Customer service intention recognition method based on natural language processing
Through multi-granularity segmentation, adaptive filtering of stop words and word form restoration processing, combined with neural network models and classification control algorithms, the intent recognition parameters are dynamically adjusted, which solves the shortcomings of traditional customer service intent recognition methods, achieves efficient and accurate intent recognition, adapts to changes in user intent, and improves customer service quality.
Patent Information
- Application Number
- CN202510917033.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-10-14
- Estimated Expiration
- 2045-07-03
AI Technical Summary
Traditional customer service intent recognition methods have shortcomings in text preprocessing, semantic feature extraction, model optimization, and control signal output timing optimization, resulting in inaccurate and inefficient intent recognition, making it difficult to meet the needs of complex application scenarios.
Adopting a method based on natural language processing, through multi-granularity segmentation, adaptive filtering of stop words, word form restoration and smoothing processing, semantic features and sentiment tendencies are extracted. Combined with neural network models and classification control algorithms, the intent recognition parameters are dynamically adjusted, the control signal output timing is optimized, and changes in user intent are monitored in real time to achieve accurate customer service intent recognition.
It improves the quality of text processing, enhances the accuracy and flexibility of intent classification, enhances the adaptability of the model and the stability of the system, ensures the accuracy and adaptability of intent recognition, and meets the diverse needs of users.
Smart Images

Figure CN120780812A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of natural language processing, in particular to a customer service intent recognition method based on natural language processing. BACKGROUND
[0002] In the modern customer service field, with the rapid development of information technology and the increasing diversification of user needs, efficient and accurate recognition of user intent has become the key to improving the quality and efficiency of customer service. Traditional customer service intent recognition methods have many limitations and cannot meet the current complex application scenario requirements.
[0003] Traditional methods have limited capabilities in text preprocessing. For user input text information, a simple word segmentation method is often used, which cannot perform multi-granularity segmentation and cannot effectively process word subsequences of different lengths and structures. Moreover, when removing stop words, a fixed filtering threshold is usually used, which cannot be adjusted adaptively according to the text content, resulting in incomplete or excessive removal of stop words, affecting subsequent semantic analysis. At the same time, there is a lack of reduction of word forms and smoothing of redundant expressions, making the preprocessed text information not standardized and coherent, making it difficult for subsequent intent recognition.
[0004] In the intent recognition process, traditional methods do not fully and deeply extract semantic features. Only some simple keywords are focused on, while important features such as sentiment orientation are ignored, making it difficult to accurately grasp the trend of user intent. Moreover, it is difficult to accurately determine whether the user intent is in the consultation or complaint stage according to the frequency and trend of user intent, resulting in inaccurate intent classification.
[0005] In terms of model optimization and parameter adjustment, traditional methods lack effective closed-loop control mechanisms. The parameter adjustment of the intent classification model often relies on human experience and cannot automatically adjust according to the deviation between the intent recognition target value and the actual recognition confidence, resulting in poor adaptability and accuracy of the model. At the same time, the parameter adjustment of the word segmentation algorithm is not intelligent enough to dynamically optimize according to the feedback of intent recognition accuracy, affecting the effectiveness of text preprocessing.
[0006] Traditional methods have deficiencies in the optimization of output timing of control signals and dynamic updating of preset parameters. They cannot consider the response time of the intent classification model and the word segmentation algorithm to optimize the output timing of the control signal, making the semantic analysis and classification process not smooth and stable. Moreover, they cannot monitor the trend of user text information and intent recognition parameters in real time, dynamically update the preset parameters, and match the changes in user intent state, resulting in low accuracy of intent recognition.
[0007] With the development of artificial intelligence technology, natural language processing technology is being used more and more widely in the customer service field. However, the defects of the above-mentioned traditional methods seriously restrict the accuracy and efficiency of customer service intention recognition. There is an urgent need for a more advanced and intelligent customer service intention recognition method and system to solve the problems existing in the existing technology and improve the quality and level of customer service. Summary of the Invention
[0008] The purpose of the present invention is to provide a customer service intention recognition method based on natural language processing to solve the problems raised in the above background technology.
[0009] To achieve the above objectives, the present invention provides the following technical solution: a customer service intention recognition method based on natural language processing, the method comprising: Obtain original user text information from the customer service dialogue system, pre-process the text using a word segmentation algorithm, remove stop words, and perform lemmatization to obtain standardized user text information; Based on standardized user text information, the semantic features of the text are calculated, the frequency of occurrence and change 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. If the intent is in the consultation stage, the current required intent recognition confidence threshold is calculated based on the preset intent recognition parameters. Combined with the semantic change trend of the user text information, the intent recognition target value is dynamically adjusted; Using a classification control algorithm, the parameter adjustment of the intent classification model is calculated based on the deviation between the intent recognition target value and the current actual recognition confidence level, and a model optimization signal is generated and output 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 real-time monitoring of the changing trends of user text information and intent recognition parameters, the preset parameters are dynamically updated to match the changes in user intent status and accurately identify customer service intentions.
[0010] Preferably, a word segmentation algorithm is used to pre-process the text, remove stop words and perform lemmatization to obtain standardized user text information, specifically including: Input the user text information data into the word segmentation algorithm model, and then use the word segmentation algorithm to perform multi-granularity segmentation on the user text information, decomposing the user text information into word subsequences of different lengths; According to the decomposition of each word sub-sequence, the stop word component features are extracted, the extracted stop word component features are input into the stop word filtering model, the filtering threshold is adaptively adjusted according to the input stop word component features, and the stop word components in the word sub-sequence are adaptively filtered and removed. The filtered stop word of each word sub-sequence is subjected to morphological reduction processing to obtain standardized user text information, and the standardized user text information is subjected to smoothing processing to eliminate residual redundant expressions in the text to obtain coherent user text information. The user text information after smoothing processing is output as the final user text information processing result, which is used for semantic and frequency analysis of user intent.
[0011] Preferably, according to the standardized user text information, the semantic features of the text are calculated, the occurrence frequency and change trend of the user intent are extracted, and the user intent is determined by an intent classification algorithm to be a consultation or complaint intent stage, which specifically includes: According to the standardized user text information, word embedding processing is performed to obtain semantic vector representation of the user text information, and then the semantic features of the user text information are extracted, including intent keywords and sentiment tendency features; According to the extracted intent keyword features, the occurrence frequency of the user intent is calculated to obtain the start time and end time of each user intent period, and then according to the extracted sentiment tendency features, the change trend of the user intent is calculated to obtain the change curve of the intent type over time; According to the occurrence frequency and change trend of the user intent, it is determined whether the current user intent state is in a consultation stage or a complaint stage. When the user intent shows a consultation tendency and is in the first half of the user intent period, it is determined to be in a consultation stage. When the user intent shows a complaint tendency and is in the second half of the user intent period, it is determined to be in a complaint stage. According to 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.
