Intelligent dialogue scene analysis method based on AI large model

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an intelligent dialogue scene analysis method based on an AI large model, and the method comprises the steps: recognizing background noise and voice accent in voice data of a user, and converting the voice data into initial text information; performing standardized adjustment on the initial character information based on a standard language database to obtain adjusted character information; analyzing the voice data to obtain emotional features, analyzing the adjusted text information to obtain vocabulary features, and determining user dialogue features according to the emotional features and the vocabulary features; constructing a dialogue scene model according to the dialogue features; generating initial answer information for the voice data based on the dialogue scene model; and receiving feedback data of the user on the initial answer information, and optimizing the dialogue scene model according to the feedback data, so that the target answer information is superior to the initial answer information. According to the invention, the accuracy of the intelligent dialogue result of the AI large model is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent dialogue scenario analysis method based on an AI big model. Background Art

[0002] With the development of artificial intelligence and Internet information technology, intelligent dialogue systems have been widely used in many scenarios such as intelligent customer service, smart speakers, smart cars, smart Internet of Things, and smart industrial services.

[0003] The patent document with Chinese patent application publication number CN116662511A discloses an intelligent dialogue scene analysis method and AI dialogue system using AI model recognition. By obtaining the service interaction dialogue text generated by the target user in the current intelligent dialogue scene, and using the AI ​​model recognition strategy to perform interest mining on the service interaction dialogue text, at least one interest dialogue sentence cluster is generated, and the interest tag corresponding to the interest dialogue sentence cluster with a number of interest dialogue sentences greater than a set number is determined as the target interest tag. Based on the target interest tag, the corresponding dialogue feedback page content is loaded to the intelligent dialogue scene page where the target user is located, thereby combining the number of interest dialogue sentences with different interest tags in the intelligent dialogue scene to mine the interest of the target user.

[0004] Although existing technologies can mine user interests through AI model recognition strategies, they are often unable to accurately convert speech due to different user accents, resulting in loss or misunderstanding of conversation information and an inability to provide accurate answers. Summary of the Invention

[0005] To this end, the present invention provides an intelligent dialogue scene analysis method based on an AI large model. By deeply analyzing the user's voice accent and text information and combining the characteristics of the dialogue scene, the problem of inaccurate intelligent dialogue results of the AI ​​large model can be solved.

[0006] To achieve the above objectives, the present invention provides an intelligent conversation scenario analysis method based on an AI big model, the method comprising:

[0007] Receive the user's initial voice data, analyze the initial voice data, and identify background noise and voice accent based on the analysis results;

[0008] Processing the background noise, obtaining processed speech data, analyzing the speech accent, and determining calibration parameters based on the analysis results;

[0009] Adjusting initial conversion parameters in the initial speech conversion model based on the calibration parameters to obtain an adjusted speech conversion model, and converting the processed speech data based on the adjusted speech conversion model to obtain initial text information;

[0010] Standardize the initial text information based on a standard language database to obtain standard text information;

[0011] Analyzing the standard text information, determining vocabulary features based on the text information analysis results, analyzing the processed voice data, determining emotional features based on the voice data analysis results, and determining conversation features based on the vocabulary features and the emotional features;

[0012] Analyzing the conversation features, determining scene labels based on the conversation feature analysis results, and constructing an initial conversation scene model based on the scene labels;

[0013] generating initial answer information based on the initial dialogue scenario model and the processed voice data;

[0014] Collect feedback information on the initial answer information, judge the degree of model answer deviation based on the feedback information, and adjust the speech conversion model or the initial dialogue scene model based on the judgment result, so that the target answer information obtained after adjustment is better than the initial answer information.

[0015] Furthermore, the step of identifying background noise and voice accent according to the analysis results includes:

[0016] Identifying energy features, prosodic features, and phoneme features of the initial speech data;

[0017] identifying the background noise based on the energy signature;

[0018] The speech accent is recognized based on the prosodic features and the phonemic features.

[0019] Furthermore, the step of determining the calibration parameters according to the analysis results includes:

[0020] determining a deviation value between the speech data and a standard speech based on the prosodic feature and the phoneme feature;

[0021] The calibration parameter is determined based on the offset value.

