Customer service robot interaction control method and system based on voice recognition

Abstract

The invention relates to the technical field of message interaction, in particular to a customer service robot interaction control method and system based on voice recognition. Loading a plurality of user voices interacting with the customer service robot, and executing semantic analysis to obtain a plurality of semantic recognition results; similar semantic clustering is conducted on the semantic recognition results, and multi-cluster semantic recognition results are obtained; for each cluster of semantic recognition results of the multi-cluster semantic recognition results, randomly taking any semantic recognition result for networking indexing to obtain a plurality of retrieval results of the plurality of data sources; traversing a plurality of data sources, executing historical matching sample retrieval based on the semantic association field, counting an accurate sample proportion, and setting the accurate sample proportion as a plurality of confidence degrees; based on the multiple confidence coefficients, selecting a selected retrieval result with the confidence coefficient larger than or equal to a confidence coefficient threshold value from the multiple retrieval results, executing reply generation, and obtaining a target reply to execute unified interaction control on all the users in the cluster. The stability and reliability of the reply content are enhanced, and the reply efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of message interaction technology, and in particular to a method and system for interactive control of customer service robots based on speech recognition. Background Technology

[0002] Currently, the most advanced customer service robot solutions are based on generative AI combined with online knowledge bases. These solutions perform semantic analysis on user voice messages to interpret their needs, then match the information in the knowledge base to generate responses and complete the interaction. However, this approach has significant shortcomings. Firstly, due to the lack of a mechanism for batch processing similar needs, semantic analysis and information matching must be repeatedly performed on each user's voice message, consuming substantial resources, resulting in low processing efficiency and difficulty adapting to high-concurrency scenarios. Secondly, it cannot verify the validity and accuracy of information in the online knowledge base. Faced with content from complex sources of varying quality, the generated responses are prone to inaccuracies and mismatches with the user's needs, leading to poor response stability and negatively impacting the user experience. Summary of the Invention

[0003] This invention addresses the technical problems of low processing efficiency and inaccurate generated responses in existing technologies by providing a voice recognition-based customer service robot interaction control method and system.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a customer service robot interaction control method based on speech recognition, comprising: loading several user voices interacting with the customer service robot, performing semantic analysis to obtain several semantic recognition results; performing similar semantic clustering on the several semantic recognition results to obtain multi-cluster semantic recognition results; randomly selecting any semantic recognition result from each cluster of the multi-cluster semantic recognition results for network indexing to obtain multiple retrieval results from multiple data sources; traversing the multiple data sources, performing historical matching sample retrieval based on semantic association domains, statistically analyzing the proportion of accurate samples, and setting multiple confidence levels; based on the multiple confidence levels, selecting a selected retrieval result with a confidence level greater than or equal to a confidence level threshold from the multiple retrieval results, performing response generation, and obtaining a target response to perform unified interaction control on all users within the cluster.

[0005] Optionally, several user voice recordings interacting with the customer service robot are loaded, and semantic analysis is performed to obtain several semantic recognition results, including: loading several user voice recording location information; traversing the several user voice recording location information, retrieving a voice noise detection sample set, performing frequent pattern analysis, and obtaining several sets of high-frequency noise spectrum features whose noise detection support is greater than or equal to the detection support threshold; performing noise reduction processing on the several sets of high-frequency noise spectrum features to obtain several enhanced user voice recordings; traversing the several enhanced user voice recordings, performing semantic analysis, and obtaining several semantic recognition results.

[0006] The process involves traversing the user voice input location information, retrieving a voice noise detection sample set, performing frequent pattern analysis, and obtaining several sets of high-frequency noise spectrum features with noise detection support greater than or equal to a detection support threshold. This includes: extracting first user voice input location information from the user voice input location information; when the first user voice input location information is empty, reloading the first user voice input location information to obtain reloaded first user voice input location information; when the reloaded first user voice input location information is still empty, performing general software noise reduction on the first user voice to obtain enhanced first user voice; when the first user voice input location information or the reloaded first user voice input location information is not empty, retrieving a voice noise detection sample set based on the first user voice input location information or the reloaded first user voice input location information, performing frequent pattern analysis, obtaining a first set of high-frequency noise spectrum features with noise detection support greater than or equal to a detection support threshold, and adding these features to the set of high-frequency noise spectrum features.

[0007] Specifically, when the first user's voice input location information or the second-loaded first user's voice input location information is not empty, based on the first user's voice input location information or the second-loaded first user's voice input location information, a voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a first set of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold is obtained. These features are then added to the plurality of sets of high-frequency noise spectrum features, including: when the first set of high-frequency noise spectrum features is empty, general software noise reduction is performed on the first user's voice to obtain the first user's enhanced voice; when the first set of high-frequency noise spectrum features is not empty, the first set of high-frequency noise spectrum features is added to the plurality of sets of high-frequency noise spectrum features.

[0008] The process involves denoising the user speech based on the aforementioned sets of high-frequency noise spectral features to obtain several enhanced user speech samples. This includes: extracting a first set of high-frequency noise spectral features from the aforementioned sets of high-frequency noise spectral features; extracting a first user speech spectral feature from the aforementioned user speech samples based on the first set of high-frequency noise spectral features; inputting the first set of high-frequency noise spectral features and the first user speech spectral feature into a noise separation model to obtain a first enhanced user speech sample, which is then added to the aforementioned enhanced user speech samples. The noise separation model is generated through machine learning training using multiple sets of data, and each set of data includes noise spectral feature recording data, speech spectral feature recording data, and labels identifying the enhanced user speech sample after noise separation.

[0009] Optionally, similar semantic clustering is performed on the several semantic recognition results to obtain multi-cluster semantic recognition results, including: extracting a first semantic recognition text and a second semantic recognition text from the several semantic recognition results; processing the first semantic recognition text and the second semantic recognition text respectively through a word embedding model to obtain a first text vector and a second text vector; calculating the vector distance between the first text vector and the second text vector; when the vector distance is less than the vector distance threshold, adding the first semantic recognition text and the second semantic recognition text into the same cluster; otherwise, considering them as different clusters.

