Customer service intent recognition method based on natural language processing

Abstract

This invention relates to the field of natural language processing (NLP) technology and discloses a customer service intent recognition method based on NLP. The method first obtains user text information from a customer service dialogue system, and then preprocesses it through word segmentation, stop word removal, and lemmatization to obtain standardized text. Next, semantic features are calculated to extract intent frequency and trends, determine the intent stage, and dynamically adjust the intent recognition confidence threshold. Then, the intent classification model parameters are adjusted through a classification control algorithm, and the word segmentation algorithm parameters are optimized based on accuracy feedback. Finally, the timing of the control signal output is optimized, and preset parameters are dynamically updated to achieve accurate recognition. The system includes modules for text processing, intent calculation, classification control, intent adjustment, and dynamic optimization. This method and system can improve the accuracy and efficiency of customer service intent recognition, adapt to different user intent state changes, and have good application prospects.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a customer service intent recognition method based on natural language processing. Background Technology

[0002] In the modern customer service field, with the rapid development of information technology and the increasing diversification of user needs, efficiently and accurately identifying user intent has become crucial for improving the quality and efficiency of customer service. Traditional customer service intent identification methods have many limitations and are difficult to meet the needs of today's complex application scenarios.

[0003] Traditional methods have limited capabilities in text preprocessing. For user-input text, they often employ simple word segmentation, failing to perform multi-granular segmentation and struggling to effectively handle word subsequences of varying lengths and structures. Furthermore, stop word removal typically uses fixed filtering thresholds that cannot adaptively adjust based on text content, leading to incomplete or excessive stop word removal, impacting subsequent semantic analysis. Simultaneously, the lack of lexical reconstruction and smoothing of redundant expressions results in preprocessed text that is neither standardized nor coherent, posing challenges for subsequent intent recognition.

[0004] In the process of intent recognition, traditional methods do not extract semantic features comprehensively and deeply enough. They only focus on some simple keywords, while ignoring important features such as sentiment, and cannot accurately grasp the changing trends of user intent. Moreover, it is difficult to accurately determine whether a user's intent is at the consultation stage or the complaint stage based on the frequency and changing trends of the intent, resulting in inaccurate intent classification.

[0005] In terms of model optimization and parameter tuning, traditional methods lack effective closed-loop control mechanisms. Parameter tuning for intent classification models often relies on human experience and cannot be automatically adjusted based on the deviation between the intent recognition target value and the actual recognition confidence level, resulting in poor model adaptability and accuracy. Simultaneously, the parameter tuning for word segmentation algorithms is not intelligent enough, failing to dynamically optimize based on feedback from intent recognition accuracy, thus affecting the effectiveness of text preprocessing.

[0006] Traditional methods have shortcomings in optimizing the output timing of control signals and dynamically updating preset parameters. They cannot comprehensively consider the response time of the intent classification model and word segmentation algorithm to optimize the output timing of control signals, resulting in a less smooth and stable semantic analysis and classification process. Moreover, they cannot monitor the changing trends of user text information and intent recognition parameters in real time, nor can they dynamically update preset parameters, making it difficult to match changes in user intent states, leading to low accuracy in intent recognition.

[0007] With the development of artificial intelligence technology, natural language processing technology is being used more and more widely in the field of customer service. However, the shortcomings of the traditional methods mentioned above seriously restrict the accuracy and efficiency of customer service intent recognition. There is an urgent need for a more advanced and intelligent customer service intent recognition method and system to solve the problems existing in the current technology and improve the quality and level of customer service. Summary of the Invention

[0008] The purpose of this invention is to provide a customer service intent recognition method based on natural language processing to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a customer service intent recognition method based on natural language processing, the method comprising:

[0010] The original user text information is obtained from the customer service dialogue system. The text is preprocessed using a word segmentation algorithm to remove stop words and perform word form restoration to obtain standardized user text information.

[0011] Based on standardized user text information, the semantic features of the text are calculated, the frequency and trend of user intent are extracted, and the intent classification algorithm is used to determine whether the user intent is in the consultation or complaint stage. When it is in the consultation stage, the required intent recognition confidence threshold is calculated according to the preset intent recognition parameters, and the intent recognition target value is dynamically adjusted in combination with the semantic change trend of user text information.

[0012] A classification control algorithm is adopted to calculate the parameter adjustment amount of the intent classification model based on the deviation between the intent recognition target value and the current actual recognition confidence level, generate a model optimization signal and output it to the intent classification model execution module;

[0013] The current intent recognition accuracy value is obtained through the accuracy feedback signal of the intent feedback module. Combined with the preset accuracy range, it is determined whether the parameters of the word segmentation algorithm need to be adjusted. If the current accuracy value exceeds the preset range, the parameter optimization algorithm is used to calculate the parameter adjustment amount of the word segmentation algorithm based on the accuracy deviation value, and a control signal is generated and output to the word segmentation algorithm execution module.

[0014] Based on the response time of the intent classification model and word segmentation algorithm, the output timing of the control signal is optimized to make the semantic analysis and classification process of user intent recognition smooth and stable. Then, by monitoring the changing trends of user text information and intent recognition parameters in real time, the preset parameters are dynamically updated to match the changes in user intent state and to accurately identify customer service intent.

[0015] Preferably, a word segmentation algorithm is used to preprocess the text, removing stop words and performing lemmatization to obtain standardized user text information, specifically including:

[0016] User text information data is input into the word segmentation algorithm model, and then the word segmentation algorithm is used to perform multi-granular segmentation of user text information, decomposing user text information into word subsequences of different lengths;

[0017] Based on the word subsequences obtained from the decomposition, stop word component features are extracted. The extracted stop word component features are input into the stop word filtering model. Based on the input stop word component features, the filtering threshold is adaptively adjusted to perform adaptive filtering on the word subsequences and remove the stop word components.

[0018] After filtering out stop words, each word subsequence is subjected to lemmatization to obtain standardized user text information. The standardized user text information is then smoothed to eliminate residual redundant expressions in the text, resulting in coherent user text information.

[0019] The smoothed user text information is output as the final user text information processing result, which is used for semantic and frequency analysis of user intent.

[0020] Preferably, based on standardized user text information, the semantic features of the text are calculated, the frequency and trend of user intent are extracted, and an intent classification algorithm is used to determine whether the user intent is for consultation or complaint purposes. Specifically, this includes:

[0021] Based on standardized user text information, word embedding is performed to obtain the semantic vector representation of the user text information. Then, the semantic features of the user text information are extracted, including intent keywords and sentiment features.

[0022] Based on the extracted intent keyword features, the frequency of user intent occurrence is calculated to obtain the start and end times of each user intent cycle. Then, based on the extracted sentiment characteristics, the trend of user intent change is calculated to obtain the curve of intent type change over time.

[0023] Based on the frequency and trend of user intent, determine whether the current user intent is in the consultation stage or the complaint stage. When the user intent is consultation-oriented and is in the first half of the user intent cycle, it is judged as consultation-oriented. When the user intent is complaint-oriented and is in the second half of the user intent cycle, it is judged as complaint-oriented.

[0024] Based on the determined user intent state, the current user intent state is output, and the user intent state is associated with the corresponding dialogue time point to obtain a complete user intent state sequence.

[0025] Preferably, after determining whether the user's intent is for consultation or complaint through the intent classification algorithm, the process further includes:

[0026] Based on the preset intent recognition parameters and the real-time detected semantic vector data, these are used as input to the neural network model. Through the trained neural network model, the intent recognition confidence threshold required for the current consultation stage is calculated.