[0022] Furthermore, the step of standardizing the initial text information based on the standard language database includes:

[0023] Comparing the standard words in the standard language database with the initial text information one by one, determining abnormal words, calculating the similarity between the abnormal words and the standard words, and replacing the abnormal words corresponding to the similarity greater than the preset word similarity with the standard words;

[0024] Identify actual grammatical rules in the initial text information, compare the actual grammatical rules with standard grammatical rules in the standard language database, and determine and correct incorrect grammar based on the comparison result.

[0025] Furthermore, the step of determining vocabulary features based on the text information analysis results includes:

[0026] Performing word segmentation processing on the standard text information to obtain a number of word segments;

[0027] The occurrence frequencies of several segmented words are identified, and the semantic features corresponding to the segmented words with the highest occurrence frequency are determined as the lexical features.

[0028] Furthermore, the step of determining the emotional characteristics based on the speech data analysis results includes:

[0029] Extracting timbre features and speech rate features of the speech data;

[0030] Constructing a speech feature vector based on the timbre feature and the speech rate feature;

[0031] The emotional feature is determined based on the speech feature vector and a sentiment analysis model.

[0032] Furthermore, the step of determining the dialogue feature based on the vocabulary feature and the emotion feature includes:

[0033] Constructing a feature vector based on the vocabulary feature and the sentiment feature;

[0034] The conversation feature is determined based on the feature vector.

[0035] Furthermore, the step of determining scene labels and establishing an initial dialogue scene model based on the dialogue feature analysis results includes:

[0036] Determining the scene label based on matching the conversation features with a preset scene classification model;

[0037] The initial dialogue scenario model is constructed according to the scenario label and the preset scenario template.

[0038] Furthermore, the step of generating initial answer information based on the initial dialogue scenario model and the processed voice data includes:

[0039] Determining a target dialogue scene to which the processed speech data belongs based on the initial dialogue scene model;

[0040] Generate initial response information based on the target dialogue scenario and the preset scenario response strategy;

[0041] The preset scenario response strategy includes response templates and response content under different dialogue scenarios.

[0042] Furthermore, the step of determining the degree of deviation of the model answer based on the feedback information includes:

[0043] Collect user feedback data on initial responses, including satisfaction ratings and response content evaluations;

[0044] Extracting keywords from the response content evaluation, analyzing the keywords, and determining text satisfaction based on the analysis results;

[0045] determining actual satisfaction based on the satisfaction rating and textual satisfaction;

[0046] The degree of deviation is determined based on the actual satisfaction level and a preset satisfaction level.

[0047] Compared with the prior art, the present invention has the following advantages: by analyzing initial speech data, it can accurately identify and process background noise, effectively reducing the interference of noise on speech recognition accuracy. The parameters of the speech conversion model are adjusted based on the analysis results of the speech accent, making the model more adaptable to speech recognition with different accents and improving recognition accuracy. The initial text information is standardized through a standard language database to ensure that the output text information is more standardized and accurate. The analysis of lexical features and emotional features helps to more deeply understand the user's intention and emotional state, making intelligent conversations more natural and smooth. Based on the analysis of conversation features, scene labels can be automatically determined and an initial conversation scene model can be constructed, enabling the system to provide personalized conversation services based on different scenarios. By optimizing the scene model, user needs in different scenarios can be more accurately identified and understood, improving the relevance and effectiveness of conversations. By collecting feedback on the initial answer information and judging the degree of model answer deviation based on the feedback information, the system can understand user feedback in real time, identify potential problems, and adjust the speech conversion model or the initial conversation scene model based on the judgment results, achieving continuous learning and optimization of the system and continuously improving the accuracy of the conversation system.

