the e-commerce platform's customer service system via network protocols (such as HTTP / HTTPS, WebSocket), continuously monitoring interface data pushes or actively polling to obtain dialogue data; the received data stream (such as JSON, XML format) is parsed to extract metadata such as the text content of user inquiry messages and customer service reply messages, sender identifiers (user ID / customer service ID), and message type; the parsed mixed text stream is stored in a temporary buffer, and data integrity is verified (such as text content not being empty, timestamp format being correct), and invalid data fragments are filtered out.

[0074] Step S1.2 above involves segmenting the mixed text stream based on punctuation marks (periods, question marks, exclamation marks, etc.) and semantic sentence segmentation rules (such as short sentence segmentation models), breaking down continuous text into independent sentence units; marking each sentence with a role attribute ("user" or "customer service") based on the sender identifier (user ID / customer service ID); if the interface does not directly provide role information, the role is inferred through preset rules (such as message prefixes and dialogue context logic); and attaching a system timestamp accurate to milliseconds to each sentence. If the interface returns a timestamp, it is standardized and calibrated (such as time zone unification) to ensure the accuracy of the time sequence of the dialogue.

[0075] In the above step S1.3, the advertisement texts automatically sent by the system are matched through a preset keyword library (such as "click to purchase", "time-limited discount") or regular expressions (such as promotional links, event pushes), and are removed from the conversation sequence; based on an abbreviation-standard vocabulary mapping dictionary (such as "dear" → "hello", "what to do" → "how to handle") and dialect-Mandarin conversion rules (such as "nong" → "you"), non-standard expressions are converted into standard texts; regular expressions are used to filter out emoji, special characters (such as @#$), garbled codes, and format markers (such as HTML tags), and only the pure text content is retained.

[0076] In a preferred embodiment of the present invention, the above step S2: performs semantic understanding on the conversation text stream and outputs a semantic understanding result, where the semantic understanding result includes: the user's consultation intention, key problem elements, and the entity integrity of the customer service's inquiry information, which may include:

[0077] Step S2.1, extracts user statements from the conversation text stream, analyzes the user's intention based on the domain knowledge graph, outputs the category of the user's consultation intention, and activates the corresponding structured extraction template according to the intention category;

[0078] Step S2.2, based on the structured extraction template, locates the nodes matching the template in the dependency tree and extracts the values of the key problem elements;

[0079] Step S2.3, based on the values of the key problem elements, calls the rule library to generate a list of necessary inquiry entities, that is, each value of the key problem element triggers a set of supplementary entity sets that the customer service must confirm; performs named entity recognition on the customer service's reply text, extracts the customer service entity set, compares the extracted customer service entity set with the list of necessary inquiry entities, and outputs an entity integrity boolean flag;

[0080] Step S2.4, integrates the user's consultation intention, the values of the key problem elements, and the entity integrity boolean flag, and outputs a structured semantic understanding result.

[0081] In the embodiment of the present invention, the knowledge graph combines the semantic relationships in the e-commerce field to avoid generalization understanding, automatically matches a dedicated extraction template according to the intention, ensures that the subsequent element extraction focuses on the core problem, and reduces the interference of irrelevant information; the dependency tree disassembles the sentence from the grammatical level to ensure the accuracy of the element extraction; converts natural language into a key-value pair form for subsequent rule matching and quantitative analysis; forcibly validates the necessary entities through the rule library to avoid the customer service omitting key information due to negligence; the entity integrity verification is used as a standardized process to ensure that the customer service response complies with the business rule requirements and reduces the risk of subsequent disputes; integrates three core types of information, namely intention, elements, and integrity, to provide multi-dimensional data support for the subsequent service quality evaluation.

[0082] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0083] In step S2.1 above, key entity categories are sorted out from the e-commerce business, such as "user", "product", "order", "logistics" and "after-sales". Each entity contains detailed attributes (such as the "product" entity containing attribute nodes such as "model", "color", "price" and "specification"); historical customer service dialogue data, e-commerce platform product database, order system fields, after-sales rule documents, etc. are integrated to extract entity instances (such as "product model = iPhone15" "order status = shipped").

[0084] Establish logical relationships between entities, for example:

[0085] The "ownership relationship" between "users" and "orders" (user A owns order B);

[0086] The "triggering relationship" between "products" and "after-sales service" (product quality issues trigger return requests);

[0087] The "status association" between "logistics" and "order" (order logistics status = in transit).

[0088] Construct an intent classification tree, such as "Consultation Intent" → "Product Inquiry" → "Specification Inquiry" → "Size Inquiry", forming hierarchical nodes from abstract to concrete.

[0089] Entity nodes and relationship edges are stored using graph databases such as Neo4j, with each node having an attribute label attached (e.g., "Intent node: Product Inquiry" labeled with keywords such as "size", "color", and "material"). Entity relationships are updated regularly based on new business scenarios (e.g., new product launches, promotional activities), and graph nodes are supplemented through manual annotation and semi-supervised learning.

[0090] The user's statement (e.g., "What is the chest measurement of this garment?") is segmented to extract the keywords "garment", "chest measurement", and "how much". These are then matched against the "specifications" attribute node of the "product" entity in the graph, as well as the "size inquiry" sub-node under "inquiry intent". The thesaurus in the graph (e.g., "chest measurement" → "size" → "specifications") is used to expand the query scope and avoid missed matches due to variations in user statements (e.g., "garment size" is equivalent to "size inquiry"). If a direct match fails, the query backtracks to higher-level nodes (e.g., if a user asks "Where is the package?" and "tracking number query" is not matched, then the parent node "logistics progress inquiry" is matched) to ensure comprehensive intent recognition. Through the hierarchical relationship of the graph, keywords are aggregated upwards from the bottom layer to determine the final intent category. For example:

[0091] Keywords "mobile phone" → "product" entity;

[0092] Keyword "Battery life" → "Specification" attribute;

[0093] The combination is mapped to the intent node "Product consultation → Specification consultation → Battery life consultation".

[0094] The template library is indexed by intent nodes, and each intent corresponds to an exclusive template. For example:

[0095] The "Logistics progress query" template contains the required fields "Order number" and "Logistics company";

[0096] The "Return application" template contains fields such as "Order number", "Return reason", and "Product status".

[0097] Each template field is marked with "Required / Optional", "Data type" (e.g., the order number is a string), and "Semantic hint" (e.g., "Return reason" needs to be selected from the preset options).

[0098] A one-to-one mapping relationship between intent nodes and templates is established and stored in a configuration file (such as JSON format). When the knowledge graph analyzes the intent category (such as "Product size consultation"), the system directly retrieves the corresponding template (such as "Product specification extraction template").

[0099] If the user's intent contains multiple sub-categories (such as consulting "Product price" and "Logistics timeliness" simultaneously), the system supports parallel activation of multiple templates and merges the fields to form a composite extraction structure to ensure that multi-dimensional information is extracted without omission.

[0100] For the above step S2.2, taking the user's statement "What is the battery capacity of this mobile phone" as an example:

[0101] The sentence is split into independent words, and the word types are marked (such as nouns, verbs, pronouns, etc.).

[0102] Word segmentation result: "This" (pronoun), "mobile phone" (noun), "of" (particle), "battery" (noun), "capacity" (noun), "is" (verb), "how much" (interrogative pronoun);

[0103] Grammatical role recognition: Analyze the dependency relationship between words to determine the subject-predicate-object, attributive-adverbial-complement, etc. components;

[0104] Subject: "The battery capacity of this mobile phone" (a modified structure connected by "of", the core word is "capacity");

[0105] Predicate: "is";

[0106] Object: "how much".

