Conference data intelligent analysis and decision support system

Abstract

The invention relates to a conference data intelligent analysis and decision support system. Audio data in a conference process are collected in real time; extracting conference information of the audio data by adopting a natural language processing technology, wherein the conference information comprises conference participant positions, conference themes, conference contents and voting results; and adopting a big data analysis model to predict an expected effect of meeting decision execution according to the meeting information. According to the invention, the problem of low working efficiency caused by the fact that the prior art depends on manual recording and summarization, effective information of the conference is difficult to extract and the execution effect of the conference decision is predicted according to the conference information is solved. The conference information variable set having significant influence on the execution effect is selected through the random forest model, and the efficiency and accuracy of conference information analysis are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of data analysis and relates to a conference data intelligent analysis and decision support system. Background Art

[0002] In the operation of modern enterprises and organizations, meetings are an important way to communicate, make decisions and coordinate. However, traditional meeting management often relies on manual recording and summarization, which makes it difficult to effectively analyze and utilize the large amount of data in the meeting, resulting in certain limitations on the scientificity and efficiency of meeting decision-making.

[0003] With the rapid development of information technology, although some digital conference management tools have emerged, most of them only focus on basic functions such as conference scheduling and document sharing, and lack the ability to deeply mine and analyze conference data. Therefore, it is necessary to invent a system that can conduct comprehensive and in-depth analysis of conference data and provide strong support for decision-making. Summary of the invention

[0004] The present invention provides a conference data intelligent analysis and decision support system, which aims to use natural language processing technology and artificial intelligence models to integrate various types of data in the conference system, use big data analysis technology to perform analysis and processing, extract valuable information during the conference, provide scientific and accurate decision support for conference organizers and decision makers, and improve the efficiency and quality of the conference.

[0005] The purpose of the present invention can be achieved by the following technical solutions: The present application provides a conference data intelligent analysis and decision support system, including a conference data acquisition module, a conference data processing module, a conference data analysis module and a conference data storage module, wherein the conference data acquisition module, the conference data processing module, the conference data analysis module and the conference data storage module are communicatively connected, wherein: The conference data collection module is used to collect audio data during the conference in real time; The conference data processing module uses natural language processing technology to extract conference information from audio data, where the conference information includes positions of participants, conference topics, conference content, and voting results; The conference data analysis module uses a big data analysis model to predict the expected effect of conference decision execution based on conference information; The conference data storage module is used to store and process audio data and conference information using distributed storage and computing technology.

[0006] Furthermore, the real-time collection of audio data during the meeting is specifically the use of professional audio collection equipment to collect audio signals in the conference space according to a preset sampling frequency and quantization bit number, using a distributed microphone array layout or a centralized multi-point layout, and then using a high-speed data transmission interface to convert the collected analog audio signals into digital audio data in real time and transmit them to a subsequent conference data processing module.

[0007] Furthermore, the method of extracting conference information from audio data using natural language processing technology comprises the following steps: Speech recognition: Use acoustic models and deep neural network algorithms to train historical audio data samples to learn different voice features, accents, and speech speed patterns of audio data; Part-of-speech tagging: Use hidden Markov models or conditional random fields to tag the part of speech of each word in the audio data: Syntactic analysis: Use formal grammar rules to analyze the structure of sentences in audio data, and use dependency syntactic analysis to determine the dependency relationship between the various components in the sentence; Entity recognition: Use recurrent neural networks or long short-term memory networks to capture contextual information and sequence features in text to identify entities; Semantic understanding: semantic analysis technology is used to analyze the text, associate entities in the text with information in the knowledge graph, and use natural language processing technology to further understand the semantics of the text; Information extraction: Extract specific meeting information from the text based on rules and machine learning algorithms, and use machine learning algorithms to classify and verify the extracted information.

[0008] Furthermore, the use of a big data analysis model to predict the expected effect of conference decision execution based on conference information includes the following steps: Collect meeting information and historical data samples of known performance; The random forest model is used to select the meeting information variable set that has a significant impact on the execution effect; Establish a BP neural network model of conference information variable set and execution effect; The meeting information to be predicted is input into the BP neural network model and the expected effect is output.