[0010] Optionally, the process involves traversing the multiple data sources, performing historical matching sample retrieval based on semantic association domains, and calculating the proportion of accurate samples, which is set as multiple confidence levels. This includes: extracting a first data source from the multiple data sources; extracting a first semantic association domain as a constraint based on the semantic recognition results, retrieving several historical knowledge selected response feedback records from the first data source, which are set as historical matching samples of the first semantic association domain. Each of these historical knowledge selected response feedback records includes a user like / dislike flag, where 0 indicates dislike and 1 indicates like, and a selected response indicates that a response has been selected. The ratio of the number of records with a flag of 1 to the total number of historical knowledge selected response feedback records is set as the proportion of accurate samples in the first semantic association domain. This process continues until the proportion of accurate samples in the Qth semantic association domain is obtained, where Q represents the total number of semantic association domains, and Q≥1. The average of the proportions of accurate samples in the first semantic association domain up to the proportion of accurate samples in the Qth semantic association domain is calculated and set as the first data source confidence level, which is then added to the multiple confidence levels.

[0011] Secondly, the present invention provides a voice recognition-based customer service robot interaction control system, comprising: The user voice analysis module is used to load several user voices that interact with the customer service robot, perform semantic analysis, and obtain several semantic recognition results. The similar semantic clustering module is used to perform similar semantic clustering on the several semantic recognition results to obtain multi-cluster semantic recognition results; The multi-source semantic indexing module is used to randomly select any semantic recognition result from each cluster of the multi-cluster semantic recognition results and perform network indexing to obtain multiple search results from multiple data sources. The semantic confidence assessment module is used to traverse the multiple data sources, perform historical matching sample retrieval based on semantic association domain, count the proportion of accurate samples, and set multiple confidence levels. The response content generation module is used to select a search result with a confidence level greater than or equal to a confidence level threshold from the multiple search results based on the multiple confidence levels, execute response generation, and obtain the target response to perform unified interactive control for all users within the cluster.

[0012] By implementing this invention, it is possible to load several user voices interacting with a customer service robot, perform semantic analysis, obtain several semantic recognition results, and convert the voices into semantic recognition results, thereby realizing the conversion of voice information into machine-understandable data and laying the foundation for subsequent clustering, retrieval and other operations. By implementing this invention, it is possible to perform similar semantic clustering on the aforementioned semantic recognition results to obtain multi-cluster semantic recognition results, classify a large number of scattered semantic recognition results according to similarity, avoid processing each independent semantic result one by one, reduce the number of processing objects in subsequent steps, and improve overall processing efficiency. By implementing this invention, it is possible to randomly select any semantic recognition result from each cluster of the multi-cluster semantic recognition results for network indexing, thereby obtaining multiple retrieval results from multiple data sources. Instead of searching all samples within a cluster one by one, the random selection of a single sample for indexing significantly reduces the number of network searches, reduces system resource consumption, and improves the efficiency of the retrieval process. By implementing this invention, it is possible to traverse the multiple data sources, perform historical matching sample retrieval based on semantic association domain, statistically analyze the proportion of accurate samples, set multiple confidence levels, and calculate the confidence level through user "dislike" and "like" feedback data. This can intuitively reflect the accuracy and reliability of the retrieval results in each data source, providing a quantitative basis for subsequent screening of high-quality retrieval results. By implementing this invention, it is possible to select search results with a confidence level greater than or equal to a confidence threshold from the multiple search results based on the multiple confidence levels, execute response generation, obtain the target response, perform unified interactive control on all users within the cluster, and filter search results based on confidence levels, which can prioritize the selection of information with high accuracy to generate responses, effectively improving the stability and reliability of response content and reducing user dissatisfaction caused by inaccurate information.

[0013] In summary, by implementing this invention, the stability and reliability of responses can be effectively enhanced, solving the core problems of low processing efficiency and poor response stability in traditional customer service robots, and providing users with a more efficient and reliable customer service interaction experience. Attached Figure Description

[0014] Figure 1 A flowchart illustrating a voice recognition-based customer service robot interaction control method provided by the present invention; Figure 2 This is a schematic diagram of the structure of a voice recognition-based customer service robot interactive control system provided by the present invention.

[0015] In the attached diagram, the components represented by each number are as follows: The module includes a user voice analysis module 11, a similar semantic clustering module 12, a multi-source semantic indexing module 13, a semantic confidence evaluation module 14, and a response content generation module 15. Detailed Implementation

[0016] 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.

[0017] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0019] Example 1, as Figure 1 As shown, this embodiment of the invention provides a voice recognition-based customer service robot interaction control method, including: S100: Load several user voices that interact with the customer service robot, perform semantic analysis, and obtain several semantic recognition results; S200: Perform similar semantic clustering on the several semantic recognition results to obtain multi-cluster semantic recognition results; S300: For each cluster semantic recognition result of the multi-cluster semantic recognition results, randomly select any semantic recognition result for network indexing to obtain multiple search results from multiple data sources; S400: Traverse the multiple data sources, perform historical matching sample retrieval based on semantic association domain, calculate the proportion of accurate samples, and set multiple confidence levels; S500: Based on the multiple confidence levels, select the search results with a confidence level greater than or equal to the confidence level threshold from the multiple search results, execute the response generation, and obtain the target response to perform unified interactive control on all users within the cluster.

[0020] In step S100 of this application embodiment, several user voice recordings interacting with the customer service robot are loaded, semantic analysis is performed, and several semantic recognition results are obtained, including: Load the voice recording location information of the aforementioned users; Traverse the several user voice input location information, retrieve the voice noise detection sample set, perform frequent pattern analysis, and obtain several sets of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold. Based on the aforementioned high-frequency noise spectrum characteristics, the aforementioned user speech is subjected to noise reduction processing to obtain several enhanced user speech. The enhanced user speech is traversed, semantic analysis is performed, and several semantic recognition results are obtained.

