Intelligent text review method and device integrating big data and small data models

Abstract

The present invention discloses an intelligent text review method and device integrating big data and small data models, belonging to the field of information processing technology. The method comprises: acquiring text data of multiple data modes; performing feature extraction on the text data to obtain feature data; inputting the feature data into a pre-trained hybrid model, extracting keywords of general concern and generating corresponding content summaries; according to the keywords and content summaries, an intelligent review rule engine dynamically adjusts and matches rules, extracts content paragraphs that need to be reviewed and sends them to a review interface; the present invention integrates the breadth advantage of big data models and the deep understanding ability of small data models to realize efficient keyword extraction, accurate content comparison, intelligent matching of review rules for text files, and displaying review results through an interactive interface, thereby significantly improving the efficiency and accuracy of text review.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to an intelligent text review method and device integrating big data and small data models. Background Art

[0002] With the deepening of the Internet and digital transformation, the speed of text information generation and dissemination is accelerating, and the demand for text review is becoming more urgent, especially in industries such as law, finance, and news media, which have strict requirements for the accuracy and compliance of information. Traditional manual review methods are not only time-consuming and labor-intensive, but also prone to omissions and errors. In recent years, although the advancement of natural language processing technology has made automated text review possible, how to ensure the quality of review while taking into account processing speed and flexibility is still a problem that needs to be solved.

[0003] Most existing text review systems are based on keyword matching or simple machine learning models. These methods show obvious limitations when dealing with complex contexts and specific domain knowledge. Although big data models can process large amounts of data, they lack generalization capabilities when dealing with specific domains or professional terms. On the contrary, small data models excel in in-depth analysis of specific domains, but their generalization capabilities are limited due to data volume constraints. Therefore, how to effectively integrate the advantages of big data and small data has become the key to improving the performance of text review systems. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an intelligent text review method and device that integrates big data and small data models, so as to integrate the breadth advantages of big data models and the deep understanding capabilities of small data models, realize efficient keyword extraction of text files, accurate content comparison, intelligent matching of review rules, and display of review results through an interactive interface, thereby significantly improving the efficiency and accuracy of text review.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] On the one hand, the present invention provides an intelligent text review method integrating big data and small data models, comprising:

[0007] Acquire text data in multiple data modes, wherein the text data includes public big data resources and small data samples in the audit field;

[0008] Extracting features from the text data to obtain representative feature data;

[0009] Input representative feature data into the pre-trained hybrid model to extract commonly-used keywords and generate corresponding content summaries;

[0010] Based on the keywords and content summary, the intelligent review rule engine dynamically adjusts and matches the rules, extracts the content paragraphs that need to be reviewed and sends them to the review interface;

[0011] The right side of the review interface integrates a review information panel, wherein the review information includes review points, review content, and the reviewer's opinions or decisions. The review point position is pre-integrated with a link for the user to click and jump to the corresponding document location.

[0012] On the other hand, the present invention provides an intelligent text review device integrating big data and small data models, comprising:

[0013] An acquisition module, used to acquire text data in multiple data modes, wherein the text data includes public big data resources and small data samples in the audit field;

[0014] An extraction module, used for performing feature extraction on the text data to obtain representative feature data;

[0015] The generation module is used to input representative feature data into the pre-trained hybrid model, extract commonly concerned keywords and generate corresponding content summaries;

[0016] A sending module, which is used to dynamically adjust and match rules based on the keywords and content summary of the intelligent review rule engine, extract the content paragraphs that need to be reviewed, and send them to the review interface;

[0017] The right side of the review interface integrates a review information panel, wherein the review information includes review points, review content, and the reviewer's opinions or decisions. The review point position is pre-integrated with a link for the user to click and jump to the corresponding document location.

[0018] On the other hand, the present invention provides an electronic device, comprising: a shell, a processor, a memory, a circuit board and a power supply circuit, wherein the circuit board is placed inside the space enclosed by the shell, and the processor and the memory are arranged on the circuit board; the power supply circuit is used to supply power to various circuits or devices of the above-mentioned electronic device; the memory is used to store executable program code; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute the above-mentioned method.

[0019] In yet another aspect, the present invention provides a computer-readable storage medium storing one or more programs, wherein the one or more programs can be executed by one or more processors to implement the above method.