[0048] In particular, energy features reflect the distribution and changes of energy in speech signals, and background noise can be accurately identified by identifying energy features, thereby effectively reducing the impact of noise on subsequent speech recognition and processing. Prosodic features and phoneme features reflect the pronunciation characteristics and accent differences of speech. Accent recognition based on these features can more accurately judge the user's accent type and provide an accurate basis for the subsequent adjustment of the speech conversion model. Identifying background noise by identifying energy features enables the system to adapt to different noise environments and improve the robustness of the system. Accent recognition based on prosodic features and phoneme features can enable the system to better adapt to users with different accents, improve the adaptability and robustness of the system, and provide more accurate and personalized services.

[0049] In particular, by analyzing the prosodic and phonemic features of speech data and comparing them with standard speech, the deviation value between the speech data and the standard speech can be accurately determined. Based on this deviation value, the parameters of the initial speech conversion model are precisely adjusted so that the model can better adapt to and recognize different speech data, thereby improving the accuracy of speech recognition.

[0050] In particular, by comparing the standard vocabulary in the standard language database with the vocabulary in the initial text information one by one, abnormal vocabulary can be identified. By calculating the similarity and replacing abnormal vocabulary with a similarity higher than a preset threshold, the incorrect vocabulary in the text can be significantly reduced and the accuracy of the text can be improved. The replacement of abnormal vocabulary helps to ensure that the text follows the standard language specifications and makes the text clearer and more accurate. By identifying the actual grammatical rules in the initial text information and comparing them with the standard grammatical rules in the standard language database, grammatical errors can be accurately discovered and corrected, which helps to optimize the structure of the text, improve the grammatical correctness of the text, and make comprehensive standardized adjustments to the initial text information, effectively improving the overall quality of the text and improving processing efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flowchart of an intelligent conversation scenario analysis method based on an AI big model provided by an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the process of identifying background noise and voice accents in the intelligent conversation scene analysis method based on the AI ​​large model provided in an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the process of standardizing initial text information in the intelligent conversation scenario analysis method based on the AI ​​large model provided in an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the flow of determining vocabulary features in the intelligent dialogue scenario analysis method based on the AI ​​large model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention.

[0056] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0057] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be understood as a limitation on the present invention.

[0058] Furthermore, it should be noted that, in the description of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0059] See also Figure 1 As shown, an embodiment of the present invention provides an intelligent dialogue scene analysis method based on an AI big model, the method comprising:

[0060] Step S100: receiving initial voice data of a user, analyzing the initial voice data, and identifying background noise and voice accent based on the analysis results;

[0061] Step S200, processing the background noise, obtaining processed speech data, analyzing the speech accent, and determining calibration parameters based on the analysis results;

[0062] Step S300, adjusting initial conversion parameters in the initial speech-to-speech conversion model based on the calibration parameters to obtain an adjusted speech-to-speech conversion model, and performing conversion processing on the processed speech data based on the adjusted speech-to-speech conversion model to obtain initial text information;

[0063] Step S400, standardizing the initial text information based on a standard language database to obtain standard text information;

[0064] Step S500: analyzing the standard text information, determining vocabulary features based on the text information analysis results, analyzing the processed voice data, determining emotional features based on the voice data analysis results, and determining conversation features based on the vocabulary features and the emotional features;

[0065] Step S600: analyzing the conversation features, determining scene labels based on the conversation feature analysis results, and constructing an initial conversation scene model based on the scene labels;

[0066] Step S700, generating initial answer information based on the initial dialogue scene model and the processed voice data;

[0067] Step S800: Collect feedback information on the initial answer information, determine the degree of deviation of the model answer based on the feedback information, and adjust the speech conversion model or the initial dialogue scene model based on the determination result, so that the target answer information obtained after adjustment is better than the initial answer information.

[0068] Specifically, embodiments of the present invention accurately identify and process background noise through analysis of initial speech data, effectively reducing the impact of noise on speech recognition accuracy. The parameters of the speech conversion model are adjusted based on the analysis results of the speech accent, making the model more adaptable to speech recognition with different accents and improving recognition accuracy. Initial text information is standardized using a standard language database to ensure that the output text information is more standardized and accurate. Analysis of lexical and emotional features facilitates a deeper understanding of user intent and emotional state, making intelligent conversations more natural and fluent. Analysis of conversation features automatically determines scene labels and constructs an initial conversation scene model, enabling the system to provide personalized conversation services based on different scenarios. By optimizing the scene model, user needs in different scenarios can be more accurately identified and understood, improving the relevance and effectiveness of conversations. By collecting feedback on initial responses and determining the degree of model response deviation based on the feedback, the system can understand user feedback in real time, identify potential problems, and adjust the speech conversion model or the initial conversation scene model based on the judgment results, achieving continuous learning and optimization of the system and continuously improving the accuracy of the conversation system.