[0027] Based on the calculated confidence threshold for intent recognition, combined with semantic vector change trend data, a dynamic adjustment algorithm is used to dynamically adjust the output threshold of the intent classification model so that it gradually approaches the target value.

[0028] During the adjustment process, changes in semantic vectors and intent recognition confidence are detected in real time. When the semantic vectors change significantly or the intent recognition confidence deviates greatly from the target value, the neural network model is triggered to recalculate the intent recognition confidence threshold.

[0029] The new intent recognition confidence threshold calculated by the neural network model is input into the dynamic adjustment algorithm to dynamically adjust the output threshold of the intent classification model again until the intent recognition confidence stabilizes near the new target value.

[0030] Preferably, a classification control algorithm is used to calculate the parameter adjustment amount of the intent classification model based on the deviation between the intent recognition target value and the current actual recognition confidence level, generate a model optimization signal, and output it to the intent classification model execution module, specifically including:

[0031] Obtain the target value and the actual recognition confidence value for intent recognition, calculate the deviation between the two, input the deviation value into the classification control algorithm, and calculate the parameter adjustment amount of the intent classification model based on the weight, offset, and learning rate parameters of the classification control algorithm.

[0032] Based on the calculated adjustment amount of the intent classification model parameters, a corresponding model optimization signal is generated, and the generated model optimization signal is output to the intent classification model execution module to control the parameter update of the intent classification model.

[0033] The intent classification model adjusts its internal parameters based on the received model optimization signal, thereby changing the actual confidence level of intent recognition. By continuously acquiring real-time intent recognition confidence values ​​and comparing them with target values, the deviation value is calculated to form a closed-loop control.

[0034] Preferably, the current intent recognition accuracy value is obtained through the accuracy feedback signal of the intent feedback module, and combined with a preset accuracy range, it is determined whether the parameters of the word segmentation algorithm need to be adjusted, specifically including:

[0035] The accuracy value data is transmitted to the control system and compared with the preset accuracy threshold range to determine whether the current accuracy is within a reasonable range. When the accuracy value is lower than the lower limit threshold of the preset range, the control system issues an instruction to increase the segmentation granularity of the word segmentation algorithm through the actuator.

[0036] When the accuracy value is higher than the upper limit of the preset range, the control system issues an instruction to reduce the segmentation granularity of the word segmentation algorithm through the actuator, thereby optimizing the preprocessing effect of the user's text information. After the parameters of the word segmentation algorithm are adjusted, the standardized text information output by it is obtained again until the accuracy value is stable within the preset range.

[0037] During parameter adjustment, historical accuracy data is analyzed using machine learning algorithms to establish a mapping model between accuracy values ​​and word segmentation parameters. Based on the current accuracy value, the optimal word segmentation parameters for the next moment are predicted using the established mapping model, and feedforward control of the word segmentation algorithm is performed.

[0038] Preferably, a parameter optimization algorithm is used to calculate the parameter adjustment amount of the word segmentation algorithm, generate a control signal, and output it to the word segmentation algorithm execution module, specifically including:

[0039] The current system accuracy value is obtained in real time through the intent feedback module, and the accuracy value is compared with the preset accuracy range to determine whether the accuracy value exceeds the preset range.

[0040] When the accuracy value exceeds the preset range, the deviation between the current accuracy value and the value in the preset accuracy range is calculated. The deviation value is used as the input of the parameter optimization algorithm. Based on the accuracy deviation value, the parameter adjustment amount of the word segmentation algorithm is calculated through rule reasoning using a pre-established parameter optimization rule library.

[0041] The calculated word segmentation algorithm parameter adjustment amount is converted into a corresponding control signal. The digital signal is converted into an analog signal through a digital-to-analog converter. The generated analog control signal is output to the word segmentation algorithm execution module to drive the execution module to adjust the word segmentation algorithm parameters.

[0042] The word segmentation algorithm changes the segmentation parameters according to the adjustment actions of the execution module, thereby optimizing the preprocessing effect of user text information and bringing the accuracy value back to the preset range.

[0043] Preferably, based on the response time of the intent classification model and word segmentation algorithm, the output timing of the control signal is optimized to make the semantic analysis and classification process of user intent recognition smooth and stable, specifically including:

[0044] Obtain the response time parameters of the intent classification model and word segmentation algorithm, establish a dynamic model of the model, obtain the relationship curves between model parameters and accuracy and confidence, and calculate the target parameters of the model and the timing of control signal output through the dynamic model of the model based on the target values ​​of intent recognition accuracy and confidence.

[0045] Then, using the model parameter deviation and accuracy confidence deviation as inputs, the optimized adjustment amount of the control signal is obtained through rule reasoning. The optimized control signal is then sent to the intent classification model and word segmentation algorithm in sequence according to the calculated output timing to dynamically adjust the model parameters.

[0046] During the dynamic adjustment of the model, text information and intent recognition results from the customer service dialogue system are collected in real time and compared with the target value to obtain the deviation of accuracy and confidence. When the accuracy deviation or confidence deviation exceeds the preset threshold, the deviation is used as a feedback signal and input into the parameter optimization algorithm for further optimization and adjustment of the control signal.

[0047] Preferably, by monitoring the changing trends of user text information and intent recognition parameters in real time, preset parameters are dynamically updated to match changes in user intent states, thereby achieving accurate identification of customer service intent. Specifically, this includes:

[0048] The system acquires the user's real-time text information and current intent recognition parameters, compares them with preset normal range thresholds, and determines whether they exceed the normal range. When the real-time text information or intent recognition parameters exceed the normal range, an alert is triggered, and the magnitude of the intent recognition parameters that need to be adjusted is determined based on the degree of deviation from the normal range.

[0049] Using machine learning algorithms, based on the changing trends of users' historical text information and intent recognition parameters, the system predicts the changing trends of users' intent states over a future period of time; then, based on users' real-time text information, current intent recognition parameters, and predicted future intent state changing trends, the system dynamically adjusts the intent recognition parameters to generate new intent recognition parameter settings.

[0050] The dynamically adjusted intent recognition parameter settings are sent to the intent recognition system, and the control device recognizes the user's intent in real time according to the new parameter settings to match the user's real-time intent status.

[0051] After adjusting the intent recognition parameters, continuously monitor changes in the user's text information to determine whether the user's intent state has returned to normal. If it has not returned to normal, return to continue adjusting.

[0052] The user's real-time text information, intent recognition parameters, and adjusted results are stored in a database to optimize the machine learning model.

[0053] Preferably, the present invention includes a customer service intent recognition system based on natural language processing, used to implement the customer service intent recognition method based on natural language processing described above, the system comprising:

[0054] The text processing module obtains user text information from the customer service dialogue system, uses word segmentation algorithms to segment it at multiple granularities, extracts and filters out stop word components, performs word form restoration and smoothing processing, and finally outputs coherent user text information.

[0055] The intent calculation module calculates semantic features based on standardized user text information, extracts the frequency and trend of user intent occurrence, and determines the current user intent state. When in the consultation stage, it combines preset parameters and real-time semantic vector data to calculate the intent recognition confidence threshold through a neural network model, and uses a dynamic adjustment algorithm to dynamically adjust the output threshold of the intent classification model to meet the intent recognition needs of users at different stages.