[0020] The present invention has the following beneficial effects:

[0021] The present invention integrates the breadth advantages of big data models and the deep understanding capabilities of small data models to achieve efficient keyword extraction, accurate content comparison, intelligent matching of audit rules for text files, and display of audit results through an interactive interface, which significantly improves the efficiency and accuracy of text auditing. By comprehensively solving a series of technical challenges such as data processing, information extraction, rule application, difference comparison, interactive design, and system evolution, it provides a comprehensive, efficient and sustainable development solution for the field of text auditing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0023] Figure 1 It is a flow chart of the intelligent text review method integrating big data and small data models of the present invention;

[0024] Figure 2 It is a schematic diagram of the principle of the intelligent text review method integrating big data and small data models of the present invention;

[0025] Figure 3 It is a structural schematic diagram of the intelligent text review device integrating big data and small data models of the present invention;

[0026] Figure 4 It is a schematic structural diagram of the electronic device of the present invention. DETAILED DESCRIPTION

[0027] In order to more clearly illustrate the overall concept of the present invention, a detailed description is given below in an exemplary manner in conjunction with the accompanying drawings.

[0028] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0029] On the one hand, the present invention provides an intelligent text review method integrating big data and small data models, such as Figure 1-2 As shown, including:

[0030] S101: Acquire text data in multiple data modes, where the text data includes public big data resources and small data samples in the audit field;

[0031] In this step, text data is collected from multiple channels, where the multiple channels may include: city open data platforms, academic databases, social media APIs, small-scale data sets collected and published within enterprises or organizations, etc.

[0032] S102: extracting features from the text data to obtain representative feature data;

[0033] In this step, there are four ways to extract text features.

[0034] As an optional embodiment, the feature extraction method may include:

[0035] Transforming the original features into new features with a smaller number than the original features by mapping or transforming methods;

[0036] In this step, we first standardize or normalize the data and process missing values; then, we design new features based on business needs and data characteristics; finally, we reduce the dimension of the data to obtain new features.

[0037] and / or, selecting some of the most representative features from the original features;

[0038] In this step, the most representative features are the features that best meet the business needs. After selecting such features, new features can be obtained.

[0039] and / or, selecting the most influential features based on expert knowledge;

[0040] In this step, first, the opinions of experts in related fields are collected; then, all features are preliminarily screened based on the experts' opinions; finally, the remaining features are new features.

[0041] And / or, select based on mathematical methods to find the features that are most informative for classification.

[0042] This step is used to measure the degree of mutual dependence between two random variables. It can be applied to discrete and continuous variables. Therefore, the result is more accurate and less interfered by human factors. It is especially suitable for text applications. The main process is as follows:

[0043] First, by constructing a mutual information function (evaluation function), each feature in the feature set is evaluated and scored, so that each word obtains an evaluation value, also known as a weight; then all features are sorted by weight; finally, a predetermined number of optimal features are extracted as a feature subset of the extraction result to obtain a new feature. The calculation process of the mutual information function can be as follows:

[0044] I(X;Y) is a measure of the dependence between random variables XX and YY and is defined as follows:

[0045] I(X;Y)=∑x∈X∑y∈Yp(x,y)logp(x,y)p(x)p(y);

[0046] I(X;Y)=∑x∈X∑y∈Yp(x,y)logp(x)p(y)p(x,y);

[0047] Among them, p(x,y) is the joint probability distribution of XX and YY, and p(x) and p(y) are the marginal probability distributions of XX and YY respectively.

[0048] In the specific implementation, first, you can load a text classification dataset, then use the TF-IDF vectorizer to convert the text into numerical features, and then calculate the mutual information score between these features and the category label, and get the top 10 features with the highest scores. Use the mutual_info_classif function in the scikit-learn library to calculate the mutual information. This function accepts the feature matrix XX and the target vector yy, and returns an array containing the mutual information score between each feature and the target variable.

