Speech text generation method, device, system, electronic device and storage medium

Abstract

The present invention discloses a method, device, system, electronic device and storage medium for generating speech text, the method comprising: obtaining characteristic data of a user when the user is giving a speech, the characteristic data comprising: data on the user's speech process behavior and data on the evaluation of the speech audience; performing feature mining analysis on the characteristic data using a mining knowledge tree to obtain mining data; then calling a deep learning model to generate basic text information according to the mining data, and calling a preset real-time content adjustment model to adjust the content of the basic text information to obtain the speech text. The present invention can generate text that conforms to the user's current scene in real time in combination with the characteristic data of the user during the speech, thereby fitting the user's actual scene, avoiding the situation where the text deviates from the real-time scene, and can improve the accuracy of the generated text data.

Description

Background Art

[0002] With the development of economy and technology, there are more and more promotion channels in various industries, and the demand for speech text data is also increasing. In order to meet the needs of users for speech text data, natural language generation (Neural Language Generation, NLG) technology can be used to automatically generate text data for users to view.

[0003] The commonly used method at present is: pre-train an end-to-end text generation model, input the individual keyword values ​​of user needs into the trained text generation model, let the text generation model make predictions based on the keywords, and finally combine the predicted text with the keywords to construct the required text data.

[0004] However, the currently commonly used methods have the following technical problems: in real-time speeches or live broadcasts, the scope of the fields involved is wide and many keywords are required, while the text data generated by a single-dimensional end-to-end text generation model is difficult to adapt to multiple keywords, resulting in deviations between the text data and the text data required for real-time speeches or live broadcasts, and the accuracy of the generated text data is low. Summary of the invention

[0005] The present invention provides a method, device, system, electronic device and storage medium for generating speech text, and the method can solve one or more of the above-mentioned technical problems.

[0006] A first aspect of an embodiment of the present invention provides a method for generating a speech text, the method comprising:

[0007] Acquire characteristic data of the user when the user is giving a speech, wherein the characteristic data includes: data on the user's behavior during the speech and data on the evaluation of the speech audience;

[0008] Performing feature mining analysis on the feature data using a preset mining knowledge tree to obtain mining data, wherein the preset mining knowledge tree is a knowledge tree constructed by using topic data after acquiring multiple different market topic data;

[0009] After calling the preset deep learning model to generate basic text information based on the mined data, the preset real-time content adjustment model is called to adjust the content of the basic text information to obtain the speech text, wherein the preset deep learning model is a model that extracts data about speech eloquence and style from a text training set and performs deep learning training, and the preset real-time content adjustment model is a time series model that receives feedback data about user speech feedback from multiple different social platforms and performs training and optimization respectively.

[0010] In a possible implementation of the first aspect, the preset deep learning model includes: an eloquence deep learning model for generating eloquence content text and a style decoder for generating style content text;

[0011] The calling of a preset deep learning model to generate basic text information according to the mining data includes:

[0012] The mined data are respectively input into the eloquence deep learning model and the style decoder, so that the eloquence deep learning model generates eloquence text based on the mined data, and the style decoder generates style text based on the mined data, and the eloquence text and the style text are combined to obtain basic text information.

[0013] In a possible implementation manner of the first aspect, the eloquence deep learning model generates an eloquence text based on the mining data, including:

[0014] Extracting a plurality of eloquence dimension data from the mining data;

[0015] Calling the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values;

[0016] Each eloquence dimension score value is used to calculate the eloquence weighted value corresponding to each eloquence dimension data, and multiple eloquence weighted values ​​and multiple eloquence dimension data are input into the eloquence deep learning model to obtain the eloquence text.

[0017] In a possible implementation manner of the first aspect, the style decoder generates a style text based on the mined data, including:

[0018] A plurality of eloquence dimension data are extracted from the mining data, and a weighting factor corresponding to each of the eloquence dimension data is calculated to obtain a style weighting factor.

[0019] Each of the eloquence dimension data is converted into style dimension data using the weighting factor, and a preset style decoder is called to generate a style text according to the style dimension data.

[0020] In a possible implementation manner of the first aspect, calling a preset real-time content adjustment model to adjust the content of the basic text information to obtain the speech text includes:

[0021] After obtaining the user's interactive data, pre-processing the interactive data to obtain processed data, wherein the interactive data is data about the user's comments and interactive feedback during the real-time speech, and the pre-processing includes cleaning processing, encoding processing and embedding processing;

[0022] The preset real-time content adjustment model is called to generate text adjustment information according to the processed data, and a speech text is generated based on the text adjustment information and the basic text information.

[0023] In a possible implementation manner of the first aspect, the preset process of constructing the mining knowledge tree includes:

[0024] Collecting a plurality of market data, each of the market data is data collected from a different open source media platform;

[0025] Using machine learning technology to identify and process each of the market data to obtain popular topic data about popular market topics, and using natural language processing technology to perform topic analysis on each of the market data to obtain popular topic data about popular market topics;

[0026] The popular topic data and the popular subject data are used to construct a preset mining knowledge tree.

[0027] A second aspect of an embodiment of the present invention provides a device for generating a speech text, the device comprising:

[0028] A data acquisition module, used to acquire characteristic data of the user when the user is giving a speech, wherein the characteristic data includes: data on the user's behavior during the speech and data on the evaluation of the speech audience;

[0029] A data mining module, used for performing feature mining analysis on the feature data using a preset mining knowledge tree to obtain mining data, wherein the preset mining knowledge tree is a knowledge tree constructed by obtaining topic data on multiple different market topics and using the topic data;

[0030] The text generation module is used to call a preset deep learning model to generate basic text information according to the mining data, and then call a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text, wherein the preset deep learning model is a model that extracts data about speech eloquence and style from a text training set and performs deep learning training, and the preset real-time content adjustment model is a time series model that receives feedback data about user speech feedback from multiple different social platforms and performs training and optimization respectively.

[0031] In a possible implementation of the second aspect, the preset deep learning model includes: an eloquence deep learning model for generating eloquence content text and a style decoder for generating style content text;

[0032] The calling of a preset deep learning model to generate basic text information according to the mining data includes:

[0033] The mined data are respectively input into the eloquence deep learning model and the style decoder, so that the eloquence deep learning model generates eloquence text based on the mined data, and the style decoder generates style text based on the mined data, and the eloquence text and the style text are combined to obtain basic text information.

[0034] In a possible implementation of the second aspect, the eloquence deep learning model generates an eloquence text based on the mining data, including:

[0035] Extracting a plurality of eloquence dimension data from the mining data;

[0036] Calling the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values;

[0037] Each eloquence dimension score value is used to calculate the eloquence weighted value corresponding to each eloquence dimension data, and multiple eloquence weighted values ​​and multiple eloquence dimension data are input into the eloquence deep learning model to obtain the eloquence text.

[0038] In a possible implementation manner of the second aspect, the style decoder generates a style text based on the mined data, including:

[0039] A plurality of eloquence dimension data are extracted from the mining data, and a weighting factor corresponding to each of the eloquence dimension data is calculated to obtain a style weighting factor.

[0040] Each of the eloquence dimension data is converted into style dimension data using the weighting factor, and a preset style decoder is called to generate a style text according to the style dimension data.

[0041] In a possible implementation manner of the second aspect, the calling a preset real-time content adjustment model to adjust the content of the basic text information to obtain the speech text includes:

[0042] After obtaining the user's interactive data, pre-processing the interactive data to obtain processed data, wherein the interactive data is data about the user's comments and interactive feedback during the real-time speech, and the pre-processing includes cleaning processing, encoding processing and embedding processing;

[0043] The preset real-time content adjustment model is called to generate text adjustment information according to the processed data, and a speech text is generated based on the text adjustment information and the basic text information.

[0044] In a possible implementation of the second aspect, the preset process of constructing the mining knowledge tree includes:

[0045] Collecting a plurality of market data, each of the market data is data collected from a different open source media platform;

[0046] Using machine learning technology to identify and process each of the market data to obtain popular topic data about popular market topics, and using natural language processing technology to perform topic analysis on each of the market data to obtain popular topic data about popular market topics;

[0047] The popular topic data and the popular subject data are used to construct a preset mining knowledge tree.