[0056] The classification control module employs a classification control algorithm. Based on the deviation between the target value of intent recognition and the actual recognition confidence level, it calculates the parameter adjustment amount of the intent classification model, generates a model optimization signal, and outputs it to the intent classification model execution module. This controls the update of the intent classification model parameters, adjusts the model output to change the actual confidence level of intent recognition, and stabilizes the intent recognition confidence level at the target value through closed-loop control, thereby achieving accurate customer service intent recognition.

[0057] The intent adjustment module obtains the current intent recognition accuracy value through the accuracy feedback signal of the intent feedback module, and makes a judgment based on the preset accuracy range. When the accuracy value exceeds the preset range, it calculates the parameter adjustment amount of the word segmentation algorithm based on the accuracy deviation value, generates a control signal and outputs it to the execution module to drive it to adjust the word segmentation algorithm parameters, optimize the text preprocessing effect, and stabilize the intent recognition accuracy within the preset range.

[0058] The dynamic optimization module integrates the response time of the intent classification model and word segmentation algorithm, optimizes the timing of control signal output, and makes the semantic analysis and classification process of user intent recognition smooth and stable. By monitoring the changes in user text information and intent recognition parameters in real time, it dynamically updates preset parameters, matches changes in user intent state, accurately identifies customer service intent, and stores relevant data in the database for optimizing machine learning models.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] In the text preprocessing stage, through multi-granularity segmentation, adaptive filtering of stop words, word form restoration and smoothing, standardized and coherent user text information can be accurately obtained, laying a solid foundation for subsequent semantic analysis and intent recognition, and effectively improving the quality of text processing.

[0061] In terms of semantic feature extraction and intent judgment, semantic vectors are obtained through word embedding, and intent keywords and sentiment features are extracted. This can accurately calculate the frequency and trend of user intent occurrence, accurately determine whether user intent is in the consultation or complaint stage, provide a basis for subsequent targeted intent recognition, and improve the accuracy of intent classification.

[0062] When in the consultation stage, the system combines preset parameters and real-time semantic vector data to calculate the confidence threshold for intent recognition through a neural network model. The system also uses a dynamic adjustment algorithm to dynamically adjust the output threshold of the intent classification model, which can meet the intent recognition needs of users at different stages and make intent recognition more flexible and accurate.

[0063] In terms of model optimization, a classification control algorithm is adopted. Based on the deviation between the target value of intent recognition and the actual recognition confidence, the parameter adjustment amount of the intent classification model is calculated, and a model optimization signal is generated. Through closed-loop control, the intent recognition confidence is stabilized at the target value, thereby achieving accurate customer service intent recognition and improving the adaptability and accuracy of the model.

[0064] Regarding the adjustment of word segmentation algorithm parameters, the accuracy feedback signal is obtained through the intent feedback module and judged in combination with the preset accuracy range. When the accuracy exceeds the range, the parameter optimization algorithm is used to calculate the parameter adjustment amount of the word segmentation algorithm, optimize the text preprocessing effect, stabilize the intent recognition accuracy within the preset range, and further improve the accuracy of intent recognition.

[0065] In terms of optimizing the timing of control signal output, the response time of the intent classification model and the word segmentation algorithm are combined to optimize the timing of control signal output, making the semantic analysis and classification process of user intent recognition smooth and stable, and improving the stability and reliability of the system.

[0066] By monitoring the changes in user text information and intent recognition parameters in real time, the preset parameters are dynamically updated to match changes in user intent state, enabling accurate identification of customer service intent. The relevant data is stored in the database to optimize the machine learning model, allowing the system to continuously learn and evolve, thereby continuously improving the accuracy and adaptability of intent recognition and better meeting the ever-changing needs of users. Attached Figure Description

[0067] Figure 1 This is a schematic diagram illustrating the working principle of the customer service intent recognition method based on natural language processing described in this invention.

[0068] Figure 2 Design diagram for text preprocessing;

[0069] Figure 3 Design diagram for dynamically adjusting confidence level for intent recognition;

[0070] Figure 4Design diagram of the classification control algorithm;

[0071] Figure 5 Design diagram for the execution of parameter optimization algorithm. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figures 1-5 The present invention relates to a customer service intent recognition method based on natural language processing, the specific implementation steps of which are as follows:

[0074] The original user text information is obtained from the customer service dialogue system. The text is preprocessed using a word segmentation algorithm to remove stop words and perform word form restoration to obtain standardized user text information.

[0075] Based on standardized user text information, semantic features of the text are calculated, and the frequency and trend of user intent are extracted. An intent classification algorithm is used to determine whether the user intent is at the consultation or complaint stage. If it is at the consultation stage, the required intent recognition confidence threshold is calculated based on preset intent recognition parameters, and the intent recognition target value is dynamically adjusted in conjunction with the semantic change trend of the user text information.

[0076] A classification control algorithm is adopted to calculate the parameter adjustment amount of the intent classification model based on the deviation between the target value of intent recognition and the current actual recognition confidence level, generate a model optimization signal and output it to the intent classification model execution module.

[0077] The current intent recognition accuracy value is obtained through the accuracy feedback signal from the intent feedback module. Combined with the preset accuracy range, it is determined whether the parameters of the word segmentation algorithm need to be adjusted. If the current accuracy value exceeds the preset range, the parameter optimization algorithm is used to calculate the parameter adjustment amount of the word segmentation algorithm based on the accuracy deviation value, and a control signal is generated and output to the word segmentation algorithm execution module.

[0078] Based on the response time of the intent classification model and word segmentation algorithm, the output timing of the control signal is optimized to ensure a smooth and stable semantic analysis and classification process for user intent recognition. Furthermore, by monitoring the changing trends of user text information and intent recognition parameters in real time, preset parameters are dynamically updated to match changes in user intent state, enabling accurate recognition of customer service intent.

[0079] Example 1: In the customer service intent recognition method based on natural language processing, text preprocessing is a crucial step to ensure the accuracy of subsequent semantic analysis and intent classification. This example details how to standardize the original user text information through a word segmentation algorithm, including multi-granularity segmentation, stop word filtering, word form reduction, and smoothing processing, and finally generate coherent and structured text data to provide high-quality input for intent recognition.

[0080] The original text information generated during the interaction between users and the customer service system usually contains colloquial expressions, redundant words, and meaningless stop words, which can affect the accuracy of intent recognition. Therefore, the original text needs to be input into the word segmentation algorithm model first. This word segmentation algorithm adopts a multi-granularity segmentation strategy, which can decompose the text into word subsequences of different lengths according to the context. For example, for longer compound words or proper nouns in specific domains, the algorithm will preferentially retain their complete semantic units, while for regular words, they will be segmented according to the smallest semantic unit. This flexible segmentation method ensures that the text can retain key information and adapt to the semantic analysis requirements of different scenarios.

[0081] After the initial segmentation, the system extracts the stop word component features from the generated word subsequences. The stop word filtering model is based on a predefined stop word library and dynamically adjusts the filtering threshold in combination with the context relevance. Specifically, the model will analyze the grammatical role and semantic contribution of each word in the sentence. If it belongs to high-frequency but low-information words (such as "de", "le", "ma", etc.), it will be marked as a stop word and filtered out. In addition, the model can also identify invalid expressions in specific scenarios, such as common polite expressions or repetitive statements in customer service conversations, further improving the purity of the text.