[0107] Tree structure generation: Taking the predicate "is" as the root node and the subject and object as the child nodes, a hierarchical tree is constructed.

[0108] Sub - node refinement: "This" is the attributive of "mobile phone", and "battery" is the attributive of "capacity", forming a nested structure of "attributive - headword".

[0109] Each node in the tree contains words, word - classes, and grammatical roles (such as "subject", "object", "attributive"), and the edges represent dependency relationships (such as "attributive - headword relationship", "verb - object relationship").

[0110] Example tree structure:

[0111] is (predicate)

[0112] ├─ The battery capacity of this mobile phone (subject)

[0113] │├─ This (attributive)

[0114] │├─ Mobile phone (headword, object of the attributive modification)

[0115] │├─ of (auxiliary word)

[0116] │├─ Battery (attributive)

[0117] │└─ Capacity (headword, core of the subject)

[0118] └─ How much (object)

[0119] Taking the "battery capacity" field in the "product consultation template" as an example, the semantic definition of this field in the template includes:

[0120] Core words: "battery capacity", "battery life", etc.;

[0121] Grammatical role: The core noun in the object or subject.

[0122] First layer: The words of the "battery" and "capacity" nodes in the dependency tree coincide with the keyword of the template field, initially locating the candidate nodes;

[0123] Second layer: Confirm that the "capacity" node is the core of the object or subject in the tree (in the above example, "capacity" is the core of the subject), meeting the requirements of the template field for the grammatical role (such as "battery capacity" is usually the object of consultation, corresponding to the object or subject);

[0124] Third layer: Through the attributive - headword relationship between "battery" and "capacity" in the dependency tree ("battery" modifies "capacity"), confirm that the combination of the two is a complete semantic unit, avoiding semantic deviation caused by separately extracting "battery" or "capacity".

[0125] If the user's statement involves multiple elements (such as "What are the battery capacity and price of this mobile phone"), the "battery capacity" and "price" fields in the template will simultaneously match the corresponding nodes in the dependency tree ("battery capacity" is one object, and "price" is another object), achieving parallel extraction of multiple elements.

[0126] If the object of the dependency tree is an interrogative word (such as "how much", "what", "where"), then the interrogative word is extracted as the value to be filled, prompting customer service to provide more specific information; if the statement is a declarative sentence (such as "I need a 128G phone"), then the specific description in the subject or object (such as "128G") is extracted as the element value.

[0127] When an element value contains multiple layers of modification relationships (such as "red long-sleeved pure cotton T-shirt"), the dependency tree will retain the complete modification chain ("red" → "long-sleeved" → "pure cotton" → "T-shirt"). During extraction, the elements will be combined in the order of the words to form the complete value "red long-sleeved pure cotton T-shirt" to avoid missing modification information. If a template field node is matched in the dependency tree, but there is no specific value at the corresponding position (such as the statement "phone cannot be charged" without mentioning the model), the element value will be marked as "not retrieved", triggering the subsequent list of entities that customer service must ask for (such as "please provide the phone model").

[0128] Example of handling complex scenarios:

[0129] User message: "My XPro laptop, which I bought yesterday, is now showing a black screen when I turn it on. How can I fix this?"

[0130] Dependency tree processing: split the main clause "the laptop computer has a black screen when it boots up" and the subordinate clause "the XPro model I bought yesterday";

[0131] The template field "Product Model" matches "XPro" in the subordinate clause, and "Problem Description" matches "Black Screen on Startup" in the main clause. The corresponding values ​​are extracted respectively.

[0132] User comments: "How is the sound quality of these headphones? Can the battery last up to 8 hours?"

[0133] The dependency tree generates two independent clauses, which are matched with the template fields "sound quality consultation" and "battery life consultation" respectively, and "sound quality" and "8 hours" are extracted as the corresponding element values.

[0134] In step S2.3 above, based on the key issue element values, the rule base is called (e.g., entities such as "order number," "reason for return," and "product status" are required for the "return" intent) to generate a supplementary entity list that customer service must inquire about; Named Entity Recognition (NER) is performed on the customer service response text, and entities (such as order number and product name) are extracted through dictionary matching and machine learning models (such as LSTM-CRF) to form a customer service entity set; The customer service entity set is compared one by one with the required entity list. If all required entities are mentioned in the customer service response, it is marked as "complete," otherwise it is marked as "missing," and a Boolean value result is output.

[0135] Step S2.4 above merges the user's consultation intent category, key-value pairs of key issue elements (such as "consultation type = logistics, express tracking number = 12345"), and entity integrity boolean flags (true / false) into a unified data structure; the integration result is converted into a structured format such as JSON or XML, and timestamps and dialogue identifiers are added to facilitate direct calls by subsequent modules (such as sentiment analysis and path planning).

[0136] In a preferred embodiment of the present invention, step S3 above: performing sentiment analysis on the dialogue text stream based on the semantic understanding result and outputting structured sentiment analysis results, may include:

[0137] Step S3.1: Based on the user's consultation intent category and key question element values ​​in the semantic understanding results, extract the user's statements from the dialogue text stream, and then obtain the complete statement of the user's last feedback question according to the timestamp of the dialogue text stream; load the sentiment dictionary in the e-commerce domain, match the sentiment words in the complete statement of the user's last feedback question and obtain the basic sentiment score, and then calculate these sentiment word scores by weight; detect the transition conjunctions in the complete statement of the user's last feedback question and segment the statement into sub-units, and perform attenuation or reverse propagation on the calculated sentiment word scores according to the order of the sub-units to generate the user's emotional state value;

[0138] Step S3.2: Based on the user's emotional state value, extract the emotional adjustment keywords from the customer service reply text. When at least one keyword is detected, mark the customer service representative as having an intention to soothe the user. Obtain the next sentence of the user's reply after marking the customer service representative as having an intention to soothe the user. For the obtained next sentence, repeat the emotional state value calculation to generate a new emotional state value. Calculate the emotional change value based on the new emotional state value and the previous user emotional state value. Determine whether the emotional change value exceeds the threshold. If it does, and the customer service representative has an intention to soothe the user, mark the soothing as effective and obtain the soothing effect mark.

[0139] Step S3.3: Integrate user emotional state values ​​and soothing effect markers to output structured sentiment analysis results.

[0140] In this embodiment of the invention, user text is converted into quantifiable emotion values, helping customer service staff to grasp user emotional fluctuations in real time. Transitional processing avoids misjudgment of emotions and improves analysis accuracy. The effectiveness of customer service responses is quantified by marking reassurance effects, helping to optimize training direction. Tracking user emotional changes can identify potential complaint risks in advance, reducing the probability of disputes. A structured format ensures that emotion analysis results can be seamlessly integrated into downstream processes, improving system collaboration efficiency. After the integrated results are stored in the database, it supports searching for emotion trends by time, intent category, and other dimensions, providing data support for service strategy optimization.

[0141] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0142] In step S3.1 above, the user's inquiry intent category (such as "logistics complaint") and key issue element value (such as "lost package") are obtained from the semantic understanding results. Combined with the dialogue timestamp, the complete statement of the user's last feedback is selected (such as "Your courier was not only late, but also damaged the package!").