[0048] A third aspect of an embodiment of the present invention provides a system for generating speech texts, the system comprising: an expanded user data collection module, a dynamic knowledge tree construction module, a deep learning driven content generation module, and a real-time content adjustment module;

[0049] The extended user data collection module is used to obtain user behavior data and context data;

[0050] The dynamic knowledge tree construction module is used to construct a knowledge tree and use the knowledge tree to perform data mining based on user data;

[0051] The deep learning driven content generation module is used to generate basic text information based on the mined data;

[0052] The real-time content adjustment module is used to generate text adjustment information according to the user's behavior data and context data, and generate a speech text using the text adjustment information and basic text information.

[0053] Compared with the prior art, the embodiments of the present invention provide a method, device, system, electronic device and storage medium for generating a speech text, which has the following beneficial effects: the present invention can obtain the characteristic data of a user when the user is giving a speech; use a mining knowledge tree to perform feature mining analysis on the characteristic data to obtain mining data; then call a deep learning model to generate basic text information based on the mining data, call a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text, and combine the characteristic data of the user during the speech to generate text that conforms to the user's current scenario in real time, thereby fitting the user's actual scenario and avoiding the situation where the text deviates from the real-time scenario, which can improve the accuracy of the generated text data. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of a method for generating a speech text provided by an embodiment of the present invention;

[0055] Figure 2 It is a structural schematic diagram of a speech text generation device provided by an embodiment of the present invention;

[0056] Figure 3 It is a structural diagram of a speech text generation system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] In order to solve the above problems, a method for generating a speech text provided in an embodiment of the present application will be introduced and explained in detail through the following specific embodiments.

[0059] Reference Figure 1 , which shows a flow chart of a method for generating a speech text provided by an embodiment of the present invention.

[0060] In one embodiment, the method is applicable to a smart device, which can be a smart device for users to conduct online live broadcasts or online meetings, and can also be applicable to background control devices at the speech site. The smart device can quickly generate the text required by the user, making it convenient for the user to conduct online live speeches, online meetings, or offline speeches.

[0061] As an example, the method for generating the speech text may include:

[0062] S11. Acquire characteristic data of the user when the user is giving a speech, wherein the characteristic data includes data on the user's behavior during the speech and data on the evaluation of the speech audience.

[0063] In one embodiment, the present invention can generate text before a speech or during a speech. During a speech, the user's characteristic data can be obtained, and the characteristic data can include the user's personal behavior data during the speech, and can also include the evaluation data of the audience listening to the speech.

[0064] In one embodiment, the data can be collected by automatically collecting data from the user's social media account and web browsing history, collecting data from social media, speech feedback, online activities, etc. Sensor technology, user devices (such as smartphones, computers) and online platforms are used to collect data about the user's current situation, such as location, time, type of activity, etc. The user's environment and activities can also be monitored in real time, such as meetings, social activities or being alone, to obtain the user's behavioral data.

[0065] After acquiring the data, the data can be preprocessed. For example, the collected data can be cleaned and standardized and converted into a format that can be analyzed to obtain characteristic data.

[0066] In one embodiment, an adaptive algorithm may be set up to perform self-optimization according to constantly changing user data and feedback.

[0067] In one implementation, the self-optimization process using an adaptive algorithm in the context of analyzing user behavior and context in a speech or presentation scenario may be as follows:

[0068] First, continuous learning: It can continuously learn from new user data, including social media interactions, browsing habits, and feedback from speeches or demonstrations. This learning process is dynamic and can adapt to changes in user behavior and preferences.

[0069] Second, predictive analytics: The algorithm predicts user needs and preferences based on learned patterns. It uses these predictions to adjust content generation strategies for voice or presentation contexts.

[0070] Third, feedback incorporation: user feedback on the effectiveness of generated content is collected and analyzed. This feedback is used to refine the predictions and adjustments made by the algorithm.

[0071] Fourth, iterative improvement: The algorithm iteratively improves its predictions and strategies based on continuous analysis of user data and feedback, ensuring that the generated content becomes more and more in line with user needs and context.

[0072] Adaptive algorithms ensure that the recommendations and generated content remain relevant and effective over time.

[0073] In an optional embodiment, after acquiring the characteristic data, the characteristic data can be used to analyze the user's psychological characteristics and emotional state. For example, existing user behavior tracking technology and psychological analysis algorithms can be used to include emotion recognition and psychological tendency analysis in addition to standard social media interaction and browsing habit analysis. Through deep learning and big data analysis technology, deeper psychological characteristics and emotional states can be extracted from the user's online behavior.

[0074] Emotion recognition can apply voice and expression recognition technology to analyze users’ emotional expressions on different social media. Psychological tendency analysis can use psychology principles and machine learning to analyze users’ personality traits, motivations, and preferences.

[0075] In actual implementation, the specific means and process of analyzing psychological characteristics and emotional states are as follows:

[0076] First, the means and process of extracting psychological characteristics and emotional states: Natural language processing (NLP) technology can be used to analyze users' speeches and interactions on social media to identify emotional tendencies and psychological characteristics. Using applied text mining technology, keywords and topics can be extracted from users' social media speeches to analyze emotions and psychological states. Combined with machine learning algorithms such as support vector machines (SVM) and random forests (RF), deep mining of big data can be carried out to identify user characteristics and behavioral trends.

[0077] Second, psychological characteristics include: the user's personality traits, such as openness, responsibility, extroversion, etc. Motivations and preferences, such as the intensity of the user's interest in certain topics or activities.

[0078] Third, emotional states include: basic emotions, such as joy, sadness, anger, surprise, etc. Emotional tendencies, such as positive, negative or neutral emotional expressions.

[0079] Through the above methods, we can have a deeper understanding of the user's psychological characteristics and emotional state, thereby generating speech content that is more in line with the user's personality and needs.

[0080] In an optional embodiment, after acquiring the feature data, the feature data can be used to perform behavioral analysis on the user. In a specific implementation, the behavioral analysis can use text mining and natural language processing (NLP) technology to analyze the user's speech, interaction and sharing behaviors on the social platform to extract emotional tendencies, points of interest and behavioral patterns.

[0081] The specific process of analyzing users’ speech, interaction and sharing behaviors on social media to obtain emotional tendencies, interests and behavior patterns is as follows:

[0082] First, social media analysis: Use text mining and natural language processing (NLP) technology to analyze users' social media posts. This includes keyword extraction, topic analysis, and sentiment analysis. Identify emotional expressions in user posts, such as joy, sadness, anger, etc., as well as attitudes towards specific topics or activities.

[0083] It should be noted that the use of NLP and data mining techniques to extract key features, such as keywords, topics, emotions, etc. This technical means is to use natural language processing (NLP) and data mining techniques to extract key features, which is an existing technology. These technologies are widely used in text analysis, sentiment analysis, and user behavior analysis. NLP technology can process and understand human language and is used to identify keywords, topics, and emotions. Data mining technology identifies patterns and trends by analyzing large amounts of data. These technologies have application examples in areas such as social media analysis, web content analysis, and consumer behavior research. The application of these technologies is to gain a deeper understanding of users' interests, emotions, and behavior patterns, so as to generate more personalized content.

[0084] Second, interaction and sharing behavior analysis: Evaluate the way users interact with others, such as likes, comments, and reposts, to identify users’ social preferences and behavior patterns. Analyze the types of content users share to understand their interests and behavior trends.

[0085] Third, data mining and machine learning: Apply advanced data mining techniques and machine learning algorithms, such as cluster analysis and decision trees, to process and interpret the large amounts of data collected. Identify user behavior characteristics from the data, such as the user's interest in certain topics, frequency of activity participation, and interaction patterns in social networks.

[0086] Through this process, we can deeply understand the user's psychological characteristics and emotional state, and provide support for the generation of personalized speech content. In addition, we can analyze the user's behavior patterns through algorithms and output the user's behavior characteristics and preference analysis results.

[0087] In one implementation, the specific process of analyzing user behavior patterns to obtain behavior characteristics and preferences is as follows:

[0088] First, data collection and preprocessing: Automatically collect data from users’ social media accounts and web browsing records, and clean and standardize the collected data.

[0089] Second, feature extraction: Use natural language processing (NLP) and data mining techniques to extract key features from the data, such as keywords, topics, and emotions.

[0090] Third, behavioral pattern analysis: Apply machine learning algorithms to analyze user behavior data and identify behavioral patterns and trends. Use cluster analysis, decision trees and other methods to identify user behavior characteristics, such as user interest in specific topics and participation in social activities.