[0021] In this embodiment, the goal of step S100 is to improve the accuracy of user speech semantic recognition results, providing high-quality foundational data for subsequent steps such as similar semantic clustering and network indexing. Because user speech is easily affected by environmental noise during recording, direct semantic analysis may lead to recognition errors; furthermore, noise characteristics differ at different recording locations, and targeted processing can further optimize speech quality. Therefore, step S100 reduces noise interference with semantic recognition from the source by loading recording location information, extracting noise features, performing targeted noise reduction, and then performing semantic analysis, ensuring that subsequent steps are based on accurate semantic data.

[0022] Specifically, the first step is to load the location information of the user voice recordings. This involves acquiring the location information at the time of user voice recording, providing a basis for subsequent extraction of specific noise features corresponding to that location. Since environmental noise varies significantly between different locations such as offices, streets, and subway stations, location information allows for site-specific noise reduction processing, improving the targeting of noise reduction efforts. In practice, the system can automatically load the recording location information bound to each user voice recording. This location information is typically collected synchronously by the user's terminal device, such as a mobile phone or computer, during voice recording and uploaded to the customer service robot system.

[0023] Furthermore, it is necessary to traverse the several user voice input location information, retrieve the voice noise detection sample set, perform frequent pattern analysis, and obtain several sets of high-frequency noise spectral features with noise detection support greater than or equal to the detection support threshold.

[0024] In step S100 of this application embodiment, the plurality of user voice input location information are traversed, a voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a plurality of high-frequency noise spectral features with noise detection support less than the detection support threshold are obtained, including: Extract the location information of the first user's voice recording from the location information of the plurality of users' voice recordings; When the first user's voice input location information is empty, the first user's voice input location information is loaded a second time to obtain the second-loaded first user's voice input location information. When the second-loaded first user's voice input location information is still empty, general software noise reduction is performed on the first user's voice to obtain the first user's enhanced voice. When the first user voice input location information or the second loading of the first user voice input location information is not empty, based on the first user voice input location information or the second loading of the first user voice input location information, the voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a first set of high-frequency noise spectrum features with noise detection support less than the detection support threshold is obtained and added to the several sets of high-frequency noise spectrum features.

[0025] In step S100 of this application embodiment, the purpose of the above step is to accurately extract the high-frequency noise spectrum features that match the user's voice input position.

[0026] First, it is necessary to extract the first user's voice input location information from the aforementioned user voice input location information. That is, from the loaded "user voice input location information", the location information corresponding to the single user voice to be processed is selected and defined as the "first user voice input location information". This means that independent location matching and noise extraction are performed on a single voice to ensure the targeted processing.

[0027] However, in actual operation, due to reasons such as the user not granting location permissions, device acquisition failure, or network transmission loss, the location information of the first user's voice recording may be empty. At this time, the system automatically triggers the "secondary loading" operation to retrieve the location information corresponding to the user's voice again, and obtains the "secondary loading of the first user's voice recording location information" to try to fill in the missing location information.

[0028] If the "secondary loading of the first user's voice input location information" is still empty, it indicates that the location information cannot be completed, and targeted noise reduction cannot be achieved. In this case, general software noise reduction is directly performed on the user's voice, i.e., the first user's voice. The general software is a software tool or module that does not rely on the user's voice input location information, does not require optimization for specific scene noise features, and only uses general audio noise reduction algorithms to achieve basic noise filtering. For example, algorithms such as spectral subtraction, wavelet transform, and Wiener filtering are used to denoise the first user's voice, or software such as Audacity and Adobe Audition are used directly for noise reduction processing. Through the above method, the denoised enhanced first user voice is generated, skipping the subsequent noise feature extraction steps and directly entering the semantic analysis stage.

[0029] In step S100 of this application embodiment, when the first user voice input location information or the second-loaded first user voice input location information is not empty, based on the first user voice input location information or the second-loaded first user voice input location information, a voice noise detection sample set is retrieved, frequent pattern analysis is performed, a first set of high-frequency noise spectrum features with noise detection support less than the detection support threshold is obtained, and these features are added to the plurality of sets of high-frequency noise spectrum features, including: When the first set of high-frequency noise spectrum features is empty, general software noise reduction is performed on the first user's speech to obtain the first user's enhanced speech. When the first group of high-frequency noise spectrum features is not empty, the first group of high-frequency noise spectrum features is added to the plurality of groups of high-frequency noise spectrum features.

[0030] Specifically, the system first uses the valid location information of the first user's voice input as a retrieval condition to search for historical noise data corresponding to that location in the system's preset voice noise detection sample set. The voice noise detection sample set stores the correlation data between different locations and corresponding environmental noise characteristics, such as "subway station - traffic noise spectrum" and "office - keyboard typing + air conditioner operation noise spectrum".

[0031] Then, frequent pattern analysis is performed on the noise data at the first user's voice input location, which involves statistically analyzing the detection support of various noise types. The detection support is a quantitative indicator such as the frequency and intensity percentage of a certain type of noise at that location. Noise types with "noise detection support ≥ detection support threshold" are selected, and their high-frequency noise spectral features are extracted and defined as "the first group of high-frequency noise spectral features." The detection support threshold is preset by the system and is used to exclude occasional, low-interference noise.

[0032] For example, suppose a customer service robot receives a user's voice message, and the recording location is marked as "3rd floor office of an office building". The system needs to complete frequent pattern analysis, detection support calculation, and threshold filtering through the following steps: First, historical noise data for an office on the 3rd floor of a certain office building was retrieved from the system's "voice noise detection sample set," and a total of 100 noise samples were collected. Through spectrum analysis, the noise in the 100 samples was divided into three categories: Noise A (air conditioner operation sound), frequency 500-800Hz, intensity 60-70dB; Noise B (keyboard typing sound), frequency 1000-1500Hz, intensity 50-55dB; and Noise C (occasional conversation sound), frequency 2000-3000Hz, intensity 55-65dB.