[0082] The word subsequences after filtering out the stop words need to be further processed for word form reduction. The goal of word form reduction is to restore words in different forms to their basic forms. For example, "in operation" and "ran away" are respectively restored to "run" and "run". This process depends on dictionary matching and a rule engine, and at the same time eliminates ambiguity in combination with the context semantics. For example, for the polysemous word "play", the system will judge its specific meaning according to the adjacent words (such as "make a phone call" is restored to "dial", "play basketball" is restored to "play"), ensuring the accuracy of the restoration result. After completing the word form reduction, the words in the text are presented in a standardized form, facilitating subsequent feature extraction and intent analysis.

[0083] To further improve text coherence, the system performs smoothing on standardized text. The smoothing operation primarily targets two types of issues: first, removing residual redundant expressions, such as when a user repeatedly mentions the same intent (e.g., "I want to ask, I just want to consult..."), the system will merge semantically repetitive segments; second, correcting sentences with loose grammatical structures or incomplete expressions, such as adding missing subjects or objects to make them conform to standard expressions. The smoothed text is semantically more compact and logically clearer, providing a higher-quality data foundation for intent recognition.

[0084] Finally, the system outputs user text information that has undergone a complete preprocessing workflow. The result has the following characteristics: First, the lexical granularity is appropriate, preserving key semantic units while avoiding semantic fragmentation caused by excessive segmentation; second, stop words and redundant information are effectively filtered out, reducing noise interference; and third, the word forms are consistent and conform to grammatical norms, facilitating subsequent vectorization representation and feature calculation. This standardized text data can be directly input into the intent calculation module for semantic feature extraction and intent state determination.

[0085] Throughout the preprocessing process, the system employs a dynamic adjustment mechanism to optimize the word segmentation algorithm parameters. For example, when detecting specialized terminology from a specific domain, the algorithm automatically relaxes the segmentation granularity to ensure terminology integrity; while when processing colloquial expressions, it activates stricter filtering rules to reduce the impact of informal expressions. Furthermore, the system monitors the preprocessing effect in real time. If it detects a decrease in the semantic coherence of the output text, it triggers a self-checking mechanism to recalibrate the stop word filtering threshold or lexical restoration rules, ensuring the stability of the preprocessing module.

[0086] Example 2: In a customer service intent recognition system, accurately understanding the semantic features of user text and determining the intent type is the core of achieving intelligent interaction. This example details how to extract semantic features based on standardized text information, analyze intent frequency and trends, and ultimately achieve accurate classification of inquiry and complaint intents.

[0087] The preprocessed standard text information first enters the semantic feature extraction stage. The system uses word embedding technology to transform discrete words into continuous vector representations. This process can capture the semantic relationships and contextual information between words. Each word is mapped to a vector point in a high-dimensional space according to its distribution characteristics in the corpus, with words with similar semantics being close in distance in the vector space. Through word vector superposition and weighted calculation, the system generates a semantic vector representation of the entire text, which comprehensively reflects the core meaning and sentiment of the text.

[0088] Building upon semantic vectors, the system performs dual feature extraction. Intent keyword features are determined by analyzing salient points in the word vector space; these points correspond to recurring or domain-specific words in the text. The system establishes a keyword weight model, scoring words based on their frequency of occurrence, positional distribution, and relevance to intent categories in historical data, selecting the most indicative word combinations for intent judgment. Simultaneously, sentiment characteristics are obtained by analyzing the overall distribution pattern of semantic vectors. The system identifies regions in the vector space associated with positive and negative emotions, calculates the projection intensity of text vectors into these regions, and thus quantifies the emotional polarity expressed by the user.

[0089] The intent frequency analysis module processes the extracted keyword features to construct an intent evolution map over time. The system uses a sliding window technique to segment and scan continuous dialogues, counting the frequency of specific intent keywords within each time window, and identifies periodic patterns in intent through time series analysis algorithms. For each detected intent cycle, the system records its start time, duration, and intensity change curve; this time-series data reflects the shift in user attention. Simultaneously, sentiment characteristics are input into the trend analysis engine, which calculates the rate and direction of intent sentiment evolution by comparing the difference in sentiment intensity between adjacent time windows, generating quantifiable trend indicators.

[0090] Based on a comprehensive analysis of frequency and trend data, the system establishes an intent state determination model. This model defines consultation intent as a behavioral pattern characterized by the following features: intent keywords are concentrated on consultative topics such as product functions and service processes; the sentiment tendency remains neutral or slightly positive; and it appears in the early stages of the conversation or the rising phase of the intent cycle. Correspondingly, complaint intent is characterized by: keywords involving negative content such as problem descriptions and expressions of dissatisfaction; a significantly negative sentiment tendency; and it mostly appears in the later stages of the conversation or the peak phase of the intent cycle. The system distinguishes between these two states through a decision boundary established by machine learning. When a user's expression simultaneously meets multiple conditions of consultation intent characteristics, the consultation intent determination is triggered; otherwise, the complaint intent determination process begins.

[0091] In real-time dialogue scenarios, the system continuously tracks the dynamic changes in intent states. Each new user input updates the semantic vector and feature data. The system compares this incremental information with the current intent cycle to determine if a critical point for state transition has occurred. For example, when it detects a shift in sentiment from positive to negative accompanied by a surge in complaint keywords, the system immediately labels the state as a complaint and adjusts its subsequent response strategy. All judgments are associated with specific timestamps, forming a complete sequence of intent states. This sequence not only records the intent category at each moment but also includes metadata such as intensity and duration, providing historical data for subsequent confidence calculations and parameter adjustments.

[0092] To ensure the accuracy of state determination, the system employs multiple verification mechanisms. First, the semantic feature extraction stage utilizes an ensemble learning approach, combining the results of multiple word embedding models for voting decisions to avoid bias from a single model. Second, intent frequency analysis incorporates an anomaly detection algorithm, capable of identifying and filtering noisy data caused by user expression habits or system errors. Finally, the state determination model possesses online learning capabilities, continuously optimizing its decision rules based on newly generated dialogue data to adapt to the linguistic characteristics of different industries and user groups.

[0093] The technical advantages of this implementation method are reflected in three aspects: First, through the collaborative analysis of semantic vectors and dual features, a multi-dimensional characterization of user intent is achieved; second, the temporal modeling method can capture the dynamic evolution of intent and avoid misjudgment caused by static analysis; and third, the adaptive mechanism of the judgment model ensures that the system maintains stable recognition performance in long-term use.

[0094] Example 3: In the dynamic adjustment mechanism of the customer service intent recognition system, the precise setting of the intent recognition confidence threshold plays a decisive role in the reliability of the classification results. This example details how to achieve intelligent calculation and real-time optimization of the confidence threshold for the consultation intent stage based on a neural network model and dynamic adjustment algorithm, ensuring that the system can adaptively match the evolution of user intent.

[0095] Once the system determines that the current intent is consultation-related, it immediately initiates the confidence threshold calculation process. Pre-set intent recognition parameters and real-time semantic vector data together constitute the input feature set of the neural network model. These parameters include system-level indicators such as historical accuracy statistics, domain knowledge weights, and dialogue context relevance, while the semantic vector data carries the complete semantic information of the current user's expression. The neural network model employs a deep network architecture design, containing multiple hidden layers and non-linear activation units, capable of capturing high-order interaction relationships between input features. During the training phase, the model has learned the complex mapping relationship between the optimal confidence threshold and input features in different consultation scenarios; therefore, during the inference phase, it can predict the most appropriate threshold suggestion value based on the current dialogue features.