[0143] Load the e-commerce sentiment dictionary (containing words such as "satisfied", "terrible", "thank you" and their corresponding base scores, such as "damaged" = -3 points, "satisfied" = +2 points), traverse the words in the user's sentence, match sentiment words and extract scores; adjust the scores according to the position of sentiment words in the sentence (such as sentiment words at the end of the sentence have higher weight) and modifiers (such as "very" to strengthen the sentiment) (such as "very terrible" = "terrible" score × 1.5).

[0144] Detect conjunctions of transition ("but", "however", "however", etc.) and divide the sentence into sub-units (e.g., "although the price is cheap, the quality is poor" is divided into "although the price is cheap" and "but the quality is poor"); process them in the order of sub-units: the sentiment score of the previous sub-unit decreases sequentially (e.g., the score of the first unit is ×0.5), and the score of the sub-unit after the transition is reversed or strengthened (e.g., in "but the quality is poor", the score of "poor" is -4, which is directly used as the dominant sentiment).

[0145] The weighted scores of all sentiment words are aggregated and normalized to generate the user's emotional state value (usually ranging from -1 to 1, such as -0.8 indicating strong negative emotions and 0.3 indicating slight positive emotions).

[0146] Step S3.2 above involves the preset process of the emotion regulation keyword library:

[0147] We analyzed a massive amount of historical e-commerce customer service dialogues, extracted frequently occurring reassuring expressions (such as "sorry," "we'll handle it immediately," and "please rest assured"), and statistically analyzed their frequency of occurrence in successful reassurance scenarios. Referring to e-commerce industry customer service standards (such as the "E-commerce After-Sales Customer Service Script Guide"), we compiled standardized reassurance phrases (such as "we will follow up and handle it for you" and "we are very sorry for the inconvenience"). We invited experienced customer service teams to compile effective scripts for high-frequency reassurance scenarios (such as logistics delays and product quality issues), and extracted core keywords (such as "reissue," "refund," and "compensation").

[0148] Classified by semantic function:

[0149] Apology-related: includes words expressing remorse ("sorry", "excuse me", "I apologize", "I'm sorry") and compound phrases ("I'm very sorry", "I'm really sorry");

[0150] Action commitments: Verb phrases involving solutions ("process immediately", "reissue as soon as possible", "verify immediately", "arrange immediately");

[0151] Reassuring and reassuring statements: statements that give users confidence ("Please rest assured", "It will definitely be resolved", "Satisfaction guaranteed", "Guaranteed processing");

[0152] Compensation-related terms: Terms involving the redress of losses ("compensation", "refund", "coupon", "extra gift").

[0153] Adaptive tagging by scenario: Mark the applicable scenario for each keyword (e.g., "reissue" is mainly used for logistics loss scenarios, and "refund" is suitable for return scenarios) to improve matching accuracy.

[0154] Every quarter, based on the latest dialogue data, new high-frequency reassurance words are added (such as emerging internet slang like "truly" and "arranged"), while low-frequency or outdated words are removed (such as the old phrase "please wait a moment"). When the system detects an effective reassurance expression that is not covered by the keyword database (such as "don't worry, I'll handle it for you right away"), it is automatically marked and submitted for manual review before being added to the database.

[0155] The correspondence between the customer service reply and the user's next statement is determined based on the timestamp (e.g., if the customer service replies at 14:30:00 and the user replies at 14:30:15, it is considered as a set of interactions); the emotional state value calculation of S3.1 is performed on the user statement before the customer service reply (old statement) and the user statement after the reply (new statement) to obtain the old emotional value (e.g., -0.7) and the new emotional value (e.g., -0.2).

[0156] The new emotional value is subtracted from the old emotional value to obtain the emotional change value (e.g., -0.2 - (-0.7) = 0.5). A positive number indicates emotional improvement, and a negative number indicates deterioration. A preset basic improvement threshold (e.g., 0.3) is set and adjusted according to the complexity of the scenario (e.g., the threshold is set to 0.4 for after-sales rights protection scenarios and 0.2 for product consultation scenarios). "Effective reassurance" is marked only when both "emotional change value > threshold" and "customer service representative shows reassurance intention" are met. If the customer service representative has no reassurance intention but the emotional state improves (e.g., the user relieves it on their own), it is marked as "natural improvement" and is not included in the reassurance effect statistics.

[0157] Handling complex scenarios, such as:

[0158] Multi-round reassurance scenario: If customer service replies multiple times in a row (e.g., "Sorry" → "We will resend it immediately" → "We will follow up for you"), only the last reply from the user will be used to calculate the emotional change, avoiding double counting.

[0159] Scenario of fluctuating emotions: After a user's emotion value changes from -0.7 to -0.2 (a change of 0.5, which is effective) and then drops to -0.6, the system records the first effective reassurance. Subsequent fluctuations are separately marked as "fluctuating emotions".

[0160] The purpose of the keyword database is to quickly determine whether customer service representatives should take reassurance actions through a pre-set database, avoiding subjective judgment bias (e.g., "We will handle it" clearly matches "action commitment" keywords); to analyze high-frequency effective keywords (e.g., the reassurance effectiveness rate of "compensation" reaches 75%), and to guide customer service training to focus on high-value phrases (e.g., prioritizing the use of expressions containing "compensation").

[0161] The abstract concept of "soothing effect" is transformed into a calculable numerical indicator, supporting horizontal comparisons (e.g., Customer Service A's soothing effectiveness rate is 80% vs. Customer Service B's is 60%). When soothing is repeatedly ineffective (e.g., the change value is <0.3 more than 3 times), the supervisor is automatically triggered to intervene and remind, so as to prevent the user's emotions from continuing to deteriorate and leading to complaints.

[0162] In step S3.3 above, the user's emotional state value (e.g., -0.5), the comfort effect flag ("effective" or "ineffective"), the corresponding timestamp (e.g., 2025-06-2014:30:22), and the dialogue identifier (e.g., session ID-12345) are integrated into a unified data structure and converted into JSON format (example: `{"emotion_value":-0.5,"comfort_effect":"effective","timestamp":"2025-06-2014:30:22"}`), which is convenient for subsequent modules (e.g., dynamic path planning) to call and for database storage.

[0163] In a preferred embodiment of the present invention, step S4 above involves: setting a starting state point and an ending state point in the dialogue processing path based on the semantic understanding results and sentiment analysis results; the starting state point is a semantic vector that integrates the core question elements of the user's initial feedback and the current user's emotional state value; the ending state point is a semantic vector that integrates the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold; connecting the starting state point and the ending state point to form a reference straight line vector, which serves as the expected processing path; mapping the real-time matched customer service solution to the current trajectory vector; calculating the directional deviation angle and magnitude difference value of the current trajectory vector relative to the reference straight line vector, and generating a dynamic correction coefficient, which may include:

[0164] Step S4.1: Based on the semantic understanding results, extract the core problem elements from the user's initial feedback, process the extracted core problem elements, and convert them into semantic vector representations to obtain the problem semantic vector; input the sentiment analysis results to extract the current user's emotional state value; fuse the problem semantic vector and the current user's emotional state value to generate an initial state vector; input the initial state vector into the platform's business rule base to obtain a standardized solution matching the current problem type; process the standardized solution to convert it into a semantic vector representation to obtain a solution semantic vector; input the sentiment analysis association rules to extract the target user's emotional threshold; fuse the solution semantic vector and the target user's emotional threshold to generate a termination state vector.