[0012] Preferably, after determining the user intent to be in a consultation or complaint intent stage by the intent classification algorithm, it further includes: According to the preset intent recognition parameters and the real-time detected semantic vector data, the neural network model is input, and the trained neural network model is used to calculate the intent recognition confidence threshold required for the current consultation stage; According to the calculated intent recognition confidence threshold, combined with the 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; In the adjustment process, the semantic vector change and the intention recognition confidence change are detected in real time. When the semantic vector changes significantly or the intention recognition confidence deviates greatly from the target value, the neural network model is triggered to recalculate the intention recognition confidence threshold value; The new intention recognition confidence threshold value calculated by the neural network model is input into the dynamic adjustment algorithm to dynamically adjust the output threshold value of the intention classification model again until the intention recognition confidence stabilizes around the new target value.
[0013] Preferably, a classification control algorithm is used to calculate the parameter adjustment amount of the intention classification model according to the deviation between the intention recognition target value and the current actual recognition confidence, generate a model optimization signal and output it to the intention classification model execution module, which specifically includes: The target value and the actual recognition confidence value of the intention recognition are obtained, the deviation value between them is calculated, the deviation value is input into the classification control algorithm, and the parameter adjustment amount of the intention classification model is calculated according to the weight, offset and learning rate parameters of the classification control algorithm; According to the calculated intention classification model parameter adjustment amount, a corresponding model optimization signal is generated, and the generated model optimization signal is output to the intention classification model execution module to control the parameter update of the intention classification model; The intention classification model adjusts the internal parameters according to the received model optimization signal, thereby changing the actual confidence of the intention recognition. The real-time intention recognition confidence value is continuously obtained, compared with the target value, and the deviation value is calculated to form a closed-loop control.
[0014] Preferably, the current intention recognition accuracy value is obtained through the accuracy feedback signal of the intention feedback module, and whether the parameters of the word segmentation algorithm need to be adjusted is judged in combination with the preset accuracy range, which specifically includes: 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 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 threshold of the preset range, the control system issues an instruction to reduce the segmentation granularity of the word segmentation algorithm through the actuator to optimize the preprocessing effect of the user text information. After the parameter adjustment of the word segmentation algorithm is completed, the standardized text information output by the word segmentation algorithm is obtained again until the accuracy value stabilizes within the preset range; In the parameter adjustment process, the historical accuracy data is analyzed through a machine learning algorithm to establish a mapping model of the accuracy value and the word segmentation parameters. The established mapping model is used to predict the optimal word segmentation parameters at the next moment according to the current accuracy value for feedforward control of the word segmentation algorithm.
[0015] Preferably, the parameter adjustment amount of the word segmentation algorithm is calculated by using a parameter optimization algorithm, a control signal is generated and output to the word segmentation algorithm execution module, and specifically includes: The current system accuracy value is obtained in real time by the intention 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 value of the current accuracy value from the median value of the preset accuracy range is calculated, the deviation value is taken as the input of the parameter optimization algorithm, and the parameter adjustment amount of the word segmentation algorithm is calculated by using the pre-established parameter optimization rule library and rule-based reasoning according to the accuracy deviation value; The calculated word segmentation algorithm parameter adjustment amount is converted into a corresponding control signal, a digital signal is converted into an analog signal by a digital-to-analog converter, and the generated analog control signal is output to the execution module of the word segmentation algorithm to drive the execution module to adjust the parameters of the word segmentation algorithm; The word segmentation algorithm changes the segmentation parameters according to the adjustment action of the execution module, thereby optimizing the preprocessing effect of the user text information and making the accuracy value return to the preset range.
[0016] Preferably, according to the response time of the intention classification model and the word segmentation algorithm, the output timing of the control signal is optimized to make the semantic analysis and classification process of user intention recognition smooth and stable, and specifically includes: The response time parameters of the intention classification model and the word segmentation algorithm are obtained, a dynamic model is established, the relationship curve of the model parameters and the accuracy and confidence is obtained, the target parameters of the model and the output timing of the control signal are calculated by the dynamic model according to the target values of the intention recognition accuracy and confidence; The model parameter deviation and the accuracy and confidence deviation are taken as inputs, the optimization adjustment amount of the control signal is obtained by rule-based reasoning, and the optimized control signal is sent to the intention classification model and the word segmentation algorithm in turn according to the calculated output timing for dynamic adjustment of the model parameters; During the dynamic adjustment of the model, the text information and the intention recognition result data of the customer service dialogue system are collected in real time, which are compared with the target values to obtain the accuracy and confidence deviations; when the accuracy deviation or the confidence deviation exceeds the preset threshold, the deviation is taken as a feedback signal and input into the parameter optimization algorithm for further optimization adjustment of the control signal.
[0017] Preferably, by monitoring the change trend of the user text information and the intention recognition parameters in real time, the preset parameters are dynamically updated to match the change of the user intention state for accurate identification of the customer service intention, and specifically includes: The real-time text information and the current intention recognition parameter of the user are acquired, compared with a preset normal range threshold, and it is judged whether the normal range is exceeded; when the real-time text information or the intention recognition parameter exceeds the normal range, a pre-warning is triggered, and according to the degree of deviation from the normal range, the magnitude of the intention recognition parameter that needs to be adjusted is determined; Through a machine learning algorithm, based on the historical text information of the user and the change trend of the intention recognition parameter, the intention state change trend of the user in the future period of time is predicted; and then according to the real-time text information of the user, the current intention recognition parameter and the predicted future intention state change trend, the intention recognition parameter is dynamically adjusted to generate a new intention recognition parameter setting; The intention recognition parameter after dynamic adjustment is issued to the intention recognition system, and the device controls the real-time recognition of the user intention according to the new parameter setting to match the real-time intention state of the user; After adjusting the intention recognition parameter, the text information of the user is continuously monitored to judge whether the intention state of the user returns to normal, and if not, the adjustment is continued; The real-time text information of the user, the intention recognition parameter and the data of the effect after adjustment are stored in a database for optimizing the machine learning model.
[0018] Preferably, the application includes a customer service intention recognition system based on natural language processing, which is used to realize the customer service intention recognition method based on natural language processing as described above, and the system includes: A text processing module acquires user text information from a customer service dialogue system, uses a word segmentation algorithm to perform multi-granularity segmentation, extracts and filters out stop word components, then performs morphological restoration and smoothing processing, and finally outputs coherent user text information; An intention calculation module calculates the semantic features based on the standardized user text information, extracts the user intention occurrence frequency and change trend, judges the current user intention state, when in the consultation stage, combines the preset parameters and real-time semantic vector data, calculates the intention recognition confidence threshold through a neural network model, and dynamically adjusts the output threshold of the intention classification model through a dynamic adjustment algorithm to meet the intention recognition needs of the user in different stages; A classification control module uses a classification control algorithm to calculate the parameter adjustment amount of the intention classification model according to the deviation of the intention recognition target value and the actual recognition confidence, generates a model optimization signal output to the intention classification model execution module, controls the parameter update of the intention classification model, adjusts the model output to change the actual confidence of the intention recognition, and through closed-loop control, the intention recognition confidence is stabilized at the target value for accurate customer service intention recognition; The intention adjustment module obtains a current intention recognition accuracy value through the accuracy feedback signal of the intention feedback module, judges in combination with a preset accuracy range, when the accuracy value exceeds the preset range, calculates a parameter adjustment amount of the word segmentation algorithm by using a parameter optimization algorithm according to an accuracy deviation value, generates a control signal and outputs the control signal to the execution module, drives the execution module to adjust the word segmentation algorithm parameter, optimizes the text preprocessing effect, and makes the intention recognition accuracy stable within the preset range. The dynamic optimization module optimizes the output timing of the control signal in combination with the response time of the intention classification model and the word segmentation algorithm, makes the semantic analysis and classification process of user intention recognition smooth and stable, dynamically updates the preset parameter by monitoring the user text information and intention recognition parameter change trend in real time, matches the user intention state change, performs accurate identification of the customer intention, and stores the related data into the database for optimizing the machine learning model.