[0069] See also Figure 2 As shown, the step of identifying background noise and speech accent according to the analysis results includes:

[0070] Step S110, identifying energy features, prosodic features, and phoneme features of the initial speech data;

[0071] Step S120, identifying the background noise based on the energy characteristics;

[0072] Step S130: identifying the speech accent based on the prosodic features and the phonemic features.

[0073] Specifically, the step of identifying the energy features, prosodic features, and phoneme features of the initial speech data in the embodiment of the present invention includes:

[0074] Converting the initial voice data into an initial digital signal, preprocessing the initial digital signal, segmenting the preprocessed voice signal to obtain a number of frames, calculating the sum of squares of a number of sampling points in any frame, and obtaining the energy feature;

[0075] extracting the fundamental frequency of the initial speech data by an autocorrelation method, and determining the prosodic feature based on changes in a number of fundamental frequencies;

[0076] The initial speech data is converted into a phoneme sequence by a phoneme recognition algorithm, and the phoneme features are extracted based on the phoneme sequence.

[0077] Specifically, the embodiments of the present invention may perform preprocessing on the initial digital signal, including filtering, denoising, pre-emphasis, and other operations to improve the quality of the signal;

[0078] The pre-processed speech signal is usually segmented based on a length of 20-30 milliseconds.

[0079] Specifically, the step of identifying the background noise based on the energy feature in the embodiment of the present invention includes:

[0080] Determining frames corresponding to energy greater than a preset energy based on the energy feature, and treating the initial speech data corresponding to several frames as background noise;

[0081] The background noise is suppressed based on a noise suppression algorithm.

[0082] Specifically, the embodiments of the present invention reflect the distribution and changes of energy in the speech signal through energy characteristics, and accurately identify background noise by identifying energy characteristics, thereby effectively reducing the impact of noise on subsequent speech recognition and processing. Prosodic characteristics and phoneme characteristics reflect the pronunciation characteristics and accent differences of speech. Accent recognition based on these characteristics can more accurately judge the user's accent type and provide an accurate basis for the subsequent adjustment of the speech conversion model. By identifying background noise through identifying energy characteristics, the system can adapt to different noise environments and improve the robustness of the system. Accent recognition based on prosodic characteristics and phoneme characteristics can enable the system to better adapt to users with different accents, improve the adaptability and robustness of the system, and provide more accurate and personalized services.

[0083] Specifically, the step of determining the calibration parameters according to the analysis results includes:

[0084] determining a deviation value between the speech data and a standard speech based on the prosodic feature and the phoneme feature;

[0085] The calibration parameter is determined based on the offset value.

[0086] Specifically, the step of calculating the deviation value in the embodiment of the present invention includes:

[0087] Comparing the prosodic features with standard prosodic features of standard speech to obtain prosodic feature similarity;

[0088] Comparing the phoneme features with the standard phoneme features of standard speech to obtain phoneme feature similarity;

[0089] Calculating a total similarity based on the prosodic feature similarity and the phoneme feature similarity;

[0090] The result of subtracting the total similarity from the complete similarity corresponding to the complete similarity is used as the deviation value.

[0091] Specifically, the calibration parameter in the embodiment of the present invention is the deviation value divided by the complete similarity.

[0092] Specifically, the standard speech in the embodiment of the present invention is a speech sample without an accent or with a standard pronunciation.

[0093] Specifically, embodiments of the present invention may utilize speech processing technology, such as Dynamic Time Warping (DTW) or acoustic models, to calculate the similarity between the input speech and the standard speech.