[0009] Furthermore, the random forest model includes the following construction steps: Data preparation: remove missing values ​​and outliers from the data, and standardize or normalize the data; Determine hyperparameters: Based on experience and experiments, set the number of decision trees and maximum depth hyperparameters of the random forest model; Feature and sample sampling: Perform self-service sampling on the original data, extract samples with replacement, and randomly select some features as the basis for splitting the decision tree nodes; Construct a decision tree: Based on the selected features and samples, determine the best split features according to information gain or Gini coefficient, construct a decision tree, and repeat this process until the stopping condition is met; Evaluate feature importance: By calculating the average importance of each feature in all decision trees, evaluate the importance of meeting information variables and screen out the set of variables that have a significant impact on the execution effect.

[0010] Furthermore, the establishment of the BP neural network model of the conference information variable set and the execution effect includes the following steps: Data preprocessing and division: preprocess the selected meeting information variable set data and the corresponding execution effect data, and divide the entire data set into training set, test set and validation set; Network structure design: Design the number of nodes of BP neural network neurons according to the dimensions of the input meeting information variable set and the specific form of the expected output execution effect. The neurons include input layer, hidden layer and output layer. Parameter initialization: Use random initialization method to initialize the weights and thresholds of the connections between neurons in the neural network; Forward propagation: Input the meeting information variable data in the training set into the input layer, and then calculate through each hidden layer in turn according to the network structure. Each neuron performs nonlinear transformation through the activation function according to the input value, weight and threshold, and finally obtains the corresponding execution effect prediction value in the output layer; Calculate the error: Compare the predicted execution effect value obtained by forward propagation with the actual execution effect value in the training set, and calculate the error between the two; Back propagation and parameter update: Based on the calculated error, starting from the output layer, the error is back propagated back to each neuron in the network in the opposite direction of the forward propagation, and the weights and thresholds are updated using the gradient descent method; forward propagation, error calculation, back propagation and parameter update are repeated continuously until the preset training stop condition is reached; Model evaluation and optimization: Use the test set data to evaluate the trained BP neural network model, and determine the accuracy of the model prediction by calculating the corresponding evaluation indicators.

[0011] Furthermore, the conference data analysis module also evaluates the conference participation of the participants according to the speech content of the participants, including the following steps: Speech content acquisition and text preprocessing: The speech content of each participant is obtained from the conference audio data, and the colloquial expressions and abbreviations are standardized and converted to make the format standardized and unified; Constructing a participation evaluation index system: Based on the characteristics and goals of the conference, a participation evaluation index system is constructed. The participation evaluation index includes speech frequency index, speech duration index, speech relevance index, viewpoint innovation index and interactivity index; Feature extraction and quantification: Based on the participation evaluation index system, corresponding features are extracted from the speech content and related text information, and quantified; Weight allocation and comprehensive evaluation: Determine the corresponding weight of each indicator based on its importance to the overall meeting participation, multiply the quantitative value of each indicator by the corresponding weight, and then sum them up to get the comprehensive evaluation score of each participant's meeting participation.

[0012] Furthermore, the comprehensive evaluation score of the conference participation is calculated as follows: , In the formula, H sum Provide a comprehensive evaluation score for meeting participation; H i For the i The scores of the participation evaluation indicators; w i For the i The weight of each participation evaluation indicator; n is the number of participation evaluation indicators.

[0013] Furthermore, the conference data analysis module also visualizes the expected results and proposes corresponding improvement suggestions based on the historical information of the conference data storage module.

[0014] Furthermore, the conference data storage module uses HDFS or Ceph storage system for distributed storage, builds a master-slave architecture, manages metadata by name node, stores data blocks in a dispersed manner by data nodes, and stores audio data redundantly after being cut into blocks according to rules. Structured data in conference information is stored in a distributed database, and unstructured data is stored in the file system and located by metadata management. In terms of computing technology, Spark and Hadoop frameworks are introduced to process data.