[0033] Next, each of the 100 samples was identified, and the number of samples for each type of noise was counted. It was found that noise A appeared in 85 samples; noise B appeared in 40 samples; and noise C appeared in 15 samples.

[0034] Then, the detection support for the three types of noise is calculated according to the formula: Detection Support = (Number of occurrences of a noise in the sample / Total number of samples) × 100%. Therefore, the following can be calculated: Detection support for noise A = (85 / 100) × 100% = 85%; Detection support for noise B = (40 / 100) × 100% = 40%; Detection support for noise C = (15 / 100) × 100% = 15%.

[0035] Based on customer service voice recognition requirements, the system presets a detection support threshold of 60%.

[0036] The detection support of the three types of noise is compared with the detection support threshold: Noise A (85%) ≥ 60%: it is determined as "the first group of high-frequency noise spectrum features". At this time, the first group of high-frequency noise spectrum features is not empty. Its spectrum features (500~800Hz, 60-70dB) are extracted and added to the above groups of high-frequency noise spectrum features for subsequent targeted noise reduction.

[0037] Noise B (40%) and Noise C (15%) < 60% are judged as "occasional low-interference noise". Their spectral characteristics are not extracted, and they can be slightly processed by general software noise reduction.

[0038] Furthermore, when the first group of high-frequency noise spectrum features is empty, the first user's voice to be processed is input into general software for noise reduction.

[0039] In step S100 of this application embodiment, based on the several sets of high-frequency noise spectrum features, noise reduction processing is performed on the several user voices to obtain several enhanced user voices, including: Extract the first set of high-frequency noise spectrum features from the aforementioned sets of high-frequency noise spectrum features; Based on the first set of high-frequency noise spectrum features, extract the first user speech spectrum features from the several user speech samples; The first set of high-frequency noise spectrum features and the first user speech spectrum features are input into the noise separation model to obtain the first user enhanced speech, which is then added to the plurality of enhanced user speech. The noise separation model is generated by machine learning training using multiple sets of data. Each set of data includes noise spectrum feature recording data, speech spectrum feature recording data, and labels that identify the user's enhanced speech after noise separation.

[0040] In step S100 of this application embodiment, the purpose of the above steps is to optimize the user's voice quality through targeted noise reduction, and then convert the clear voice into accurate semantic recognition results, so as to provide a high-quality data foundation for subsequent steps such as similar semantic clustering and multi-source indexing.

[0041] First, it is necessary to retrieve the set of high-frequency noise spectrum features corresponding to the first user's speech to be processed from several sets of high-frequency noise spectrum features, i.e., the first set of high-frequency noise spectrum features.

[0042] Next, based on the first set of high-frequency noise spectral features, it is necessary to extract the first user speech spectral features from the several user speech samples. Specifically, firstly, it is necessary to locate the "first user speech" to be processed, and then use audio signal processing techniques such as Fast Fourier Transform to perform spectral analysis on the first user speech, converting the time-domain speech signal into frequency-domain spectral data. Then, combined with the frequency range of the "first set of high-frequency noise spectral features," the focus is on extracting the spectral information of the speech in the corresponding frequency band and the entire frequency band, including the signal strength and duration at each frequency point, to form the "first user speech spectral features," ensuring that the extracted features can cover the noise interference frequency band and the effective speech frequency band. The purpose of this step is to obtain the "raw spectral data" of the speech to be processed, providing a basis for subsequent comparison with noise features and noise separation.

[0043] Furthermore, the first set of high-frequency noise spectrum features and the first user speech spectrum features need to be input into the noise separation model to obtain the first user enhanced speech, which is then added to the several enhanced user speech features.

[0044] The noise separation model is generated by machine learning training using multiple sets of data. Based on the task type of the noise separation model, the U-Net model can be used as the noise separation model.

[0045] The noise separation model adopts an encoder-decoder architecture, mainly consisting of an encoder, a decoder, and an output layer. The encoder comprises four convolutional blocks, each containing a convolutional layer, a normalization layer, and an activation function, used to progressively extract key information from the input spectral features and compress the feature dimensionality. Skip connections are established between corresponding layers in the encoder and decoder, passing the shallow features extracted by the encoder to the decoder to retain more detailed information. The decoder also consists of four convolutional blocks, progressively restoring the feature dimensionality through transposed convolution operations. Combined with the information passed through skip connections, it reconstructs the enhanced speech spectral features after noise separation. The output layer uses convolutional layers to output enhanced speech spectral features of the same size as the input, achieving a mapping from noisy speech to clear speech.

[0046] In the parameter settings of the noise separation model, the convolution kernel size is 3×3; the number of convolution kernels in each convolution block is 64, 128, 256, and 512 respectively; the activation function is ReLU; the batch size is 16; the learning rate is 0.001; the optimizer is Adam optimizer; and the loss function is mean squared error loss function. The training samples for the noise separation model are the aforementioned "speech noise detection sample set," which includes multiple sets of noise spectrum feature recording data, speech spectrum feature recording data, and labels indicating the enhanced user speech after noise separation. Specifically, this includes historical noise data from different user speech recording locations, user speech data containing this noise, and manually labeled, clear speech data after noise separation. The training sample size must be no less than 10,000 sets; the training rounds are 50.

[0047] During the training of the noise separation model, when the change in the loss function value of the validation set is less than 0.0001 for 5 consecutive rounds, the model is considered to have converged, training is stopped, and the noise separation model is obtained.

[0048] By inputting the first set of high-frequency noise spectrum features and the first user speech spectrum features into the noise separation model, the first user enhanced speech can be obtained and added to the plurality of enhanced user speech.

