A document review contrast method and system pretrained through document understanding

Abstract

The application discloses a document review comparison method and system pre-trained through document understanding, relates to the technical field of text comparison, and comprises the following steps: identifying a document to extract technical project technical content data, and pre-processing the obtained data to extract features; calculating entity relationship strength in the extracted feature vector, and constructing a knowledge graph model; performing deep semantic analysis on the collected document by using natural language processing and deep learning technology, and performing multi-document correlation evaluation through the knowledge graph. The document review comparison method pre-trained through document understanding can more deeply understand document content by constructing a knowledge graph, can reveal hidden correlations between entities, can provide more accurate analysis in multi-document comparison and review, and can complete visual operation supported by the knowledge graph, so that manual analysis can be supplemented by an analyst. The application achieves better accuracy and deep-level analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of text comparison, in particular to a document review comparison method and system pre-trained through document understanding. BACKGROUND

[0002] In the modern information technology era, document understanding and review have become increasingly important in many fields, especially in the extraction and analysis of technical content data in technology projects. In recent years, the rapid development of natural language processing (NLP) and deep learning technology has greatly promoted the progress of document understanding technology. These technologies can extract valuable information from vast amounts of text data and conduct in-depth semantic analysis.

[0003] The core of document understanding is to understand and interpret the content of the text, which involves extracting entities from the text, understanding the relationship between entities, and building a knowledge graph. As an effective data organization method, the knowledge graph has been widely used in multi-document correlation analysis. With the continuous optimization of machine learning models and the progress of algorithms, especially the emergence of pre-training models such as BERT and GPT, document understanding technology has been able to more accurately capture and analyze the deep semantic information of documents.

[0004] At the same time, with the development of big data and cloud computing, the demand for automated review and comparison of a large number of documents is growing. This is mainly reflected in the fields of technology project management, legal review, academic research, etc., where key information needs to be quickly extracted from a large number of documents and effectively compared and analyzed.

[0005] Although existing technologies have made significant progress in document understanding and review, there are still some limitations and challenges. First, existing document understanding technologies mainly rely on surface-level text analysis, which is difficult to deeply mine the implicit relationships and deep semantics in the document. Especially when dealing with complex technical documents, this surface-level analysis often cannot accurately reflect the core content and technical features of the document.

[0006] Second, although the knowledge graph provides an effective tool for correlation analysis between documents, constructing a high-quality and high-precision knowledge graph still faces many challenges in practical applications. For example, the construction of a knowledge graph often requires a lot of manual intervention, which is particularly difficult when dealing with large-scale document sets. In addition, existing technologies often lack sufficient depth and flexibility in assessing the correlation between multiple documents, making it difficult to adapt to complex and varied document content and format.

[0007] Finally, although deep learning models show great potential in the field of document understanding, the training and application of these models still require a large amount of data support and computing resources. In addition, the explainability and adaptability of deep learning models are also one of the main challenges currently facing.

[0008] In summary, although the prior art has made significant achievements in document understanding and review, there is still room for improvement in deep semantic understanding, efficient construction and application of knowledge graph in processing complex documents, and improving the explainability and adaptability of deep learning models. SUMMARY

[0009] In view of the above problems, the present application is proposed.

[0010] Therefore, the technical problem solved by the present application is that the existing document comparison and review method has low automation degree, large resource demand, and how to analyze deeply when processing complex technical documents to reflect the core content and technical features of the document.

[0011] To solve the above technical problems, the present application provides the following technical scheme: a document review comparison method pre-trained by document understanding, including identifying documents to extract technical content data of scientific and technological projects, and pre-processing the extracted data to extract features; calculating the entity relationship strength in the extracted feature vector to construct a knowledge graph model; using natural language processing and deep learning technology to perform deep semantic analysis on the collected documents, and performing multi-document correlation evaluation through the knowledge graph.

[0012] As a preferred scheme of the document review comparison method pre-trained by document understanding, the document information is identified to extract the text, metadata and auxiliary data of the technical content of the scientific and technological projects; the text includes research reports, paper descriptions, paper abstracts, problem-solving of thesis reports and high-frequency words; the metadata includes author information, publication date and citation data; the auxiliary data includes charts, formulas and annotations.