[0091] Fourth, preference analysis and output: Based on the results of behavioral pattern analysis, identify user preferences, such as preferences for specific topics and tendencies in interaction methods. Output detailed user behavior characteristics and preference analysis results to provide a basis for personalized content generation.

[0092] Through this process, we can gain in-depth understanding of users’ behavioral characteristics and preferences, providing important information for generating personalized speech content.

[0093] In another optional embodiment, the above-mentioned characteristic data can also be used to track browsing behavior, wherein the browsing behavior tracking can apply web page tracking technology to collect the user's browsing history and activity data on the Internet, so as to understand the user's interests and preferences. At the same time, the characteristic data can also be used for data mining and machine learning, and advanced data mining technology and machine learning algorithms (such as cluster analysis, decision tree) can be combined to analyze the large amount of collected data to identify user characteristics and behavior trends.

[0094] The specific process of analyzing data to identify user characteristics and behavior trends is as follows:

[0095] First, data collection: collect user data on social media, web browsing, online activities, etc.

[0096] Second, data preprocessing: clean the data, remove irrelevant or erroneous information, and standardize the data to ensure that the data format is consistent.

[0097] Third, feature extraction: Use natural language processing (NLP) and data mining techniques to extract key features such as keywords, topics, user behavior patterns, etc.

[0098] Fourth, machine learning analysis: Apply machine learning algorithms such as cluster analysis and decision trees to conduct in-depth analysis of features. Identify and classify user behavior trends and personality characteristics.

[0099] Fifth, result output: Based on the analysis results, build a user behavior and psychological characteristic model. Output a detailed report on user characteristics and behavior trends.

[0100] Through this process, we can also gain in-depth understanding of the user's personality, interests, emotional state, and behavioral habits, providing support for the generation of personalized speech content.

[0101] In another optional embodiment, after acquiring the feature data, the feature data may also be used for real-time context analysis. The real-time context analysis may use real-time data stream analysis technology, combined with a context-aware algorithm, to dynamically evaluate the user's speech expression needs in a specific context. The real-time context analysis may be to monitor the user's speech environment and audience feedback in real time, and dynamically adjust the content generation strategy. In addition, a multi-dimensional context model may be constructed based on the analysis results of the real-time context analysis, and the multi-dimensional context model may be a comprehensive context model.

[0102] Real-time contextual analysis can apply real-time data processing technology to quickly analyze the speech or communication situation that the user is currently in. Then use deep learning models (such as convolutional neural networks) to conduct in-depth analysis of the situational data to understand the user's possible needs and reactions in a specific situation.

[0103] Optionally, real-time context analysis can also process feature data, including data cleaning, denoising, and format conversion, to ensure data quality. The processed data is then used to predict demand, and the user's speech expression needs in specific situations are identified based on the processed data, such as a formal speech style in a formal meeting, or a relaxed conversation style in an informal social setting. Speech style prediction can then be performed, combining user historical data and current context analysis to predict the most suitable speech style for the situation. For example, an authoritative and rigorous speech style may be recommended for formal occasions. Finally, based on the prediction results, speech expression suggestions for users in specific situations can be generated, including suggestions on content, tone, style, etc.

[0104] For the output suggestions, we can first combine the user's behavior analysis and situation analysis results, and use deep learning and data mining technology to comprehensively evaluate the user's personality characteristics and current situation needs. Then, based on these comprehensive analysis results, we use natural language processing (NLP) technology and user behavior pattern analysis to generate personalized speech expression suggestions. For the rules and conditions for matching suggestions and the matching process: matching suggestions can be based on user behavior characteristics (such as interests, speech experience) and current situations (such as audience characteristics, activity types). The matching process includes analyzing user data, identifying key features, and then comparing these features with preset eloquence expression models to determine the most appropriate suggestions. Historical feedback can be considered to continuously optimize the accuracy and applicability of suggestions.

[0105] By processing the feature data in the above manner, the accuracy and efficiency of behavior analysis can be improved. Moreover, it can be tailored to the user's current actual speech environment and user needs, thereby improving the accuracy of subsequent speech text generation.

[0106] S12. Perform feature mining analysis on the feature data using a preset mining knowledge tree to obtain mining data. The preset mining knowledge tree is a knowledge tree constructed by obtaining topic data on multiple different market topics and using the topic data.

[0107] In one embodiment, the feature data is subjected to feature mining analysis using a preset mining knowledge tree, which can track market trends, the latest research, technology and strategies related to speech and eloquence. Advanced data analysis and artificial intelligence prediction models can be combined, such as using sentiment analysis to understand the popularity of speeches and speech content.

[0108] Among them, feature mining analysis uses data analysis and machine learning techniques (such as time series analysis and clustering algorithms) to analyze market data and identify popular topics and trends. It usually combines data analysis and machine learning techniques, such as time series analysis and clustering algorithms, to identify patterns, trends, and potential associations from large amounts of data. The use of the above technologies can help analyze and predict market dynamics, user behavior, product trends, etc., thereby providing support for decision-making. In the field of speech and eloquence, these technologies can be used to analyze and predict popular trends in speech content, style, strategy, etc.

[0109] When conducting feature mining analysis, user group dynamic analysis can also be performed based on the above feature data. User group dynamic analysis can analyze the dynamic needs of different user groups in terms of eloquence, including speech style, content preference, interaction method, etc. Group behavior pattern recognition technology based on machine learning can be used to automatically identify and adapt to changes in different user groups.

[0110] The operations for dynamic analysis of user groups can be as follows:

[0111] First, collect user interaction data on the platform, such as browsing, likes, comments, and sharing.

[0112] Second, use statistical analysis and machine learning methods (such as factor analysis and predictive modeling) to analyze the behavioral patterns of user groups. Specifically, statistical analysis and machine learning methods, such as factor analysis and predictive modeling, can usually be combined to analyze and explain patterns and trends in large amounts of user data. These methods enable researchers to identify and understand the behavioral characteristics, preferences, and dynamic changes of specific user groups, thereby providing valuable insights for decision-making.

[0113] Third, create a user group behavior model to predict the changing trends and preferences of different groups. The process of creating a user group behavior model involves the following steps:

[0114] 1. Data basis: Model construction is based on user behavior data on the platform, such as browsing history, likes, comments, sharing, etc. In addition, it may also include user personal information (such as age, gender, geographic location) and user response to specific content.

[0115] 2. Model building process: Data collection: Collect user behavior and feedback data from different platforms and social media. Data preprocessing: Clean and standardize the data to make it suitable for analysis. Feature extraction: Analyze the data and extract key features, such as user interests, interaction patterns, etc. Model training: Use machine learning algorithms (such as clustering, predictive modeling) to train feature data and create a model.

[0116] 3. Trend prediction process: Data analysis: Use the constructed model to analyze user behavior and preferences. Trend identification: Based on the model output, identify the behavior trends and preferences of different user groups. Prediction and adaptation: Predict future trends and adjust services or content based on these trends to meet the needs of different groups.

[0117] The entire process relies on big data analysis and machine learning techniques to ensure the accuracy and adaptability of the model.

[0118] In an optional embodiment, the preset process of constructing the mining knowledge tree may include the following sub-steps:

[0119] S21. Collect multiple market data, each of which is collected from different open source media platforms.

[0120] S22. Using machine learning technology to identify and process each of the market data to obtain popular topic data about popular market topics, and using natural language processing technology to perform topic analysis on each of the market data to obtain popular topic data about popular market topics.

[0121] S23: constructing a preset mining knowledge tree using the popular topic data and the popular subject data.

[0122] First, you can collect multiple market data. You can use crawler technology and API interfaces to collect data from multiple sources (such as social media, news portals, industry reports, speeches, and training platforms). Then, perform data synchronization to continuously synchronize the latest market data and trend information. Next, you can perform real-time analysis or identification processing on the data to extract market trends and hot topics and obtain popular topic data. You can also use natural language processing (NLP) to perform sentiment analysis, keyword extraction, and topic model analysis on the collected text data to understand popular topics in the market.

[0123] When performing natural language processing (NLP) analysis on the collected data, the following methods are usually used:

[0124] First, sentiment analysis: using machine learning-based algorithms, such as support vector machines (SVM) or deep neural networks, to identify emotional tendencies in text, such as positive, negative, or neutral.

[0125] Second, keyword extraction: Use techniques such as TF-IDF (term frequency-inverse document frequency) or neural network-based word embedding methods to extract keywords or phrases from the text to highlight the main topic or content.

