Pre-training method, pre-training system and pre-training equipment for online intelligent interactive question-answering system

Abstract

The invention relates to a pre-training method, system and device for an online intelligent interactive question-answering system, and the method comprises the steps: data collection: constructing a multivariate data collection mechanism, collecting the data of each large website platform, and forming a data flow; data preprocessing: identifying and cleaning advertisement content in the data stream, and removing irrelevant information; data cleaning: pasting corresponding labels on data from different sites in the data stream, removing parts with poor quality in the data stream at the same time, and integrating according to the types of the labels to form a plurality of data sets; quantitative analysis: calculating the data set and evaluating a deviation index to identify and remove inconsistent texts in the data set; and prejudice detection and proofreading: utilizing a plurality of data sets to train a classifier, searching an optimal classification threshold value, and outputting the optimal classification threshold value on an original data set to obtain a prediction result so as to identify and remove texts possibly with prejudice. According to the method, the fairness and the accuracy of data analysis are ensured through fine data prejudice removal processing.

Description

Technical Field

[0001] The present invention relates to the technical field of AI intelligent systems, and in particular to a pre-training method, system and device for an online intelligent interactive question-and-answer system. Background Art

[0002] The existing online intelligent interactive question-and-answer system, or AI question-and-answer system, can achieve the effect of dependency syntactic analysis and semantic role labeling, and understand the structure and intention of user questions through NLP technology, that is, natural language processing technology. For example, when a user asks "What is the latest model of Apple mobile phone?", the system not only recognizes the keywords "Apple mobile phone" and "latest model", but also understands that the question is about product information. Such a system has a high level of language understanding ability and can accurately capture user needs. At the same time, in the existing technology, using pre-trained models such as BERT and GPT, AI question-and-answer systems can also generate more natural and accurate answers. These models are pre-trained on a large amount of text, learn the deep features of language, and can understand the contextual meaning of each word in the question, thereby providing more accurate answers. As for the search engine integration of the system, through search engine technology, it is possible to access massive information on the Internet. For example, when a user asks about the details of a historical event, the system may need to extract information from online resources such as Wikipedia and combine it with data in the database to generate a comprehensive answer. In addition, in the AI ​​question-answering system, the user's intentions and conversation context can be tracked to provide personalized suggestions. It relies on a structured or semi-structured knowledge base to provide accurate answers. Therefore, the system needs to be able to provide clear logic and data support and support multilingual and cross-cultural interactions.

[0003] However, the AI ​​question-answering system based on Internet big data in the existing technology still has certain defects:

[0004] 1. Data dependency: AI assistants rely on large amounts of data for learning and prediction. The quality and diversity of data directly affect their performance. If the data is biased or erroneous, AI predictions and decisions may be affected.

[0005] 2. Information quality control: Search engines may return information from a variety of sources. AI assistants need to be able to distinguish the reliability of information, otherwise they may provide inaccurate answers.

[0006] 3. Deep understanding limitations: AI may have difficulty understanding complex or multi-faceted problems, especially those that require deep analysis or critical thinking;

[0007] 4. Real-time and dependency issues: The information indexed by search engines may not be the latest, and AI assistants may not be able to provide the latest data or event progress. Summary of the invention

[0008] In order to solve the problems existing in the above-mentioned prior art, one of the purposes of the present invention is to provide a pre-training method for an online intelligent interactive question-and-answer system, which method comprehensively integrates the data diversity and depth of Internet big data, and can accurately debias and improve the fairness of the data to ensure the fairness and accuracy of the data. A second purpose of the present invention is to provide a pre-training system for an online intelligent interactive question-and-answer system, which mainly completes the above-mentioned pre-training method for the classifier, and can be combined with the online intelligent interactive question-and-answer system to run, to ensure the authenticity and objectivity of the data analysis results. A third purpose of the present invention is to provide a pre-training device for an online intelligent interactive question-and-answer system, so that the system can apply this method to train the classifier in an AI question-and-answer system based on Internet big data to ensure the authenticity and objectivity of the AI's answers to questions.