[0094] Specifically, the embodiment of the present invention analyzes the prosodic features and phonemic features of the speech data and compares them with the standard speech to accurately determine the deviation value between the speech data and the standard speech. Based on this deviation value, the parameters of the initial speech conversion model are precisely adjusted so that the model can better adapt to and recognize different speech data, thereby improving the accuracy of speech recognition.

[0095] See also Figure 3 As shown, the step of standardizing the initial text information based on the standard language database includes:

[0096] Step S410, comparing the standard words in the standard language database with the initial text information one by one, determining abnormal words, calculating the similarity between the abnormal words and the standard words, and replacing the abnormal words corresponding to the similarity greater than the preset word similarity with the standard words;

[0097] Step S420: identifying actual grammatical rules in the initial text information, comparing the actual grammatical rules with standard grammatical rules in the standard language database, and determining and correcting incorrect grammar based on the comparison result.

[0098] Specifically, the embodiments of the present invention may correct grammatical errors in the initial text information, including tense consistency, subject-verb agreement, and article usage.

[0099] Specifically, the preset vocabulary similarity in the embodiment of the present invention is 80%. Based on historical comparison data, when the preset vocabulary similarity is set to 80%, it can be determined whether abnormal vocabulary needs to be replaced with standard vocabulary to reduce subsequent impact.

[0100] Specifically, the embodiment of the present invention can identify abnormal words by comparing the standard words in the standard language database with the words in the initial text information one by one. By calculating the similarity and replacing the abnormal words with a similarity higher than a preset threshold, the incorrect words in the text are significantly reduced and the accuracy of the text is improved. The replacement of abnormal words helps to ensure that the text follows the standard language specifications and makes the text clearer and more accurate. By identifying the actual grammatical rules in the initial text information and comparing them with the standard grammatical rules in the standard language database, grammatical errors can be accurately discovered and corrected, which helps to optimize the structure of the text, improve the grammatical correctness of the text, and make comprehensive standardized adjustments to the initial text information, effectively improving the overall quality of the text and improving processing efficiency and accuracy.

[0101] See also Figure 4 As shown, the step of determining vocabulary features based on the text information analysis results includes:

[0102] Step S510, performing word segmentation processing on the standard text information to obtain a number of word segments;

[0103] Step S520 , identifying the frequencies of occurrence of several segmented words, and determining the semantic features corresponding to the segmented words with the highest frequency of occurrence as the lexical features.

[0104] Specifically, the semantic features can be realized by querying a dictionary, using word vectors (such as Word2Vec, G loVe, etc.), or combining context information.

[0105] Specifically, the embodiment of the present invention uses word segmentation processing to decompose the text into vocabulary units with practical meaning, which helps to more accurately capture the key information in the text. The most frequently occurring word segments are selected as vocabulary features, which can highlight the core vocabulary in the text, thereby improving the accuracy of subsequent text analysis. The model trained based on these features can better adapt to different text data and improve the generalization performance of the model.

[0106] Specifically, the step of determining the emotional characteristics based on the speech data analysis results includes:

[0107] Extracting timbre features and speech rate features of the speech data;

[0108] Constructing a speech feature vector based on the timbre feature and the speech rate feature;

[0109] The emotional feature is determined based on the speech feature vector and a sentiment analysis model.

[0110] Specifically, the sentiment analysis model described in the embodiment of the present invention is constructed based on a number of speech feature vectors and their corresponding sentiment features. The sentiment analysis model is generated by training a number of speech feature vectors and their corresponding sentiment features through a machine learning algorithm. The algorithm can classify speech data and determine its sentiment label, such as positive, negative, neutral, etc.

[0111] Specifically, the timbre features of the embodiment of the present invention may include pitch, timbre loudness, etc., which can be extracted using methods such as short-time Fourier transform (STFT) and Mel-frequency cepstral coefficients (MFCC). The speech rate features may include speech speed, pause frequency, etc., which can be obtained by calculating the number of syllables or words per unit time in the speech signal.