[0015] Beneficial effects of the present invention: (1) By collecting audio data in real time during the meeting, natural language processing technology is used to extract meeting information from the audio data, including the positions of participants, meeting topics, meeting content, and voting results; a big data analysis model is used to predict the expected effect of the execution of meeting decisions based on meeting information. The present invention solves the problem that the existing technology relies on manual recording and summarization, making it difficult to extract effective information from the meeting, and predict the execution effect of meeting decisions based on meeting information, resulting in low work efficiency.

[0016] (2) The random forest model is used to select a set of meeting information variables that have a significant impact on the execution effect, thereby improving the efficiency and accuracy of meeting information analysis.

[0017] (3) A BP neural network model of the conference information variable set and execution effect was established, which improved the accuracy of predicting the effects of different conference information and solved the problem that the nonlinear relationship between conference information and execution effect is difficult to quantify. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0019] Figure 1 This is a structural diagram of a conference data intelligent analysis and decision support system in the present invention.

[0020] Figure 2 The present invention is a flowchart of extracting conference information from audio data using natural language processing technology in one embodiment of the present invention.

[0021] Figure 3 This is a flow chart of using a big data analysis model to predict the expected effect of conference decision execution based on conference information in one embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0023] See also Figure 1-Figure 3 The present application provides a conference data intelligent analysis and decision support system, including a conference data acquisition module, a conference data processing module, a conference data analysis module and a conference data storage module, wherein the conference data acquisition module, the conference data processing module, the conference data analysis module and the conference data storage module are communicatively connected, wherein: The conference data collection module is used to collect audio data during the conference in real time; Furthermore, the real-time collection of audio data during the meeting is specifically to use highly sensitive, multi-directional professional audio collection equipment, according to a preset sampling frequency (for example, 44.1kHz or 48kHz, etc.) and a quantization bit number (such as 16 bits or 24 bits), using a distributed microphone array layout or a centralized multi-point layout, to accurately capture the audio signals emitted by sound sources in all directions in the conference space, and then with the help of a high-speed data transmission interface (such as USB, Ethernet, etc.), the collected analog audio signals are converted and transmitted in real time into digital audio data, and transmitted to subsequent processing modules. At the same time, it has automatic gain control, noise reduction and other functions to ensure that the collected audio data quality is reliable and fully reflects the conference voice situation.

[0024] The conference data processing module uses natural language processing technology to extract conference information from audio data, where the conference information includes positions of participants, conference topics, conference content, and voting results; Furthermore, the method of extracting conference information from audio data using natural language processing technology comprises the following steps: Speech recognition: After the conference data acquisition module obtains the audio data, the speech recognition system uses advanced acoustic models and deep neural network algorithms. The acoustic model is trained based on a large amount of speech sample data to learn different speech features, accents, speech speed and other patterns. The deep neural network uses a complex neuron structure to perform nonlinear transformation and feature extraction on the speech signal. In the actual recognition process, the audio signal is converted into a series of feature vectors, which are then iterated and optimized many times to generate accurate text transcription. For example, in noisy environments, the system can improve the accuracy of speech recognition through adaptive noise suppression technology.

[0025] Part-of-speech tagging: Part-of-speech tagging is done with the help of statistically based machine learning algorithms, such as Hidden Markov Models (HMM) or Conditional Random Fields (CRF). First, the text in the training corpus is preprocessed to annotate the part-of-speech of each word. These algorithms determine the part-of-speech of each word based on the contextual information and statistical laws of the word. For example, in the sentence "During the meeting, everyone actively discussed", "meeting" is annotated as a noun and "discussion" is annotated as a verb. The model is trained with a large amount of annotated data so that it can accurately identify patterns of different parts of speech.

[0026] Syntactic analysis: Syntactic analysis uses a combination of rule-based and statistical methods. On the one hand, formal grammatical rules, such as Chomsky's grammatical system, are used to analyze the structure of sentences; on the other hand, statistical methods, such as dependency syntactic analysis, are used to determine the dependency relationship between the various components in the sentence. For example, in the sentence "He raised a question at the meeting", syntactic analysis determines that "he" is the subject, "raised" is the predicate, "question" is the object, and "at the meeting" is the adverbial that modifies the predicate. In this way, the grammatical structure of the sentence is constructed, providing a basis for semantic understanding.