[0049] Furthermore, it is necessary to traverse the aforementioned enhanced user speech, perform semantic analysis, and obtain several semantic recognition results.

[0050] Specifically, automatic speech recognition technology is needed to process the enhanced user's speech. An acoustic model is used to convert the acoustic features of the speech, such as frequency, amplitude, and duration, into phoneme sequences. Then, a language model is used to convert these phoneme sequences into corresponding text content, resulting in the text transcription of the enhanced speech.

[0051] Next, semantic analysis needs to be performed on the transcribed text content. Specifically, preprocessing such as word segmentation and part-of-speech tagging is performed first to break the text into meaningful word units and tag them with parts of speech, such as nouns and verbs. Then, the core semantic information of the text is extracted through existing semantic understanding models, including the user's intent, the key entities involved, and the relationships between entities. The final result is a semantic recognition result corresponding to the enhanced user speech, which is presented in the form of structured text and clearly reflects the core meaning expressed by the user through speech.

[0052] In step S200 of this application embodiment, similar semantic clustering is performed on the several semantic recognition results to obtain multi-cluster semantic recognition results, including: From the aforementioned semantic recognition results, extract the first semantic recognition text and the second semantic recognition text; The first semantic recognition text and the second semantic recognition text are processed by a word embedding model to obtain a first text vector and a second text vector. Calculate the vector distance between the first text vector and the second text vector; If the vector distance is less than the vector distance threshold, the first semantic recognition text and the second semantic recognition text are added to the same cluster; otherwise, they are considered to be in different clusters.

[0053] In this embodiment, the purpose of step S200 is to cluster several semantic recognition results according to semantic similarity to form multi-cluster semantic recognition results. Through clustering, semantic texts expressing the same or similar user intent can be grouped into the same cluster, providing a basis for batch processing in subsequent steps and reducing repetitive operations.

[0054] First, it is necessary to extract the first semantic recognition text and the second semantic recognition text from the aforementioned semantic recognition results. That is, the system selects two different semantic recognition texts from the set of "several semantic recognition results" in sequence or randomly, and defines them as "the first semantic recognition text" and "the second semantic recognition text" respectively.

[0055] Then, the first semantic recognition text and the second semantic recognition text need to be processed separately using a word embedding model to obtain a first text vector and a second text vector. The word embedding model is an existing model such as Word2Vec or GloVe. The "first semantic recognition text" is input into the word embedding model. The word embedding model first segments the text into words, such as splitting "consultation order logistics status" into "consultation," "order," and "logistics status." Then, each word is mapped to a corresponding word vector. Finally, the word vectors are aggregated into a single vector, i.e., the "first text vector," through averaging, weighting, or other methods. The same method is used to process the "second semantic recognition text" to obtain the "second text vector". For example, after processing, "inquiry about order logistics status" and "query about express delivery progress" will generate two vectors that are close in position in the vector space.

[0056] Furthermore, it is necessary to calculate the vector distance between the first text vector and the second text vector. Specifically, cosine similarity or Euclidean distance can be used to calculate the vector distance between the "first text vector" and the "second text vector".

[0057] Taking cosine similarity as an example, its value ranges from [-1, 1]. The closer the value is to 1, the more consistent the directions of the two vectors (the more semantically similar), and the smaller the corresponding vector distance. The closer the value is to -1, the more opposite the directions of the two vectors (the greater the semantic difference), and the larger the corresponding vector distance. For example, the vector cosine similarity between "inquire about order logistics status" and "check express delivery progress" is 0.85, and the corresponding vector distance is small, indicating that the semantics are highly similar.

[0058] Next, based on the accuracy requirements of clustering, a "vector distance threshold" needs to be preset in the system. For example, the vector distance threshold based on cosine similarity transformation is set to 0.3. That is, when the distance between two vectors is ≤0.3, they are judged to be semantically similar and classified into one cluster. The vector distance based on cosine similarity transformation is calculated as follows: vector distance = 1 - cosine similarity.

[0059] If the vector distance is less than the vector distance threshold, then the "first semantic recognition text" and the "second semantic recognition text" are added to the same cluster; If the vector distance is greater than or equal to the vector distance threshold, the two texts are considered to be in different clusters and are assigned to different clusters, or a new cluster is created for one of the texts. Repeat the above steps to perform pairwise comparisons and clustering operations on all texts in several semantic recognition results, ultimately forming multi-cluster semantic recognition results.

[0060] In step S300 of this application embodiment, it is necessary to randomly select any semantic recognition result from each cluster of the multi-cluster semantic recognition results and perform network indexing to obtain multiple search results from multiple data sources.

[0061] Specifically, the system first needs to randomly select one semantic recognition result from all semantic recognition results within a given cluster as a retrieval sample for the current processing cluster. For example, the cluster "Inquire about order logistics" contains multiple semantic results such as "When will my order arrive?", "Check delivery progress", and "How to check order logistics status", and "Check delivery progress" is randomly selected as the retrieval sample for this cluster.

[0062] Then, randomly selected semantic recognition result samples are used as search keywords. Multiple pre-defined customer service-related data sources are connected via a network, and information retrieval is performed in each data source to obtain search results semantically related to the sample. These relevant data sources include knowledge websites, enterprise order management system databases, logistics information query interfaces, and customer service FAQ knowledge bases.

[0063] For each cluster in the multi-cluster semantic recognition results, repeat the above process of "random sampling - network indexing - obtaining results" to ensure that each cluster corresponds to retrieval results from multiple data sources, thus preparing for subsequent steps to filter high-quality results based on confidence and generate a unified response.