[0013] As a preferred scheme of the document review comparison method pre-trained by document understanding, the pre-processing of the extracted data to extract features includes constructing a document set D = {d1, d2, …, d n}, where each document d i contains document collection data, a document-term frequency matrix A is constructed, and the elements a ij in the matrix represent the frequency of word j in document d i , which is represented as:

[0014]

[0015] where f ij represents the number of occurrences of word j in document d i , and f ik represents the frequency of high-frequency words in document di The number of occurrences in the set of documents D, the TF-IDF transformation is applied to the matrix A, denoted as:

[0016]

[0017] Where b ij denotes the element in the transformed matrix B, |D| denotes the size of the document set; for each document d i The feature vector is extracted denoted as:

[0018]

[0019] Where, denotes the pre-trained word embedding vector of word j.

[0020] As a preferred scheme of the document review comparison method pre-trained by document understanding, wherein: the entity relationship strength calculation in the extracted entity in the document comprises identifying the entity in the feature vector and mapping to Where m is the embedding dimension, for entities (e1, e2), the entity relationship strength is calculated denoted as:

[0021]

[0022] Where E(e) denotes the high-dimensional space embedding of entity e, e' denotes the standard entity, U e denotes the entity set.

[0023] As a preferred scheme of the document review comparison method pre-trained by document understanding, wherein: the construction of the knowledge graph model comprises constructing a knowledge network of interconnected technology content data nodes of the scientific and technological project, denoted as:

[0024]

[0025]

[0026] Where E(d i ) denotes the entity set extracted from a single document d i , R(e, d i ) denotes the relationship set of entity e in document d i , Ω denotes the definition domain of the mapping, and γ denotes the attenuation coefficient.

[0027] As a preferred scheme of the document review comparison method pre-trained by document understanding, wherein: the deep semantic analysis of the collected documents by natural language processing and deep learning technology comprises receiving the document d iAs input, and output the semantic representation S of the document i , is expressed as:

[0028]

[0029] Wherein, alpha j Indicates the activation function weighting parameter, W jl Indicates the weight matrix b j Indicates the bias term, L indicates the feature dimension of d i ; Construct a double document similarity identification model, expressed as:

[0030]

[0031] Wherein, w k Indicates the weight coefficient, S ik Indicates the semantic representation of document d i In k dimension, S jk Indicates the semantic representation of document d j In k dimension, K indicates the number of dimensions involved in the document when calculating similarity.

[0032] As a preferred scheme of the document review comparison method pre-trained by document understanding according to the application, wherein: the multi-document correlation evaluation through the knowledge graph comprises multi-document similarity identification combining the double document similarity identification model and the knowledge graph, expressed as:

[0033]

[0034] Wherein, n indicates the total number of documents, omega ij Indicates the weight coefficient of document d i And d j , Phi (G, d i , d j ) indicates the relationship strength between entities in document d i And d j And knowledge graph.

[0035] Another object of the application is to provide a document review comparison system pre-trained by document understanding, which can understand the document content more deeply by constructing a knowledge graph, reveal the hidden association between entities, and solve the problem that the current technology document comparison review method can only identify and compare on the surface and cannot understand the paper technology deeply.

[0036] As a preferred scheme of the document review comparison system pre-trained by document understanding provided by the application, wherein: a data processing module, a knowledge graph module, and a correlation analysis module are included; the data processing module is used for identifying documents to extract technology project technology content data, and pre-processing and extracting features of the obtained data; the knowledge graph module is used for extracting entities in the feature vector to calculate entity relationship strength and construct a knowledge graph model; and the correlation analysis module is used for performing deep semantic analysis on collected documents by using natural language processing and deep learning technology, and performing multi-document correlation evaluation by using the knowledge graph.

[0037] A computer device includes a memory and a processor, the memory stores a computer program, and the processor executes the computer program to implement the steps of the document review comparison method pre-trained by document understanding.

[0038] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the document review comparison method pre-trained by document understanding.