[0027] Entity recognition: Entity recognition uses deep learning neural network models, such as recurrent neural networks (RNNs) or long short-term memory networks (LSTMs). These models can capture contextual information and sequence features in text to identify entities. During the training process, a large amount of text data is annotated to allow the model to learn the characteristic patterns of different entities. For example, entities such as "Zhang San", "Manager", and "Company" can be identified in meeting texts. Through model training and optimization, the accuracy and efficiency of entity recognition can be improved.

[0028] Semantic understanding: Semantic understanding is achieved through semantic analysis technology combined with knowledge graph and natural language processing technology. The knowledge graph builds a database containing various entities and their relationships. Through semantic analysis of the text, the entities in the text are associated with the information in the knowledge graph. For example, for the meeting text "The project is progressing smoothly, and some problems need to be solved", semantic analysis is used to match entities such as "project" and "problem" with relevant information in the knowledge graph to understand the meaning of the text. At the same time, natural language processing technologies such as sentiment analysis and logical reasoning are used to further understand the semantics of the text.

[0029] Information extraction: Information extraction is based on rules and machine learning algorithms to extract specific meeting information from text. By setting rules and templates, the text is matched and filtered to extract information such as meeting topics, meeting content, and voting results. For example, "Meeting topic: project progress discussion; meeting content: project progress, problems and solutions; voting results: agreed" are extracted from the meeting text. At the same time, machine learning algorithms are used to classify and verify the extracted information to improve the accuracy and reliability of information extraction.

[0030] The conference data analysis module uses a big data analysis model to predict the expected effect of conference decision execution based on conference information; Furthermore, the use of a big data analysis model to predict the expected effect of conference decision execution based on conference information includes the following steps: Collect meeting information and historical data samples of known execution results: Meeting information includes many contents, such as meeting time, location, list of participants (including their departments, positions, etc.), topics discussed at the meeting, details of the proposed plan, etc. The historical data samples of known execution results are the actual results records after the execution of similar meeting decisions in the past, such as whether sales increased, whether the project was completed on time, etc. These historical data are like "reference answers" to help the model learn the relationship between meeting information and actual execution results.

[0031] Select the set of meeting information variables that have a significant impact on the execution effect through the random forest model: Random forest is an integrated learning algorithm that consists of multiple decision trees. In this step, the collected meeting information is used as input features, and the random forest model is used to evaluate the importance of each feature (that is, the meeting information variable) to the execution effect. For example, the model may find that the variable of the opinions of key decision makers in the meeting has a great impact on the execution effect, while the variable of the meeting location may be relatively less important. Then select those variables that have a significant impact on the execution effect to form a variable set, which can reduce the amount of calculation of subsequent models and focus on key factors.

[0032] Furthermore, the random forest model includes the following construction steps: (1) Data preparation: Collect historical data on meeting information and known execution effects, clean up missing values ​​and outliers in the data, and standardize the data (both meeting information and execution effect data need to be converted into continuous variables or categorical variables) or normalize the data to ensure data quality and consistency, thus providing a solid foundation for subsequent model training.

[0033] (2) Determine hyperparameters: Based on experience and experiments, set hyperparameters such as the number of decision trees and maximum depth of the random forest model to balance the complexity and performance of the model and avoid overfitting or underfitting.

[0034] (3) Feature and sample sampling: Perform self-service sampling on the original data, extract samples with replacement, and randomly select some features as the basis for splitting decision tree nodes to increase the diversity of the decision tree.

[0035] (4) Construct a decision tree: Based on the selected features and samples, determine the best splitting features according to the splitting criteria such as information gain or Gini coefficient, construct a decision tree, and repeat this process until the stopping condition is met.