[0126] Third, topic model analysis: use algorithms such as Latent Dirichlet Allocation (LDA) to identify hidden topic structures in the data set and obtain popular topic data.

[0127] It should be noted that the above operation method can adopt existing technologies and is widely used in many fields such as text analysis, market research and user behavior analysis. The above technology can help understand the patterns and trends in large amounts of text data, which is particularly useful for market trend analysis and user interest insights.

[0128] Finally, use popular topic data and popular subject data to build or update the knowledge tree. Based on the analysis results, dynamically update the knowledge tree to ensure that the content is consistent with current market trends and user interests.

[0129] In one embodiment, dynamically updating the knowledge tree based on popular topic data and popular subject data generally involves the following steps:

[0130] First, data analysis: Machine learning and NLP techniques can be used to analyze the latest market trends and user behavior data. This includes sentiment analysis, trend detection, and behavioral pattern recognition.

[0131] Second, identify key insights: From the analysis, key insights and trends related to speeches and presentations were identified, such as trending topics, emerging strategies, and audience preferences.

[0132] Third, update the knowledge tree: The knowledge tree can then be updated to incorporate these insights. This includes adding new topics, adjusting the importance of existing topics, and modifying the connections between different nodes based on the latest trends and user interests.

[0133] Fourth, continuous monitoring and feedback integration: New data and user feedback can be continuously monitored and used to further improve and update the knowledge tree.

[0134] The above operations ensure that the knowledge tree remains up-to-date and accurately reflects the latest market trends and user needs in the field of speech and presentation.

[0135] In an optional embodiment, when performing dynamic analysis of user groups, feature data may also be aggregated: data may be aggregated from user interactions and feedback to build a user behavior database.

[0136] The fused data is then used to perform behavioral pattern recognition to analyze user data and identify the behaviors and preference patterns of different groups. Optionally, the knowledge tree can also be adjusted based on the analysis results of the behaviors of different groups to meet the recognition needs of different user groups.

[0137] Optionally, you can also add eloquence dimension indicators to update the knowledge tree. For example, you can introduce eloquence dimension indicators in market trend analysis, such as speech influence, audience participation, content attractiveness, etc. After the update, the knowledge tree not only reflects market trends and user interests, but also dynamically adjusts content based on eloquence dimension indicators. Then analyze the effect of speech expression through user feedback and behavior data on the platform.

[0138] After the knowledge tree is updated, a performance benchmark test may be performed on it, specifically, the update speed and adaptability of the knowledge tree may be tested.

[0139] S13. After calling a preset deep learning model to generate basic text information according to the mined data, call a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text, wherein the preset deep learning model is a model that extracts data about speech eloquence and style from a text training set and performs deep learning training, and the preset real-time content adjustment model is a time series model that receives feedback data about user speech feedback from multiple different social platforms and performs training and optimization respectively.

[0140] In one embodiment, after obtaining the mining data, the preset deep learning model can be directly called to generate basic text information according to the mining data, and the basic text information can be the text information of the speech content. On this basis, the preset real-time content adjustment model is called to adjust the content of the basic text information, and the speech text can be obtained.

[0141] It should be noted that before a speech, the user can view the speech text on his smart device and make a speech based on the text. If the speech is in progress, the user can use the speech text as a reference to adjust the original text.

[0142] Among them, the preset deep learning model is a model obtained by pre-acquiring multiple text training sets, extracting data about speaking eloquence and style from the text training sets and performing deep learning training.

[0143] The preset real-time content adjustment model is a time series model that receives feedback data from listeners on multiple different social platforms after listening to a user's speech, and then uses the feedback data for training and optimization.

[0144] In one of the embodiments, the preset deep learning model includes: an eloquence deep learning model for generating eloquence content text and a style decoder for generating style content text;

[0145] Wherein, as an example, the calling of a preset deep learning model to generate basic text information according to the mining data may include the following sub-steps:

[0146] S31. Input the mined data into the eloquence deep learning model and the style decoder respectively, so that the eloquence deep learning model generates eloquence text based on the mined data, and the style decoder generates style text based on the mined data, and combines the eloquence text and the style text to obtain basic text information.

[0147] Specifically, the mining data can be input into the eloquence deep learning model and the style decoder respectively, so that the eloquence deep learning model can generate eloquence text according to the mining data, and the style decoder can generate style text according to the mining data; finally, the two texts can be combined to obtain basic text information, and specifically, the two texts can be spliced ​​front and back to form a basic text information.

[0148] Wherein, as an example, the eloquence deep learning model generates eloquence text based on the mining data, which may include the following sub-steps:

[0149] S41. Extracting a plurality of eloquence dimension data from the mining data.

[0150] S42. Call the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values.

[0151] S42. Use each of the eloquence dimension scoring values ​​to calculate the eloquence weighted value corresponding to each of the eloquence dimension data, and input multiple of the eloquence weighted values ​​and multiple of the eloquence dimension data into the eloquence deep learning model to obtain the eloquence text.

[0152] Specifically, feature extraction and analysis can be performed first, and multiple eloquence dimension data can be extracted from the mining data, wherein the eloquence dimension data can include indicators related to the eloquence dimension. Then, the eloquence deep learning model is used to analyze these features and predict the score of each eloquence dimension.

[0153] In the process of "using deep learning models to analyze these features and predict the score of each eloquence dimension", the steps usually include the following aspects:

[0154] First, feature extraction: extract features related to eloquence dimensions from user data, such as persuasiveness, emotional expression, etc. These features may include data on user interactions, style and structure of text content, etc.

[0155] Second, model training: Use these features to train a deep learning model of eloquence. This process may include using text data to train a neural network to recognize and predict different aspects of the eloquence dimension.

[0156] Third, score prediction: Use the trained model to predict the score of each eloquence dimension. The model will analyze the extracted features and give a score for each dimension, such as persuasiveness, emotional expression, etc.

[0157] Fourth, optimize and iterate: Based on user feedback and additional data, continuously optimize the accuracy and predictive ability of the model. This may involve adjusting model parameters or adding new data to improve prediction results.

[0158] After calculating the score value, the weighted factor of the eloquence dimension can be calculated, and the weighted factor can be used as the eloquence weighted value. Finally, the eloquence deep learning model can generate an eloquence text based on the eloquence weighted value and the eloquence dimension data.

[0159] In one embodiment, the eloquence dimension data may be expressed as follows:

[0160] Y=Activation(W n ·(LayerNorm(Attention(Q,K,V)))+b n .

[0161] In the above formula, Y is the output of the model, which is the eloquence dimension data and can represent the content adjusted based on the eloquence dimension. n and b n are the weights and biases of the Eloquent deep learning model. LayerNorm is the layer normalization function used to stabilize the training process. Attention(Q, K, V) is a self-attention mechanism used to capture key information in the text. Q, K, and V are query, key, and value, respectively, representing different text features.

[0162] In one embodiment, the acquired eloquence dimension data Y may include the following data:

[0163] Persuasion (P): measures the ability of a text to persuade its audience.

[0164] Emotional expression (E): Evaluate the effectiveness of the text in conveying emotions and mobilizing the audience's emotions.

[0165] Clarity (C): reflects the ability of the text to convey clear and understandable information.

[0166] Personalization (G): Indicates the degree to which the text matches the speaker's personal style.

[0167] Interactivity (I): evaluates the effectiveness of the text in promoting audience participation and interaction.

[0168] In one embodiment, the calculation of the eloquence dimension score value may be as follows:

[0169] D score=Softmax(W d ·Y+b d ).

[0170] In the above formula, D score is the score value of the eloquence dimension, which can be based on the scores of different eloquence dimensions. d and b d are the weights and biases used for eloquence dimension analysis

[0171] In one embodiment, the eloquence weighted value may be calculated as follows:

[0172]

[0173] In the above formula, W di is the weighting factor of the i-th eloquence dimension. i is the score of the i-th eloquence dimension.

[0174] After weighting, it can be expressed as follows:

[0175]

[0176] Z is the weighted eloquence dimension data, which is used to guide content adjustment. is the eloquence dimension score value of the i-th eloquence dimension. Embed(X i ) is to convert the i-th eloquence dimension data into an embedding vector.

[0177] In one embodiment, the training process of the eloquence deep learning model may be as follows:

[0178] First, feature extraction and model training: Extract features related to eloquence dimensions, such as persuasiveness, emotional expression, etc. Use a large amount of text data to pre-train the deep learning model.