[0019] Compared with the prior art, the beneficial effects of the present application are: In the text preprocessing link, the multi-granularity segmentation, adaptive filter stop words, morphological restoration and smoothing processing can accurately obtain standardized and coherent user text information, lay a solid foundation for subsequent semantic analysis and intention recognition, and effectively improve the text processing quality.
[0020] In the semantic feature extraction and intention judgment aspect, the semantic vector is obtained through the word embedding processing, the intention keyword and the sentiment tendency feature are extracted, the user intention occurrence frequency and the change trend can be accurately calculated, the user intention is accurately judged to be in the consultation type or the complaint type stage, the basis is provided for the subsequent targeted intention recognition, and the accuracy of the intention classification is improved.
[0021] When being in the consultation type stage, the intention recognition confidence threshold is calculated by the neural network model in combination with the preset parameter and the real-time semantic vector data, the output threshold of the intention classification model is dynamically adjusted by using the dynamic adjustment algorithm, the intention recognition demand of the user in different stages can be met, the intention recognition is more flexible and accurate.
[0022] In the model optimization aspect, the classification control algorithm is adopted, the parameter adjustment amount of the intention classification model is calculated according to the deviation between the intention recognition target value and the actual recognition confidence, the model optimization signal is generated, the intention recognition confidence is stabilized at the target value through the closed-loop control, the accurate customer intention recognition is realized, and the adaptability and accuracy of the model are improved.
[0023] In the word segmentation algorithm parameter adjustment aspect, the accuracy feedback signal is obtained through the intention feedback module, the judgment is performed in combination with the preset accuracy range, when the accuracy exceeds the range, the parameter adjustment amount of the word segmentation algorithm is calculated by using the parameter optimization algorithm, the text preprocessing effect is optimized, the intention recognition accuracy is stabilized within the preset range, and the accuracy of the intention recognition is further improved.
[0024] In terms of control signal output timing optimization, the response time of the comprehensive intent classification model and the segmentation algorithm is considered, the control signal output timing is optimized, the semantic analysis and classification process of user intent recognition is smooth and stable, and the stability and reliability of the system are improved.
[0025] By monitoring the user text information and the trend of the intent recognition parameter in real time, the preset parameter is dynamically updated, the user intent state change is matched, the customer service intent is accurately identified, and the related data is stored in the database for optimizing the machine learning model, so that the system can continuously learn and evolve, continuously improve the accuracy and adaptability of intent recognition, and better meet the changing needs of users. BRIEF DESCRIPTION OF DRAWINGS
[0026] Figure 1 The working principle diagram of the customer service intent recognition method based on natural language processing is described. Figure 2 The design diagram of text preprocessing is described. Figure 3 The design diagram of intent recognition confidence dynamic adjustment is described. Figure 4 The design diagram of the classification control algorithm is described. Figure 5 The design diagram of the parameter optimization algorithm execution is described. DETAILED DESCRIPTION
[0027] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.
[0028] Please refer to Figures 1-5 The customer service intent recognition method based on natural language processing is described. The original user text information is obtained from the customer service dialogue system, the text is preprocessed by using the segmentation algorithm, the stop words are removed and the word shape is restored, and the standardized user text information is obtained.
[0029] According to the standardized user text information, the semantic features of the text are calculated, the occurrence frequency and the trend of the user intent are extracted, and the user intent is determined to be a consultation type or a complaint type by using the intent classification algorithm. When it is a consultation type, the required intent recognition confidence threshold is calculated according to the preset intent recognition parameter, and the target value of the intent recognition is dynamically adjusted in combination with the semantic trend of the user text information.
[0030] The classification control algorithm is adopted to calculate the parameter adjustment amount of the intention classification model according to the deviation between the target value of the intention recognition and the current actual recognition confidence, generate a model optimization signal and output the signal to an intention classification model execution module.
[0031] The current intention recognition accuracy value is obtained through the accuracy feedback signal of the intention feedback module, and it is judged whether the parameters of the word segmentation algorithm need to be adjusted in combination with the preset accuracy range. If the current accuracy value exceeds the preset range, the parameter adjustment amount of the word segmentation algorithm is calculated using the parameter optimization algorithm according to the accuracy deviation value, a control signal is generated and output to a word segmentation algorithm execution module.
[0032] According to the response time of the intention classification model and the word segmentation algorithm, the output timing of the control signal is optimized to make the semantic analysis and classification process of user intention recognition smooth and stable. Then, by monitoring the change trend of user text information and intention recognition parameters in real time, the preset parameters are dynamically updated to match the change of user intention state, and the accurate identification of customer service intention is carried out.
[0033] Embodiment 1: In the customer service intention recognition method based on natural language processing, text preprocessing is a key link to ensure the accuracy of subsequent semantic analysis and intention classification. This embodiment describes in detail how to standardize the original user text information through the word segmentation algorithm, including multi-granularity segmentation, stop word filtering, morphological restoration and smoothing processing, and finally generate coherent and structured text data to provide high-quality input for intention recognition.
[0034] The original text information generated in the interaction process between the user and the customer service system usually contains colloquial expressions, redundant words and useless stop words, which will affect the accuracy of intention recognition. Therefore, the original text needs to be input into the word segmentation algorithm model first. The word segmentation algorithm adopts a multi-granularity segmentation strategy, which can divide the text into word subsequences of different lengths according to the context. For example, for long compound words or specific domain-specific terms, the algorithm will preferentially retain their complete semantic units, while for regular words, it will be segmented according to the smallest semantic unit. This flexible segmentation method ensures that the text can retain key information while adapting to the semantic analysis needs of different scenarios.
[0035] After the initial segmentation is completed, 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 according to the context relevance. Specifically, the model analyzes the grammatical role and semantic contribution of each word in the sentence, and if it is a high-frequency but low-information word (such as "of", "of", "of", etc.), it is marked as a stop word and filtered out. In addition, the model can also identify invalid expressions in specific scenarios, such as polite language or repetitive expressions commonly used in customer service conversations, further improving the purity of the text.