[0112] Specifically, the embodiments of the present invention capture key information closely related to emotional state in speech signals by extracting timbre features and speech rate features, thereby improving the accuracy of emotion recognition. The use of the emotion analysis model further classifies and identifies the extracted features, improving the accuracy and efficiency of emotion recognition. By adjusting the model parameters and feature extraction methods, it can flexibly adapt to different application scenarios.

[0113] Specifically, the step of determining the dialogue feature based on the vocabulary feature and the emotion feature includes:

[0114] Constructing a feature vector based on the vocabulary feature and the sentiment feature;

[0115] The conversation feature is determined based on the feature vector.

[0116] Specifically, for lexical features, the embodiments of the present invention adopt methods such as TF-IDF (term frequency - inverse document frequency) and word embedding (such as Word2Vec) to convert the words in the text into numerical vectors. For example, the word vector obtained by training the Word2Vec model is used to represent each word segment, and these word vectors are aggregated (such as averaging, summing, etc.) to form a vector representation of the lexical features.

[0117] Specifically, the embodiments of the present invention use a sentiment analysis model (such as a deep learning-based model) to perform sentiment classification on the text for sentiment features, and convert the classification results (such as positive, negative, neutral, etc.) into numerical representations, which are directly used as vectors of sentiment features, or further encoded (such as independent encoding, embedded encoding, etc.) to adapt to the construction of feature vectors.

[0118] Specifically, embodiments of the present invention use various machine learning methods (such as classification, clustering, regression, etc.) to determine the characteristics of a conversation based on the constructed feature vector. For example, if our goal is to identify the topic of a conversation, we can train a classifier to predict the topic category of the conversation based on the feature vector.

[0119] Specifically, the embodiments of the present invention capture the information in the conversation more comprehensively by combining lexical features and sentiment features, thereby improving the accuracy of conversation understanding. The combination of lexical features and sentiment features enables a more accurate understanding of the intention and meaning of the conversation.

[0120] Specifically, the steps of determining scene labels based on the results of conversation feature analysis and establishing an initial conversation scene model include:

[0121] Determining the scene label based on matching the conversation features with a preset scene classification model;

[0122] The initial dialogue scenario model is constructed according to the scenario label and the preset scenario template.

[0123] Specifically, the preset scene classification model described in the embodiment of the present invention is a model obtained by training a large amount of scene data through a machine learning algorithm. The model can identify the corresponding scene category based on the input dialogue features. During the matching process, the embodiment of the present invention will calculate the similarity between the dialogue features and the various scene categories in the preset scene classification model, and select the scene category with the highest similarity as the scene label.

[0124] Specifically, the embodiments of the present invention can accurately identify the scene category of the conversation by matching conversation features with preset scene classification models, providing an important basis for subsequent conversation understanding and processing. By using preset scene templates to construct an initial conversation scene model, it can quickly generate conversation content that meets the scene requirements, thereby improving the processing efficiency and accuracy of the conversation system. The establishment of the initial conversation scene model provides a basic framework for subsequent conversation generation and response, which helps to achieve a more natural and smooth conversation interaction experience.

[0125] Specifically, the step of generating initial answer information based on the initial dialogue scene model and the processed voice data includes:

[0126] Determining a target dialogue scene to which the processed speech data belongs based on the initial dialogue scene model;

[0127] Generate initial response information based on the target dialogue scenario and the preset scenario response strategy;

[0128] The preset scenario response strategy includes response templates and response content under different dialogue scenarios.

[0129] Specifically, the preset scenario response strategy described in the embodiment of the present invention may include some common dialogue scenarios and their corresponding response templates and response content. For example, in a shopping scenario, the response template may include product introduction, price inquiry, purchase suggestions, etc., and the response content is filled in according to the specific product information and user needs.

[0130] Specifically, the embodiment of the present invention utilizes the initial dialogue scene model to perform scene recognition on the processed voice data to determine the target dialogue scene to which it belongs. It can quickly and accurately identify the scene category to which the voice data belongs. After determining the target dialogue scene, it generates initial answer information according to the preset scene reply strategy, which provides a basis for subsequent dialogue generation and response. At the same time, since the preset scene reply strategy is configurable, it can be adjusted and optimized according to actual needs to adapt to different application scenarios and needs.