[0179] Second, model optimization and adjustment: adjust and optimize the model structure and parameters according to the eloquence dimension score. Then use cross-validation and hyperparameter adjustment to improve model performance.

[0180] Among them, the model structure and parameters can be adjusted and optimized according to the eloquence dimension score value, and the deep learning model output (Y) and specific weights (W d ) and bias (b d ) Calculate the eloquence dimension score value (D score ). The eloquence dimension scores reflect the performance of the text in different eloquence dimensions (such as persuasiveness, emotional expression, etc.). These eloquence dimension scores can then be used in the optimization and adjustment process of the model to ensure that the generated content better meets the user's preferences in the eloquence dimension.

[0181] In the process of "Using cross-validation and hyperparameter tuning to improve model performance", the steps are generally as follows:

[0182] 1. Cross-validation: This is a method to evaluate the generalization ability of a model. The dataset is usually divided into several parts, one of which is used to test the model and the rest is used for training. By alternating between different test and training sets, a more comprehensive evaluation of the model's performance can be performed.

[0183] 2. Hyperparameter adjustment: Deep learning models contain multiple hyperparameters, such as learning rate, number of layers, number of neurons, etc. These parameters are set before model training. Hyperparameter adjustment refers to finding the parameter combination that optimizes model performance by systematically changing the values ​​of these parameters. The combination of these two steps helps improve the accuracy and generalization ability of the model, so that it can better adapt and predict the eloquence dimension indicators.

[0184] Third, content generation strategy adjustment: adjust the content generation strategy based on the model output to better meet the user's eloquence dimension preferences.

[0185] Fourth, continuous iteration and optimization: continue to collect new user data. Regularly update and optimize the model to adapt to changes in user preferences.

[0186] After training, the eloquence deep learning model needs to be verified. The verification can include the following tests:

[0187] First, the model prediction accuracy test: evaluating the prediction accuracy of the model on different eloquence dimensions.

[0188] Second, content adaptability evaluation: analyzing whether the adjusted content is more in line with the user's eloquence preferences.

[0189] Third, user satisfaction monitoring: monitor user satisfaction with the adjusted content.

[0190] After completion, the eloquence deep learning model can be shown as follows:

[0191]

[0192] In the above formula, T adj is the adjusted text output, i.e., the eloquence text. n is the number of eloquence dimension data. di is the weight of the i-th eloquence dimension data. NN(X, W ti , b ti ) is the output of the neural network adjusted for the i-th eloquence dimension data, X is the input feature, W ti and b ti are the weights and biases of the Eloquent deep learning model.

[0193] After training the eloquence deep learning model, you can continue to optimize and iterate, collect user feedback, and continuously optimize the model and adjust the strategy. Optimization and iteration can be verified through the following projects:

[0194] First, content adaptability assessment: analyzing whether the adjusted content is more in line with the user's eloquence dimension preferences.

[0195] Second, user satisfaction testing: evaluate the effectiveness of content adjustments through user surveys and feedback.

[0196] Third, model performance monitoring: monitor the accuracy and response speed of the model.

[0197] Wherein, as an example, the style decoder generates style text based on the mining data, which may include the following sub-steps:

[0198] S51. Extract multiple eloquence dimension data from the mining data, and calculate the weighting factor corresponding to each eloquence dimension data to obtain a style weighting factor.

[0199] S52: using the weighting factor to convert each of the eloquence dimension data into style dimension data, and calling a preset style decoder to generate a style text according to the style dimension data.

[0200] In one embodiment, feature extraction and analysis may be performed first, and multiple eloquence dimension data may be extracted from the mining data, wherein the eloquence dimension data may include indicators related to the eloquence dimension. Then, the eloquence deep learning model is used to analyze these features and predict the score of each eloquence dimension.

[0201] In one embodiment, the eloquence dimension data may be as follows:

[0202] Persuasion (P): measures the ability of a text to persuade its audience.

[0203] Emotional expression (E): Evaluate the effectiveness of the text in conveying emotions and mobilizing the audience's emotions.

[0204] Clarity (C): reflects the ability of the text to convey clear and understandable information.

[0205] Personalization (I): Indicates how well the text matches the speaker’s personal style.

[0206] Next, the weighting factor corresponding to each of the eloquence dimension data is calculated to obtain the style weighting factor.

[0207] As an example, the style weighting factor may be as shown in the following formula:

[0208]

[0209] In the above formula, α i It is the weighting factor of the i-th eloquence dimension data.

[0210] Score i is the eloquence dimension score value of the i-th eloquence dimension data. For details, please refer to the above calculation formula.

[0211] Then, each eloquence dimension data may be converted into style dimension data. In one embodiment, the style dimension data may be expressed as follows:

[0212]

[0213] In the above formula, S is the style dimension data, which is the stylized text feature representation. s and b s are the weights and biases of the style simulator. Embed(X i ) is the embedding representation of the i-th eloquence dimension data. i is the weighting factor of the ith eloquence dimension data. n is the number of eloquence dimension data.

[0214] The weighting factor of each eloquence dimension data is calculated by the above formula. The weighting factor is calculated based on the eloquence dimension score value of each eloquence dimension data. This formula actually applies an exponential function to the score of each eloquence dimension data and normalizes all eloquence dimension score values ​​to obtain a set of weighting factors for adjusting the contribution of each eloquence dimension in generating stylized content.

[0215] In one embodiment, the calculation of the weighting factor has the following two functions:

[0216] First, personalized emphasis: weighting factors are used to determine which eloquence dimensions should be emphasized in content generation for a specific user. For example, if a user attaches particular importance to persuasiveness, the weighting factor on this dimension will be relatively high.

[0217] Second, content adjustment guidance: These factors guide the style simulator on how to balance and integrate different eloquence dimensions (such as persuasiveness, emotional expression) when generating content.

[0218] Optionally, the weighting factor can also be calculated based on the user's historical interaction data and feedback. By analyzing the user's response to different types of content, the eloquence dimension of the user's preference can be identified.

[0219] In one implementation, after obtaining the style dimension data, a preset style decoder may be called to generate a style text according to the style dimension data. The specific execution process of using the preset style decoder to generate the final style text based on the style dimension data may be divided into the following steps:

[0220] First, feature extraction and weight calculation: Extract features about eloquence dimensions from user interactions and feedback, such as persuasiveness, emotional expression, etc. Calculate the weighting factors of these eloquence dimension features to determine the impact of each dimension on text generation.

[0221] Second, style feature transformation: Use a transformation model (such as a deep learning network) to process the extracted features and transform the eloquence dimension features into a form suitable for text generation. This transformation process will be combined with a weighting factor to ensure that the influence of each dimension is appropriately reflected.

[0222] Third, text generation: Use the preset style decoder to generate text based on the transformed style dimension data. The style decoder can generate stylized content based on these style dimension data, ensuring that the text not only conforms to the predetermined eloquence dimension, but also has fluency and coherence.

[0223] Specifically, the style text can be as follows:

[0224] T style =Dec(S).

[0225] In the above formula, T style is the style text, i.e. the final text output after stylization. Dec is the decoder based on the style dimension data S.

[0226] The present invention applies a preset style decoder in which weighting factors are combined with features (such as embedding vectors) of each eloquence dimension to generate text with specific style features. This fusion ensures that the generated content is not only linguistically fluent, but also matches the user's preferences in the eloquence dimension. Moreover, as the user's interaction pattern changes, these weighting factors can be dynamically adjusted to reflect the latest changes in user preferences. Such dynamic adjustment allows the generated content to always keep pace with the user's current needs.

[0227] By integrating the calculation of weighted factors of eloquence dimensions into the preset style decoder, the patent of this invention can not only generate highly personalized eloquence content, but also flexibly adapt to changes in user needs, thus providing an innovative solution in the field of personalized eloquence expression.

[0228] In order to further improve the accuracy of the preset style decoder in converting text, the preset style decoder can be continuously iterated and optimized. Specifically, the feature transformation model and text generation strategy can be continuously adjusted and optimized based on the continuous feedback and interaction data of users. Through the iterative process, it can better learn and adapt to user preferences, thereby improving the quality of generated content and user satisfaction.

[0229] This process combines deep learning and natural language processing technologies to generate high-quality, personalized eloquence content that meets the user's specific speaking style and skill needs.