[0036] (5) Evaluate feature importance: By calculating the average importance of each feature in all decision trees, the importance of meeting information variables is evaluated and the set of variables that have a significant impact on the execution effect is screened out.

[0037] Establish a BP neural network model of conference information variable set and execution effect: BP neural network (Back-Propagation Neural Network) is a multi-layer feedforward neural network. The conference information variable set selected above is used as the input layer of the neural network, and the execution effect is used as the output layer. A hidden layer is set in the middle, and the error between the output and the actual execution effect is calculated by forward propagation of input data, and then the error is back-propagated to adjust the connection weights between neurons. It is like building a "black box" model to let it learn how to accurately output the execution effect based on the input conference information variables. This process is the key stage of model "learning". The network parameters are adjusted through a large amount of training data so that the model can well fit the relationship between conference information and execution effect.

[0038] Furthermore, the establishment of the BP neural network model of the conference information variable set and the execution effect includes the following steps: (1) Data preprocessing and division: Further preprocess the selected meeting information variable set data and the corresponding execution effect data. For example, for possible outliers, use statistical methods such as box plots to identify and reasonably correct them to ensure data stability. Then, divide the entire data set into training sets, test sets, and possible validation sets according to a certain ratio (common ones are 7:3, 8:2, etc.). The training set is used for model parameter learning, the test set is used to evaluate the model's generalization ability on unseen data, and the validation set can assist in adjusting model hyperparameters and optimizing model performance.

[0039] (2) Network structure design: Based on the dimensions of the input meeting information variable set and the specific form of the expected output execution effect, the structure of the BP neural network is reasonably designed. Determine the number of input layer nodes to match the number of meeting information variables to ensure that the input information can be fully received. At the same time, according to the complexity of the problem and the characteristics of the data, select the appropriate number of hidden layers and the number of nodes in each hidden layer. Generally speaking, more complex relationships may require more hidden layers and nodes to fit, but over-complexity that leads to overfitting problems should also be avoided. Finally, the number of output layer nodes is clarified, which is usually the number of nodes corresponding to one or more indicators representing the execution effect.

[0040] (3) Parameter initialization: Initialize the weights and thresholds of the connections between neurons in the neural network. Common initialization methods include random initialization, such as using uniform distribution or normal distribution to randomly assign initial values ​​to weights and thresholds within a certain interval, and initialization strategies such as Xavier initialization and He initialization that are optimized for different activation functions. The purpose is to help the network have better gradient propagation characteristics in the early stages of training, speed up training convergence, and avoid problems such as gradient vanishing or gradient exploding.

[0041] (4) Forward propagation: The meeting information variable data in the training set is input into the input layer of the designed BP neural network in sequence, and then calculated through each hidden layer in sequence according to the network structure. Each neuron performs nonlinear transformation through the activation function (such as Sigmoid, ReLU, Tanh and other commonly used activation functions) according to the input value, its own weight and threshold, and finally obtains the corresponding execution effect prediction value in the output layer. This process is the forward information transmission from input to output, which simulates the process of generating prediction results based on input information.

[0042] (5) Calculation error: Compare the execution effect prediction value obtained by forward propagation with the actual execution effect value in the training set, and calculate the error between the two using appropriate error metrics. This error reflects the degree of deviation between the current network model prediction result and the actual situation, and is an important basis for subsequent adjustment of network parameters to optimize the model.

[0043] (6) Back propagation and parameter update: Based on the calculated error, the error is back propagated back to each neuron in the network, starting from the output layer, in the opposite direction of the forward propagation. In this process, the weights and thresholds are updated based on the derivative (i.e., gradient) of the error with respect to the weights and thresholds of each neuron connection, so that the network can reduce the error in the next round of forward propagation. By continuously repeating the steps of forward propagation, error calculation, back propagation, and parameter update, the preset training stop condition is reached, such as reaching the specified training rounds or the error converges to below a certain threshold.