[0049] In summary, all methods of text feature extraction use words as feature items, called feature words. These feature words are used as the intermediate representation of documents to calculate the similarity between documents and documents, and between documents and user goals. If all words are used as feature items, the dimension of the feature vector will be too large, which will put great pressure on the computing performance of the classification system. In this case, it is almost impossible to complete text classification. Seeking an effective feature dimensionality reduction method can not only reduce the computational complexity, but also improve the efficiency and accuracy of classification. Therefore, it is an important technology in automatic text classification.

[0050] The main function of the above four text feature extraction methods is to reduce the dimension of the vector space without damaging the core information. By using one or a combination of several of them, new features suitable for use can be obtained, which can simplify calculations and improve the speed and efficiency of text processing.

[0051] S103: inputting the feature data into a pre-trained hybrid model, extracting commonly concerned keywords and generating corresponding content summaries;

[0052] In this step, deep analysis refers to multi-level analysis of data to reveal hidden patterns, trends, and insights. In practice, in the context of text analytics, deep analysis includes natural language processing (NLP), machine learning, and deep learning, which can, for example:

[0053] Topic modeling: identifying the main topics or themes of a text;

[0054] Entity recognition: Identify specific types of entities from text, such as names of people and places;

[0055] Intent recognition: understanding the intent or purpose behind the text;

[0056] Sensitive word filtering: Detect and mark content containing sensitive words.

[0057] As an optional embodiment, the training method of the hybrid model may include:

[0058] A1: Obtain public big data resources, obtain a big data resource set, select a deep learning architecture suitable for the problem type, train the big data model, and optimize the model's hyperparameters;

[0059] In this step, the deep learning architecture suitable for the problem type may include a variety of learning architectures that can be easily thought of by technicians in this field, such as convolutional neural networks, recurrent neural networks, bidirectional recurrent neural networks, deep Q networks, and multimodal fusion networks. During specific implementation, it can be selected according to the direction of the application.

[0060] A2: Obtain small data samples in the audit field, obtain a small data sample set, train a small data model, and obtain prediction results;

[0061] Preferably, this step may include:

[0062] A21: Use data enhancement technology to increase data diversity and finely annotate the small data sample set;

[0063] In this step, data enhancement techniques such as rotation, flipping, and cropping can be used to finely label small data sample sets to ensure data quality.

[0064] A22: Initialize the small data model using the pre-trained weights of the big data model, freeze the bottom layer of the model at the beginning of training, and only train the top layer classifier. After the model converges, gradually unfreeze more layers and continue to fine-tune the model to adapt to the data in the audit domain. Use domain adaptation techniques to reduce the difference between the big data model (source domain) and the small data model (target domain).

[0065] In this step, the process of transfer learning can reduce the number of training iterations and complete the model training process faster. At the same time, through transfer learning, useful feature representations can be learned from a large amount of data in the source domain, thereby reducing the demand for target domain data.

[0066] A23: Use the big data model to predict samples in the small data sample set, generate soft labels, construct a loss function, and use the small data sample set and its corresponding hard labels and soft labels to train the small data model.

[0067] In this step, the smoothness of the soft label is controlled by adjusting the temperature parameter T. The larger T is, the smoother the soft label is. The constructed loss function usually includes two parts: one is the cross entropy loss of the hard label, and the other is the cross entropy loss of the soft label.

[0068] A3: Fuse the prediction results of the big data model (teacher model) and the small data model (student model), fine-tune the fused model using the small data sample set, and further optimize the model based on domain characteristics.

[0069] After completing the above training process, a trained hybrid model can be obtained. At this time, the accuracy, adaptability and stability of the model have been enhanced to a certain extent, and it can more efficiently extract commonly concerned keywords and generate corresponding content summaries.

[0070] S104: Based on the keywords and content summary, the intelligent review rule engine dynamically adjusts and matches the rules, extracts the content paragraphs that need to be reviewed, and sends them to the review interface;

[0071] The right side of the review interface integrates a review information panel, wherein the review information includes review points, review content, and the reviewer's opinions or decisions. The review point position is pre-integrated with a link for the user to click and jump to the corresponding document location.

[0072] In this step, a dynamically updated review rule library is established, covering multiple dimensions such as laws and regulations, industry standards, and sensitive information. Through the rule engine, the content paragraphs that need to be reviewed can be automatically matched and extracted based on keywords and content summaries; the review information panel is integrated on the right side of the review interface, which lists in detail the location, reason, and suggested modifications of each review point. Users can click on the review information on the right to achieve an instant link with the corresponding review point in the document on the left, which is convenient for quick browsing and processing.