[0064] In step S400 of this application embodiment, the multiple data sources are traversed, historical matching sample retrieval is performed based on semantic association domain, the proportion of accurate samples is statistically analyzed, and multiple confidence levels are set, including: Extract the first data source from the multiple data sources; Based on the semantic recognition results, the first semantic association domain is extracted as a constraint, and several historical knowledge selected response feedback records of the first data source are retrieved and set as historical matching samples of the first semantic association domain. Among them, any historical knowledge selected response feedback record includes a user like / dislike identifier. When the identifier is 0, it represents dislike; when the identifier is 1, it represents like; and a selected response means that the response has been selected. The ratio of the number of records marked with a statistical identifier of 1 to the total number of selected response feedback records in several historical knowledge domains is set as the proportion of accurate samples in the first semantic association domain. Until the accurate sample proportion of the Qth semantic association domain is obtained, where Q represents the total number of semantic association domains, and Q≥1; The average percentage of accurate samples in the first semantic association domain up to the Qth semantic association domain is calculated and set as the first data source confidence score, and added to the plurality of confidence scores.

[0065] In this embodiment of the application, the core objective of step S400 is to quantitatively evaluate the reliability of the information provided by each data source. By statistically analyzing the proportion of user-accepted responses in historical feedback, the confidence level of each data source is generated, providing a quantitative basis for the subsequent step (S500) to select high-quality search results.

[0066] First, from the multiple data sources, a primary data source needs to be extracted. This primary data source is one of the multiple data sources obtained from S300, such as the enterprise order management system database, logistics information query interface, and customer service FAQ knowledge base. This primary data source is then selected sequentially as the current evaluation object and defined as the "primary data source." For example, the "logistics information query interface" can be selected as the primary data source for confidence calculation.

[0067] Then, based on the semantic recognition results, the first semantic association domain is extracted as a constraint, and several historical knowledge selected response feedback records from the first data source are retrieved and set as historical matching samples of the first semantic association domain. That is, from the semantic recognition results of the currently processed cluster, the corresponding "first semantic association domain" is extracted, such as the domain corresponding to logistics query for the "consultation order logistics" cluster.

[0068] Then, using this domain as a constraint, all "historical knowledge selected answer feedback records" are retrieved in the "first data source". These records are the user feedback data when the data source was selected to answer users in the past. Each record contains user like / dislike indicators: 0 means "dislike", that is, the user does not agree with the answer; 1 means "like", that is, the user agrees with the answer. All relevant records retrieved are defined as historical matching samples in the first semantic association domain. For example, in the "Logistics Information Interface", retrieving all historical feedback records under the "Logistics Inquiry" domain yields 500 records containing 0 or 1 user likes or dislikes.

[0069] Furthermore, it is necessary to calculate the ratio of the number of records marked with '1' to the total number of selected response feedback records in the historical knowledge domain, and set this ratio as the accurate sample proportion of the first semantic association domain. That is, count the records in the "historical matching samples of the first semantic association domain": count the number of records marked with '1' (user approval) and the total number of records in this sample set; then calculate the ratio between the two. For example, if 420 out of 500 records are marked with '1', then the accurate sample proportion of the first semantic association domain is 420 ÷ 500 = 84%.

[0070] If the current semantic association domain has multiple sub-domains, i.e. Q≥2, such as "logistics query" can be subdivided into "domestic logistics" and "international logistics", repeat the above steps to calculate the accurate sample proportion of each sub-domain until the accurate sample proportion of the Qth semantic association domain is obtained.

[0071] Then, calculate the arithmetic mean of the percentage of accurate samples in all Q domains, and define this mean as the "first data source confidence level". For example, if the percentage of accurate samples for "logistics information interface" is 84% ​​in the "domestic logistics" domain and 78% in the "international logistics" domain, then its confidence level is (84% + 78%) ÷ 2 = 81%.

[0072] Finally, add the "First Data Source Confidence" to the "Multiple Confidence" set, repeat the above steps, perform the same operation on all data sources, and finally obtain the confidence corresponding to each data source, providing complete quantitative evaluation data for filtering retrieval results based on confidence in S500.

[0073] In step S500 of this application embodiment, based on the multiple confidence levels, it is necessary to select a selected search result with a confidence level greater than or equal to the confidence level threshold from the multiple search results, perform response generation, and obtain the target response to perform unified interactive control on all users within the cluster.

[0074] In this embodiment, the core objective of step S500 is to generate a high-quality, unified response and efficiently address all users within the cluster. By filtering reliable search results based on confidence levels, the generated target response is ensured to be accurate and effective. Simultaneously, unified interaction control is implemented for all users within the same cluster to avoid repeatedly processing similar requests, significantly improving the response efficiency of the customer service robot and ensuring that users with the same needs receive a consistent service experience.

[0075] First, a confidence threshold needs to be preset, such as 70%. This confidence threshold is determined based on historical customer service interaction data and user satisfaction assessments, and is used to determine whether the search results provided by the data source are sufficiently reliable. When the confidence level of the data source is greater than or equal to this confidence threshold, the corresponding search results are considered to have high accuracy and can be used as candidate materials for generating responses.

[0076] Then, the system iterates through multiple search results and their corresponding data source confidence levels, filtering out search results with confidence levels greater than or equal to the confidence threshold and defining them as selected search results. For example, if the confidence threshold is set to 70%, and a data source has a confidence level of 81%, search results such as "Order expected to be delivered tomorrow, currently located at XX transit station" provided by that data source will be selected; while search results provided by data sources with a confidence level of 65% will be excluded.

[0077] Next, the selected search results are integrated and optimized to generate a target response. Specifically, the system extracts key information from the selected search results, such as logistics nodes, processing procedures, and solutions, and organizes this information into fluent and easy-to-understand natural language text according to a preset response template and natural language generation rules. For example, after integrating information from multiple high-confidence data sources, a target response is generated: "Your order is currently at XX transit station and is expected to arrive before 18:00 tomorrow. You can view the logistics details in real time through the 'My Orders' section of the app."