[0036] The word subsequences after stop word filtering need to be further processed for word reduction. The goal of word reduction is to reduce different forms of words to their basic forms, for example, "running" and "ran" are reduced to "run" and "run" respectively. This process relies on dictionary matching and rule engine, while combining context semantics to eliminate ambiguity. For example, for the polysemous word "hit", the system will determine its specific meaning according to the adjacent words (e.g. "hit the phone" is reduced to "call", and "hit the basketball" is reduced to "play"), to ensure the accuracy of the reduction result. After completing the word reduction, the words in the text are in standardized form, which facilitates subsequent feature extraction and intent analysis.
[0037] To further improve the coherence of the text, the system performs smoothing processing on the standardized text. The smoothing operation mainly aims at two problems: one is to remove residual redundant expressions, for example, the user may repeatedly mention the same intent (e.g. "I want to ask, that is, I want to consult…"), the system will combine the fragments with repeated semantics; the second is to correct the sentence with loose syntax structure or incomplete expression, for example, to supplement the missing subject or object, so as to make it conform to the standard expression. The text after smoothing processing is more compact in semantics and clearer in logic, providing a higher quality data basis for intent recognition.
[0038] Finally, the system outputs the user text information after the complete preprocessing process. The result has the following characteristics: one is that the vocabulary granularity is moderate, which retains the key semantic units and avoids semantic fragmentation caused by excessive segmentation; two is that stop words and redundant information are effectively filtered out, reducing noise interference; three is that the word form is uniform and conforms to the grammar standard, which is convenient for 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 judgment.
[0039] During the entire preprocessing process, the system uses a dynamic adjustment mechanism to optimize the segmentation algorithm parameters. For example, when detecting specific domain professional terms, the algorithm will automatically relax the segmentation granularity to ensure the integrity of the terms; while processing colloquial expressions, it will enable stricter filtering rules to reduce the impact of informal expressions. In addition, the system monitors the preprocessing effect in real time, and if it finds that the semantic coherence of the output text has decreased, it will trigger a self-checking mechanism to recalibrate the stop word filtering threshold or word reduction rules, ensuring the stability of the preprocessing module.
[0040] Embodiment 2: In the customer service intent recognition system, accurately understanding the semantic features of user text and judging the intent type is the core link to realize intelligent interaction. This embodiment details how to extract semantic features based on standardized text information, analyze intent frequency and trend, and finally complete the accurate classification of consultation and complaint intents.
[0041] After preprocessing, standard text information first enters the semantic feature extraction stage. The system uses word embedding technology to convert discrete words into continuous vector representations. This process captures the semantic connections and contextual information between words. Each word is mapped to a vector point in a high-dimensional space based on its distribution characteristics in the corpus. Words with similar semantics are close to each other in the vector space. By superimposing word vectors and performing weighted calculations, the system generates a semantic vector representation of the entire text. This vector comprehensively reflects the core meaning and emotional orientation of the text.
[0042] Based on the semantic vector, the system performs dual feature extraction. The intent keyword feature is determined by analyzing the significant points in the word vector space. These points correspond to words that appear repeatedly or are domain-specific in the text. The system establishes a keyword weight model, scores the words based on their frequency of occurrence, position distribution, and relevance to the intent category in historical data, and selects the word combinations that are most indicative of intent judgment. At the same time, the sentiment tendency feature is obtained by analyzing the overall distribution pattern of the semantic vector. The system identifies areas in the vector space associated with positive and negative emotions, calculates the projection strength of the text vector in these areas, and thus quantifies the emotional polarity expressed by the user.
[0043] The intent frequency analysis module processes the extracted keyword features and constructs a time-dependent intent evolution graph. The system uses sliding window technology to segment continuous conversations, counting the frequency of occurrence of specific intent keywords within each time window, and identifying the periodic patterns of intent through a time series analysis algorithm. For each detected intent cycle, the system records its start time, duration, and intensity change curve. These time series data can reflect the shift in user focus. At the same time, emotional tendency features are input into the trend analysis engine, which calculates the evolution rate and direction of intent emotion by comparing the difference in emotional intensity between adjacent time windows, generating quantifiable trend indicators.
[0044] Based on a comprehensive analysis of frequency and trend data, the system establishes an intent state determination model. This model defines consulting intent as a behavioral pattern with the following characteristics: intent keywords focus on consulting topics such as product features and service processes; emotional tendencies remain neutral or slightly positive; and appear in the early stages of a conversation or in the rising phase of the intent cycle. Correspondingly, complaint intent is manifested as: keywords involve negative content such as problem descriptions and expressions of dissatisfaction; emotional tendencies are significantly biased towards negativity; and appear more often in the late stages of a conversation or at the peak of the intent cycle. The system distinguishes these two states through a decision boundary established by machine learning. When a user's expression simultaneously meets multiple conditions of consulting characteristics, the consulting intent determination is triggered; otherwise, the complaint determination process is entered.
[0045] In real-time dialogue scenarios, the system continuously tracks the dynamic changes of the intent state. Each new user input updates the semantic vector and feature data, and the system compares these incremental information with the current intent cycle to determine whether a critical point of state transition has occurred. For example, when detecting that the sentiment orientation has turned from positive to negative and is accompanied by an increase in complaint keywords, the system will immediately mark the state as complaint and adjust the subsequent response strategy. All determination results are associated with specific timestamps to form a complete intent state sequence, which not only records the intent category at each time, but also contains intensity, duration, and other metadata, providing historical basis for subsequent confidence calculation and parameter adjustment.
[0046] To ensure the accuracy of state determination, the system sets up multiple verification mechanisms. First, the semantic feature extraction stage uses an ensemble learning method to combine the results of multiple word embedding models for voting decision, avoiding the bias of a single model. Second, the intent frequency analysis introduces an anomaly detection algorithm that can identify and filter noise data caused by user expression habits or system errors. Finally, the state determination model has online learning ability, which can continuously optimize its decision rules according to newly generated dialogue data, adapting to the language characteristics of different industries and user groups.
[0047] The technical advantages of this implementation are reflected in three aspects: first, through the collaborative analysis of semantic vectors and double features, multi-dimensional description of user intent is achieved; second, the time series modeling method can capture the dynamic evolution process of intent, avoiding misjudgment caused by static analysis; third, the adaptive mechanism of the determination model ensures that the system maintains stable recognition performance in long-term use.
[0048] Embodiment 3: In the dynamic adjustment mechanism of the customer service intent recognition system, the accurate setting of the intent recognition confidence threshold plays a decisive role in the reliability of the classification results. This embodiment details how to realize intelligent calculation and real-time optimization of the confidence threshold of the consultation intent phase based on neural network models and dynamic adjustment algorithms, ensuring that the system can adaptively match the evolution process of user intent.
[0049] When the system determines that it is currently in the consultation intent phase, the confidence threshold calculation process is immediately started. The 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 historical accuracy statistics, domain knowledge weights, dialogue context correlation, and other system-level indicators, while the semantic vector data carries the complete semantic information of the current user expression. The neural network model uses a deep network architecture design, including multiple hidden layers and non-linear activation units, which can capture the high-order interaction relationships between input features. The model has learned the complex mapping relationship between the optimal confidence threshold and the input features in different consultation scenarios during the training phase, so it can predict the most suitable threshold suggestion value based on the current dialogue features during the inference phase.