[0230] Subsequently, the effect of the preset style decoder can also be verified, where the effect verification can include the following tests:

[0231] First, style suitability assessment: analyzing whether the generated content conforms to the specified eloquence style and dimensions.

[0232] Second, user satisfaction testing: evaluate the acceptance of stylized content through user surveys and feedback.

[0233] Third, model performance monitoring: monitoring the response speed and accuracy of the style simulator.

[0234] As an example, the calling of a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text may include the following sub-steps:

[0235] S61. After obtaining the user's interactive data, pre-process the interactive data to obtain processed data. The interactive data is data about the user's comments and interactive feedback during the real-time speech. The pre-processing includes cleaning processing, encoding processing and embedding processing.

[0236] S62: calling the preset real-time content adjustment model to generate text adjustment information according to the processed data, and generating a speech text based on the text adjustment information and the basic text information.

[0237] In one embodiment, user interaction data may be obtained, and the interaction data may include user interaction frequency, feedback type, content preference, etc. Then, the interaction data may be preprocessed to obtain processed data. The preprocessing may include cleaning, normalization, encoding, and embedding, and then a preset real-time content adjustment model is called to generate text adjustment information according to the processed data.

[0238] In one embodiment, the preset real-time content adjustment model can be a comprehensive neural network model. This model combines neural networks and time series analysis, and can analyze the user's historical behavior and real-time feedback, and predict the user's future demand changes. This prediction model generates text adjustment information by analyzing and processing data, using the output of the activation function and the time series model, and then adds the text adjustment information to the basic text information to generate a speech text. After the speech text is generated, the user can read it directly or use it as a speech script, or convert it into speech as needed.

[0239] In one embodiment, the structure of the preset real-time content adjustment model may include: a data analysis engine (combining time series analysis and user behavior analysis), an adaptive content adjustment algorithm, and a decision logic closely integrated with eloquence dimension data (such as persuasiveness, emotional expression, and clarity).

[0240] The preset real-time content adjustment model can learn user preferences and needs in real time by analyzing and processing data (user interaction with generated content (such as questions and answers, comments, direct feedback)), and then generate text adjustment information.

[0241] The structure of the preset real-time content adjustment model can also integrate an interaction data processor (for capturing and analyzing user interaction data) and a machine learning-driven feedback analyzer (for learning and extracting useful information from the interaction).

[0242] In one embodiment, the preset real-time content adjustment model may be pre-trained by a user, and after the training is completed, cross-validation and hyperparameter adjustment may be applied to optimize the model performance.

[0243] In one embodiment, the process of applying cross-validation and hyperparameter adjustment to optimize model performance mainly includes the following steps:

[0244] First, cross-validation: This is a technique for evaluating the generalization ability of a model, usually done by dividing the dataset into multiple small training and test sets. In each round, a portion of the data is used to train the model, and the other portion is used to test the performance of the model. Through cross-validation, the performance of the model on unseen data can be more accurately evaluated.

[0245] Second, hyperparameter tuning: The hyperparameters of a model are the parameters that are set before the training process, such as learning rate, number of hidden layers, etc. Adjusting these hyperparameters can significantly affect the performance of the model. Usually, this is done by experimenting with different parameter combinations, with the goal of finding the best combination of parameters to improve the accuracy and efficiency of the model.

[0246] The combination of these two methods can help find the optimal model configuration, thereby improving the performance of the model in practical applications.

[0247] Specifically, the formula of the preset real-time content adjustment model can be as follows:

[0248] Y=ReLU(W1·(tanh(W2·X+b2))+b1)+α·ARIMA(X ts , p, d, q).

[0249] In the above formula, Y is the predicted change in user demand, specifically the text adjustment information generated based on the processed data. X is the input feature vector, i.e. the processed data, including user interaction frequency, feedback type, content preference, etc. W1, W2, b1, b2 are the weights and biases of the preset real-time content adjustment model. ReLU and tanh are activation functions. α is the weight for adjusting the time series analysis in the prediction. ARIMA(X ts , p, d, q) is the output of the time series model, where X ts It is time series data.

[0250] In one embodiment, the text adjustment information may be expressed as follows:

[0251] C adj =W3·D+b3.

[0252] C adj It is the content strategy adjusted based on the user's eloquence dimension data or eloquence dimension indicator preference, that is, text adjustment information. W3 and b3 are the weights and biases used for content adjustment. D is a function for adjusting content strategy.

[0253] In one embodiment, the integration of eloquence dimension data or eloquence dimension indicators may be as shown in the following formula:

[0254] D = softmax(W d ·Z+b d )

[0255] In the above formula, D is the user preference prediction based on the eloquence dimension. d and b d are the weights and biases used for eloquence dimension analysis.

[0256] The softmax function is used to transform the output into a probability distribution, indicating the importance of different eloquence dimensions.

[0257] It should be noted that the integration of eloquence dimension indicators in the pre-trained deep learning model is mainly to train the model to understand and identify text features of different eloquence dimensions. This is a basic step to ensure that the model can accurately identify and process data related to a specific eloquence dimension.

[0258] In generating text adjustment information so that real-time content adjustment can be carried out, the integration of the eloquence dimension involves more real-time data analysis and immediate feedback. The content is adjusted mainly based on the real-time interaction and feedback of users, ensuring that the generated content reflects the current user needs and preferences.

[0259] In order to further improve the accuracy of the text adjustment information generated by the preset real-time content adjustment model, the preset real-time content adjustment model can be mobilized to analyze the processed data. The preset real-time content adjustment model can include SVM and RF models.

[0260] In one embodiment, the preset real-time content adjustment model analyzes the processed data as shown in the following formula:

[0261] S = β·SVM(X sv )+γ·RF(X rf ).

[0262] In the above formula, S is the score based on the processing data analysis, through which the user's preference for the content of eloquence expression can be evaluated. sv , X rf are the feature sets prepared for the support vector machine (SVM) and random forest (RF) models respectively. β and γ are weight coefficients used to adjust the influence of the two models. sv ) and RF(X rf ) are the outputs of the SVM and RF models respectively.

[0263] In one embodiment, the score obtained by the preset real-time content adjustment model after analyzing the processed data can use the support vector machine (SVM) and random forest (RF) model to calculate the eloquence dimension score according to the user interaction data. The score reflects the user's performance in different eloquence dimensions (such as persuasiveness, emotional expression, and clarity).

[0264] For high score analysis: If a user scores high on a dimension, it means that the user performs well in that dimension. For example, a high score on the persuasion dimension means that the user is effective in persuading the audience.

[0265] For low score analysis: If the user scores low on a dimension, it means there is room for improvement in that dimension. For example, a low score on the clarity dimension may mean that the user needs to be clearer and more accurate in expressing information.

[0266] Finally, based on these scoring results, content generation strategies can be adjusted and optimized as information for adjusting texts to better suit the strengths and weaknesses of users and improve the effectiveness of eloquent expression.

[0267] Specifically, the generation strategy as the operation of text adjustment information can adjust the content generation strategy in real time according to the prediction and eloquence dimension analysis results. At the same time, user feedback and interaction are continuously monitored to optimize the model and adjust the strategy.

[0268] In an optional embodiment, the SVM and RF models may be used to analyze the eloquence dimension of the user data. Specifically, the implementation process of using the SVM and RF models to analyze the eloquence dimension of the user data is as follows:

[0269] First, feature extraction: extract features related to eloquence dimensions from user interaction and feedback data. These features may include user responses to content, interaction frequency, feedback type, etc. These data reflect the user's eloquence dimension preferences, such as persuasiveness, emotional expression, clarity, etc.

[0270] Second, model application: Support Vector Machine (SVM): Use the SVM model to analyze these features. SVM is an effective classification algorithm that can handle nonlinear data and distinguish different eloquence dimensions by building a decision boundary. In the formula, SVM(Xsv) represents the output of the SVM model, Xsv is the feature set prepared for the SVM model, and β is the weight coefficient.

[0271] Random Forest (RF): At the same time, the RF model is used for analysis. RF is an ensemble learning method based on decision trees, which can handle a large number of features and provide an evaluation of feature importance. In the formula, RF(Xrf) represents the output of the RF model, Xrf` is the feature set prepared for the RF model, and γ is the weight coefficient.

[0272] Third, comprehensive scoring: Combine the outputs of the SVM and RF models for comprehensive scoring, that is, the formula:

[0273] S = β·SVM(X sv )+γ·RF(X rf ).