[0165] Step S4.2: Connect the initial state vector and the final state vector to generate the reference line vector;

[0166] Step S4.3: Based on the resolution semantic vector in the termination state vector, use it as the semantic transformation benchmark; map the real-time customer service solution to the same semantic space and output the current trajectory vector;

[0167] Step S4.4: Based on the reference line vector and the current trajectory vector, calculate the directional deviation angle between the two vectors: calculate the angle deviation value using the cosine theorem of the vector angle; calculate the difference in magnitude between the two vectors: take the absolute difference between the magnitude of the reference line vector and the magnitude of the current trajectory vector; generate a dynamic correction coefficient by combining the directional deviation angle and the difference in magnitude.

[0168] In this embodiment of the invention, the starting vector integrates the semantics of the problem and the user's emotions, while the ending vector combines the standard solution with the target emotions. This allows dialogue path planning to consider both business needs and user experience simultaneously. It provides clear starting and ending points for customer service, ensuring that the solution does not deviate from the core problem and the goal of emotional reassurance. It transforms the abstract service process into a quantifiable vector path, facilitating real-time comparison of the deviation between the actual service and the ideal state, providing a unified reference for subsequent deviation calculations of the trajectory vector. It converts the real-time response of customer service into dynamic points in the vector space, intuitively demonstrating whether the service process deviates from the standard path. It maps the trajectory vector based on the standard solution, ensuring that the customer service solution always revolves around the core steps required by the business rules. Finally, it converts service deviations into specific values, facilitating automatic system adjustment and human understanding.

[0169] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0170] In step S4.1 above, key issue elements from the user's initial feedback (such as "order number" and "reason for return" in the "return application" scenario) are selected from the semantic understanding results to form an element set; the element set is input into a pre-trained semantic model in the e-commerce domain (such as Word2Vec), and each element is mapped to a dense vector in a high-dimensional space (such as "order number = 12345" is converted into a vector containing numerical features), and merged into an issue semantic vector; the user's emotional state value (such as -0.5) in the sentiment analysis results is extracted and concatenated with the issue semantic vector as an independent dimension (scalar) to generate an initial state vector containing semantics and emotion (such as [semantic vector dimension, emotion value dimension]).

[0171] Based on the current problem type (e.g., "return"), retrieve the standard solution from the business rule base (e.g., "verify order → confirm return reason → generate return address → refund processing"); convert the key steps of the standard solution (e.g., "verify order" and "generate return address") into semantic vectors, using the same method as the problem semantic vector, to form a solution semantic vector; obtain the expected sentiment value through sentiment analysis association rules (e.g., "the target sentiment threshold for the return scenario is -0.2"), and concatenate it with the solution semantic vector to generate a termination state vector.

[0172] In step S4.2 above, each dimension of the high-dimensional space (e.g., 100 dimensions) corresponds to a semantic feature in the e-commerce customer service scenario, for example:

[0173] Dimensions 1-10: Product attributes (model, color, price, etc.);

[0174] Dimensions 11-20: Order information (order number, order time, logistics status, etc.);

[0175] Dimensions 21-30: Problem types (returns, exchanges, repairs, etc.);

[0176] Dimensions 31-99: Solution steps (verify information, generate address, refund processing, etc.);

[0177] 100th dimension: Emotion value dimension (standardized value from -1 to 1).

[0178] All vectors (start, end, trajectory) follow the same dimension definition to ensure that the semantic meaning of each dimension is consistent (e.g., the 21st dimension always represents "problem type").

[0179] Each vector can be viewed as a coordinate point in space. The distance and direction between two points reflect semantic similarity and differences in processing paths. For example, the starting point of a return scenario and the starting point of an exchange scenario have different coordinates in the "problem type" dimension, and their spatial locations are different. The termination point of a standard solution is always located at the complete coordinates in the "solution steps" dimension (e.g., all steps correspond to a dimension of 1).

[0180] In 100-dimensional space, starting from the initial point A`(x1, x2, ..., x... 100 To the termination point B (y1, y2, ..., y... 100 Draw a directed line segment with the starting point A and the ending point B; the coordinates of each dimension of the reference line vector are the difference between the ending vector and the starting vector (e.g., the coordinate of the i-th dimension is yi-xi), representing the amount of change in that dimension from the starting point to the ending point.

[0181] The vector direction indicates the optimal path from the user's problem to the standard solution. For example, the vector direction from the starting point to the ending point of "screen scratch return" requires first verifying the order (35th dimension from 0 to 1), then confirming the reason (36th dimension from 0.9 to 1), and finally completing the refund (38th dimension from 0 to 1). At the same time, the sentiment value increases from -0.7 to -0.2. Dimensions with larger changes (such as solution steps) dominate in direction, prompting customer service to prioritize these core steps.

[0182] The vector magnitude is the square root of the sum of the squares of the changes in each dimension (a non-formulaic description), reflecting the expected processing complexity from problem to solution; Case study:

[0183] Simple returns (if the order has been verified, only the address needs to be confirmed) have a smaller mold length (small changes in each dimension);

[0184] Complex returns (such as missing order information requiring multiple rounds of verification) have a larger module length (numerous and significant dimensional changes).

[0185] By transforming textual semantics (such as "reason for return"), numerical indicators (emotional value), and process steps (solutions) into unified spatial coordinates, customer service is transformed from "unstructured information" to "structured vectors." The baseline straight line vector provides a geometric reference for calculating subsequent trajectory vector deviations (direction deviation angle, magnitude difference), for example:

[0186] If customer service fails to verify the order (the 35th dimension is still 0), the current trajectory vector differs from the baseline vector in this dimension, resulting in an increased directional deviation angle.

[0187] If customer service skips the "Generate Return Address" step (37th dimension is 0), the modulus difference value increases, indicating insufficient processing completeness.

[0188] Whether the user's problem is "product inquiry" or "after-sales complaint", a unified path planning and deviation assessment are achieved through high-dimensional space mapping, avoiding the problem of inconsistent assessment standards in different scenarios.

[0189] Step S4.3 above, taking the customer service reply "Please provide your order number, and we will process the return and apply for a refund for you" as an example:

[0190] The sentence is broken down into independent semantic units. Modal words ("please" and "and") and conjunctions are filtered out to extract core action phrases: "provide order number"; "process return"; "apply for refund". According to e-commerce business rules, the actions are classified as: information collection ("provide order number") and process execution ("process return" and "apply for refund").

[0191] Remove invalid expressions (such as comforting phrases like "I'm very sorry," which do not directly correspond to solution steps) and retain only actions directly related to solving the problem (such as "processing returns" being a valid action, and "sorry" is not included in the analysis).

[0192] Taking the termination state vector of the "return" scenario as an example, its semantic vector includes the following core dimensions:

[0193] Dimension A: Order Verification (corresponding to the standard step "Verify Order Information");

[0194] Dimension B: Confirmation of the reason for return;

[0195] Dimension C: Return address generation;

[0196] Dimension D: Refund Processing.

[0197] Each dimension has a pre-defined standard action mapping table (non-formulaic), for example:

[0198] Dimension A (Order Verification) can be mapped to the actions: "Provide Order Number", "Verify Order", and "Confirm Order Information".

[0199] Dimension D (Refund Processing) can be mapped to the actions: "Request a Refund", "Process a Refund", and "Initiate a Refund".

[0200] Keyword matching means that the customer service action "provide order number" contains the keyword "order number", and directly matches the keyword library of dimension A ("order", "order number", "OD", etc.); semantic similarity judgment means that if the action expression is varied (such as "please inform me of the order number"), the association with dimension A is confirmed through semantic expansion in the e-commerce domain ("number" = "number"); business logic verification means confirming the position of the action in the business process (such as "provide order number" is a prerequisite step for "verify order", belonging to the same dimension) and avoiding mismapping (such as "provide address" should not be mapped to the order verification dimension).