[0050] The dynamic adjustment algorithm receives the initial threshold suggestion from the neural network model output and initiates the threshold optimization process. The algorithm employs a gradual approximation strategy to adjust the output threshold of the current intent classification model towards the target value step by step. The adjustment step size is dynamically determined based on the semantic vector's change trend: when rapid semantic content evolution is detected, the algorithm increases the step size to speed up the response; when the semantics tend to be stable, the step size is reduced to improve the 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 user intent changes while avoiding system oscillation caused by excessive adjustments.
[0051] The system sets up dual trigger conditions to monitor the effectiveness of threshold adjustment. On the one hand, the semantic change monitoring module continuously analyzes the distance measure between the newly generated semantic vector and the historical vector, and immediately triggers threshold recalculation when significant concept drift is detected. This drift may manifest as the user suddenly introducing a new consultation topic or changing the expression method. On the other hand, the confidence deviation analysis module tracks the deviation of the actual recognition result from the expected target in real time, and also triggers recalculation when the deviation exceeds the pre-set tolerance range. This dual safeguard mechanism ensures that the system can respond to various unexpected situations in a timely manner and maintain the rationality of the threshold setting.
[0052] When the trigger condition is met, the system starts the threshold recalculation process. The latest semantic vector data and updated system parameters are input into the neural network model again, and the model generates a new threshold suggestion based on the changed context. The dynamic adjustment algorithm immediately responds to this change and replans the adjustment path. It is worth noting that the new adjustment process does not start from zero, but inherits the accumulated experience data from the previous adjustment, including the regular characteristics of semantic changes, user response patterns, and other implicit knowledge, which enables faster convergence to the optimal value in subsequent adjustments. The system also sets a maximum number of adjustment limits, and when the limit is reached and still not stable, it will start an exception handling process, including expanding the input feature range of the neural network model or introducing an artificial intervention mechanism.
[0053] During the adjustment process, the system maintains a complete threshold evolution log, recording the time point of each adjustment, adjustment amplitude, trigger cause, and adjustment effect. These data are not only used for real-time monitoring, but also regularly fed back to the training process of the neural network model to achieve continuous optimization of the model. At the same time, the system establishes a threshold stability evaluation mechanism to judge whether the current threshold has entered a stable state by analyzing the amplitude change trend of the last few adjustments. When detecting continuous fine-tuning and confidence deviation remaining within a very small range, the system determines that the stable condition has been reached, temporarily suspending active adjustment and switching to monitoring mode to save computing resources.
[0054] The technical implementation of this embodiment has three significant features: first, the combination of the neural network model and the dynamic adjustment algorithm ensures both the intelligence of threshold calculation and the smoothness of the adjustment process; second, the design of the double trigger mechanism enables the system to balance the response speed and stability requirements; finally, the complete knowledge accumulation and migration mechanism enables the system to continuously improve. These characteristics work together to enable the customer intent recognition system to maintain accurate judgment ability in the face of complex and changing consultation scenarios, providing users with a high-quality interactive experience.
[0055] In actual operation, the system exhibits good environmental adaptability. For regular consultation scenarios, threshold adjustment usually converges quickly within 3-5 iterations; in the face of sudden theme changes or changes in expression methods, the system can complete re-adaptation within an acceptable time. The entire adjustment process is completely transparent to the end user and does not cause interaction delays or service interruptions. System administrators can observe the curve of threshold dynamic changes through a visual monitoring interface to understand the current adaptation status of the system, but in most cases, no manual intervention is required. This highly autonomous intelligent adjustment mechanism significantly reduces the system's operational costs while ensuring the stability of service quality.
[0056] Embodiment 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 consults "how to apply for increasing 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 which point the classification control algorithm initiates the parameter adjustment process. After accurately calculating a deviation of 0.05, the system determines that the weight parameters of the "limit adjustment" class keywords need to be highlighted by analyzing the current dialogue context features, and slightly adjusts the node parameters in the neural network that handle application process-related expressions. After three iterations of adjustment, the actual recognition confidence is steadily increased to 0.91, and the deviation from the target value is reduced to an acceptable range, at which point the system automatically enters the maintenance monitoring mode.
[0057] The core of the classification control algorithm is to build a multi-level parameter adjustment system. When detecting that an e-commerce customer asks "the newly purchased mobile phone has a screen flicker problem", the system finds that the actual recognition confidence (0.85) is significantly lower than the target value (0.94) of the return consultation. The algorithm first decomposes the composition elements of the deviation: 0.04 comes from the insufficient recognition of technical descriptions such as "screen failure", 0.03 comes from the emotional judgment deviation of "return intention", and the remaining 0.02 belongs to the basic semantic understanding error. In view of this structured analysis result, the system adjusts the model parameters of the three dimensions respectively: enhances the recognition weight of hardware failure terms, optimizes the threshold parameter of negative sentiment analysis, and fine-tunes the connection strength of the basic semantic understanding layer. This precise targeted adjustment avoids the side effects that may be brought by "one-size-fits-all" parameter changes.
[0058] The generation and execution process of the model optimization signal embodies the adaptive technical characteristics. Taking the aviation customer service scene as an example, when the passenger queries "how to apply for a flight delay certificate", the system recognizes that the request belongs to the "certificate acquisition" subcategory, but the confidence (0.89) does not reach the standard threshold (0.93) of this subcategory. The parameter adjustment amount calculation module will comprehensively consider multiple factors: the current dialogue round (the 3rd round), historical interaction data (the user has confirmed the flight delay), business priority (delay certificate belongs to high timeliness demand), etc., and finally generates an optimization signal package containing 17 specific parameter adjustment instructions. These instructions are synchronized to the various functional modules of the intent classification model through a dedicated model execution channel, including the word vector layer, the attention mechanism layer, and the classification output layer, etc., to achieve precise distributed parameter updating.
[0059] In the insurance claim consultation scene, the closed-loop control mechanism shows unique value. When the user describes "what materials are needed for insurance claim after a car accident", the system initially recognizes the confidence as 0.86, which is improved to 0.90 after the first parameter adjustment, but still does not reach the target value of 0.93. At this time, the control system automatically starts the secondary adjustment strategy: on the one hand, it retains the original parameter adjustment direction, and on the other hand, it introduces the enhanced recognition module of claim professional terms. When the confidence reaches 0.92, the system detects that the improvement amplitude begins to decay, and intelligently switches to the fine-tuning mode, only adjusting the most critical three parameters at the nanoscale (the adjustment amplitude is 1 / 10 of the initial value), and finally stabilizes the confidence at 0.928 in the fifth iteration. The whole adjustment process forms a complete "monitoring-analysis-adjustment-verification" cycle, and each cycle is optimized based on the previous result.