[0096] The dynamic adjustment algorithm receives initial threshold suggestions from the neural network model and initiates a threshold optimization process. This algorithm employs a progressive approximation strategy, gradually adjusting the output threshold of the current intent classification model towards the target value. The adjustment step size is dynamically determined based on the changing trend of the semantic vector: when rapid changes in semantic content are detected, the algorithm increases the step size to accelerate the response; when the semantics tend to stabilize, the step size decreases to improve adjustment accuracy. After each adjustment, the system collects the latest intent recognition confidence data, calculates its deviation from the target value, and feeds this deviation back to the adjustment algorithm to form a closed-loop control. This mechanism ensures that the threshold adjustment process maintains sensitivity to changes in user intent while avoiding system oscillations caused by over-adjustment.

[0097] The system employs a dual-trigger mechanism to monitor the effectiveness of threshold adjustments. Firstly, the semantic change monitoring module continuously analyzes the distance between the newly generated semantic vector and historical vectors, immediately triggering a threshold recalculation upon detecting significant concept drift. This drift might manifest as a user suddenly introducing a new consultation topic or changing their expression. Secondly, the confidence bias analysis module tracks the deviation between the actual recognition results and the expected target in real time, also triggering a recalculation when the deviation exceeds a preset tolerance range. This dual-protection mechanism ensures the system can respond promptly to various unexpected situations, maintaining the rationality of the threshold settings.

[0098] When the triggering conditions are met, the system initiates a threshold recalculation process. The latest semantic vector data and updated system parameters are re-input into the neural network model, which generates new threshold suggestions based on the changed context. The dynamic adjustment algorithm immediately responds to this change and replans the adjustment path. It's important to note that the new adjustment process does not start from scratch but inherits accumulated experience data from previous adjustments, including implicit knowledge such as the regularity of semantic changes and user response patterns. This allows subsequent adjustments to converge to the optimal value more quickly. The system also sets a maximum adjustment limit. If the system remains unstable after reaching this limit, an exception handling process will be initiated, including expanding the input feature range of the neural network model or introducing manual intervention mechanisms.

[0099] During the adjustment process, the system maintains a complete threshold evolution log, recording the time, magnitude, triggering reason, and effect of each adjustment. This data is not only used for real-time monitoring but also periodically fed back into the training process of the neural network model, enabling continuous model optimization. Simultaneously, the system establishes a threshold stability assessment mechanism. By analyzing the trend of recent adjustments, it determines whether the current threshold has reached a stable state. When multiple consecutive fine-tunings are detected and the confidence deviation remains within a very small range, the system determines that stability has been achieved, temporarily suspends active adjustments, and switches to monitoring mode to conserve computing resources.

[0100] The technical implementation of this method has three significant characteristics: First, the combination of the neural network model and the dynamic adjustment algorithm ensures both the intelligence of the threshold calculation and the stability of the adjustment process; second, the dual triggering mechanism allows the system to balance response speed and stability requirements; and finally, a complete knowledge accumulation and transfer mechanism enables the system to continuously improve. These characteristics work together to ensure that the customer service intent recognition system maintains accurate judgment in the face of complex and ever-changing consultation scenarios, providing users with a high-quality interactive experience.

[0101] In actual operation, the system demonstrates excellent environmental adaptability. For routine consultation scenarios, threshold adjustments typically converge quickly within 3-5 iterations; faced with sudden topic shifts or changes in expression, the system can readjust within an acceptable timeframe. The entire adjustment process is completely transparent to end users, causing no interaction delays or service interruptions. System administrators can observe the dynamic changes in thresholds through a visual monitoring interface to understand the system's current adaptation status, but in most cases, no manual intervention is required. This highly autonomous intelligent adjustment mechanism significantly reduces system maintenance costs while ensuring service quality stability.

[0102] Example 4: In the actual operation of the intelligent customer service system, the dynamic adjustment mechanism of the classification control algorithm directly affects the accuracy and stability of intent recognition. Taking a bank credit card customer service scenario as an example, when a user inquires about "how to apply for an increase in credit card limit," the system first sets the intent recognition target value to 0.92 (based on historical data of this business scenario), but the actual recognition confidence is only 0.87. At this point, the classification control algorithm immediately initiates the parameter adjustment process. After accurately calculating the deviation value of 0.05, the system analyzes the current dialogue context features to determine the weight parameters of keywords related to "limit adjustment" that need to be emphasized, while fine-tuning the node parameters in the neural network that process application process-related expressions. After three iterations of adjustment, the actual recognition confidence steadily increases to 0.91, and the deviation from the target value is reduced to an acceptable range. At this point, the system automatically switches to maintenance and monitoring mode.

[0103] The core of the classification control algorithm lies in constructing a multi-layered parameter adjustment system. When a customer inquires about a "screen flickering problem with a newly purchased phone" on an e-commerce platform, the system finds that the actual recognition confidence level (0.85) is significantly lower than the target value (0.94) for return-related inquiries. The algorithm first decomposes the components of the bias: 0.04 comes from insufficient recognition of technical descriptions like "screen malfunction," 0.03 stems from the bias in sentiment judgment regarding "return intention," and the remaining 0.02 belongs to the basic semantic understanding error. Based on this structured analysis result, the system adjusts the model parameters in three dimensions: enhancing the recognition weight of hardware malfunction terms, optimizing the threshold parameters for negative sentiment analysis, and fine-tuning the connection strength of the basic semantic understanding layer. This precise targeted adjustment avoids the potential side effects of a "one-size-fits-all" parameter change.

[0104] The generation and execution process of the model optimization signal demonstrates adaptive technical characteristics. Taking the aviation customer service scenario as an example, when a passenger inquires about "how to issue a flight delay certificate," the system identifies the request as belonging to the "certificate acquisition" subcategory, but the confidence level (0.89) does not reach the standard threshold (0.93) for this subcategory. The parameter adjustment calculation module comprehensively considers multiple factors: the current dialogue round (round 3), historical interaction data (the user has confirmed the flight delay), business priority (delay certificates are a high-time-sensitivity requirement), etc., and finally generates an optimization signal package containing 17 specific parameter adjustment instructions. These instructions are synchronously sent to various functional modules of the intent classification model through a dedicated model execution channel, including the word vector layer, attention mechanism layer, and classification output layer, to achieve accurate distributed parameter updates.

[0105] In insurance claims consultation scenarios, closed-loop control mechanisms demonstrate unique value. When a user describes "what materials are needed for insurance claims after a car accident," the system initially recognizes a confidence level of 0.86. After the first parameter adjustment, this increases to 0.90, but still falls short of the target value of 0.93. At this point, the control system automatically initiates a secondary adjustment strategy: on the one hand, it retains the original parameter adjustment direction, and on the other hand, it introduces an enhanced recognition module for claims-related terminology. When the confidence level reaches 0.92, the system detects that the improvement rate is beginning to decay and intelligently switches to fine-tuning mode, making nano-level adjustments only to the three most critical parameters (the adjustment range is 1 / 10 of the initial value). Finally, in the fifth iteration, the confidence level stabilizes at 0.928. The entire adjustment process forms a complete "monitoring-analysis-adjustment-verification" cycle, with each cycle optimizing based on the previous results.