[0009] The pre-training method of an online intelligent interactive question-and-answer system described in the present invention is applied to an AI question-and-answer system based on Internet big data, and the method comprises:

[0010] Data collection: Build a multi-dimensional data collection mechanism to collect data from major website platforms and form a data flow;

[0011] Data preprocessing: identifying and cleaning the advertising content in the data stream and removing irrelevant information;

[0012] Data cleaning: labeling the data from different sites in the data stream with corresponding labels, removing the poor quality parts of the data stream, and integrating them according to the label type to form multiple data sets;

[0013] Quantitative analysis: calculating and evaluating deviation indicators for the data sets respectively to identify and remove inconsistent texts in the data sets;

[0014] Bias detection and proofreading: Use multiple datasets to train classifiers, find the optimal classification threshold and output the prediction results on the original dataset to identify and remove text that may be biased.

[0015] In one embodiment, the bias detection and correction method specifically includes:

[0016] Dataset initialization: divide the multiple data sets into training sets and test sets;

[0017] Data purification and configuration: Using natural language processing technology, set the structure of the training set, define the positive and negative labels of each data set in the training set, clarify the names of protected attributes in the training set, assign privileged and non-privileged attribute values ​​to the text in the training set, and mark the corresponding feature names of the data set;

[0018] Training and evaluation of the classifier: calculating the classification threshold and average rate difference of the training set, and continuously adjusting the assignment parameters of the training set to make the classification threshold and the average rate difference as small as possible and close to 0;

[0019] Classifier prediction: The classification threshold is output on the test set to obtain the prediction result.

[0020] In one embodiment, the classifier prediction method specifically includes:

[0021] Prediction execution: using the trained classifier to predict each of the data sets in the test set to generate preliminary prediction results;

[0022] Threshold exploration: Calculate and evaluate the impact of each classification threshold on different evaluation indicators;

[0023] Result display: Detailed display of the results under each threshold value to identify the optimal classification threshold setting.

[0024] In one embodiment, the classification threshold is calculated as follows: F=1-abs(d), where F is the classification threshold and d is the difference influence, i.e., the significant difference value of the classification results of different groups;

[0025] The calculation method of the mean rate difference is: P = 0.5 [(F u -F p )+(T u -T p )], where P is the average rate difference, F u is the false positive rate in the non-privileged group, F p is the false positive rate in the privileged group, T u is the true probability in the non-privileged group, T p is the true rate among privileged groups.

[0026] In one embodiment, the evaluation indicators include accuracy, precision, recall, F1 value, ROC and AUC curve indicators.

[0027] In one embodiment, the deviation indicators include deviation metrics, variance analysis and correlation tests.

[0028] In one embodiment, the method for collecting data from major website platforms includes: dynamically adjusting the collection frequency according to the website update frequency and the historical data change pattern.

[0029] A pre-training system for an online intelligent interactive question-answering system of the present invention comprises a human-computer interaction module and a data storage and processing module connected to each other, the system further comprises a data optimization module and a data debiasing module connected in sequence, the data debiasing module is connected to the data storage and processing module,

[0030] The data optimization module includes:

[0031] Data collection unit, collecting data from different sources on major website platforms;

[0032] A data integration unit, which sets a unified framework to integrate the data collected by the data collection unit to form a data flow;

[0033] A dynamic processing unit, which adjusts the collection frequency of the data collection unit according to the update frequency of different websites;

[0034] The data debiasing processing module includes:

[0035] A data preprocessing unit, for identifying and cleaning advertisement contents in the data stream transmitted from the data integration unit;

[0036] A data cleaning unit, which affixes corresponding labels to the data processed by the data preprocessing unit, removes unrecognizable content, and integrates the data according to the type of the label to form multiple data sets;

[0037] A quantitative analysis unit, which calculates and evaluates deviation indicators to identify and remove inconsistent texts in each of the data sets of the data cleaning unit;

[0038] A bias detection and correction unit, using the plurality of data sets processed by the quantitative analysis unit to establish a model training classifier, and determining an optimal classification threshold through learning;

[0039] A classifier is interconnected with the bias detection and correction unit and connected with the quantitative analysis unit, and classifies the multiple data sets processed by the quantitative analysis unit respectively, and sends the prediction results to the data storage and processing module.