[0044] (7) Model evaluation and optimization: Use the test set data to evaluate the trained BP neural network model, and judge the generalization ability of the model on unseen data by calculating the corresponding evaluation indicators (such as accuracy, recall rate, F1 value, etc., depending on the specific measurement method of the execution effect). If the model performance does not meet the expected requirements, the model can be retrained by adjusting the network structure, hyperparameters (such as learning rate, number of hidden layer nodes, etc.) or further optimizing data preprocessing, and the model can be continuously optimized until a satisfactory prediction effect is obtained.

[0045] The meeting information to be predicted is input into the BP neural network model and the expected effect is output.

[0046] When the model training is completed, for new meetings, the corresponding meeting information variables are extracted (the format of the variable set should be consistent with that when the model was trained), and this information is input into the BP neural network model. The model will calculate and output the expected effect of the execution of this meeting decision based on the previously learned knowledge and patterns. For example, it is expected that the market share of the product will increase by how much after the execution of this meeting decision, or whether the project progress will be advanced or delayed, etc.

[0047] Furthermore, the conference data analysis module also evaluates the conference participation of the participants according to the speech content of the participants, including the following steps: (1) Speech content acquisition and text preprocessing: The speech content of the participants is obtained from relevant sources such as meeting records and voice transcription texts. After acquisition, preprocessing work is carried out on these original texts to remove irrelevant punctuation, special characters and redundant spaces, and to standardize colloquial expressions and abbreviations to make them standardized and unified in format. Subsequently, lexical analysis is used to segment the text into independent words, paving the way for subsequent in-depth analysis of its semantics, ensuring that the speech content enters the next stage in a state that is conducive to analysis.

[0048] (2) Construct a participation evaluation index system: Based on the characteristics and goals of the meeting, a multi-dimensional participation evaluation index system is constructed. Among them, the speaking frequency index is used to count the number of speeches of participants, reflecting their level of activity in the meeting; the speaking duration index calculates the cumulative time of speeches and measures the time it takes to express their views; the speech relevance index uses natural language processing technology to analyze the degree of fit between the speech content and the theme and topics of the meeting; the viewpoint innovation index evaluates whether the views raised in the speech are novel and unique by comparing the conventional ideas of the industry with the views of other participants; the interactivity index considers the interaction between participants and others, such as responding, asking questions, refuting or supplementing others' views, etc., and comprehensively reflects the participation of participants from different aspects.

[0049] (3) Feature extraction and quantification: Based on the constructed evaluation indicators, the corresponding features are extracted from the speech content and related text information, and quantified. The frequency of speech can be directly counted by the specific number of times; the speech duration can be quantified by the corresponding duration of the voice recording or by the number of words in the text combined with the normal speaking speed, using time units; the relevance of the speech is converted into a numerical value by calculating the cosine similarity with the conference theme word vector, and the value is between 0 and 1; the innovation of the viewpoint is quantified by setting the innovation level score, inviting professionals or according to established rules to score; the interactivity is quantified by counting the number of times other participants are mentioned and the number of times others' views are responded to, so that each indicator can reflect its characteristics with a specific value.

[0050] (4) Weight allocation and comprehensive evaluation: Based on the importance of each evaluation indicator to the overall meeting participation, the corresponding weights of each indicator are determined by using expert scoring method, hierarchical analysis method, or analyzing a large amount of historical meeting data, using machine learning model training, etc. After the weights are determined, the quantitative value of each indicator is multiplied by the corresponding weight, and then the comprehensive evaluation score of each participant's meeting participation is obtained through comprehensive operations such as summation. Finally, the participation of the participants is ranked or graded according to the scores, such as high, medium, and low, or reflected according to specific score ranges, thereby achieving a comprehensive and objective evaluation of the meeting participation of the participants.

[0051] The participation evaluation results obtained can provide strong data support for subsequent decision-making in conference organization, personnel management and other aspects. For example, participants with high participation can be given more opportunities to participate or given more important roles in subsequent important meetings; for participants with low participation, targeted communication and counseling can be carried out to understand the reasons and help them improve their participation ability. At the same time, through the accumulation and analysis of participation evaluation data from multiple meetings, common problems in the conference organization process can also be discovered, such as whether the conference agenda setting is reasonable (if the relevance indicators of most participants are generally low), and then continuous improvements can be made to the planning and organization of the meeting to improve the overall meeting quality.