[0073] As an optional embodiment, the method for creating the intelligent audit rule engine may include:

[0074] B1: Obtain and organize historical audit data, analyze historical audit cases, and extract key audit points, common violation types and keywords;

[0075] In this step, the audit results are analyzed in depth, including cases that passed and failed the audit. A dynamic update mechanism for key audit points, common violation types and keywords is established. By learning from a large amount of historical audit data, the model can automatically identify potential violation patterns and characteristics, providing early warnings and references for auditors.

[0076] B2: Based on the extracted keywords and legal terms, design an audit rule template, which includes the focus of the audit (such as privacy protection, antitrust, advertising compliance, etc.), relevant legal basis and keyword matching logic;

[0077] In this step, the audit results and reasons for both passed and failed content must be fed back to relevant personnel to help auditors understand the audit standards and improve content quality. It can also provide reference for the improvement of audit rules.

[0078] B3: Generate intelligent rules, use natural language processing technology (such as TF-IDF, word vector model) to learn keywords and phrases from historical operations, and automatically generate or optimize audit rules.

[0079] In this step, an audit report is generated for historical operations to analyze the execution effect of the rules, existing problems and improvement directions, and provide data support for rule optimization. During specific implementation, for example, when a keyword in a historical operation is mistakenly reviewed as illegal content, the keyword matching logic or audit standards can be adjusted.

[0080] As an optional embodiment, the method may include:

[0081] S105: Based on historical operation data, analyze the auditor's operation records and automatically identify the audit areas where each auditor is good at and efficient;

[0082] In this step, an adaptive learning mechanism is introduced to establish a user operation habit model through clustering or machine learning algorithms to automatically identify the audit areas that each auditor is good at and efficient. The model can automatically adjust parameters and weights according to the auditor's performance and feedback. During specific implementation, the user's historical operation data can be annotated to clarify the audit task type, difficulty level, importance and other attributes corresponding to each operation, and automatically adjust the model parameters and weights according to the above attributes.

[0083] S106: Analyze the content characteristics of the files to be reviewed, intelligently assign review tasks based on the file content characteristics, give priority to reviewers who are good at this field, and mark their historical review preferences.

[0084] In this step, the historical operation data is annotated, and the audit tasks are prioritized according to the characteristics of the file content, the auditor's areas of expertise, and the annotated historical operation data. Based on the auditor's feedback and task completion status, the intelligent allocation strategy is dynamically adjusted to achieve personalized audit task allocation.

[0085] As another optional embodiment, the method may include:

[0086] S107: Receive the manual reviewer's operations on the audit results, and perform feedback operations on the audit results based on the above operations. The feedback operations include confirmation, cancellation of confirmation, ignore, and cancellation of ignore. The training parameters of the hybrid model are adjusted according to the feedback operation data, and the feedback operation data is used for incremental learning.

[0087] In this step, feedback operations can be divided into accuracy issues, completeness issues, timeliness issues, etc. If the feedback shows that the model performs poorly in certain specific scenarios, parameters such as the learning rate and number of iterations can be adjusted specifically to improve the performance of the model in these scenarios. At the same time, new feedback data is regularly added to the training set to retrain the model so that the model can continuously adapt to new data and problems.

[0088] In summary, the problems solved by the method of the present invention are as follows:

[0089] 1. The balance between data size and quality

[0090] Traditional text review systems are often limited to a single data model, either relying on a large amount of general data with limited generalization capabilities, or focusing on a small amount of professional data that lacks broad representation.

[0091] This invention effectively solves this balance problem by integrating big data models with small data models. The big data model can learn a wide range of language rules and patterns and provide universal text understanding capabilities, while the small data model focuses on the audit field and masters the professional knowledge and terminology in the field through transfer learning and fine tuning, thereby maintaining generalization capabilities while enhancing in-depth understanding of audit scenarios.