[0040] A pre-training device for an online intelligent interactive question-and-answer system of the present invention comprises a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores the pre-training system for the online intelligent interactive question-and-answer system, and the processor executes the pre-training method for the online intelligent interactive question-and-answer system by executing instructions of the system.

[0041] A computer-readable storage medium of the present invention stores a pre-training system for the above-mentioned online intelligent interactive question-and-answer system, and the system is suitable for executing the pre-training method for the above-mentioned online intelligent interactive question-and-answer system.

[0042] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0043] 1. Advanced multivariate data collection system: This invention has developed an efficient multivariate data collection system that can capture information from multiple channels such as the Internet, social media, and user feedback in real time. The design purpose of this system is to ensure that the data of our AI intelligent question-answering system is always up-to-date, so as to provide users with the most timely and accurate information. Through this multi-dimensional data collection, we can not only keep up with the pulse of the times, but also enrich our data resources from different angles and levels, laying a solid foundation for deep learning and intelligent analysis of AI.

[0044] 2. Sophisticated data debiasing process: To ensure the fairness and accuracy of data analysis, the present invention uses a series of data preprocessing techniques, including thorough data cleaning and standardization, to improve the overall quality of the data. In addition, the present invention also establishes a bias detection and correction mechanism, using statistical tests and cutting-edge machine learning algorithms to identify and reduce bias in the data. The implementation of this process enables us to proactively discover and correct imbalances and injustices in the data, ensuring the authenticity and objectivity of the data analysis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a pre-training method of an online intelligent interactive question-answering system of the present invention;

[0046] Figure 2 It is a structural schematic diagram of a pre-training system of an online intelligent interactive question-answering system of the present invention;

[0047] Figure 3 is a schematic diagram of deviation index calculation in an embodiment of the present invention;

[0048] Figure 4 is a schematic diagram of an untrained classifier searching for an optimal classification threshold in an embodiment of the present invention;

[0049] Figure 5 is a schematic diagram of a classifier after learning in an embodiment of the present invention searching for an optimal classification threshold;

[0050] Figure 6 It is a comparison diagram before and after optimization of the classification thresholds of the present invention under different data sets;

[0051] Figure 7It is a comparison chart before and after the present invention is applied to the AI ​​question-answering system based on Internet big data. DETAILED DESCRIPTION

[0052] The drawings are only used for illustrative purposes and should not be construed as limitations on this patent. In order to better illustrate this embodiment, some parts of the drawings may be omitted, enlarged, or reduced, and do not represent the size of the actual product. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0053] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installation" and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. It can be said that the internal connection of two components is connected. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The technical solution of the present invention is further explained below in conjunction with the accompanying drawings and embodiments.

[0054] According to the online intelligent interactive question-answering system in the prior art, the inventors have found that the following problems need to be solved in view of its existing shortcomings:

[0055] 1. Multivariate data collection: Enhance data diversity, design and implement strategies to collect and integrate data from different sources, backgrounds, and situations, ensure the diversity of data sets, and avoid overfitting during model training;

[0056] 2. Ensuring data quality: Ensuring the quality of the diverse data collected is another key issue. The data may contain errors, biases, or limitations, which will affect the training effect and generalization ability of the model. How to ensure the quality of each data source and how to process and clean the data to reduce noise and bias is a major challenge in technical implementation;

[0057] 3. Data preprocessing and integration: Diverse data requires different processing methods and tools. For example, image data requires image processing technology, and voice data requires voice recognition technology. It is a complex problem to effectively label and integrate these different types of data.

[0058] 4. Data dependency and bias issues: AI question-answering systems rely heavily on large amounts of high-quality data for learning and prediction. However, a major drawback of existing technologies is the bias that may exist in the data collection process. If the data set is unrepresentative in terms of gender, race, age, etc., or is affected by some bias during the collection process, the decisions and predictions of the AI ​​question-answering system may tend to reflect these biases, leading to unfair or discriminatory results. In addition, incompleteness and errors in the data may also lead to inaccurate predictions of the AI ​​question-answering system, affecting its performance and credibility.