[0078] Finally, the generated target response is synchronously sent to all users within the cluster, achieving unified interaction control. This means that regardless of the initial semantic expression of a user within the cluster, such as "When will my order arrive?" or "Check delivery progress," they will all receive the same target response. This approach avoids the repetitive operation of generating responses individually for each user, significantly improving customer service response speed while ensuring that users with the same needs receive consistent information, thus enhancing service standardization and user experience.

[0079] Example 2, as Figure 2 As shown, based on the same inventive concept as the voice recognition-based customer service robot interaction control method provided in Embodiment 1, this embodiment of the invention also provides a voice recognition-based customer service robot interaction control system, including: User voice analysis module 11 is used to load several user voices that interact with the customer service robot, perform semantic analysis, and obtain several semantic recognition results; Similar semantic clustering module 12 is used to perform similar semantic clustering on the several semantic recognition results to obtain multi-cluster semantic recognition results; The multi-source semantic indexing module 13 is used to randomly select any semantic recognition result from each cluster of the multi-cluster semantic recognition results and perform network indexing to obtain multiple search results from multiple data sources. The semantic confidence evaluation module 14 is used to traverse the multiple data sources, perform historical matching sample retrieval based on semantic association domain, count the proportion of accurate samples, and set multiple confidence levels; The response content generation module 15 is used to select a selected search result with a confidence level greater than or equal to a confidence level threshold from the multiple search results based on the multiple confidence levels, perform response generation, and obtain a target response to perform unified interactive control on all users within the cluster.

[0080] Furthermore, the user voice analysis module 11 includes the following execution steps: Load the voice recording location information of the aforementioned users; Traverse the several user voice input location information, retrieve the voice noise detection sample set, perform frequent pattern analysis, and obtain several sets of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold. Based on the aforementioned high-frequency noise spectrum characteristics, the aforementioned user speech is subjected to noise reduction processing to obtain several enhanced user speech. The enhanced user speech is traversed, semantic analysis is performed, and several semantic recognition results are obtained.

[0081] Specifically, the process involves traversing the various user voice input location information, retrieving a voice noise detection sample set, performing frequent pattern analysis, and obtaining several sets of high-frequency noise spectral features whose noise detection support is greater than or equal to the detection support threshold, including: Extract the location information of the first user's voice recording from the location information of the plurality of users' voice recordings; When the first user's voice input location information is empty, the first user's voice input location information is loaded a second time to obtain the second-loaded first user's voice input location information. When the second-loaded first user's voice input location information is still empty, general software noise reduction is performed on the first user's voice to obtain the first user's enhanced voice. When the first user voice input location information or the second loading of the first user voice input location information is not empty, based on the first user voice input location information or the second loading of the first user voice input location information, the voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a first set of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold is obtained and added to the plurality of sets of high-frequency noise spectrum features.

[0082] Wherein, when the first user voice input location information or the second-loaded first user voice input location information is not empty, based on the first user voice input location information or the second-loaded first user voice input location information, a voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a first set of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold is obtained, which are then added to the plurality of sets of high-frequency noise spectrum features, including: When the first set of high-frequency noise spectrum features is empty, general software noise reduction is performed on the first user's speech to obtain the first user's enhanced speech. When the first group of high-frequency noise spectrum features is not empty, the first group of high-frequency noise spectrum features is added to the plurality of groups of high-frequency noise spectrum features.

[0083] Specifically, based on the aforementioned sets of high-frequency noise spectrum features, the aforementioned user speech is subjected to noise reduction processing to obtain several enhanced user speech, including: Extract the first set of high-frequency noise spectrum features from the aforementioned sets of high-frequency noise spectrum features; Based on the first set of high-frequency noise spectrum features, extract the first user speech spectrum features from the several user speech samples; The first set of high-frequency noise spectrum features and the first user speech spectrum features are input into the noise separation model to obtain the first user enhanced speech, which is then added to the plurality of enhanced user speech. The noise separation model is generated by machine learning training using multiple sets of data. Each set of data includes noise spectrum feature recording data, speech spectrum feature recording data, and labels that identify the user's enhanced speech after noise separation.

[0084] Furthermore, the similarity semantic clustering module 12 includes the following execution steps: From the aforementioned semantic recognition results, extract the first semantic recognition text and the second semantic recognition text; The first semantic recognition text and the second semantic recognition text are processed by a word embedding model to obtain a first text vector and a second text vector. Calculate the vector distance between the first text vector and the second text vector; If the vector distance is less than the vector distance threshold, the first semantic recognition text and the second semantic recognition text are added to the same cluster; otherwise, they are considered to be in different clusters.

[0085] Furthermore, the semantic confidence assessment module 14 includes the following execution steps: Extract the first data source from the multiple data sources; Based on the semantic recognition results, the first semantic association domain is extracted as a constraint, and several historical knowledge selected response feedback records of the first data source are retrieved and set as historical matching samples of the first semantic association domain. Among them, any historical knowledge selected response feedback record includes a user like / dislike identifier. When the identifier is 0, it represents dislike; when the identifier is 1, it represents like; and a selected response means that the response has been selected. The ratio of the number of records marked with a statistical identifier of 1 to the total number of selected response feedback records in several historical knowledge domains is set as the proportion of accurate samples in the first semantic association domain. Until the accurate sample proportion of the Qth semantic association domain is obtained, where Q represents the total number of semantic association domains, and Q≥1; The average percentage of accurate samples in the first semantic association domain up to the Qth semantic association domain is calculated and set as the first data source confidence score, and added to the plurality of confidence scores.

[0086] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0087] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for interactive control of a customer service robot based on speech recognition, characterized in that, include: Load several user voice recordings that interact with the customer service robot, perform semantic analysis, and obtain several semantic recognition results; Perform similar semantic clustering on the aforementioned semantic recognition results to obtain multi-cluster semantic recognition results; For each cluster of semantic recognition results, randomly select any semantic recognition result and perform online indexing to obtain multiple search results from multiple data sources; Traverse the multiple data sources, perform historical matching sample retrieval based on semantic association domain, calculate the proportion of accurate samples, and set multiple confidence levels; Based on the multiple confidence levels, select the search results with a confidence level greater than or equal to the confidence threshold from the multiple search results, execute the response generation, and obtain the target response to perform unified interactive control on all users within the cluster.