[0060] This implementation shows good adaptability in cross-industry applications. In the education agency consultation scenario, when a parent asks about "summer English course arrangements", the system needs to handle multiple features simultaneously: course type (English), time attribute (summer), and query object (arrangement). The classification control algorithm establishes a multi-dimensional parameter correlation matrix to ensure that adjusting the English professional term recognition parameters does not affect the processing accuracy of time expressions. In the medical consultation scenario, when a patient asks "Can aspirin and ibuprofen be taken at the same time?", the algorithm will particularly strengthen the parameter weight of professional knowledge such as drug interactions, while maintaining the recognition stability of regular parameters such as dosage and usage. This field-oriented parameter adjustment strategy enables the system to maintain high-precision intent recognition capability in different professional scenarios.
[0061] Specific to the technical implementation of parameter adjustment, the system adopts a hierarchical and progressive adjustment strategy. Taking the cross-border e-commerce scenario as an example, when a foreign buyer consults "tariff calculation method", the first round of adjustment mainly targets the basic semantic understanding layer, addressing the recognition of professional terms such as "tariff"; the second round shifts to the business logic layer, strengthening the correlation analysis of operation-related words such as "calculate" and "method"; the third round focuses on the context understanding layer, making comprehensive judgments based on previously confirmed information such as product price and destination country. Each adjustment stage has an independent evaluation mechanism, and only when the current stage's goal is achieved will it proceed to the next stage, avoiding system instability caused by excessive adjustment.
[0062] During model parameter updating, the system uses "hot switching" technology to ensure service continuity. When a user on a travel website consults the change and cancellation policy for a certain situation, the conversation flow will not be interrupted even if the system is performing parameter adjustment. New and old parameter versions run in parallel in memory, and after the new parameters pass integrity checks, the system will complete the switch within milliseconds, and the user will not feel the adjustment process. For complex consultations such as "international flight connecting ticket change and cancellation rules", the system also supports partial module gray update, prioritizing the adjustment of key module parameters that most affect the current conversation, and silently updating the remaining modules in the background. This design significantly improves the system's response speed and service quality.
[0063] The abnormal handling mechanism is an important part of this implementation. When encountering rare consultations such as "antique auction deposit payment methods", the system will start a special processing flow after detecting sustained low confidence: first, expand the parameter adjustment range and activate the backup model path; at the same time, automatically retrieve similar historical cases for adjustment reference; finally, retain complete debugging logs for subsequent analysis. This mechanism ensures that the system can maintain basic service levels when facing edge cases, while accumulating valuable data for subsequent model optimization. The entire process is fully automated and does not require human intervention, demonstrating the self-repairing ability of intelligent systems.
[0064] Through continuous observation of actual dialogue scenes, it can be found that this classification control mechanism can intelligently distinguish key parameters that need to be adjusted and basic parameters that should be kept stable. In the logistics query case, when the user asks "cold chain transportation temperature record query method", the system will focus on adjusting the identification parameters of professional fields such as "cold chain" and "temperature record", while keeping the processing parameters of ordinary query words (such as "how" and "query") unchanged. This precise differential adjustment strategy not only solves the problem of recognition accuracy in specific scenarios, but also maintains the overall stability of the system, avoiding global fluctuations caused by local adjustments. With the accumulation of system running time, the accuracy and efficiency of this parameter adjustment will continue to improve, forming a virtuous cycle of self-evolution.
[0065] In the actual operation of the intelligent customer service system, dynamic maintenance of intent recognition accuracy is a key link to ensure service quality. Taking the customer service system of a certain telecom operator as an example, when a large number of users concentrate on consulting "5G package upgrade" business, the system detects that the recent intent recognition accuracy is showing a downward trend, triggering the word segmentation algorithm parameter adjustment mechanism. The system first analyzes the problem characteristics: the user consultation frequently appears "5G+", "SA / NSA" and other professional term combinations, and the existing word segmentation granularity causes these key information to be incorrectly segmented. The control system immediately issues instructions to adjust the word segmentation granularity of technical term related fields from the default fine granularity to medium granularity, ensuring that "5G package", "network speed improvement" and other business keywords remain complete. After adjustment, the system collects new accuracy data and finds that the recognition accuracy for package type consultation has improved significantly, but ordinary business queries have decreased slightly, so a balance fine-tuning is started, and finally the optimal parameter combination that takes into account all types of queries is found.
[0066] The process of accuracy monitoring and parameter adjustment forms a complete adaptive cycle. In the customer service system of an online education platform, when the summer course consultation peak period arrives, the system detects that the accuracy rate of "course arrangement" type queries is lower than the preset standard. Analysis shows that the problem is caused by the user's extensive use of short expressions such as "junior three mathematics summer and autumn class time", and the original word segmentation strategy cannot correctly handle such compact phrases. The system executes a hierarchical adjustment strategy: first, appropriately increase the basic word segmentation granularity to ensure that "summer and autumn class" and other specific combinations remain complete; second, activate the education field special dictionary to strengthen the recognition of "junior three" and "mathematics" and other discipline terms; finally, optimize the stop word filtering rules to retain "class", "period" and other words that have practical meaning in the education scene. After several rounds of progressive adjustment, the system restores the recognition accuracy of course arrangement type intent to the ideal level without affecting other types of queries.
[0067] Intelligent decision-making mechanisms are particularly important during parameter adjustment. In a customer service case study involving an airline, the system observed significant fluctuations in recognition accuracy for queries related to "baggage check-in." In-depth analysis revealed significant variations in how users phrased the same question, ranging from the standard "checked baggage allowance inquiry" to the more colloquial "Can a suitcase be checked for free?" Rather than adopting a single adjustment strategy, the system implemented a differentiated strategy: maintaining existing word segmentation rules for formal expressions while focusing on optimizing the grammatical analysis module; while adjusting word segmentation flexibility for colloquial expressions, allowing for cross-grammatical unit combinations such as "suitcase" and "checked baggage." Furthermore, a correlation mapping was established between expression variants to ensure that different expressions accurately refer to the same intent category. This multi-dimensional adjustment approach effectively improved the system's adaptability to diverse expressions.
[0068] Historical data analysis plays a key role in parameter optimization. A bank's credit card customer service system, which had accumulated nearly three years of user query data, noticed a decline in the accuracy of inquiries related to "installment payments." The system first traced the historical data to identify evolving trends in user expressions: Early expressions favored standardized terminology like "bill installment application," while more recent expressions like "divide the repayment into several months" became popular. Based on this trend analysis, the system dynamically adjusted its word segmentation strategy, maintaining accurate segmentation for specialized terms while strengthening its ability to recognize colloquial numerical expressions (such as "divide into 12 months"). Furthermore, the system implemented a seasonal adjustment model to pre-optimize consumer credit-related word segmentation parameters during specific periods, such as shopping festivals, to ensure proactive performance maintenance.