[0201] Each dimension is assigned a coordinate value of 0-1 based on the degree of completion of the action: if the customer service only requests "provide the order number" (verification not yet completed), the coordinate value of dimension A is set to 0.5 (in progress); if the customer service has confirmed that the order verification is complete, the coordinate value of dimension A is set to 1 (completed).

[0202] If customer service mentions multiple actions at the same time (such as "process return and apply for refund"), after mapping to the corresponding dimensions, the maximum value of each dimension is taken as the coordinate value (such as dimension B=1, dimension D=1); if there is a dependency relationship between actions (such as "verify the order before processing the return"), if the preceding action is not completed, the dimension value of the subsequent action is set to 0 (such as dimension B=0 if the order is not verified).

[0203] Example of dynamic update of real-time trajectory vector:

[0204] Dialogue Phase 1:

[0205] Customer service replied: "Please provide the order number" → Mapped to dimension A = 0.5 (only a request was made, verification was not completed), other dimensions = 0; trajectory vector: [0.5, 0, 0, 0, ...] (assuming dimension A is the first dimension).

[0206] Dialogue Phase 2:

[0207] After the user provides the order number, customer service replies: "The order has been verified and we are processing your return" → Dimension A = 1 (verification completed), Dimension B = 0.5 (processing); the trajectory vector is updated to: [1, 0.5, 0, 0, ...].

[0208] Dialogue Phase 3:

[0209] Customer service replied: "Return application has been submitted and refund will be processed for you" → Dimension B = 1 (processing completed), Dimension D = 0.5 (refund in progress); trajectory vector updated to: [1, 1, 0, 0.5, ...].

[0210] Each dimension value in the vector represents the progress of the corresponding solution step, and the overall coordinate point represents the customer service's current position on the standard path. For example, the trajectory vector [1, 1, 0, 0.5] means "order has been verified, return has been processed, address not generated, refund application in progress". Compared with the termination vector [1, 1, 1, 1], it can be seen that "generating return address" and "refund completed" have not yet ended.

[0211] The system transforms customer service responses into coordinates in a semantic space, visually displaying the completion status of each solution step. Based on the dimension of the termination vector, it automatically detects whether customer service has missed key steps (e.g., if the dimension of "generate return address" in the trajectory vector is always 0, it indicates that the customer service has not performed this step). The trajectory vector is dynamically updated with each round of customer service responses, forming a path curve from start to finish, supporting backtracking analysis (e.g., if the trajectory deviates from the baseline path after a certain round of responses, the specific deviation can be located).

[0212] In step S4.4 above, the angle between the reference line vector and the current trajectory vector is calculated. The larger the angle, the further the customer service solution deviates from the standard path (e.g., an angle of 90° indicates a complete deviation). The angle is converted into a deviation coefficient of 0-1 (e.g., 0° corresponds to 0, 180° corresponds to 1). The larger the angle, the higher the coefficient.

[0213] Calculate the absolute difference between the baseline vector magnitude (representing the completeness of the ideal processing flow) and the current trajectory vector magnitude (representing the actual progress of processing). The larger the difference, the more incomplete the solution is (e.g., baseline magnitude 5, current magnitude 3, difference 2). Weight the deviation angle coefficient and the magnitude difference (e.g., deviation angle weight 60%, magnitude difference weight 40%) to generate a dynamic correction coefficient of 0-1 (e.g., 0.7 indicates a large deviation that needs to be corrected).

[0214] In a preferred embodiment of the present invention, step S5 above: weighting and calibrating the business specification compliance score based on the dynamic correction coefficient to generate a quality assessment index, may include:

[0215] Step S5.1: Input the dynamic correction coefficient, which includes the path deviation value, and obtain the initial score of real-time matching business specification compliance.

[0216] Step S5.2: Based on the path deviation value in the dynamic correction coefficient, when the path deviation is ≤5%, the initial score of business specification compliance remains unchanged. When the path deviation is >5%, the score deduction ratio is calculated. The compensation factor is obtained by subtracting the deduction ratio from 1. The initial score of business specification compliance is multiplied by the compensation factor to obtain the calibrated business specification compliance score. If the calibrated business specification compliance score is lower than 60 points, it is forcibly corrected to 60 points.

[0217] Step S5.3: Input the calibration business specification compliance score, the information inquiry integrity score converted from the entity integrity Boolean mark, and the emotional reassurance effectiveness score converted from the reassurance effect mark. Calculate the weighted values ​​of the calibration business specification compliance score multiplied by 50%, the information inquiry integrity score multiplied by 30%, and the emotional reassurance effectiveness score multiplied by 20%, respectively, according to the preset weights of 50% for business specification compliance, 30% for information integrity, and 20% for emotional reassurance. Then, add the three together and output the total dialogue quality score.

[0218] Step S5.4: Generate quality assessment indicators based on the calibrated business specification compliance score, information inquiry completeness score, emotional reassurance effectiveness score, and total dialogue quality score.

[0219] In this embodiment of the invention, the dynamic correction coefficient is directly related to the degree of deviation between the service trajectory and the baseline path, upgrading the evaluation basis from static rules to dynamic deviation, ensuring that the evaluation results fit the real-time service scenario, and generating a unified initial score through preset rules to provide a benchmark for subsequent dynamic calibration; the dynamic calibration of business specification scores realizes deviation quantification and precise punishment based on path deviation, sets a lower limit for scores to ensure the robustness of the evaluation, and uses a deduction mechanism to force customer service to converge to the baseline path; the multi-dimensional weighted comprehensive score integrates three dimensions: business specifications, information integrity, and emotional comfort, guides the optimization of customer service resources through weight allocation, and transforms Boolean labels into quantitative indicators to enhance interpretability; the structured evaluation indicator output realizes the transparency of results and provides data support for management decisions.

[0220] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0221] Step S5.1 above, the calculation of the vector deviation of the dynamic correction coefficient, specifically includes the comprehensive result of the direction deviation angle and the difference in magnitude. For example, when the angle between the trajectory vector of the customer service solution and the reference straight line vector is 30° and the difference in magnitude is 20% of the reference, the path deviation value is 0.3 (i.e. 30%) after weighted calculation; 0 indicates that it completely matches the reference path and 1 indicates that it completely deviates (such as the customer service does not respond according to any standard steps).

[0222] Deviation directly reflects the degree of discrepancy between customer service response and standard solution, for example:

[0223] Deviation degree 0.1: Customer service only missed one minor step (such as not proactively informing the customer of the refund arrival time), and the overall process is basically compliant;

[0224] Deviation 0.7: Customer service skips core steps (such as rejecting returns without verifying the order), which is a serious deviation from the standard procedure.

[0225] The process of building a pre-defined business rule base (taking e-commerce returns as an example):

[0226] Extract the official after-sales policies of e-commerce platforms (such as the "Return Processing Guidelines") and break them down into executable rules:

[0227] Order validity must be verified (Rule 1);

[0228] The reasons for returns must be recorded and categorized (Rule 2);

[0229] A return address must be generated within 24 hours (Rule 3).

[0230] Analyzing high-satisfaction return conversations, we identified frequently occurring compliant behaviors: customer service proactively explaining how to check return progress (Rule 4); providing compensation options (Rule 5).

[0231] Industry standards and regulations are integrated: relevant clauses of the Consumer Rights Protection Law are incorporated (such as the "seven-day no-reason return" rule 6).