2. The method as described in claim 1, characterized in that, Load several user voice recordings from interactions with the customer service chatbot, perform semantic analysis, and obtain several semantic recognition results, including: Load the voice recording location information of the aforementioned users; Traverse the several user voice input location information, retrieve the voice noise detection sample set, perform frequent pattern analysis, and obtain several sets of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold. Based on the aforementioned high-frequency noise spectrum characteristics, the aforementioned user speech is subjected to noise reduction processing to obtain several enhanced user speech. The enhanced user speech is traversed, semantic analysis is performed, and several semantic recognition results are obtained.

3. The method as described in claim 2, characterized in that, The system iterates through the user voice input location information, retrieves the voice noise detection sample set, performs frequent pattern analysis, and obtains several sets of high-frequency noise spectral features with noise detection support greater than or equal to the detection support threshold, including: Extract the location information of the first user's voice recording from the location information of the plurality of users' voice recordings; When the first user's voice input location information is empty, the first user's voice input location information is loaded a second time to obtain the second-loaded first user's voice input location information. When the second-loaded first user's voice input location information is still empty, general software noise reduction is performed on the first user's voice to obtain the first user's enhanced voice. When the first user voice input location information or the second loading of the first user voice input location information is not empty, based on the first user voice input location information or the second loading of the first user voice input location information, the voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a first set of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold is obtained and added to the plurality of sets of high-frequency noise spectrum features.

4. The method as described in claim 3, characterized in that, When the first user voice input location information or the second-loaded first user voice input location information is not empty, based on the first user voice input location information or the second-loaded first user voice input location information, a voice noise detection sample set is retrieved, frequent pattern analysis is performed, and a first set of high-frequency noise spectrum features with noise detection support greater than or equal to the detection support threshold is obtained. These features are then added to the plurality of sets of high-frequency noise spectrum features, including: When the first set of high-frequency noise spectrum features is empty, general software noise reduction is performed on the first user's speech to obtain the first user's enhanced speech. When the first group of high-frequency noise spectrum features is not empty, the first group of high-frequency noise spectrum features is added to the plurality of groups of high-frequency noise spectrum features.

5. The method as described in claim 2, characterized in that, Based on the aforementioned sets of high-frequency noise spectral characteristics, noise reduction processing is performed on the aforementioned user speech to obtain several enhanced user speech, including: Extract the first set of high-frequency noise spectrum features from the aforementioned sets of high-frequency noise spectrum features; Based on the first set of high-frequency noise spectrum features, extract the first user speech spectrum features from the several user speech samples; The first set of high-frequency noise spectrum features and the first user speech spectrum features are input into the noise separation model to obtain the first user enhanced speech, which is then added to the plurality of enhanced user speech. The noise separation model is generated by machine learning training using multiple sets of data. Each set of data includes noise spectrum feature recording data, speech spectrum feature recording data, and labels that identify the user's enhanced speech after noise separation.

6. The method as described in claim 1, characterized in that, Perform similar semantic clustering on the aforementioned semantic recognition results to obtain multi-cluster semantic recognition results, including: From the aforementioned semantic recognition results, extract the first semantic recognition text and the second semantic recognition text; The first semantic recognition text and the second semantic recognition text are processed by a word embedding model to obtain a first text vector and a second text vector. Calculate the vector distance between the first text vector and the second text vector; If the vector distance is less than the vector distance threshold, the first semantic recognition text and the second semantic recognition text are added to the same cluster; otherwise, they are considered to be in different clusters.

7. The method as described in claim 1, characterized in that, Iterate through the multiple data sources, perform historical matching sample retrieval based on semantic association domain, calculate the proportion of accurate samples, and set multiple confidence levels, including: Extract the first data source from the multiple data sources; Based on the semantic recognition results, the first semantic association domain is extracted as a constraint, and several historical knowledge selected response feedback records of the first data source are retrieved and set as historical matching samples of the first semantic association domain. Among them, any historical knowledge selected response feedback record includes a user like / dislike identifier. When the identifier is 0, it represents dislike; when the identifier is 1, it represents like; and a selected response means that the response has been selected. The ratio of the number of records marked with a statistical identifier of 1 to the total number of selected response feedback records in several historical knowledge domains is set as the proportion of accurate samples in the first semantic association domain. Until the accurate sample proportion of the Qth semantic association domain is obtained, where Q represents the total number of semantic association domains, and Q≥1; The average percentage of accurate samples in the first semantic association domain up to the Qth semantic association domain is calculated and set as the first data source confidence score, and added to the plurality of confidence scores.

8. A voice recognition-based customer service robot interactive control system, characterized in that, The system is used to implement the voice recognition-based customer service robot interaction control method as described in any one of claims 1-7, including: The user voice analysis module is used to load several user voices that interact with the customer service robot, perform semantic analysis, and obtain several semantic recognition results. The similar semantic clustering module is used to perform similar semantic clustering on the several semantic recognition results to obtain multi-cluster semantic recognition results; The multi-source semantic indexing module is used to randomly select any semantic recognition result from each cluster of the multi-cluster semantic recognition results and perform network indexing to obtain multiple search results from multiple data sources. The semantic confidence assessment module is used to traverse the multiple data sources, perform historical matching sample retrieval based on semantic association domain, count the proportion of accurate samples, and set multiple confidence levels. The response content generation module is used to select a search result with a confidence level greater than or equal to a confidence level threshold from the multiple search results based on the multiple confidence levels, execute response generation, and obtain the target response to perform unified interactive control for all users within the cluster.

Images