[0059] Therefore, if Figure 1 As shown, a pre-training method for an online intelligent interactive question-and-answer system of the present invention is applied to an AI question-and-answer system based on Internet big data, and the method comprises:

[0060] Data collection: Build a multi-dimensional data collection mechanism to collect data from major website platforms and form a data flow;

[0061] Data preprocessing: Identify and clean advertising content in data streams and remove irrelevant information;

[0062] Data cleaning: Label the data from different sites in the data stream accordingly, remove the poor quality data stream, and integrate them by label type to form multiple data sets;

[0063] Quantitative analysis: Calculate and evaluate deviation indicators for each data set to identify and remove inconsistent text in the data set. Deviation indicators include deviation measurement, variance analysis, and correlation test, which can be specifically as follows: Figure 3 As shown, statistical parity differences, equality of opportunity differences, mean rate differences, differential impacts, and the Theil index are calculated;

[0064] Bias detection and proofreading: Use multiple datasets to train classifiers, find the optimal classification threshold and output the prediction results on the original dataset to identify and remove text that may be biased.

[0065] The method of the present invention is committed to achieving real-time data convergence through an innovative multivariate data collection mechanism, and extracting information from all over the world, different fields and diverse situations. The multivariate data collection mechanism includes using distributed crawlers and deep learning-driven crawlers to collect data from sites such as social media, news portals, industry reports, etc., to enrich data dimensions and improve the comprehensiveness of data; using advanced Kafka stream processing technology to ensure the continuity and real-time nature of data flow; and through API integration strategies, dynamically obtaining data from social media and other online platforms to further enrich data resources. This technology not only broadens the source of data, but also deepens the connotation of data, significantly improving the diversity and comprehensiveness of data. This extensive and in-depth data collection strategy enables us to gain insight into problems from multiple angles, effectively avoid data bias, and provide a richer and more three-dimensional perspective for the decision-making process, thereby significantly improving the quality and reliability of decision-making. Furthermore, the method for collecting data from major website platforms includes: dynamically adjusting the collection frequency according to the website update frequency and the historical data change pattern, and using an intelligent scheduling algorithm. For example, for websites that are frequently updated, the collection frequency can be increased; for websites that are less updated, the frequency is appropriately reduced. In the field of data analysis, the present invention uses advanced data debiasing processing technology, preprocessing, cleaning and analysis, removing useless information and bias, and then completing debiasing detection through classifier training, which can accurately identify and reduce potential biases in data. The AI ​​question-answering system processed by this method can provide an unbiased data analysis environment, let the data speak for itself, truly reflect its diversity and complexity, and provide credible and fair data support for decision makers.

[0066] Furthermore, the bias detection and correction methods include:

[0067] Dataset initialization: divide multiple data sets into training sets and test sets;

[0068] Data purification and configuration: Use natural language processing technology to set the structure of the training set, define the positive and negative labels of each data set in the training set, clarify the names of protected attributes in the training set, assign privileged and non-privileged attribute values ​​to the text in the training set, and mark the feature names of the corresponding data sets;

[0069] Training and evaluation of the classifier: Calculate the classification threshold and average rate difference of the training set, and continuously adjust the assignment parameters of the training set to make the classification threshold and average rate difference as small as possible and close to 0;

[0070] Classifier prediction: Output the found classification threshold on the test set to get the prediction result.

[0071] This method ensures the fairness and accuracy of each analysis result through meticulous data analysis and correction. In particular, for classifier training, the above series of data preprocessing techniques are first used, including thorough data cleaning and standardization, to improve the overall quality of the data. Then a set of bias detection and correction mechanisms are established, using statistical tests and cutting-edge machine learning algorithms to identify and reduce bias in the data. The implementation of this process enables us to proactively discover and correct imbalances and injustices in the data, ensuring the authenticity and objectivity of the data analysis results. Through this meticulous data debiasing process, a more fair and transparent data analysis environment is built.

[0072] Moreover, the classifier prediction methods specifically include:

[0073] Prediction execution: Use the trained classifier to predict each data set in the test set and generate preliminary prediction results;

[0074] Threshold exploration: Calculate and evaluate the impact of each classification threshold on different evaluation indicators;

[0075] Result display: Detailed display of the results under each threshold value to identify the optimal classification threshold setting.