[0069] The feedforward control mechanism enables the system to make predictive adjustments. Before a policy or regulation update, a government service platform predicted a surge in related inquiries by analyzing historical patterns. The system implemented three proactive measures: importing the latest policy terminology, enhancing the segmentation rules for common conjunctions in regulatory clauses (such as "Article X, Paragraph X"), and adjusting parameters for processing long sentences unique to legal texts. When users actually inquired about "Detailed Implementation Rules for the Newly Issued XX Regulation," the system already had optimized processing capabilities, accurately identifying key elements such as "regulations" and "detailed rules" and their modifying relationships, avoiding the typical decline in recognition accuracy during policy transitions. This predictive, proactive adjustment significantly improved the system's ability to respond to unexpected query changes.
[0070] Parameter adjustment in a multilingual environment requires special considerations. An international e-commerce platform's customer service system handles mixed English and Chinese queries, such as "How do I track my order." The system uses a hierarchical adjustment strategy: first, identify language mixing characteristics and distinguish between English and Chinese boundaries; then, keep the English part intact without segmentation, and process the Chinese part according to optimized rules; finally, establish cross-language associations to ensure that "track" and "order status query" point to the same intent. When detecting that some regional users habitually use pinyin abbreviations such as "DDL" instead of "deadline," the system dynamically expands the segmentation dictionary to add such regional variants without affecting the processing rules for mainstream expressions. This fine-grained multilingual parameter management enables the system to adapt to the diverse expression habits of global users.
[0071] Parameter protection mechanisms in abnormal situations ensure system stability. When a certain government hotline system encounters a large number of malicious inquiries, the accuracy monitoring module detects abnormal fluctuations. The system activates the protection protocol: first, freeze non-critical parameter adjustments to maintain the stability of core segmentation rules; second, enhance input verification to filter meaningless character combinations; finally, activate the anti-interference mode to prioritize the processing quality of regular government consultations. After the abnormal traffic subsides, the system gradually restores the full parameter adjustment function. This self-protection mechanism avoids system performance degradation caused by improper adjustments in abnormal situations, demonstrating the robustness design required for industrial-grade applications.
[0072] Parameter adjustment in a multilingual environment requires special considerations. An international e-commerce platform's customer service system handles mixed English and Chinese queries, such as "How do I track my order." The system uses a hierarchical adjustment strategy: first, identify language mixing characteristics and distinguish between English and Chinese boundaries; then, keep the English part intact without segmentation, and process the Chinese part according to optimized rules; finally, establish cross-language associations to ensure that "track" and "order status query" point to the same intent. When detecting that some regional users habitually use pinyin abbreviations such as "DDL" instead of "deadline," the system dynamically expands the segmentation dictionary to add such regional variants without affecting the processing rules for mainstream expressions. This fine-grained multilingual parameter management enables the system to adapt to the diverse expression habits of global users.
[0073] The parameter optimization in long-term operation accumulates to form a system knowledge base. A large e-commerce customer service system continuously records each accuracy anomaly and its solution, and builds a rich adjustment case library. When a similar problem occurs again, such as a decline in the accuracy rate of "promotion rule" type queries, the system intelligently matches historical cases and finds that a similar situation occurred during a certain product promotion activity due to improper handling of "top X" and "deposit inflation" and other promotion terms. Based on these historical experiences, the system quickly locates the current problem as possibly originating from newly emerging "group purchase cashback" and "cross-store discount" and other marketing jargon, and adjusts the relevant word segmentation rules and intent mapping relationships in a targeted manner. This continuous learning mechanism enables the parameter optimization efficiency of the system to improve over time, forming a virtuous cycle of self-improvement.
[0074] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof are intended to cover non-exclusive inclusions, so that a process, method, article, or device that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed, or inherent to such a process, method, article, or device.
[0075] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, alternatives, and variations can be made to the embodiments without departing from the principles and spirit of the application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A customer service intention recognition method based on natural language processing, characterized by: The steps include: Obtain original user text information from the customer service dialogue system, pre-process the text using a word segmentation algorithm, remove stop words, and perform lemmatization to obtain standardized user text information; Based on standardized user text information, the semantic features of the text are calculated, the frequency of occurrence and change 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. If the intent is in the consultation stage, the current required intent recognition confidence threshold is calculated based on the preset intent recognition parameters. Combined with the semantic change trend of the user text information, the intent recognition target value is dynamically adjusted; Using a classification control algorithm, the parameter adjustment of the intent classification model is calculated based on the deviation between the intent recognition target value and the current actual recognition confidence level, and a model optimization signal is generated and output 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 real-time monitoring of the changing trends of user text information and intent recognition parameters, the preset parameters are dynamically updated to match the changes in user intent status and accurately identify customer service intentions.
2. The method for identifying customer service intentions based on natural language processing according to claim 1, characterized in that: Use word segmentation algorithms to preprocess the text, remove stop words, and perform lemmatization to obtain standardized user text information, including: Input the user text information data into the word segmentation algorithm model, and then use the word segmentation algorithm to perform multi-granularity segmentation on the user text information, decomposing the user text information into word subsequences of different lengths; Extract the stop word component features from each word subsequence obtained by decomposition, input the extracted stop word component features into the stop word filtering model, adaptively adjust the filtering threshold according to the input stop word component features, and perform adaptive filtering on the word subsequence to remove the stop word components; Performing lemmatization on each word subsequence after filtering out stop words to obtain standardized user text information, and then smoothing the standardized user text information to eliminate residual redundant expressions in the text and obtain coherent user text information; The smoothed user text information is output as the final user text information processing result for semantic and frequency analysis of user intent.
3. The customer service intention recognition method based on natural language processing according to claim 1, characterized in that: Based on standardized user text information, the semantic features of the text are calculated, the frequency and changing trends of user intent are extracted, and the intent classification algorithm is used to determine whether the user intent is consultation or complaint. Specifically, the following steps are involved: Perform word embedding processing based on standardized user text information to obtain the semantic vector representation of the user text information, and then extract the semantic features of the user text information, including intent keywords and sentiment tendency features; Based on the extracted intent keyword features, the frequency of user intent is calculated to obtain the start and end time of each user intent cycle. Then, based on the extracted emotional tendency features, the user intent change trend is calculated to obtain a curve of intent type change over time. Based on the frequency and changing trend of user intentions, determine whether the current user intention state is in the consultation stage or the complaint stage. When the user intention tends to be consultation-oriented and is in the first half of the user intention cycle, it is judged to be in the consultation stage. When the user intention tends to be complaint-oriented and is in the second half of the user intention cycle, it is judged to be in the complaint stage. According to the judged user intention state, the current user intention state is output, and the user intention state is associated with the corresponding dialogue time point to obtain a complete user intention state sequence.
4. The method for identifying customer service intentions based on natural language processing according to claim 3, characterized in that: After the intent classification algorithm determines whether the user's intent is consultation or complaint, the following steps are also performed: Based on the preset intent recognition parameters and the real-time detected semantic vector data, it is used as the input of the neural network model. The trained neural network model is used to calculate the intent recognition confidence threshold required for the current consultation stage. Based on the calculated intent recognition confidence threshold and combined with the 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, the semantic vector changes and intent recognition confidence changes are detected in real time. When the semantic vector changes significantly or the intent recognition confidence deviates significantly 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, and the output threshold of the intent classification model is dynamically adjusted again until the intent recognition confidence stabilizes near the new target value.