[0106] This implementation demonstrates excellent adaptability in cross-industry applications. In educational institution consultation scenarios, when parents inquire about "summer English course arrangements," the system needs to process multiple features simultaneously: course type (English), time attribute (summer), and query object (arrangement). The classification control algorithm establishes a multi-dimensional parameter association matrix to ensure that adjusting the parameters for recognizing English technical terms does not affect the processing accuracy of the time expression. In medical consultation scenarios, when patients inquire about "whether aspirin and ibuprofen can be taken simultaneously," the algorithm specifically strengthens the parameter weights for professional knowledge such as drug interactions, while maintaining the stability of recognizing conventional parameters such as dosage and usage. This domain-oriented parameter adjustment strategy enables the system to maintain high-precision intent recognition capabilities in different professional scenarios.

[0107] Specifically, regarding the technical implementation of parameter adjustments, the system adopts a layered and gradual adjustment strategy. Taking a cross-border e-commerce scenario as an example, when a foreign buyer inquires about "tariff calculation methods," the first round of adjustments primarily targets the basic semantic understanding layer, addressing the recognition of technical terms such as "tariff." The second round shifts to the business logic layer, strengthening the association analysis of operational terms such as "calculation" and "method." The third round focuses on the contextual understanding layer, making a comprehensive judgment based on information such as the product price and destination country confirmed in previous conversations. Each adjustment stage has an independent evaluation mechanism; the system only moves to the next stage after the current stage's goal is achieved, avoiding system instability caused by excessive adjustments.

[0108] During model parameter updates, the system employs "hot-switching" technology to ensure service continuity. When a user on a travel website inquires about a refund or change policy for a particular situation, the conversation flow remains uninterrupted even while the system is adjusting parameters. The old and new parameter versions run in parallel in memory, and once the new parameters pass integrity verification, the system completes the switch within milliseconds, making the adjustment process completely imperceptible to the user. For complex inquiries such as "international connecting flight ticket refund and change rules," the system also supports partial module canary updates, prioritizing adjustments to the most critical module parameters that significantly impact the current conversation, while other modules are silently updated in the background. This design significantly improves the system's response speed and service quality.

[0109] Anomaly handling mechanisms are a crucial component of this implementation. When encountering rare inquiries such as "antique auction deposit payment methods," the system initiates a special processing procedure upon detecting persistently low confidence levels: first, it expands the parameter adjustment range and activates backup model paths; simultaneously, it automatically searches for similar historical cases to find adjustment references; and finally, it retains complete debugging logs for subsequent analysis. This mechanism ensures that the system maintains a basic service level even when facing edge cases, while accumulating valuable data for subsequent model optimization. The entire process is fully automated, requiring no manual intervention, demonstrating the self-healing capabilities of the intelligent system.

[0110] Continuous observation of real-world dialogue scenarios reveals that this classification control mechanism intelligently distinguishes between critical parameters requiring significant adjustment and fundamental parameters that should remain stable. In a logistics query case, when a user inquires about "how to query cold chain transportation temperature records," the system prioritizes adjusting the recognition parameters for specialized fields such as "cold chain" and "temperature records," while keeping the processing parameters for common query terms (such as "how" and "query") unchanged. This precise, differentiated adjustment strategy solves the recognition accuracy problem in specific scenarios while maintaining the overall stability of the system, preventing global fluctuations caused by local adjustments. As the system runs longer, the accuracy and efficiency of this parameter adjustment will continue to improve, forming a virtuous cycle of self-evolution.

[0111] Example 5: In the actual operation of an intelligent customer service system, the dynamic maintenance of intent recognition accuracy is a key link in ensuring service quality. Taking a telecom operator's customer service system as an example, when a large number of users inquired about "5G package upgrades," the system detected a recent downward trend in intent recognition accuracy, triggering a word segmentation algorithm parameter adjustment mechanism. The system first analyzed the problem characteristics: user inquiries frequently included combinations of technical terms such as "5G+" and "SA / NSA," and the existing word segmentation granularity caused these key information to be incorrectly segmented. The control system then issued an instruction to adjust the word segmentation granularity of technical terminology-related fields from the default fine granularity to medium granularity, ensuring that business keywords such as "5G package" and "network speed improvement" remained intact. After the adjustment, the system collected new accuracy data and found that the recognition accuracy for package-related inquiries had significantly improved, but ordinary business queries showed a slight decline. Therefore, a balance fine-tuning was initiated, ultimately finding the optimal parameter combination that took into account various types of queries.

[0112] The process of accuracy monitoring and parameter adjustment forms a complete adaptive loop. In the customer service system of an online education platform, during the peak season for summer course inquiries, the system detected that the accuracy of queries related to "course schedules" was lower than the preset standard. Analysis showed that the problem stemmed from users' frequent use of abbreviated expressions such as "junior high school math summer and autumn class schedules," which the original word segmentation strategy struggled to handle correctly. The system implemented a tiered adjustment strategy: first, it appropriately increased the basic word segmentation granularity to ensure that specific combinations such as "summer and autumn classes" remained intact; second, it activated a dedicated dictionary for the education field to strengthen the recognition of subject-specific terms such as "junior high school" and "mathematics"; and finally, it optimized the stop word filtering rules to retain words such as "class" and "semester" that have practical significance in educational contexts. After multiple rounds of gradual adjustments, the system restored the accuracy of course schedule intent recognition to an ideal level without affecting other types of queries.

[0113] Intelligent decision-making mechanisms are particularly important during parameter adjustment. In a case study of airline customer service, the system found that the accuracy of "baggage check-in" related queries fluctuated significantly. In-depth analysis revealed that different users expressed the same question in significantly different ways: there were formal expressions like "baggage allowance inquiry" and colloquial expressions like "can my suitcase be checked in for free?" Instead of simply adopting a single adjustment solution, the system implemented a differentiated strategy: for formal expressions, the existing word segmentation rules were maintained, with a focus on optimizing the grammar analysis module; for colloquial expressions, the word segmentation flexibility was adjusted to allow for cross-grammatical unit combinations such as "suitcase" and "baggage check-in." At the same time, an association mapping of expression variants was established to ensure that different expressions accurately point to the same intent category. This multi-dimensional adjustment approach effectively improved the system's adaptability to diverse expressions.

[0114] Historical data analysis plays a crucial role in parameter optimization. A bank's credit card customer service system accumulated nearly three years of user query data. When it detected a decline in the accuracy of inquiries about "installment payments," the system first traced historical data to identify the evolution of user expression patterns: early on, standardized terms like "bill installment application" were prevalent, while more recently, natural language expressions such as "split the repayment into several months" have become popular. Based on this trend analysis, the system dynamically adjusted its word segmentation strategy: retaining accurate segmentation of technical terms while enhancing its ability to recognize colloquial numerical expressions (such as "split into 12 months"). Furthermore, the system established a seasonal adjustment model to pre-optimize word segmentation parameters related to consumer credit during specific periods such as shopping festivals, achieving proactive performance maintenance.

[0115] Feedforward control mechanisms enable systems to make predictive adjustments. A government service platform, anticipating a surge in related inquiries before policy and regulatory updates, analyzed historical patterns. The system proactively implemented three measures: importing the latest policy terminology database, enhancing word segmentation rules for common conjunctions in regulatory clauses (such as "Article X, Paragraph X"), and adjusting parameters for handling long sentences specific to legal texts. When users actually began inquiring about "the newly issued implementation details of the XX regulations," the system already possessed optimized processing capabilities, accurately identifying key elements such as "regulations" and "details" and their modifying relationships, thus avoiding the accuracy drop common during policy change periods. This proactive, prediction-based adjustment significantly improved the system's ability to respond to sudden changes in queries.