[0039] The application has the beneficial effects that: the document review comparison method pre-trained by document understanding provided by the application can more deeply understand document content by constructing a knowledge graph, reveal hidden correlations between entities, thereby providing more accurate analysis in multi-document comparison and review, and complete knowledge graph support visualization operation, which can be supplemented by manual analysis of an analyst. Deep learning model is used to process document data, extract deep semantic features, and evaluate the correlation between documents in combination with the knowledge graph, so that the subtle differences between documents can be more accurately identified and compared, and more efficient and accurate document review is achieved. The application achieves better effects in accuracy and deep analysis. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0041] Figure 1 A flowchart of a document review comparison method pre-trained by document understanding provided by the first embodiment of the application.

[0042] Figure 2 A flowchart of a document review comparison system pre-trained by document understanding provided by the third embodiment of the application. DETAILED DESCRIPTION

[0043] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0044] Example 1

[0045] Reference Figure 1 As an embodiment of the present invention, a document review and comparison method based on document understanding pre-training is provided, comprising:

[0046] S1: Identify documents to extract technical content data of science and technology projects, and preprocess the acquired data to extract features.

[0047] Furthermore, document recognition for extracting technical content data from scientific and technological projects includes identifying document information and extracting text, metadata, and auxiliary data of the technical content of scientific and technological projects; the text includes research reports, paper descriptions, paper abstracts, project proposals, problem-solving information, and high-frequency words; the metadata includes author information, publication date, and citation data; and the auxiliary data includes charts, formulas, and annotations.

[0048] It should be noted that preprocessing the acquired data to extract features includes constructing a document set D = {d1, d2, ..., d...} n}, where each document d i Using data collected from documents, construct a document-term frequency matrix A, where each element a... ij Indicates word j in document d i The frequency in is represented as:

[0049]

[0050] Among them, f ij Indicates word j in document d i The number of times it appears in f ik This indicates high-frequency words in document d. i The frequency of occurrence of each word in the document quantifies its importance and provides a foundation for subsequent TF-IDF transformation and feature vector extraction.

[0051] Applying the TF-IDF transformation to matrix A, it can be represented as:

[0052]

[0053] Among them, b ij Let |D| represent the elements in the transformed matrix B, and |D| represent the size of the document set; for each document di extracting feature vectors is represented as:

[0054]

[0055] wherein, represents the pre-trained word embedding vector of word j.

[0056] S2: Extracting entities in the feature vectors, calculating the strength of entity relationships, and constructing a knowledge graph model.

[0057] Further, extracting entities in the document to calculate the strength of entity relationships includes identifying entities in the feature vectors and mapping them to wherein m is the embedding dimension, and for entities (e1, e2), the strength of the entity relationship is calculated as is represented as:

[0058]

[0059] wherein E(e) represents the high-dimensional space embedding of entity e, e' represents the standard entity, U e represents the set of entities. The entity recognition and encoding step increases the depth of analysis, enabling the system to understand and process more complex structures in the document, such as entities and their attributes and relationships

[0060] It should be noted that constructing a knowledge graph model includes constructing a knowledge network of interconnected data nodes of the technical content of the scientific and technological project, represented as:

[0061]

[0062]

[0063] wherein E(d i ) represents the set of entities extracted from a single document d i , R(e, d i ) represents the set of relationships of entity e in document d i , Ω represents the definition domain of the mapping, and γ represents the decay coefficient. The calculation of the strength of the relationship provides a quantitative method for understanding the interaction and association between entities in the document, which is particularly useful when constructing a knowledge graph and conducting deep semantic analysis. And the knowledge graph can reflect the strength of the relationship between entities.

[0064] S3: Using natural language processing and deep learning techniques to conduct deep semantic analysis on the collected documents, and performing multi-document correlation evaluation through the knowledge graph.

[0065] Further, using natural language processing and deep learning techniques to conduct deep semantic analysis on the collected documents includes receiving the document d iAs input, and output the semantic representation S of the document i , is expressed as:

[0066]

[0067] Where, alpha j represents the activation function weighting parameter, W jl represents the weight matrix b j represents the bias term, L represents the feature dimension of d i ; Construct a double document similarity identification model, expressed as:

[0068]

[0069] Where, w k represents the weight coefficient, S ik represents the semantic representation of document d i in k dimensions, S jk represents the semantic representation of document d j in k dimensions, and K represents the number of dimensions involved in the document when calculating similarity.