[0076] The prediction after classifier training may not only find an optimal classification threshold, because it requires the difference between the classification threshold and the average rate to be as small as possible and close to 0. In this case, these near-optimal classification thresholds can be used to further judge and identify the optimal classification threshold setting based on their impact on different evaluation indicators. The evaluation indicators include accuracy, precision, recall, F1 value, ROC and AUC curve indicators, etc. At this time, applying them to the AI ​​question-answering system can further ensure the fairness and accuracy of the data.

[0077] Specifically, the classification threshold is calculated as: F = 1-abs(d), where F is the classification threshold and d is the difference effect, i.e., the significant difference value of the classification results of different groups. This threshold should be as small as possible, close to 0, to ensure that the prediction results of the classifier show a high degree of consistency among different groups, thereby achieving fairness.

[0078] The calculation method of the mean rate difference is: P = 0.5 [(F u -F p )+(T u -T p )], where P is the average rate difference, F u is the false positive rate in the non-privileged group, F p is the false positive rate in the privileged group, T u is the true probability in the non-privileged group, T pis the true positive rate in the privileged group. The value of the mean rate difference is also small, close to 0, which indicates that the classifier treats all groups fairly in positive and negative predictions, whether privileged or non-privileged.

[0079] The above classification threshold and average rate difference are made as small as possible and close to 0, which is calculated by these two formulas after assigning values ​​to the data set by existing natural language processing technology. During the classifier learning process, the training set is kept as close as possible to the real data, and the parameters are continuously adjusted by existing natural language processing technology to make the above two values ​​as small as possible to obtain the corresponding classification threshold. Then, the classification threshold is applied to the test set after the classifier model is tuned to evaluate and verify the test results. Figure 4 and Figure 5 As shown, Figure 4 The data in are the default values ​​of the unadjusted parameters, that is, the classification threshold results directly output when the classifier is classified without training, and Figure 5 The training set is assigned parameters that are adjusted to make the difference between the classification threshold and the average rate calculated by the above formula as small as possible and close to 0, so that the classification threshold result obtained by the classifier after learning and outputting it to the test set is obtained. For further reference Figure 6 By comparing multiple data sets, it can be seen that when using the original training data, the classification threshold is high, which means that the data classification is unfair; however, after the assignment weighting of natural language processing technology, the optimal classification threshold is lower and closer to each other, which can achieve higher fairness and accuracy indicators.

[0080] like Figure 2 As shown, a pre-training system of an online intelligent interactive question-answering system of the present invention comprises a human-computer interaction module and a data storage and processing module connected to each other, the system further comprises a data optimization module and a data debiasing module connected in sequence, the data debiasing module is connected to the data storage and processing module,

[0081] The data optimization module includes:

[0082] Data collection unit, collecting data from different sources on major website platforms;

[0083] The data integration unit sets up a unified framework to integrate the data collected by the data collection unit to form a data flow;

[0084] A dynamic processing unit adjusts the collection frequency of the data collection unit according to the update frequency of different websites;

[0085] The data debiasing processing module includes:

[0086] A data preprocessing unit, which identifies and cleans the advertising content in the data stream transmitted by the data integration unit;

[0087] The data cleaning unit attaches corresponding labels to the data processed by the data preprocessing unit, removes unrecognizable content, and integrates them according to the type of labels to form multiple data sets;

[0088] The quantitative analysis unit calculates and evaluates deviation indicators to identify and remove inconsistent texts in each data set of the data cleaning unit;

[0089] The bias detection and proofreading unit uses the multiple data sets processed by the quantitative analysis unit to build a model to train the classifier and determines the optimal classification threshold through learning;

[0090] The classifier is interconnected with the bias detection and proofreading unit and connected with the quantitative analysis unit, and classifies the multiple data sets processed by the quantitative analysis unit respectively, and sends the prediction results to the data storage and processing module.