[0274] This composite score represents the score based on the eloquence dimension, which combines the results of the two models and adjusts the influence of the two models through the two weight coefficients β and γ.

[0275] In this way, the evaluation of eloquence dimensions by different analysis models can be comprehensively considered to more comprehensively understand and predict users' eloquence dimension preferences, thereby generating more accurate text adjustment information, which can be used to make more accurate dynamic adjustments to the basic text information.

[0276] In another optional embodiment, the structure of the preset real-time content adjustment model may also integrate an interactive data processor (capturing and analyzing user interactive data). Therefore, the preset real-time content adjustment model may also be used to analyze the interactive characteristics of the user. Optionally, the formula of the integrated interactive data processor may be as follows:

[0277] F = ReLU(W1 · Embed(X int )+b1).

[0278] In the above formula, F is the feature representation of user interaction data. int is user interaction data (such as questions and answers, comments). Embed is a function that converts interaction data into an embedding vector. W1 and b1 are the weights and biases of the preset real-time content adjustment model. ReLU is the activation function.

[0279] In yet another optional embodiment, the structure of the preset real-time content adjustment model may also be a machine learning driven feedback analyzer (for learning and extracting useful information from interactions).

[0280] Alternatively, the formula of the deep learning model of the feedback analyzer can be as follows:

[0281] Z=σ(W f Embed(X f )+b f ).

[0282] In the above formula, Z is the feature representation of user interaction, which is used to guide the basic text information to adjust the content. f and b f It is the weight and bias of the preset real-time content adjustment model. Embed(X f ) is to convert user feedback X f is transformed into an embedding vector. σ is a nonlinear activation function, such as ReLU.

[0283] It should be noted that the training and application of the preset real-time content adjustment model: a large amount of user interaction data can be used to train the deep learning model to extract user preference features. The model is applied to analyze the interaction data and predict the user's eloquence dimension preference.

[0284] In the process of extracting user preference features from interactive data, the ReLU activation function can be used to process the embedding vector, thereby transforming user interaction data (such as questions and answers, comments) into feature representations that can be used for further analysis.

[0285] In the application of the integration of the deep learning model of the feedback analyzer and the eloquence dimension indicators, the deep learning model of the feedback analyzer: This model is used to convert user feedback (such as direct reactions or evaluations) into feature representations in order to further analyze and understand user preferences and needs. Integration of eloquence dimension indicators: This part is mainly used to convert the feature representation obtained from the feedback analyzer into a probability distribution, which reflects the user's preferences on different eloquence dimensions (such as persuasiveness, emotional expression, and clarity). In this way, the content generation strategy can be dynamically adjusted based on the user's eloquence dimension preferences.

[0286] In another optional embodiment, the preset real-time content adjustment model can also adjust the content generation strategy based on the user preference prediction. Specifically, the corresponding strategy is generated through the above text adjustment information. The whole process includes using the data extracted from user interaction (feature representation after deep learning model analysis) and the user's eloquence dimension preferences predicted based on these features. Then, this information is integrated and applied to adjust the content generation strategy to ensure that the generated content can better meet the user's personalized needs and preferences.

[0287] In one implementation, feature data may be extracted from user interaction data, and the user's eloquence dimension preferences may be predicted using the feature data. Finally, an adjustment strategy may be generated using the feature data and the user's eloquence dimension preferences as text adjustment information.

[0288] Specifically, the process of generating text adjustment information using feature data and user eloquence dimension preferences is as follows:

[0289] First, data integration: First, the user’s interactive data, such as comments, questions and answers, is analyzed through a deep learning model to extract preference features about the user’s eloquence dimensions (e.g., persuasiveness, emotional expression).

[0290] Second, preference analysis: Then, use the model output (such as dimension scores) to analyze users' preferences and determine which eloquence dimensions users are most concerned about.

[0291] Third, strategy formulation: Based on these analyses, content generation strategies are formulated, such as emphasizing the user's preferred eloquence dimension in the content. For example, if the user scores high on persuasiveness, the content generation will focus on enhancing persuasiveness.

[0292] Fourth, dynamic adjustment: The strategy will be dynamically adjusted based on users' continuous feedback and new interaction data to ensure that the content always meets the users' current preferences.

[0293] For example, if a user shows a strong interest in a certain topic, you can adjust your strategy to integrate more of that topic into your content while also incorporating the user's preferred eloquence style. Use the improved strategy to generate content that better matches the user's preferences.

[0294] The present invention can utilize a deep learning model to automatically generate high-quality speech expression content; can personalize and adjust the generated text content according to the user's eloquence dimension indicators, such as persuasiveness, emotional expression, etc.; can dynamically adjust the content generation strategy according to the user's real-time feedback and behavioral data; can comprehensively analyze and adjust the weight of the user's eloquence dimension; can apply cross-validation and hyperparameter adjustment methods to optimize the performance and accuracy of the model.

[0295] In this embodiment, the embodiment of the present invention provides a method for generating a speech text, which has the beneficial effects of: the present invention can obtain the user's feature data when the user is giving a speech; use a mining knowledge tree to perform feature mining analysis on the feature data to obtain mining data; then call the deep learning model to generate basic text information based on the mining data, and then call a preset real-time content adjustment model to adjust the content of the basic text information to obtain the speech text, and combine the feature data of the user during the speech to generate text that conforms to the user's current scene in real time, thereby fitting the user's actual scene, avoiding the situation where the text deviates from the real-time scene, and can improve the accuracy of the generated text data.

[0296] The embodiment of the present invention also provides a device for generating a speech text, see Figure 2 , showing a schematic diagram of the structure of a speech text generation device provided by an embodiment of the present invention.

[0297] Wherein, as an example, the speech text generating device may include:

[0298] The data acquisition module 201 is used to acquire the characteristic data of the user when the user is giving a speech, wherein the characteristic data includes: data on the user's behavior during the speech and data on the evaluation of the speech audience;

[0299] The data mining module 202 is used to perform feature mining analysis on the feature data using a preset mining knowledge tree to obtain mining data. The preset mining knowledge tree is a knowledge tree constructed by using topic data after acquiring multiple different market topic data;

[0300] The text generation module 203 is used to call a preset deep learning model to generate basic text information according to the mining data, and then call a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text, wherein the preset deep learning model is a model that extracts data about speech eloquence and style from a text training set and performs deep learning training, and the preset real-time content adjustment model is a time series model that receives feedback data about user speech feedback from multiple different social platforms and performs training and optimization respectively.

[0301] Optionally, the preset deep learning model includes: an eloquence deep learning model for generating eloquence content text and a style decoder for generating style content text;

[0302] The calling of a preset deep learning model to generate basic text information according to the mining data includes:

[0303] The mined data are respectively input into the eloquence deep learning model and the style decoder, so that the eloquence deep learning model generates eloquence text based on the mined data, and the style decoder generates style text based on the mined data, and the eloquence text and the style text are combined to obtain basic text information.

[0304] Optionally, the eloquence deep learning model generates an eloquence text based on the mining data, including:

[0305] Extracting a plurality of eloquence dimension data from the mining data;

[0306] Calling the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values;

[0307] Each eloquence dimension score value is used to calculate the eloquence weighted value corresponding to each eloquence dimension data, and multiple eloquence weighted values ​​and multiple eloquence dimension data are input into the eloquence deep learning model to obtain the eloquence text.

[0308] Optionally, the style decoder generates style text based on the mined data, including:

[0309] A plurality of eloquence dimension data are extracted from the mining data, and a weighting factor corresponding to each of the eloquence dimension data is calculated to obtain a style weighting factor.

[0310] Each of the eloquence dimension data is converted into style dimension data using the weighting factor, and a preset style decoder is called to generate a style text according to the style dimension data.

[0311] Optionally, the calling of a preset real-time content adjustment model to adjust the content of the basic text information to obtain the speech text includes:

[0312] After obtaining the user's interactive data, pre-processing the interactive data to obtain processed data, wherein the interactive data is data about the user's comments and interactive feedback during the real-time speech, and the pre-processing includes cleaning processing, encoding processing and embedding processing;

[0313] The preset real-time content adjustment model is called to generate text adjustment information according to the processed data, and a speech text is generated based on the text adjustment information and the basic text information.