[0092] 2. Accurate extraction of key information

[0093] It is a challenge to quickly and accurately extract key information from massive texts. This paper adopts advanced algorithms combined with hybrid models. It not only uses classic methods such as TF-IDF and TextRank to extract keywords of general concern, but also uses small data models to deeply explore the core concepts and professional vocabulary in the audit field to ensure the comprehensiveness and accuracy of key information extraction.

[0094] 3. Flexible application and intelligent matching of audit rules

[0095] In the face of ever-changing audit standards and complex and ever-changing text content, it is crucial to establish a rigorous and flexible audit rule system. The intelligent audit rule engine of the present invention can dynamically adjust and match rules to ensure the timeliness and applicability of the rule base. By deeply analyzing the extracted text content, the system can automatically identify and match the corresponding audit rules, reduce the burden of manual configuration, and improve the level of automation of auditing.

[0096] 4. Optimization of user interaction experience

[0097] Traditional audit systems often overlook the importance of user experience. This invention designs an interactive audit information display interface. Through the intuitive right-hand audit panel, users can directly view the detailed information of each audit point and directly link to the corresponding position in the document by clicking, thus achieving seamless connection between audit information and original text, greatly improving the convenience and efficiency of user operation.

[0098] 5. Continuous optimization and self-evolution capabilities

[0099] In order to cope with the continuous changes in the language environment and laws and regulations, the present invention integrates a user feedback mechanism to feed back the manually reviewed correction information into the continuous optimization process of the model and rule base, forming a closed-loop self-evolution system. This mechanism ensures that the accuracy and adaptability of the system are continuously enhanced over time, maintaining technological leadership and practicality.

[0100] On the other hand, the present invention provides an intelligent text review device integrating big data and small data models, such as Figure 3 As shown, including:

[0101] An acquisition module 11 is used to acquire text data in multiple data modes, wherein the text data includes public big data resources and small data samples in the audit field;

[0102] An extraction module 12 is used to extract features from the text data to obtain representative feature data;

[0103] A generation module 13 is used to input representative feature data into a pre-trained hybrid model, extract commonly concerned keywords and generate corresponding content summaries;

[0104] The sending module 14 is used to dynamically adjust and match the rules of the intelligent review rule engine according to the keywords and content summary, extract the content paragraphs that need to be reviewed and send them to the review interface;

[0105] The right side of the review interface integrates a review information panel, wherein the review information includes review points, review content, and the reviewer's opinions or decisions. The review point position is pre-integrated with a link for the user to click and jump to the corresponding document location.

[0106] The device of this embodiment can be used to perform Figure 1 The technical solution of the method embodiment shown has similar implementation principles and technical effects, which will not be repeated here.

[0107] Preferably, the generating module 13 includes:

[0108] The first acquisition submodule is used to acquire public big data resources, obtain a big data resource set, select a deep learning architecture suitable for the problem type, train a big data model, and optimize the model's hyperparameters;

[0109] The second acquisition submodule is used to acquire small data samples in the audit field, obtain a small data sample set, train a small data model, and obtain a prediction result;

[0110] The fusion submodule is used to fuse the prediction results of the big data model and the small data model, use the small data sample set to fine-tune the fused model, and further optimize the model according to the domain characteristics.

[0111] Preferably, the second acquisition submodule includes:

[0112] A labeling unit, used for increasing data diversity by using data enhancement technology, and performing fine labeling on the small data sample set;

[0113] The adjustment unit is used to initialize the small data model with the pre-trained weights of the big data model, freeze the bottom layer of the model at the beginning of training, and only train the top layer classifier. After the model converges, more layers are gradually unfrozen and the model is continuously fine-tuned to adapt to the data in the audit domain. Domain adaptation techniques are used to reduce the differences between the big data model and the small data model.

[0114] The training unit is used to use the big data model to predict the samples in the small data sample set, generate soft labels, construct a loss function, and train the small data model using the small data sample set and its corresponding hard labels and soft labels.

[0115] Preferably, in the extraction module 12, the feature extraction method includes:

[0116] Transforming the original features into new features with a smaller number than the original features by mapping or transforming methods;

[0117] and / or, selecting some of the most representative features from the original features;

[0118] and / or, selecting the most influential features based on the expert’s knowledge;

[0119] And / or, select based on mathematical methods to find the features that are most informative for classification.