[0232] Categorization by scenario:

[0233] Return scenario rule set: includes rules for order verification, reason recording, address generation, etc.;

[0234] Exchange scenario rule set: includes rules for product inventory inquiry, shipping cost liability, etc.

[0235] Prioritized by rule:

[0236] Core mandatory rules (high weight): such as "verify order" (rule 1), failure to execute will result in direct point deduction;

[0237] Assisted optimization rules (low weight): such as "actively explain the progress" (rule 4), which can earn points.

[0238] Each rule corresponds to an "executed / not executed" status. Some rules support graded scoring (e.g., "generate address within 24 hours": full marks for execution within 0-24 hours, deductions based on overtime percentage for executions exceeding 24 hours); the maximum score is 100 points. 10-20 points are deducted for each core rule not executed, and 5-10 points are deducted for each auxiliary rule not executed. For example: if "verify order → record reason → generate address within 24 hours" is fully executed, triggering rules 1, 2, and 3, the initial score = 100 - 0 = 100 points; if "verify order" is omitted, triggering rule 1 not executed, the initial score = 100 - 20 = 80 points.

[0239] If the platform updates its return policy (such as adding "fresh produce is not eligible for a 7-day no-reason return"), corresponding rules need to be added; if a large number of complaints are caused by "failure to inform the party responsible for return shipping costs", this matter should be upgraded to a core rule; if the order system adds an automatic verification function, rule 1 can be adjusted to "automatic system verification + manual confirmation".

[0240] If the subsequent dispute rate increases by 30% when a certain rule (such as "verify order") is not executed, its weight will be increased from 10 points to 20 points; if the execution of a certain auxiliary rule (such as "send friendly reminder") has less than 5% impact on satisfaction, its weight will be reduced or it will be removed.

[0241] Example of compliance check for initial score:

[0242] Scenario: User applies for return

[0243] Customer service response: "Please provide your order number and we will process your return" (only "verify order" was executed, no reason was recorded, and no address was generated).

[0244] Rule matching:

[0245] Rule 1 (verify order) → No points deducted;

[0246] Failure to comply with rule 2 (record the reason) → Deduct 15 points;

[0247] Failure to execute rule 3 (generated address) → Deduct 15 points;

[0248] Initial score = 100 - 15 - 15 = 70 points.

[0249] Scenario: Complete Customer Service Execution Process

[0250] Customer service response: "Your order has been verified. The reason for the return is that the size is incorrect. The return address has been sent to your mobile phone. The refund is expected to be processed within 3 business days."

[0251] Rule matching: Execute rules 1, 2, and 3 → No points deducted, initial score = 90 points (10 points reserved for process optimization bonus).

[0252] The rule base ensures the compliance of basic processes ("whether it is done"), while the dynamic correction coefficient measures the execution deviation ("how well it is done"). The combination of the two achieves a dual assessment of "compliance + accuracy". The rule base defines "what must be done", and the dynamic correction coefficient guides "how to do it better" (such as adjusting the response order according to the baseline path), forming a service improvement path from "meeting the standard" to "optimization".

[0253] In step S5.2 above, compare the path deviation value with the 5% threshold:

[0254] If the deviation is ≤5% (e.g., 0.03), the customer service response is considered to be basically consistent with the baseline path, and the initial score remains unchanged (e.g., 90 points).

[0255] If the deviation is greater than 5% (e.g., 0.2), the deduction mechanism will be triggered.

[0256] Deduction ratio determination: The portion exceeding 5% will be deducted proportionally (e.g., 5% of the score will be deducted for every 10% of the excess portion; for example, if the deviation is 0.2 (20%), the portion exceeding 5% will be 15%, and the deduction ratio = 15% × (5% / 10%) = 7.5%).

[0257] Compensation factor = 1 - deduction ratio (e.g., 1 - 0.075 = 0.925); the score after calibration is calculated as the initial score × compensation factor (e.g., 70 × 0.925 = 64.75 points; if the score after calibration is lower than 60 points (e.g., 55 points), it will be forcibly adjusted to 60 points to avoid extreme deviations that could lead to score distortion).

[0258] In step S5.3 above, the entity integrity Boolean flag (true / false) is converted into a numerical value: true (entity complete) → 100 points; false (entity missing) → 0 points (or points are deducted according to the missing proportion, such as deducting 20 points for each missing entity).

[0259] The soothing effect rating ("effective" / "ineffective") is converted into a numerical value: "effective" → 100 points; "ineffective" → 0 points (or the score is refined according to the degree of emotional improvement, such as 0.5 → 80 points).

[0260] The following weights are assigned based on preset weights: Business compliance (after calibration): 50% weight; Information integrity: 30% weight; Effectiveness of emotional reassurance: 20% weight.

[0261] Example of weighted score calculation: Business compliance score after calibration = 80 points → First weighted score = 80 × 50% = 40 points; Information integrity score = 100 points → Second weighted score = 100 × 30% = 30 points; Effective emotional reassurance = 100 points → Third weighted score = 100 × 20% = 20 points; Total score for dialogue quality = 40 + 30 + 20 = 90 points.

[0262] Step S5.4 above merges the following indicators into a unified structure:

[0263] Post-calibration business specification compliance score (e.g., 80 points);

[0264] Information query completeness score (e.g., 100 points);

[0265] Effectiveness score of emotional support (e.g., 100 points);

[0266] Total score for dialogue quality (e.g., 90 points).

[0267] The quality assessment indicators are converted into a standardized format for output. For example, when using JSON format, the content is presented as follows: In a JSON object, the value of "business_standard_score" (business standard compliance score) is 80, the value of "info_integrity_score" (information inquiry integrity score) is 100, the value of "emotion_comfort_score" (emotional comfort effectiveness score) is 100, and the value of "total_quality_score" (total dialogue quality score) is 90. If a table format is used, each assessment indicator and its corresponding score will be clearly listed in rows and columns.

[0268] In a preferred embodiment of the present invention, step S6 above: performing service quality monitoring and evaluation data storage based on quality assessment indicators, may include:

[0269] Step S6.1: Extract the total score of the dialogue quality from the quality assessment indicators. When the total score is ≥80, mark it as excellent service and activate the green light access command; when the total score is ≤60 and <80, mark it as qualified service and activate the yellow light monitoring command; when the total score is <60, mark it as poor service and activate the red light alarm command; to obtain the service marking result, wherein for dialogues marked as poor service, send a session freeze command to the customer service system;

[0270] Step S6.2: Based on the service tagging results and quality assessment indicators, generate standardized storage records, including: a unique dialogue identifier, a calibrated business specification compliance score, an information inquiry completeness score, an emotional reassurance effectiveness score, a total dialogue quality score, and service tagging results; write the storage records into the quality inspection result table of the distributed database by time partitioning;

[0271] Step S6.3: Based on the quality inspection result table data, query the quality inspection result table by customer service employee number, calculate the poor service occurrence rate for each customer service representative, query the quality inspection result table by problem type, count the number of cases with an emotional reassurance effectiveness score of <60, integrate the poor service occurrence rate and reassurance failure case data, generate a management analysis report including trend charts, and push the report to the customer service management platform.

[0272] In this embodiment of the invention, color-coded grading and command triggering enable visualized monitoring of service quality, reducing the risk of the spread of substandard services, automatically freezing substandard service sessions, preventing customer service staff from escalating problems due to inexperience, and improving the ability to prevent user complaints. Standardized storage enables long-term retention and querying of dialogue evaluation data, supporting compliance audits and dispute tracing. Time partitioning and distributed storage improve query performance under large data volumes and reduce data storage costs. Quantitative indicators and visual charts help managers accurately identify service shortcomings and develop targeted training plans. Report data feeds back into customer service management, forming a continuous optimization cycle of "evaluation-analysis-improvement" to improve overall service quality.