[0131] Specifically, the step of determining the degree of deviation of the model answer based on the feedback information includes:

[0132] Collect user feedback data on initial responses, including satisfaction ratings and response content evaluations;

[0133] Extracting keywords from the response content evaluation, analyzing the keywords, and determining text satisfaction based on the analysis results;

[0134] determining actual satisfaction based on the satisfaction rating and textual satisfaction;

[0135] The degree of deviation is determined based on the actual satisfaction level and a preset satisfaction level.

[0136] Specifically, the preset satisfaction level in the embodiment of the present invention is a satisfaction level threshold value set according to historical data.

[0137] Specifically, the embodiments of the present invention collect user feedback data on initial answer information, understand user satisfaction and specific evaluation of the model answer, extract keywords in the reply content evaluation, and conduct in-depth analysis, which helps to discover users' specific opinions and suggestions on the model answer. By analyzing keywords, the text satisfaction is determined, that is, the user's satisfaction with the content of the model answer. Further combined with the satisfaction score, the actual satisfaction is comprehensively obtained, thereby more comprehensively evaluating the quality of the model answer.

[0138] Specifically, the step of adjusting the speech conversion model or the initial dialogue scene model based on the judgment result in the embodiment of the present invention includes:

[0139] If the deviation is greater than a preset deviation, adjusting the speech-to-text model, including optimizing parameters of a speech-to-text algorithm or updating training data of the speech-to-text model;

[0140] If the deviation degree is less than or equal to the preset deviation degree, the initial dialogue scene model is adjusted, including adjusting the parameters of the scene classification model, or updating the scene template and the preset reply strategy.

[0141] The preset reply strategy includes answer length limitation and keyword replacement rules.

[0142] Specifically, in an embodiment of the present invention, when an answer is too long, it is trimmed according to the length limit, and an intelligent analysis is performed based on the key information points in the answer. The most critical parts for understanding the question are retained first, while redundant or secondary information is removed. The preset length is the average length of historical answers. When certain words in the answer are inaccurate, they are replaced according to the keyword replacement rules. The keyword replacement rules are to establish a keyword replacement list, which lists the incorrect forms and the corresponding correct forms. When the system detects these incorrect forms in the answer, it will automatically replace them, thereby ensuring the accuracy and authority of the answer. A context analysis mechanism is also incorporated, and the system will replace keywords more accurately based on the overall content of the answer and the context.

[0143] Specifically, embodiments of the present invention employ different adjustment strategies depending on the degree of deviation. When the degree of deviation is greater than a preset degree of deviation, it indicates that the speech-to-text model has a large error in converting speech to text, resulting in a large deviation in the selected dialogue features, which in turn leads to a large error in constructing the initial dialogue scene model. As a result, the generated initial answer information deviates significantly from the user's true intention. Embodiments of the present invention adjust the speech-to-text model, including optimizing the parameters of the speech-to-text algorithm or updating the training data of the speech-to-text model to improve the accuracy of speech conversion, thereby improving the accuracy of constructing the initial dialogue scene model, so that the obtained target answer information is better than the initial answer information.

[0144] When the degree of deviation is less than the preset degree of deviation, the initial dialogue scene model is adjusted, including adjusting the parameters of the scene classification model, or updating the scene template and the preset response strategy, so as to improve the accuracy of scene recognition and the effectiveness of the response strategy, further reduce the degree of answer deviation, and make the obtained target answer information better than the initial answer information.