[0120] Preferably, the sending module 14 includes:

[0121] The third acquisition submodule is used to obtain and organize historical audit data, analyze historical audit cases, and extract key audit points, common violation types and keywords;

[0122] A design submodule, for designing a review rule template based on the extracted keywords and legal clauses, wherein the template includes the focus of the review, relevant legal basis and keyword matching logic;

[0123] The optimization submodule is used to generate intelligent rules, using natural language processing technology to learn keywords and phrases from historical operations to automatically generate or optimize audit rules.

[0124] Preferably, the device comprises:

[0125] The identification module 15 is used to establish a user operation habit model based on historical operation data and identify the audit areas in which each auditor is good at and efficient;

[0126] The push module 16 is used to combine the characteristics of the file content with the user operation habit model, intelligently allocate the review task, give priority to the reviewers who are good at this field, and mark their historical review preferences.

[0127] Preferably, the device comprises:

[0128] The review module 17 is used to receive the manual reviewer's operations on the review results, and to perform feedback operations on the review results based on the above operations. The feedback operations include confirmation, cancellation of confirmation, ignore, and cancellation of ignore. The training parameters of the hybrid model are adjusted according to the feedback operation data, and the feedback operation data is used for incremental learning.

[0129] The present invention also provides an electronic device, Figure 4 The schematic diagram of the structure of the electronic device of the present invention can realize the present invention Figure 1 The process shown is as follows Figure 4 As shown, the electronic device may include: a housing 41, a processor 42, a memory 43, a circuit board 44 and a power supply circuit 45, wherein the circuit board 44 is arranged inside the space enclosed by the housing 41, and the processor 42 and the memory 43 are arranged on the circuit board 44; the power supply circuit 45 is used to supply power to various circuits or devices of the electronic device; the memory 43 is used to store executable program codes; the processor 42 runs the program corresponding to the executable program code by reading the executable program code stored in the memory 43, so as to execute the method described in any of the aforementioned method embodiments.

[0130] For details on the specific execution process of the above steps by the processor 42 and the steps further executed by the processor 42 by running the executable program code, please refer to the present invention. Figure 1 The description of the illustrated embodiment will not be repeated here.

[0131] This electronic device exists in many forms, including but not limited to:

[0132] (1) Mobile communication devices: These devices are characterized by their mobile communication functions and their main purpose is to provide voice and data communications. These terminals include: smart phones (such as iPhone), multimedia phones, functional phones, and low-end phones.

[0133] (2) Ultra-mobile personal computer devices: These devices fall into the category of personal computers, have computing and processing functions, and generally also have mobile Internet access features. These terminals include: PDA, MID and UMPC devices, such as iPad.

[0134] (3) Portable entertainment devices: These devices can display and play multimedia content. They include audio and video players (such as iPods), handheld game consoles, e-books, smart toys, and portable car navigation devices.

[0135] (4) Server: A device that provides computing services. The server consists of a processor, hard disk, memory, system bus, etc. The server is similar to a general computer architecture, but because it needs to provide highly reliable services, it has higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0136] (5) Other electronic devices with data interaction functions.

[0137] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any of the above method embodiments are implemented.

[0138] The present invention also provides an application program, which is executed to implement the method provided by any method embodiment of the present invention.

[0139] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0140] Each embodiment in this specification is described in a related manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. For the convenience of description, the description of the above device is divided into various units / modules according to their functions and described separately. Of course, when implementing the present invention, the functions of each unit / module can be implemented in the same or one or more software and / or hardware.

[0141] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An intelligent text review method integrating big data and small data models, characterized in that: include: Acquire text data in multiple data modes, wherein the text data includes public big data resources and small data samples in the audit field; Extracting features from the text data to obtain representative feature data; Input representative feature data into a pre-trained hybrid model, extract commonly concerned keywords and generate corresponding content summaries. The hybrid model training method specifically includes selecting a deep learning architecture suitable for the problem type, training a big data model, and optimizing the model's hyperparameters; training a small data model to obtain prediction results; fusing the prediction results of the big data model and the small data model, fine-tuning the fused model using a small data sample set, and further optimizing the model according to domain characteristics; According to the keywords and content summary, the intelligent review rule engine dynamically adjusts and matches the rules, extracts the content paragraphs that need to be reviewed and sends them to the review interface. The creation method of the intelligent review rule engine is specifically to analyze historical review cases, extract key review points, common violation types and keywords, and design review rule templates based on the extracted keywords and legal clauses to generate intelligent rules; The right side of the review interface integrates a review information panel, wherein the review information includes review points, review content, and the reviewer's opinions or decisions, and the review points are pre-integrated with links for users to click and jump to the corresponding document location.