[0273] In this embodiment of the invention, when applied in a specific way, it can be implemented through the following technical solutions, for example:

[0274] In step S6.1 above, the "Total Dialogue Quality Score" value (e.g., 90, 75, 55) is read from the quality assessment indicators. If the total score is ≥80, it is marked as "Excellent Service" and the green light access instruction is activated (i.e., the service process proceeds normally by default). If the total score is ≤60 and <80, it is marked as "Qualified Service" and the yellow light monitoring instruction is activated (triggering manual spot checks or system focus). If the total score is <60, it is marked as "Poor Service" and the red light alarm instruction is activated. At the same time, a conversation freeze command is sent to the customer service system (pausing the current dialogue, preventing customer service from continuing to respond, and waiting for supervisor intervention).

[0275] Based on the service tagging results, visual prompts (such as green / yellow / red lights on the interface) are pushed to the customer service workbench through the system interface, and corresponding operations are performed (such as locking the customer service input box when freezing the session).

[0276] In step S6.2 above, the following are extracted from the quality assessment indicators: unique dialogue identifier (e.g., session ID-12345), calibrated business standard score (e.g., 80 points), information integrity score (100 points), emotional reassurance score (100 points), total dialogue score (90 points), and service tag result (“high-quality service”); additional fields include: dialogue start time, end time, user ID, customer service employee number, and other metadata.

[0277] Encapsulate the above fields into a standardized record according to a preset format (such as JSON) (Example: `{"session_id":"12345","score_business":80,"score_info":100,...,"mark":"high-quality service"}`).

[0278] Records are partitioned by timestamp (e.g., year / month / day). For example, data from June 2025 is stored in the `Quality Inspection Results_202506` table. Records are stored in the Quality Inspection Results table through the batch write interface of a distributed database (e.g., HBase, MongoDB) to ensure data reliability in high-concurrency scenarios.

[0279] In step S6.3 above, query the quality inspection result table, group by customer service employee number, and count the total number of conversations and the number of poor service for each customer service representative; poor service occurrence rate = number of poor service / total number of conversations × 100% (e.g., customer service representative A handles 100 conversations, 5 of which are poor, with an occurrence rate of 5%).

[0280] Grouped by problem type (e.g., "return" or "logistics inquiry"), cases with emotional reassurance scores <60 were filtered out, and the number of cases where reassurance failed under each type was counted (e.g., 10 cases of reassurance failure in the "return" scenario).

[0281] Draw a line chart of the occurrence rate of poor service on a daily / weekly / monthly basis (e.g., showing that the occurrence rate of a certain customer service group has decreased from 8% to 3% in the past 30 days); use a bar chart to show the percentage of cases where reassurance failed for each type of problem (e.g., "returns" account for 40%, and "complaints" account for 35%).

[0282] The integrated data and charts are generated into PDF / HTML reports, which are then pushed to the customer service management platform via system messages. Scheduled pushes (e.g., daily at 9 AM) or triggered pushes (e.g., when the incidence rate of a customer service representative exceeds a threshold) are supported.

[0283] like Figure 2As shown, embodiments of the present invention also provide an NLP-based customer service dialogue quality detection system, comprising:

[0284] The acquisition module is used to acquire the real-time text stream of customer service conversations on e-commerce platforms.

[0285] The semantic understanding module is used to perform semantic understanding on the dialogue text stream and output the semantic understanding results, which include: user consultation intent, key question elements, and the completeness of customer service inquiry information entities.

[0286] The sentiment analysis module is used to perform sentiment analysis on the dialogue text stream based on the semantic understanding results and output structured sentiment analysis results.

[0287] The dynamic correction module is used to set a starting state point and an ending state point in the dialogue processing path based on semantic understanding results and sentiment analysis results. The starting state point is a semantic vector that integrates the core question elements of the user's initial feedback and the current user's emotional state value. The ending state point is a semantic vector that integrates the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold. The starting state point and the ending state point are connected to form a baseline straight line vector, which serves as the expected processing path. The real-time matched customer service solution is mapped to the current trajectory vector. The deviation angle and magnitude difference of the current trajectory vector relative to the baseline straight line vector are calculated to generate dynamic correction coefficients.

[0288] The calibration module is used to perform weighted calibration of the business specification compliance score based on dynamic correction coefficients, and generate quality assessment indicators.

[0289] The storage module is used to perform service quality monitoring and evaluation data storage based on quality assessment indicators.

[0290] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for quality inspection of customer service conversations based on NLP, characterized in that, The method comprises: Step S1: Obtain the real-time text stream of customer service conversations on the e-commerce platform; Step S2: Perform semantic understanding on the dialogue text stream and output the semantic understanding results, which include: user consultation intent, key question elements, and completeness of customer service inquiry information entities; Step S3: Based on the semantic understanding results, perform sentiment analysis on the dialogue text stream and output the structured sentiment analysis results; Step S4: Based on the semantic understanding results and sentiment analysis results, set a starting state point and an ending state point in the dialogue processing path; the starting state point is the semantic vector of the core question elements of the user's first feedback and the current user's emotional state value; the ending state point is the semantic vector of the standardized solution for the corresponding question type in the platform's business rule base and the target user's emotional threshold; connect the starting state point and the ending state point to form a reference straight line vector as the expected processing path; map the real-time matched customer service solution to the current trajectory vector; calculate the directional deviation angle and magnitude difference value of the current trajectory vector relative to the reference straight line vector, and generate a dynamic correction coefficient. Step S5: Based on the dynamic correction coefficient, perform weighted calibration on the business specification compliance score to generate quality assessment indicators; Step S6: Based on the quality assessment indicators, perform service quality monitoring and assessment data storage.

2. The NLP-based customer service dialogue quality detection method according to claim 1, characterized in that, Step S1: Obtain the real-time text stream of customer service conversations on the e-commerce platform, including: Real-time access to the dialogue data interface of the e-commerce platform's customer service system to obtain a mixed text stream containing user inquiry messages and customer service reply messages; The acquired mixed text stream is segmented into sentences and labeled with roles in real time, and a structured dialogue sequence is output; each sentence in the structured dialogue sequence is labeled with the speaking role and timestamp; The system removes automatically generated advertising text from the structured dialogue sequence, standardizes colloquial abbreviations and dialect words, filters non-textual symbols, and outputs a dialogue text stream.

3. The NLP-based customer service dialogue quality detection method according to claim 2, characterized in that, Step S2: Perform semantic understanding on the dialogue text stream and output the semantic understanding result. The semantic understanding result includes: user inquiry intent, key question elements, and the completeness of customer service inquiry information entities, including: Extract user statements from the dialogue text stream, parse user intent based on domain knowledge graph, output user consultation intent category, and activate the corresponding structured extraction template according to intent category; Based on structured extraction templates, nodes matching the template are located in the dependency tree, and key issue element values ​​are extracted. Based on the key issue element values, the rule base is called to generate a list of required inquiry entities. That is, each key issue element value triggers a set of supplementary entities that customer service must confirm. Named entity recognition is performed on the customer service reply text to extract the customer service entity set. The extracted customer service entity set is compared with the list of required inquiry entities, and a boolean flag for entity integrity is output. Integrate user consultation intent, key question element values, and entity integrity Boolean flags to output structured semantic understanding results.