[0052] The conference data storage module is used to store and process audio data and conference information using distributed storage and computing technology.

[0053] In this embodiment, the conference data storage module provides an efficient and reliable solution for the storage and processing of audio data and conference information with the help of distributed storage and computing technology. In terms of distributed storage, storage systems such as HDFS and Ceph are selected to build a master-slave architecture, with metadata managed by the name node, data nodes storing data blocks in a dispersed manner, audio data being redundantly stored after being cut into blocks according to rules, structured data in conference information being stored in a distributed database, and unstructured data being stored by the file system and located by metadata management. In terms of distributed computing technology, frameworks such as Spark and Hadoop are introduced to process data with their parallel computing advantages, such as voice transcription and feature extraction of audio data, and aggregation analysis and natural language processing of conference information. This approach brings many advantages, with strong data scalability. Even if the amount of conference data continues to grow with the increase in conferences, linear expansion of storage and computing power can be achieved by simply adding new data nodes, breaking through the bottleneck of the traditional stand-alone mode. At the same time, it can significantly improve processing efficiency, use parallel mechanisms to speed up the execution of various tasks, and allow large-scale audio data and complex conference information to be quickly and properly stored and processed, laying a solid foundation for in-depth analysis and other work related to subsequent conferences.

[0054] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A conference data intelligent analysis and decision support system, characterized by: It includes a conference data acquisition module, a conference data processing module, a conference data analysis module and a conference data storage module, which are communicatively connected to each other, wherein: The conference data collection module is used to collect audio data during the conference in real time; The conference data processing module uses natural language processing technology to extract conference information from audio data, where the conference information includes positions of participants, conference topics, conference content, and voting results; The conference data analysis module uses a big data analysis model to predict the expected effect of conference decision execution based on conference information; The conference data storage module is used to store and process audio data and conference information using distributed storage and computing technology.

2. The conference data intelligent analysis and decision support system according to claim 1, characterized in that: The real-time collection of audio data during the meeting is specifically to use professional audio collection equipment to collect audio signals in the conference space according to a preset sampling frequency and quantization bit number, using a distributed microphone array layout or a centralized multi-point layout, and then use a high-speed data transmission interface to convert the collected analog audio signals into digital audio data in real time and transmit them to a subsequent conference data processing module.

3. The conference data intelligent analysis and decision support system according to claim 1 is characterized by: The method of extracting conference information from audio data using natural language processing technology includes the following steps: Speech recognition: Use acoustic models and deep neural network algorithms to train historical audio data samples to learn different voice features, accents, and speech speed patterns of audio data; Part-of-speech tagging: Use hidden Markov models or conditional random fields to tag the part of speech of each word in the audio data: Syntactic analysis: Use formal grammar rules to analyze the structure of sentences in audio data, and use dependency syntactic analysis to determine the dependency relationship between the various components in the sentence; Entity recognition: Use recurrent neural networks or long short-term memory networks to capture contextual information and sequence features in text to identify entities; Semantic understanding: semantic analysis technology is used to analyze the text, associate entities in the text with information in the knowledge graph, and use natural language processing technology to further understand the semantics of the text; Information extraction: Extract specific meeting information from the text based on rules and machine learning algorithms, and use machine learning algorithms to classify and verify the extracted information.

4. The conference data intelligent analysis and decision support system according to claim 1, characterized in that: The method of using a big data analysis model to predict the expected effect of conference decision execution based on conference information includes the following steps: Collect meeting information and historical data samples of known performance; The random forest model is used to select the meeting information variable set that has a significant impact on the execution effect; Establish a BP neural network model of conference information variable set and execution effect; The meeting information to be predicted is input into the BP neural network model and the expected effect is output.