[0091] The system can be directly applied to online intelligent interactive question-and-answer systems, that is, AI question-and-answer systems based on Internet big data. First, the data collection unit collects data from sites such as social media, news portals, and industry reports. The dynamic processing unit can control the frequency of collection, and then transmit the data to the data integration unit, which integrates the data in a unified framework to form a data stream, such as by website type or update time. After that, the data stream enters the data preprocessing unit to identify and clean specific advertising content, such as micro-business soft articles, overseas purchasing, sales promotion, various lottery draws on Weibo, discount promotion information discounts, coupons, tourism promotion, recruitment, training, event registration, purchase calls and other advertising data, and then the data cleaning unit affixes data type labels of different sites. For example, all e-commerce related notes are uniformly marked as "e-commerce community", simplifying the label system and facilitating subsequent data analysis and processing. At the same time, for some samples with poor data quality, such as missing values ​​and outliers, the data cleaning unit will remove data with missing key fields, abnormal key fields, images without content recognition, image OCR without text recognition, and audio without content recognition to ensure data availability. The quantitative analysis unit further performs bias detection on the processed data set to identify inconsistencies and non-randomness in the data set. Through these indicators, it can be evaluated whether the data is balanced, whether there are systematic biases, and the impact these biases may have on the analysis results, so as to select appropriate data sets as subsequent training sets and test sets, and establish training and learning models for classifiers. The bias detection and proofreading unit mainly combines natural language processing technology to assign weights to data texts, train classifiers according to the above pre-training method, and obtain the optimal classification threshold after continuously adjusting the parameters. The classifier can then use the classification threshold to find the best quality text for the existing data in the original AI question-and-answer system to answer users. Figure 7As shown in the figure, the classifier that has not been trained by the pre-training system according to the above pre-training method will present a relatively brief answer and insufficient information when used in the AI ​​question-answering system. Most of the useful information will be deleted by the classifier with a higher classification threshold. After training, the optimal classification threshold is found, and the classifier will treat various types of information more accurately and fairly, and the answer text will be more detailed. At the same time, there will be no bias, deviation, or useless information such as advertising.

[0092] A pre-training device for an online intelligent interactive question-and-answer system of the present invention comprises a memory and a processor, the memory and the processor are connected to each other in communication, the memory stores the pre-training system of the online intelligent interactive question-and-answer system, and the processor executes the pre-training method of the online intelligent interactive question-and-answer system by executing the instructions of the system. The memory and the processor of the device can be connected by a bus or other means, and the processor can be a central processing unit CPU, a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other editable logic devices, discrete gates or transistor logic devices, discrete hardware components and other chips, or a combination of the above-mentioned various types of chips. The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs, non-transient computer executable programs and modules, such as program instructions / modules corresponding to the above-mentioned pre-training method. The processor executes various functional applications and data processing of the processor by running the non-transient software programs, instructions and modules stored in the memory, that is, realizing the above-mentioned pre-training method. The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0093] A computer-readable storage medium of the present invention stores a pre-training system of the above-mentioned online intelligent interactive question-and-answer system, and the system is suitable for executing the pre-training method of the above-mentioned online intelligent interactive question-and-answer system. All or part of the processes in the above-mentioned pre-training method can be completed by instructing the relevant hardware through a computer program, and the implemented program can be stored in a computer-readable storage medium, and the program, when executed, can include the processes of the above-mentioned methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory ROM, a random access memory RAM, a flash memory, a hard disk HDD or a solid-state drive SSD, etc., or it can also be a combination of the above-mentioned types of memory.

[0094] In the figure, the description of the positional relationship is only for illustrative purposes and should not be understood as a limitation on this patent; obviously, the above embodiments of the present invention are only examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made on the basis of the above description. It is not necessary and impossible to list all the implementation methods here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A pre-training method for an online intelligent interactive question-answering system, characterized in that: The method is applied to an AI question-answering system based on Internet big data, and the method includes: Data collection: Build a multi-dimensional data collection mechanism to collect data from major website platforms and form a data flow; Data preprocessing: identifying and cleaning the advertising content in the data stream and removing irrelevant information; Data cleaning: labeling the data from different sites in the data stream with corresponding labels, removing the poor quality parts of the data stream, and integrating them according to the label type to form multiple data sets; Quantitative analysis: calculating and evaluating deviation indicators for the data sets respectively to identify and remove inconsistent texts in the data sets; Bias detection and proofreading: Use multiple datasets to train classifiers, find the optimal classification threshold and output the prediction results on the original dataset to identify and remove text that may be biased.