4. The NLP-based customer service dialogue quality detection method according to claim 3, characterized in that, Step S3: Based on the semantic understanding results, perform sentiment analysis on the dialogue text stream and output structured sentiment analysis results, including: Based on the user's consultation intent category and key question element values ​​from the semantic understanding results, the user's statements are extracted from the dialogue text stream. Then, based on the timestamp of the dialogue text stream, the complete statement of the user's last feedback question is obtained. An e-commerce domain sentiment dictionary is loaded to match sentiment words in the complete statement of the user's last feedback question and obtain basic sentiment scores. Then, these sentiment word scores are calculated by weighting. The transition conjunctions in the complete statement of the user's last feedback question are detected and the statement is segmented into sub-units. The calculated sentiment word scores are decayed or reversed according to the order of the sub-units to generate the user's emotional state value. Based on the user's emotional state value, emotion adjustment keywords are extracted from the customer service reply text. When at least one keyword is detected, the customer service representative is marked as having an intention to soothe the user. After the user's reply is marked as having an intention to soothe the user, the emotional state value is calculated again for the next reply to generate a new emotional state value. Based on the new emotional state value and the previous user emotional state value, an emotional change value is calculated. It is determined whether the emotional change value exceeds the threshold. If it does, and the customer service representative has an intention to soothe the user, the soothing is marked as effective, and a soothing effect mark is obtained. Integrate user emotional state values ​​and soothing effect markers to output structured sentiment analysis results.

5. The NLP-based customer service dialogue quality detection method according to claim 4, characterized in that, Step S4: Based on the semantic understanding results and sentiment analysis results, set a starting state point and an ending state point in the dialogue processing path; the starting state point is the semantic vector of the core question elements of the user's first feedback and the current user's emotional state value; the ending state point is the semantic vector of the standardized solution for the corresponding question type in the platform's business rule library and the target user's emotional threshold; connect the starting state point and the ending state point to form a baseline straight line vector as the expected processing path; map the real-time matched customer service solution to the current trajectory vector; Calculate the difference between the directional deviation angle and the magnitude of the current trajectory vector relative to the reference straight line vector, and generate dynamic correction coefficients, including: Based on semantic understanding results, the core problem elements of the user's initial feedback are extracted. These core problem elements are processed and transformed into semantic vector representations to obtain the problem semantic vector. Sentiment analysis results are input to extract the current user's emotional state value. The problem semantic vector and the current user emotional state value are fused to generate an initial state vector. The initial state vector is input into the platform's business rule base to obtain a standardized solution matching the current problem type. The standardized solution is processed and transformed into a semantic vector representation to obtain a solution semantic vector. Sentiment analysis association rules are input to extract the target user's emotional threshold. The solution semantic vector and the target user's emotional threshold are fused to generate a termination state vector. Connect the initial state vector and the final state vector to generate the reference line vector; Based on the resolution semantic vector in the termination state vector, a semantic transformation benchmark is used; the real-time customer service solution is mapped to the same semantic space, and the current trajectory vector is output. Based on the reference line vector and the current trajectory vector, calculate the directional deviation angle between the two vectors: calculate the angle deviation value using the cosine theorem of the vector angle; calculate the difference in magnitude between the two vectors: take the absolute difference between the magnitude of the reference line vector and the magnitude of the current trajectory vector; generate a dynamic correction coefficient by combining the directional deviation angle and the difference in magnitude.

6. The NLP-based customer service dialogue quality detection method according to claim 5, characterized in that, Step S5: Based on the dynamic correction coefficient, perform weighted calibration on the business specification compliance score to generate quality assessment indicators, including: Input a dynamic correction coefficient, which includes the path deviation value, to obtain an initial score for real-time matching of business specification compliance. Based on the path deviation value in the dynamic correction coefficient, when the path deviation is ≤5%, the initial score of business specification compliance remains unchanged. When the path deviation is >5%, the score deduction ratio is calculated. The compensation factor is obtained by subtracting the deduction ratio from 1. The initial score of business specification compliance is multiplied by the compensation factor to obtain the calibrated business specification compliance score. If the calibrated business specification compliance score is lower than 60 points, it is forcibly corrected to 60 points. Input the calibration business specification compliance score, the information inquiry integrity score converted from the entity integrity Boolean mark, and the emotional reassurance effectiveness score converted from the reassurance effect mark. Calculate the weighted values ​​of the calibration business specification compliance score multiplied by 50%, the information inquiry integrity score multiplied by 30%, and the emotional reassurance effectiveness score multiplied by 20%, respectively, according to the preset weights of 50% for business specification compliance, 30% for information integrity, and 20% for emotional reassurance. Then, add the three together and output the total dialogue quality score. Quality assessment indicators are generated based on the calibration scores for business compliance, information inquiry completeness, emotional reassurance effectiveness, and total dialogue quality.

7. The NLP-based customer service dialogue quality detection method according to claim 6, characterized in that, Step S6: Based on quality assessment indicators, perform service quality monitoring and evaluation data storage, including: The total score of the conversation quality is extracted from the quality assessment indicators. When the total score is ≥80, it is marked as excellent service and the green light access command is activated; when the total score is ≤60 and <80, it is marked as qualified service and the yellow light monitoring command is activated; when the total score is <60, it is marked as poor service and the red light alarm command is activated. The service marking results are obtained. For conversations marked as poor service, a conversation freeze command is sent to the customer service system. Based on service tagging results and quality assessment indicators, standardized storage records are generated, including: a unique dialogue identifier, a calibrated business specification compliance score, an information inquiry completeness score, an emotional reassurance effectiveness score, a total dialogue quality score, and service tagging results; the storage records are then partitioned by time and written into a quality inspection result table in a distributed database. Based on the quality inspection result table data, the quality inspection result table is queried in groups by customer service employee number to calculate the occurrence rate of poor service for each customer service representative. The quality inspection result table is also queried in groups by problem type to count the number of cases with an emotional reassurance effectiveness score of <60. The data on the occurrence rate of poor service and the failure of reassurance cases are integrated to generate a management analysis report containing trend charts, and the report is pushed to the customer service management platform.

8. A customer service dialogue quality inspection system based on NLP, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire the real-time text stream of customer service conversations on e-commerce platforms. The semantic understanding module is used to perform semantic understanding on the dialogue text stream and output the semantic understanding results, which include: user consultation intent, key question elements, and the completeness of customer service inquiry information entities. The sentiment analysis module is used to perform sentiment analysis on the dialogue text stream based on the semantic understanding results and output structured sentiment analysis results. The dynamic correction module is used to set a starting point and an ending point in the dialogue processing path based on semantic understanding results and sentiment analysis results. The starting point is a semantic vector that integrates the core question elements of the user's first feedback and the current user's emotional state value. The ending point is a semantic vector that integrates the standardized solution for the corresponding question type in the platform's business rule library and the target user's emotional threshold. Connecting the starting point and the ending point forms a baseline straight line vector, which serves as the expected processing path. The real-time matched customer service solution is mapped to the current trajectory vector. Calculate the difference between the directional deviation angle and the magnitude of the current trajectory vector relative to the reference straight line vector, and generate dynamic correction coefficients. The calibration module is used to perform weighted calibration of the business specification compliance score based on dynamic correction coefficients, and generate quality assessment indicators. The storage module is used to perform service quality monitoring and evaluation data storage based on quality assessment indicators.

9. A computing device, characterized in that, include: one or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images