[0116] Parameter adjustment in multilingual environments requires special consideration. A customer service system on an international e-commerce platform handles mixed Chinese and English queries, such as "How to track my order." The system employs a layered adjustment strategy: first, it identifies the language mixing characteristics and distinguishes the boundaries between Chinese and English; then, it preserves the original English words without segmentation, while processing the Chinese words according to optimized rules; finally, it establishes cross-language associations to ensure that expressions like "track" and "order status query" point to the same intent. When it detects that users in certain regions habitually use pinyin abbreviations such as "DDL" instead of "deadline," the system dynamically expands its word segmentation dictionary to add these regional variations without affecting the processing rules for mainstream expressions. This refined multilingual parameter management enables the system to adapt to the diverse expression habits of users worldwide.

[0117] The parameter protection mechanism under abnormal conditions ensures system stability. When a government hotline system encounters a large number of malicious inquiries, the accuracy monitoring module detects abnormal fluctuations. The system activates a protection protocol: first, it freezes non-critical parameter adjustments to maintain the stability of core word segmentation rules; second, it enhances input validation to filter meaningless character combinations; and finally, it activates an anti-interference mode to prioritize the processing quality of routine government inquiries. Only after the abnormal traffic subsides does the system gradually restore its full parameter adjustment functionality. This self-protection mechanism avoids system performance degradation caused by improper adjustments under abnormal conditions, demonstrating the robust design expected of industrial-grade applications.

[0118] Collaborative management of parameter adjustments is an essential capability for complex systems. A certain automotive brand's omnichannel customer service system needed to maintain a consistent experience across different terminals, including the official website's online customer service and 4S store voice assistants. When discrepancies were found in the recognition accuracy of "maintenance appointment" queries across different channels, the system implemented a collaborative adjustment plan: first, it analyzed the expression characteristics of each channel (online text inquiries often use complete sentences, while voice interactions commonly use abbreviated expressions); then, it formulated channel-specific fine-tuning strategies; and finally, it ensured the consistency of core rules through a central parameter management system. This principle of "unified core, flexible periphery" adjustment adapts to the characteristics of different interaction scenarios while maintaining the consistency of the brand's service experience.

[0119] Long-term parameter optimization accumulates to form a system knowledge base. A large e-commerce customer service system continuously records each accuracy anomaly and its solution, building a rich library of adjustment cases. When encountering similar problems again, such as a drop in accuracy for queries related to "promotional rules," the system intelligently matches historical cases and discovers that a similar situation occurred during a product promotion due to improper handling of promotional terms such as "first X customers" and "deposit inflation." Based on this historical experience, the system quickly locates the current problem as potentially stemming from newly emerging marketing phrases such as "group purchase cashback" and "cross-store discounts," and adjusts the relevant word segmentation rules and intent mapping relationships accordingly. This continuous learning mechanism enables the system's parameter optimization efficiency to continuously improve over time, forming a virtuous cycle of self-improvement.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A customer service intent recognition method based on natural language processing, characterized in that: Includes the following steps: The original user text information is obtained from the customer service dialogue system. The text is preprocessed using a word segmentation algorithm to remove stop words and perform word form restoration to obtain standardized user text information. Based on standardized user text information, the semantic features of the text are calculated, the frequency and trend of user intent are extracted, and the intent classification algorithm is used to determine whether the user intent is in the consultation or complaint stage. When it is in the consultation stage, the required intent recognition confidence threshold is calculated according to the preset intent recognition parameters, and the intent recognition target value is dynamically adjusted in combination with the semantic change trend of user text information. A classification control algorithm is adopted to calculate the parameter adjustment amount of the intent classification model based on the deviation between the intent recognition target value and the current actual recognition confidence level, generate a model optimization signal and output it to the intent classification model execution module; The current intent recognition accuracy value is obtained through the accuracy feedback signal of the intent feedback module. Combined with the preset accuracy range, it is determined whether the parameters of the word segmentation algorithm need to be adjusted. If the current accuracy value exceeds the preset range, the parameter optimization algorithm is used to calculate the parameter adjustment amount of the word segmentation algorithm based on the accuracy deviation value, and a control signal is generated and output to the word segmentation algorithm execution module. Based on the response time of the intent classification model and word segmentation algorithm, the output timing of the control signal is optimized to make the semantic analysis and classification process of user intent recognition smooth and stable. Then, by monitoring the changing trends of user text information and intent recognition parameters in real time, the preset parameters are dynamically updated to match the changes in user intent state and to accurately identify customer service intent.

2. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: The text is preprocessed using a word segmentation algorithm to remove stop words and perform lemmatization, resulting in standardized user text information, specifically including: User text information data is input into the word segmentation algorithm model, and then the word segmentation algorithm is used to perform multi-granular segmentation of user text information, decomposing user text information into word subsequences of different lengths; Based on the word subsequences obtained from the decomposition, stop word component features are extracted. The extracted stop word component features are input into the stop word filtering model. Based on the input stop word component features, the filtering threshold is adaptively adjusted to perform adaptive filtering on the word subsequences and remove the stop word components. After filtering out stop words, each word subsequence is subjected to lemmatization to obtain standardized user text information. The standardized user text information is then smoothed to eliminate residual redundant expressions in the text, resulting in coherent user text information. The smoothed user text information is output as the final user text information processing result, which is used for semantic and frequency analysis of user intent.

3. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: Based on standardized user text information, semantic features of the text are calculated, the frequency and trend of user intent are extracted, and an intent classification algorithm is used to determine whether the user intent is for consultation or complaint purposes. Specifically, this includes: Based on standardized user text information, word embedding is performed to obtain the semantic vector representation of the user text information. Then, the semantic features of the user text information are extracted, including intent keywords and sentiment features. Based on the extracted intent keyword features, the frequency of user intent occurrence is calculated to obtain the start and end times of each user intent cycle. Then, based on the extracted sentiment characteristics, the trend of user intent change is calculated to obtain the curve of intent type change over time. Based on the frequency and trend of user intent, determine whether the current user intent is in the consultation stage or the complaint stage. When the user intent is consultation-oriented and is in the first half of the user intent cycle, it is judged as consultation-oriented. When the user intent is complaint-oriented and is in the second half of the user intent cycle, it is judged as complaint-oriented. Based on the determined user intent state, the current user intent state is output, and the user intent state is associated with the corresponding dialogue time point to obtain a complete user intent state sequence.

4. The customer service intent recognition method based on natural language processing according to claim 3, characterized in that: After determining whether a user's intent is for consultation or complaint through an intent classification algorithm, the process also includes: Based on the preset intent recognition parameters and the real-time detected semantic vector data, these are used as input to the neural network model. Through the trained neural network model, the intent recognition confidence threshold required for the current consultation stage is calculated. Based on the calculated confidence threshold for intent recognition, combined with semantic vector change trend data, a dynamic adjustment algorithm is used to dynamically adjust the output threshold of the intent classification model so that it gradually approaches the target value. During the adjustment process, changes in semantic vectors and intent recognition confidence are detected in real time. When the semantic vectors change significantly or the intent recognition confidence deviates greatly from the target value, the neural network model is triggered to recalculate the intent recognition confidence threshold. The new intent recognition confidence threshold calculated by the neural network model is input into the dynamic adjustment algorithm to dynamically adjust the output threshold of the intent classification model again until the intent recognition confidence stabilizes near the new target value.

5. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: A classification control algorithm is employed to calculate the parameter adjustment amount of the intent classification model based on the deviation between the target value of intent recognition and the current actual recognition confidence level. This generates a model optimization signal and outputs it to the intent classification model execution module. Specifically, this includes: Obtain the target value and the actual recognition confidence value for intent recognition, calculate the deviation between the two, input the deviation value into the classification control algorithm, and calculate the parameter adjustment amount of the intent classification model based on the weight, offset, and learning rate parameters of the classification control algorithm. Based on the calculated adjustment amount of the intent classification model parameters, a corresponding model optimization signal is generated, and the generated model optimization signal is output to the intent classification model execution module to control the parameter update of the intent classification model. The intent classification model adjusts its internal parameters based on the received model optimization signal, thereby changing the actual confidence level of intent recognition. By continuously acquiring real-time intent recognition confidence values ​​and comparing them with target values, the deviation value is calculated to form a closed-loop control.

6. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: By obtaining the accuracy feedback signal from the intent feedback module, the current intent recognition accuracy value is acquired. Combined with the preset accuracy range, it is determined whether the parameters of the word segmentation algorithm need to be adjusted, specifically including: The accuracy value data is transmitted to the control system and compared with the preset accuracy threshold range to determine whether the current accuracy is within a reasonable range. When the accuracy value is lower than the lower limit threshold of the preset range, the control system issues an instruction to increase the segmentation granularity of the word segmentation algorithm through the actuator. When the accuracy value is higher than the upper limit of the preset range, the control system issues an instruction to reduce the segmentation granularity of the word segmentation algorithm through the actuator, thereby optimizing the preprocessing effect of the user's text information. After the parameters of the word segmentation algorithm are adjusted, the standardized text information output by it is obtained again until the accuracy value is stable within the preset range. During parameter adjustment, historical accuracy data is analyzed using machine learning algorithms to establish a mapping model between accuracy values ​​and word segmentation parameters. Based on the current accuracy value, the optimal word segmentation parameters for the next moment are predicted using the established mapping model, and feedforward control of the word segmentation algorithm is performed.

7. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: A parameter optimization algorithm is used to calculate the parameter adjustment amount of the word segmentation algorithm, generate control signals, and output them to the word segmentation algorithm execution module. Specifically, this includes: The current system accuracy value is obtained in real time through the intent feedback module, and the accuracy value is compared with the preset accuracy range to determine whether the accuracy value exceeds the preset range. When the accuracy value exceeds the preset range, the deviation between the current accuracy value and the value in the preset accuracy range is calculated. The deviation value is used as the input of the parameter optimization algorithm. Based on the accuracy deviation value, the parameter adjustment amount of the word segmentation algorithm is calculated through rule reasoning using a pre-established parameter optimization rule library. The calculated word segmentation algorithm parameter adjustment amount is converted into a corresponding control signal. The digital signal is converted into an analog signal through a digital-to-analog converter. The generated analog control signal is output to the word segmentation algorithm execution module to drive the execution module to adjust the word segmentation algorithm parameters. The word segmentation algorithm changes the segmentation parameters according to the adjustment actions of the execution module, thereby optimizing the preprocessing effect of user text information and bringing the accuracy value back to the preset range.

8. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: Based on the response time of the intent classification model and word segmentation algorithm, the output timing of the control signal is optimized to make the semantic analysis and classification process of user intent recognition smooth and stable. Specifically, this includes: Obtain the response time parameters of the intent classification model and word segmentation algorithm, establish a dynamic model of the model, obtain the relationship curves between model parameters and accuracy and confidence, and calculate the target parameters of the model and the timing of control signal output through the dynamic model of the model based on the target values ​​of intent recognition accuracy and confidence. Then, using the model parameter deviation and accuracy confidence deviation as inputs, the optimized adjustment amount of the control signal is obtained through rule reasoning. The optimized control signal is then sent to the intent classification model and word segmentation algorithm in sequence according to the calculated output timing to dynamically adjust the model parameters. During the dynamic adjustment of the model, text information and intent recognition results from the customer service dialogue system are collected in real time and compared with the target value to obtain the deviation of accuracy and confidence. When the accuracy deviation or confidence deviation exceeds the preset threshold, the deviation is used as a feedback signal and input into the parameter optimization algorithm for further optimization and adjustment of the control signal.

9. The customer service intent recognition method based on natural language processing according to claim 1, characterized in that: By monitoring the changing trends of user text information and intent recognition parameters in real time, and dynamically updating preset parameters to match changes in user intent states, accurate identification of customer service intent is achieved, specifically including: The system acquires the user's real-time text information and current intent recognition parameters, compares them with preset normal range thresholds, and determines whether they exceed the normal range. When the real-time text information or intent recognition parameters exceed the normal range, an alert is triggered, and the magnitude of the intent recognition parameters that need to be adjusted is determined based on the degree of deviation from the normal range. Using machine learning algorithms, based on the changing trends of users' historical text information and intent recognition parameters, the system predicts the changing trends of users' intent states over a future period of time; then, based on users' real-time text information, current intent recognition parameters, and predicted future intent state changing trends, the system dynamically adjusts the intent recognition parameters to generate new intent recognition parameter settings. The dynamically adjusted intent recognition parameter settings are sent to the intent recognition system, and the control device recognizes the user's intent in real time according to the new parameter settings to match the user's real-time intent status. After adjusting the intent recognition parameters, continuously monitor changes in the user's text information to determine whether the user's intent state has returned to normal. If it has not returned to normal, return to continue adjusting. The user's real-time text information, intent recognition parameters, and adjusted results are stored in a database to optimize the machine learning model.

10. A customer service intent recognition system based on natural language processing, used to implement the customer service intent recognition method based on natural language processing as described in any one of claims 1-9, characterized in that: include: The text processing module obtains user text information from the customer service dialogue system, uses word segmentation algorithms to segment it at multiple granularities, extracts and filters out stop word components, performs word form restoration and smoothing processing, and finally outputs coherent user text information. The intent calculation module calculates semantic features based on standardized user text information, extracts the frequency and trend of user intent occurrence, and determines the current user intent state. When in the consultation stage, it combines preset parameters and real-time semantic vector data to calculate the intent recognition confidence threshold through a neural network model, and uses a dynamic adjustment algorithm to dynamically adjust the output threshold of the intent classification model to meet the intent recognition needs of users at different stages. The classification control module employs a classification control algorithm. Based on the deviation between the target value of intent recognition and the actual recognition confidence level, it calculates the parameter adjustment amount of the intent classification model, generates a model optimization signal, and outputs it to the intent classification model execution module. This controls the update of the intent classification model parameters, adjusts the model output to change the actual confidence level of intent recognition, and stabilizes the intent recognition confidence level at the target value through closed-loop control, thereby achieving accurate customer service intent recognition. The intent adjustment module obtains the current intent recognition accuracy value through the accuracy feedback signal of the intent feedback module, and makes a judgment based on the preset accuracy range. When the accuracy value exceeds the preset range, it calculates the parameter adjustment amount of the word segmentation algorithm based on the accuracy deviation value, generates a control signal and outputs it to the execution module to drive it to adjust the word segmentation algorithm parameters, optimize the text preprocessing effect, and stabilize the intent recognition accuracy within the preset range. The dynamic optimization module integrates the response time of the intent classification model and word segmentation algorithm, optimizes the timing of control signal output, and makes the semantic analysis and classification process of user intent recognition smooth and stable. By monitoring the changes in user text information and intent recognition parameters in real time, it dynamically updates preset parameters, matches changes in user intent state, accurately identifies customer service intent, and stores relevant data in the database for optimizing machine learning models.

Images