2. The pre-training method for an online intelligent interactive question-answering system according to claim 1, characterized in that: The bias detection and correction methods specifically include: Dataset initialization: divide the multiple data sets into training sets and test sets; Data purification and configuration: Using natural language processing technology, set the structure of the training set, define the positive and negative labels of each data set in the training set, clarify the names of protected attributes in the training set, assign privileged and non-privileged attribute values ​​to the text in the training set, and mark the corresponding feature names of the data set; Training and evaluation of the classifier: calculating the classification threshold and average rate difference of the training set, and continuously adjusting the assignment parameters of the training set to make the classification threshold and the average rate difference as small as possible and close to 0; Classifier prediction: The classification threshold is output on the test set to obtain the prediction result.

3. The pre-training method for an online intelligent interactive question-answering system according to claim 2, characterized in that: The classifier prediction method specifically includes: Prediction execution: using the trained classifier to predict each of the data sets in the test set to generate preliminary prediction results; Threshold exploration: Calculate and evaluate the impact of each classification threshold on different evaluation indicators; Result display: Detailed display of the results under each threshold value to identify the optimal classification threshold setting.

4. The pre-training method for an online intelligent interactive question-answering system according to claim 3, characterized in that: The classification threshold is calculated as follows: F = 1-abs(d), where F is the classification threshold and d is the difference impact, i.e., the significant difference value of the classification results of different groups; The calculation method of the mean rate difference is: P = 0.5 [(F u -F p )+(T u -T p )], where P is the average rate difference, F u is the false positive rate in the non-privileged group, F p is the false positive rate in the privileged group, T u is the true probability in the non-privileged group, T p is the true rate among privileged groups.

5. The pre-training method for an online intelligent interactive question-answering system according to claim 4, characterized in that: The evaluation indicators include accuracy, precision, recall, F1 value, ROC and AUC curve indicators.

6. The pre-training method for an online intelligent interactive question-answering system according to claim 5, characterized in that: The deviation indicators include deviation measurement, variance analysis and correlation test.

7. The pre-training method for an online intelligent interactive question-answering system according to claim 6, characterized in that: The method for collecting data from major website platforms includes: dynamically adjusting the collection frequency according to the website update frequency and the historical data change pattern.

8. A pre-training system for an online intelligent interactive question-answering system, comprising a human-computer interaction module and a data storage and processing module connected to each other, characterized in that: The system further comprises a data optimization module and a data debiasing module connected in sequence, wherein the data debiasing module is connected to the data storage and processing module. The data optimization module includes: Data collection unit, collecting data from different sources on major website platforms; A data integration unit, which sets a unified framework to integrate the data collected by the data collection unit to form a data flow; A dynamic processing unit, which adjusts the collection frequency of the data collection unit according to the update frequency of different websites; The data debiasing processing module comprises: A data preprocessing unit, for identifying and cleaning advertisement contents in the data stream transmitted from the data integration unit; A data cleaning unit, which affixes corresponding labels to the data processed by the data preprocessing unit, removes unrecognizable content, and integrates the data according to the type of the label to form multiple data sets; A quantitative analysis unit, which calculates and evaluates deviation indicators to identify and remove inconsistent texts in each of the data sets of the data cleaning unit; A bias detection and correction unit, using the plurality of data sets processed by the quantitative analysis unit to establish a model training classifier, and determining an optimal classification threshold through learning; A classifier is interconnected with the bias detection and correction unit and connected with the quantitative analysis unit, and classifies the multiple data sets processed by the quantitative analysis unit respectively, and sends the prediction results to the data storage and processing module.

9. A pre-training device for an online intelligent interactive question-answering system, characterized in that: It includes a memory and a processor, the memory and the processor are communicatively connected to each other, the memory stores a pre-training system for an online intelligent interactive question-and-answer system as described in claim 8, and the processor executes the pre-training method for an online intelligent interactive question-and-answer system as described in any one of claims 1-7 by executing instructions of the system.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a pre-training system for an online intelligent interactive question-and-answer system as described in claim 8, and the system is suitable for executing a pre-training method for an online intelligent interactive question-and-answer system as described in any one of claims 1-7.

Images