[0314] Optionally, the preset construction process of mining knowledge tree includes:

[0315] Collecting a plurality of market data, each of the market data is data collected from a different open source media platform;

[0316] Using machine learning technology to identify and process each of the market data to obtain popular topic data about popular market topics, and using natural language processing technology to perform topic analysis on each of the market data to obtain popular topic data about popular market topics;

[0317] The popular topic data and the popular subject data are used to construct a preset mining knowledge tree.

[0318] The present invention also provides a system for generating speech texts. Figure 3 , showing a structural diagram of a speech text generation system provided by an embodiment of the present invention.

[0319] In one embodiment, the speech text generation system may be installed in a smart device, so as to utilize the speech text generation system to execute the speech text generation method of the above embodiment.

[0320] Wherein, as an example, the speech text generation system may include:

[0321] Expand the user data collection module, dynamic knowledge tree construction module, deep learning-driven content generation module, and real-time content adjustment module;

[0322] The extended user data collection module is used to obtain user behavior data and context data;

[0323] The dynamic knowledge tree construction module is used to construct a knowledge tree and use the knowledge tree to perform data mining based on user data;

[0324] The deep learning driven content generation module is used to generate basic text information based on the mined data;

[0325] The real-time content adjustment module is used to generate text adjustment information according to the user's behavior data and context data, and generate a speech text using the text adjustment information and basic text information.

[0326] Those skilled in the art can clearly understand that, for the sake of convenience and brevity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0327] Furthermore, an embodiment of the present application also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for generating a speech text as described in the above embodiment is implemented.

[0328] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer-executable program, and the computer-executable program is used to enable a computer to execute the method for generating a speech text as described in the above embodiment.

[0329] Those skilled in the art will appreciate that the embodiments of the present application may also provide computer program products. Therefore, the present application may adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application may adopt the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program codes.

[0330] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), apparatuses, and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0331] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0332] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0333] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for generating a speech text, characterized in that: The method comprises: Acquire characteristic data of the user when the user is giving a speech, wherein the characteristic data includes: data on the user's behavior during the speech and data on the evaluation of the speech audience; Performing feature mining analysis on the feature data using a preset mining knowledge tree to obtain mining data, wherein the preset mining knowledge tree is a knowledge tree constructed by using topic data after acquiring multiple different market topic data; After calling a preset deep learning model to generate basic text information according to the mining data, calling a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text, wherein the preset deep learning model is a model that extracts data on speech eloquence and style from a text training set and performs deep learning training, and the preset real-time content adjustment model is a time series model that receives feedback data on user speech feedback from multiple different social platforms and performs training and optimization respectively; The preset deep learning model includes: an eloquence deep learning model for generating eloquence content text and a style decoder for generating style content text; The calling of a preset deep learning model to generate basic text information according to the mining data includes: Inputting the mining data into the eloquence deep learning model and the style decoder respectively, so that the eloquence deep learning model generates eloquence text based on the mining data, and the style decoder generates style text based on the mining data, and combining the eloquence text and the style text to obtain basic text information; The eloquence deep learning model generates eloquence text based on the mining data, including: Extracting a plurality of eloquence dimension data from the mining data, wherein the eloquence dimension data includes the following data: persuasiveness data, emotional expression data, clarity data, personalization data, and interactivity data; Calling the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values; Using each of the eloquence dimension scoring values, calculate the eloquence weighted value corresponding to each of the eloquence dimension data, and input a plurality of the eloquence weighted values ​​and a plurality of the eloquence dimension data into the eloquence deep learning model to obtain an eloquence text; The preset construction process of the mining knowledge tree includes: Collecting a plurality of market data, each of the market data is data collected from a different open source media platform; Using machine learning technology to identify and process each of the market data to obtain popular topic data about popular market topics, and using natural language processing technology to perform topic analysis on each of the market data to obtain popular topic data about popular market topics; The popular topic data and the popular subject data are used to construct a preset mining knowledge tree.

2. The method for generating a speech text according to claim 1, characterized in that: The style decoder generates a style text based on the mining data, including: Extracting a plurality of eloquence dimension data from the mining data, and calculating a weighting factor corresponding to each of the eloquence dimension data to obtain a style weighting factor; Each of the eloquence dimension data is converted into style dimension data using the weighting factor, and a preset style decoder is called to generate a style text according to the style dimension data.

3. The method for generating a speech text according to claim 1, characterized in that: The calling of a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text includes: After obtaining the user's interactive data, pre-processing the interactive data to obtain processed data, wherein the interactive data is data about the user's comments and interactive feedback during the real-time speech, and the pre-processing includes cleaning processing, encoding processing and embedding processing; The preset real-time content adjustment model is called to generate text adjustment information according to the processed data, and a speech text is generated based on the text adjustment information and the basic text information.

4. A device for generating speech text, characterized in that: The device comprises: A data acquisition module, used to acquire characteristic data of the user when the user is giving a speech, wherein the characteristic data includes: data on the user's behavior during the speech and data on the evaluation of the speech audience; A data mining module, used for performing feature mining analysis on the feature data using a preset mining knowledge tree to obtain mining data, wherein the preset mining knowledge tree is a knowledge tree constructed by obtaining topic data on multiple different market topics and using the topic data; A text generation module, for calling a preset deep learning model to generate basic text information according to the mining data, and then calling a preset real-time content adjustment model to adjust the content of the basic text information to obtain a speech text, wherein the preset deep learning model is a model that extracts data on speech eloquence and style from a text training set and performs deep learning training, and the preset real-time content adjustment model is a time series model that receives feedback data on user speech feedback from multiple different social platforms and performs training and optimization respectively; The preset deep learning model includes: an eloquence deep learning model for generating eloquence content text and a style decoder for generating style content text; The calling of a preset deep learning model to generate basic text information according to the mining data includes: Inputting the mining data into the eloquence deep learning model and the style decoder respectively, so that the eloquence deep learning model generates eloquence text based on the mining data, and the style decoder generates style text based on the mining data, and combining the eloquence text and the style text to obtain basic text information; The eloquence deep learning model generates eloquence text based on the mining data, including: Extracting a plurality of eloquence dimension data from the mining data, wherein the eloquence dimension data includes the following data: persuasiveness data, emotional expression data, clarity data, personalization data, and interactivity data; Calling the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values; Using each of the eloquence dimension scoring values, calculate the eloquence weighted value corresponding to each of the eloquence dimension data, and input a plurality of the eloquence weighted values ​​and a plurality of the eloquence dimension data into the eloquence deep learning model to obtain an eloquence text; The preset construction process of the mining knowledge tree includes: Collecting a plurality of market data, each of the market data is data collected from a different open source media platform; Using machine learning technology to identify and process each of the market data to obtain popular topic data about popular market topics, and using natural language processing technology to perform topic analysis on each of the market data to obtain popular topic data about popular market topics; The popular topic data and the popular subject data are used to construct a preset mining knowledge tree.

5. A system for generating speech text, characterized in that: The system includes: an expanded user data collection module, a dynamic knowledge tree construction module, a deep learning driven content generation module and a real-time content adjustment module; The extended user data collection module is used to obtain user behavior data and context data; The dynamic knowledge tree construction module is used to construct a knowledge tree and use the knowledge tree to perform data mining based on user data; The deep learning driven content generation module is used to generate basic text information based on the mined data; The real-time content adjustment module is used to generate text adjustment information according to the user's behavior data and context data, and generate the speech text using the text adjustment information and the basic text information; The deep learning driven content generation module is further used to input the mined data into the eloquence deep learning model and the style decoder respectively, so that the eloquence deep learning model generates eloquence text based on the mined data, and the style decoder generates style text based on the mined data, and combines the eloquence text and the style text to obtain basic text information; The eloquence deep learning model generates eloquence text based on the mined data, including: Extracting a plurality of eloquence dimension data from the mined data, wherein the eloquence dimension data includes the following data: persuasiveness data, emotional expression data, clarity data, personalization data, and interactivity data; Call the eloquence deep learning model to calculate the score of each eloquence dimension data to obtain multiple eloquence dimension score values; The eloquence weighted value corresponding to each eloquence dimension data is calculated using the scoring value of each eloquence dimension, and multiple eloquence weighted values ​​and multiple eloquence dimension data are input into the eloquence deep learning model to obtain the eloquence text.

6. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for generating a speech text as described in any one of claims 1 to 3 when executing the computer program.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer-executable program, and the computer-executable program is used to enable a computer to execute the method for generating a speech text as described in any one of claims 1 to 3.

Images