5. The conference data intelligent analysis and decision support system according to claim 4 is characterized by: The random forest model includes the following construction steps: Data preparation: remove missing values ​​and outliers from the data, and standardize or normalize the data; Determine hyperparameters: Based on experience and experiments, set the number of decision trees and maximum depth hyperparameters of the random forest model; Feature and sample sampling: Perform self-service sampling on the original data, extract samples with replacement, and randomly select some features as the basis for splitting the decision tree nodes; Construct a decision tree: Based on the selected features and samples, determine the best split features according to information gain or Gini coefficient, construct a decision tree, and repeat this process until the stopping condition is met; Evaluate feature importance: By calculating the average importance of each feature in all decision trees, evaluate the importance of meeting information variables and screen out the set of variables that have a significant impact on the execution effect.

6. The conference data intelligent analysis and decision support system according to claim 4 is characterized by: The method of establishing the BP neural network model of the conference information variable set and the execution effect includes the following steps: Data preprocessing and division: preprocess the selected meeting information variable set data and the corresponding execution effect data, and divide the entire data set into training set, test set and validation set; Network structure design: Design the number of nodes of BP neural network neurons according to the dimensions of the input meeting information variable set and the specific form of the expected output execution effect. The neurons include input layer, hidden layer and output layer. Parameter initialization: Use random initialization method to initialize the weights and thresholds of the connections between neurons in the neural network; Forward propagation: Input the meeting information variable data in the training set into the input layer, and then calculate through each hidden layer in turn according to the network structure. Each neuron performs nonlinear transformation through the activation function according to the input value, weight and threshold, and finally obtains the corresponding execution effect prediction value in the output layer; Calculate the error: Compare the predicted execution effect value obtained by forward propagation with the actual execution effect value in the training set, and calculate the error between the two; Back propagation and parameter update: Based on the calculated error, starting from the output layer, the error is back propagated back to each neuron in the network in the opposite direction of the forward propagation, and the weights and thresholds are updated using the gradient descent method; forward propagation, error calculation, back propagation and parameter update are repeated continuously until the preset training stop condition is reached; Model evaluation and optimization: Use the test set data to evaluate the trained BP neural network model, and determine the accuracy of the model prediction by calculating the corresponding evaluation indicators.

7. The conference data intelligent analysis and decision support system according to claim 1, characterized in that: The conference data analysis module also evaluates the conference participation of the participants according to the content of their speeches, including the following steps: Speech content acquisition and text preprocessing: The speech content of each participant is obtained from the conference audio data, and the colloquial expressions and abbreviations are standardized and converted to make the format standardized and unified; Constructing a participation evaluation index system: Based on the characteristics and goals of the conference, a participation evaluation index system is constructed. The participation evaluation index includes speech frequency index, speech duration index, speech relevance index, viewpoint innovation index and interactivity index; Feature extraction and quantification: Based on the participation evaluation index system, corresponding features are extracted from the speech content and related text information, and quantified; Weight allocation and comprehensive evaluation: Determine the corresponding weight of each indicator based on its importance to the overall meeting participation, multiply the quantitative value of each indicator by the corresponding weight, and then sum them up to get the comprehensive evaluation score of each participant's meeting participation.

8. The conference data intelligent analysis and decision support system according to claim 1, characterized in that: The comprehensive evaluation score of the conference participation is calculated as follows: ， In the formula, H sum Provide a comprehensive evaluation score for meeting participation; H i For the i The scores of the participation evaluation indicators; w i For the i The weight of each participation evaluation indicator; n is the number of participation evaluation indicators.

9. The conference data intelligent analysis and decision support system according to claim 1, characterized in that: The conference data analysis module also visualizes the expected results and proposes corresponding improvement suggestions based on the historical information of the conference data storage module.

10. The conference data intelligent analysis and decision support system according to claim 1, characterized in that: The conference data storage module uses HDFS or Ceph storage system in distributed storage to build a master-slave architecture, with the name node managing metadata, data nodes storing data blocks in a dispersed manner, audio data being redundantly stored after being cut into blocks according to rules, structured data in conference information being stored in a distributed database, and unstructured data being stored in the file system and located by metadata management; In terms of computing technology, Spark and Hadoop frameworks are introduced to process data.

Images