5. The method for identifying customer service intentions based on natural language processing according to claim 1, characterized in that: Using the classification control algorithm, the parameter adjustment of the intent classification model is calculated based on the deviation between the intent recognition target value and the current actual recognition confidence level. The model optimization signal is generated and output to the intent classification model execution module. Specifically, it includes: Obtain the target value and actual recognition confidence value for intent recognition, calculate the deviation between the two, input the deviation into the classification control algorithm, and calculate the parameter adjustment of the intent classification model based on the weight, offset, and learning rate parameters of the classification control algorithm; Generate a corresponding model optimization signal based on the calculated intent classification model parameter adjustment amount, and output the generated model optimization signal to the intent classification model execution module to control the parameter update of the intent classification model; The intent classification model adjusts the internal parameters of the model based on the received model optimization signal, thereby changing the actual confidence of intent recognition. By continuously obtaining the real-time intent recognition confidence value, comparing it with the target value, and calculating the deviation value, a closed-loop control is formed.
6. The customer service intention recognition method based on natural language processing according to claim 1, characterized in that: 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. Specifically, the following are performed: 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 a command to increase the segmentation granularity of the word segmentation algorithm through the actuator; When the accuracy value is higher than the upper threshold of the preset range, the control system issues an instruction to reduce the segmentation granularity of the word segmentation algorithm through the actuator to optimize the preprocessing effect of the user text information. After the parameter adjustment of the word segmentation algorithm is completed, the standardized text information output by the algorithm is obtained again until the accuracy value stabilizes within the preset range. During the parameter adjustment process, the historical accuracy data is analyzed through machine learning algorithms, and a mapping model between accuracy values and word segmentation parameters is established. Using the established mapping model, the optimal word segmentation parameters at the next moment are predicted based on the current accuracy value, and feedforward control of the word segmentation algorithm is performed.
7. The method for identifying customer service intentions based on natural language processing according to claim 1, characterized in that: The parameter optimization algorithm is used to calculate the parameter adjustment of the word segmentation algorithm, generate a control signal and output it to the word segmentation algorithm execution module, specifically including: The intent feedback module obtains the current system accuracy value in real time, compares the accuracy value with the preset accuracy range, and determines 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 median of the preset accuracy range is calculated, and the deviation value is used as the input of the parameter optimization algorithm. Then, based on the accuracy deviation value, the pre-established parameter optimization rule library is used to calculate the parameter adjustment of the word segmentation algorithm through rule reasoning; Convert the calculated word segmentation algorithm parameter adjustment amount into a corresponding control signal, convert the digital signal into an analog signal through a digital-to-analog converter, and output the generated analog control signal to the execution module of the word segmentation algorithm to drive the execution module to adjust the parameters of the word segmentation algorithm; The word segmentation algorithm changes the segmentation parameters according to the adjustment actions of the execution module, thereby optimizing the preprocessing effect of the user text information and returning the accuracy value to the preset range.
8. The method for identifying customer service intentions 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 ensure a smooth and stable semantic analysis and classification process for user intent recognition. Specifically, the following steps are performed: Obtain the response time parameters of the intent classification model and word segmentation algorithm, establish a model dynamic model, and obtain the relationship curve between model parameters, accuracy, and confidence. Based on the target values of intent recognition accuracy and confidence, calculate the target model parameters and control signal output timing through the model dynamic model; The model parameter deviation and accuracy confidence deviation are then used as inputs. Through rule reasoning, the optimized adjustment amount of the control signal is obtained. The optimized control signal is sent to the intent classification model and word segmentation algorithm in sequence according to the calculated output sequence to dynamically adjust the model parameters. During the dynamic adjustment process of the model, text information and intent recognition result data of 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 method for identifying customer service intentions 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, dynamically updating preset parameters, and matching changes in user intent status, we can accurately identify customer service intent. Specifically, we can: Obtain the user's real-time text information and current intent recognition parameters, compare them with the preset normal range threshold, and determine whether they exceed the normal range; if the real-time text information or intent recognition parameters exceed the normal range, an early warning is triggered, and the amplitude of the intent recognition parameters that need to be adjusted is determined based on the degree of deviation from the normal range; Through machine learning algorithms, based on the user's historical text information and the changing trends of intent recognition parameters, the user's intent state change trend in the future is predicted; then, based on the user's real-time text information, current intent recognition parameters and the predicted future intent state change trend, the intent recognition parameters are dynamically adjusted to generate new intent recognition parameter settings; Send the dynamically adjusted intent recognition parameter settings to the intent recognition system, and control the device to recognize the user's intent in real time based on the new parameter settings to match the user's real-time intent state; After adjusting the intent recognition parameters, continuously monitor the user's text information changes to determine whether the user's intent status has returned to normal. If not, return to continue adjusting; The user's real-time text information, intent recognition parameters, and adjusted effect data are stored in the database for optimizing the machine learning model.
10. A customer service intention recognition system based on natural language processing, for implementing the customer service intention recognition method based on natural language processing according to any one of claims 1 to 9, characterized in that: include: The text processing module obtains user text information from the customer service dialogue system, uses a word segmentation algorithm to perform multi-granular segmentation, extracts and filters out stop words, performs lemmatization and smoothing, and finally outputs coherent user text information; The intent calculation module calculates the semantic features of standardized user text information, extracts the frequency and changing trends of user intent, and determines the current user intent status. 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. It also uses a dynamic adjustment algorithm to dynamically adjust the output threshold of the intent classification model to meet the user's intent recognition needs at different stages. The classification control module uses a classification control algorithm to calculate the parameter adjustment of the intent classification model based on the deviation between the intent recognition target value and the actual recognition confidence. It generates a model optimization signal and outputs it to the intent classification model execution module. This module controls the update of the intent classification model parameters and adjusts the model output to change the actual confidence of intent recognition. Through closed-loop control, the intent recognition confidence is stabilized at the target value, enabling 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 uses the parameter optimization algorithm to calculate the parameter adjustment amount of the word segmentation algorithm based on the accuracy deviation value, and generates a control signal output 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 the word segmentation algorithm, optimizes the control signal output timing, and makes the semantic analysis and classification process of user intent recognition smooth and stable. By real-time monitoring of user text information and intent recognition parameter change trends, it dynamically updates preset parameters, matches user intent status changes, accurately identifies customer service intentions, and stores relevant data in the database for optimizing machine learning models.
Citation Information
Patent Citations
Intelligent customer service dialogue strategy optimization method based on deep learning
CN119226458A
Resource mining system based on big data
CN120123539A
Intelligent customer service dialogue model construction method and application
CN120162405A
Automation of contact workflow and automated service agents in contact center system
US10554817B1
Cited By
Multi-label classification method and system for order sending scene
CN121301576A
A multi-label classification method and system for a dispatching scene
CN121301576B
Full-text retrieval method and device, medium, equipment and computer program product
CN121434324A