2. The method according to claim 1, characterized in that Select a deep learning architecture suitable for the problem type, train a big data model, optimize the model's hyperparameters, train a small data model, and obtain prediction results, including: Using data enhancement technology to increase data diversity and finely annotate the small data sample set; Initialize the small data model using the pre-trained weights of the big data model, freeze the bottom layer of the model at the beginning of training, and only train the top classifier. After the model converges, gradually unfreeze more layers and continue to fine-tune the model to adapt to the data in the audit domain. Use domain adaptation techniques to reduce the difference between the big data model and the small data model. Use the big data model to predict the samples in the small data sample set, generate soft labels, construct a loss function, and use the small data sample set and its corresponding hard labels and soft labels to train the small data model.

3. The method according to claim 1, characterized in that The feature extraction method comprises: The original features are transformed into new features which are less in number than the original features by mapping or transformation.

4. The method according to claim 1, characterized in that: The method for creating the intelligent audit rule engine includes: Obtain and organize historical audit data, analyze historical audit cases, and extract key audit points, common violation types, and keywords; Based on the extracted keywords and legal clauses, design an audit rule template, which includes the focus of the audit, relevant legal basis, and keyword matching logic; Generate intelligent rules and use natural language processing technology to learn keywords and phrases from historical operations to automatically generate or optimize audit rules.

5. The method according to any one of claims 1 to 4, characterized in that: The method comprises: Based on historical operation data, analyze the auditor's operation records and automatically identify the audit areas where each auditor is good at and efficient; Analyze the content characteristics of the files to be reviewed, intelligently assign review tasks based on the file content characteristics, give priority to reviewers who are good at this field, and mark their historical review preferences.

6. The method according to claim 5, characterized in that The method comprises: Receive the manual reviewer's operations on the audit results, and perform feedback operations on the audit results based on the above operations. The feedback operations include confirmation, cancellation of confirmation, ignore, and cancellation of ignore. Adjust the training parameters of the hybrid model according to the feedback operation data, and use the feedback operation data for incremental learning.

7. An intelligent text review device integrating big data and small data models, characterized in that: include: An acquisition module, used to acquire text data in multiple data modes, wherein the text data includes public big data resources and small data samples in the audit field; An extraction module, used for performing feature extraction on the text data to obtain representative feature data; A generation module is used to input representative feature data into a pre-trained hybrid model, extract commonly concerned keywords and generate corresponding content summaries. The training method of the hybrid model is specifically to select a deep learning architecture suitable for the problem type, train a big data model, and optimize the hyperparameters of the model; train a small data model to obtain prediction results; fuse the prediction results of the big data model and the small data model, use the small data sample set to fine-tune the fused model, and further optimize the model according to the characteristics of the domain; A sending module is used for dynamically adjusting and matching rules of the intelligent review rule engine according to the keywords and content summary, extracting content paragraphs that need to be reviewed and sending them to the review interface. The creation method of the intelligent review rule engine is specifically to analyze historical review cases, extract key review points, common violation types and keywords, and design a review rule template based on the extracted keywords and legal clauses to generate intelligent rules; The right side of the review interface integrates a review information panel, wherein the review information includes review points, review content, and the reviewer's opinions or decisions, and the review points are pre-integrated with links for users to click and jump to the corresponding document location.

8. An electronic device, characterized in that: The electronic device comprises: a housing, a processor, a memory, a circuit board and a power supply circuit, wherein the circuit board is placed inside the space enclosed by the housing, and the processor and the memory are arranged on the circuit board; the power supply circuit is used to supply power to various circuits or devices of the above-mentioned electronic device; the memory is used to store executable program codes; the processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, so as to execute any of the methods described in claims 1-6 above.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement any of the methods described in claims 1-6.

Images