[0145] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

[0146] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that the present invention is susceptible to various modifications and variations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An intelligent dialogue scene analysis method based on AI big model, characterized by: include: Receive the user's initial voice data, analyze the initial voice data, and identify background noise and voice accent based on the analysis results; Processing the background noise, obtaining processed speech data, analyzing the speech accent, and determining calibration parameters based on the analysis results; Adjusting initial conversion parameters in the initial speech conversion model based on the calibration parameters to obtain an adjusted speech conversion model, and converting the processed speech data based on the adjusted speech conversion model to obtain initial text information; Standardize and adjust the initial text information based on a standard language database to obtain standard text information; Analyzing the standard text information, determining vocabulary features based on the text information analysis results, analyzing the processed voice data, determining emotional features based on the voice data analysis results, and determining conversation features based on the vocabulary features and the emotional features; Analyzing the conversation features, determining scene labels based on the conversation feature analysis results, and constructing an initial conversation scene model based on the scene labels; generating initial answer information based on the initial dialogue scenario model and the processed voice data; Collect feedback information on the initial answer information, judge the degree of model answer deviation based on the feedback information, and adjust the speech conversion model or the initial dialogue scene model based on the judgment result, so that the target answer information obtained after adjustment is better than the initial answer information.

2. The intelligent dialogue scene analysis method based on AI big model according to claim 1 is characterized in that: The step of identifying background noise and voice accent according to the analysis result comprises: Identifying energy features, prosodic features, and phoneme features of the initial speech data; identifying the background noise based on the energy signature; The speech accent is recognized based on the prosodic features and the phonemic features.

3. The intelligent dialogue scene analysis method based on AI big model according to claim 2 is characterized in that: The step of determining the calibration parameters according to the analysis results comprises: determining a deviation value between the speech data and a standard speech based on the prosodic feature and the phoneme feature; The calibration parameter is determined based on the offset value.

4. The intelligent dialogue scene analysis method based on AI big model according to claim 3 is characterized in that: The step of standardizing the initial text information based on the standard language database includes: Comparing the standard words in the standard language database with the initial text information one by one, determining abnormal words, calculating the similarity between the abnormal words and the standard words, and replacing the abnormal words corresponding to the similarity greater than the preset word similarity with the standard words; Identify actual grammatical rules in the initial text information, compare the actual grammatical rules with standard grammatical rules in the standard language database, and determine and correct incorrect grammar based on the comparison result.

5. The intelligent dialogue scene analysis method based on AI big model according to claim 4 is characterized in that: The step of determining vocabulary features based on the text information analysis results includes: Performing word segmentation processing on the standard text information to obtain a number of word segments; The occurrence frequencies of several segmented words are identified, and the semantic features corresponding to the segmented words with the highest occurrence frequency are determined as the lexical features.

6. The intelligent dialogue scene analysis method based on AI big model according to claim 5 is characterized in that: The step of determining the emotional characteristics according to the speech data analysis result comprises: Extracting timbre features and speech rate features of the speech data; Constructing a speech feature vector based on the timbre feature and the speech rate feature; The emotional feature is determined based on the speech feature vector and a sentiment analysis model.

7. The intelligent dialogue scene analysis method based on AI big model according to claim 6 is characterized in that: The step of determining the dialogue feature based on the vocabulary feature and the emotion feature comprises: Constructing a feature vector based on the vocabulary feature and the sentiment feature; The conversation feature is determined based on the feature vector.

8. The intelligent dialogue scene analysis method based on AI big model according to claim 7 is characterized in that: The steps of determining scene labels based on the results of conversation feature analysis and establishing an initial conversation scene model include: Determining the scene label based on matching the conversation features with a preset scene classification model; The initial dialogue scenario model is constructed according to the scenario label and the preset scenario template.

9. The intelligent dialogue scene analysis method based on AI big model according to claim 8 is characterized in that: The step of generating initial answer information based on the initial dialogue scenario model and the processed voice data includes: Determining a target dialogue scene to which the processed speech data belongs based on the initial dialogue scene model; Generate initial response information based on the target dialogue scenario and the preset scenario response strategy; The preset scenario response strategy includes response templates and response content under different dialogue scenarios.

10. The intelligent dialogue scene analysis method based on AI big model according to claim 9 is characterized in that: The step of determining the degree of deviation of the model answer based on the feedback information includes: Collect user feedback data on initial responses, including satisfaction ratings and response content evaluations; Extracting keywords from the response content evaluation, analyzing the keywords, and determining text satisfaction based on the analysis results; determining actual satisfaction based on the satisfaction rating and textual satisfaction; The degree of deviation is determined based on the actual satisfaction level and a preset satisfaction level.

Images