[0070] It should be noted that the multi-document correlation evaluation by the knowledge graph includes the multi-document similarity identification combining the double document similarity identification model and the knowledge graph, expressed as:

[0071]

[0072] Where, n represents the total number of documents, omega ij represents the weight coefficient of document d i and d j , and Phi(G, d i , d j ) represents the relationship strength between entities in document d i and d j and the knowledge graph.

[0073] Embodiment 2

[0074] An embodiment of the present application provides a document review comparison method pre-trained by document understanding. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation experiments are used for scientific demonstration.

[0075] First, the test preparation includes collecting various types of scientific project documents, including research reports, academic papers, etc. These documents are used to test the efficiency and accuracy of the invention method and the traditional document review method.

[0076] Using natural language processing tools, key information is extracted from the document, including text content, author information, publication date, etc.

[0077] The feature vectors are extracted using deep learning techniques, and a knowledge graph is constructed based on these vectors to capture complex relationships between documents.

[0078] The documents are analyzed in depth by a pre-trained deep learning model to extract their semantic representations.

[0079] Both double-document and multi-document similarity recognition are performed.

[0080] Table 1: Document Evaluation Data Comparison

[0081]

[0082] Referring to Table 1, the 95% of the present invention compared to the 80% of the traditional method shows a significant advantage in identifying document information and extracting technology project technical content. This improvement means a deeper understanding of the text and more accurate information extraction, especially when dealing with complex technical documents. The 92% accuracy of the present invention compared to the 75% of the traditional method shows higher analysis accuracy. This highlights the advantages of the present invention in deep semantic analysis and multi-document relevance evaluation, especially through deep learning models and natural language processing techniques to achieve in-depth document understanding.

[0083] Embodiment 3

[0084] Referring to Figure 2 For an embodiment of the present invention, a document review comparison system pre-trained through document understanding is provided, including a data processing module, a knowledge graph module, and a relevance analysis module.

[0085] The data processing module is used to identify documents and extract technology project technical content data, and to preprocess and extract features from the obtained data. The knowledge graph module is used to extract entities in the feature vectors and calculate entity relationship strength to construct a knowledge graph model. The relevance analysis module is used to use natural language processing and deep learning techniques to perform deep semantic analysis on collected documents and perform multi-document relevance evaluation through the knowledge graph.

[0086] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0088] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0089] It should be understood that portions of the present application can be implemented with hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented with software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, implementation can be with any or a combination of the following technologies, which are all well known in the art: a discrete logic circuit having logic gates for implementing logic functions upon an application of data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be understood that the foregoing embodiments are merely illustrative of the present application and are not to be used to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it will be apparent to those skilled in the art that various changes and modifications can be contributed to the present application without departing from the spirit and scope of the present application, and such changes and modifications should be encompassed within the scope of the appended claims.

[0090] It should be understood that the foregoing embodiments are merely illustrative of the present application and are not to be used to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it will be apparent to those skilled in the art that various changes and modifications can be contributed to the present application without departing from the spirit and scope of the present application, and such changes and modifications should be encompassed within the scope of the appended claims.

Claims

1. A document review contrast method pre-trained by document understanding, characterized in that, The method comprises the following steps: identifying documents to extract technical content data of a scientific project, pre-processing the obtained data to extract features; calculating entity relationship strength of entities in the feature vector to construct a knowledge graph model; performing deep semantic analysis on the collected documents by using natural language processing and deep learning technology, and performing multi-document correlation evaluation by using the knowledge graph; the knowledge graph model comprises a knowledge network of interconnected technical content data nodes of the scientific project, and is represented as: where E(d i ) represents the entity set extracted from a single document d i , R(e, d i ) represents the relationship set of entity e in document d i , Ω represents the definition domain of the mapping, and γ represents the attenuation coefficient; The deep semantic analysis of the collected documents using natural language processing and deep learning techniques comprises receiving the documents d i as input, and outputting a semantic representation S i of the documents, represented as: wherein, a j represents an activation function weighting parameter, W jl represents a weight matrix b j represents a bias term, L represents a d i dimension of the feature a double-document similarity identification model is constructed, and is represented as: where w k denotes the weight coefficient, S ik denotes the semantic representation of document d i in k dimensions, S jk denotes the semantic representation of document d j in k dimensions, K denotes the number of dimensions involved in the similarity computation for document the multi-document correlation evaluation by using the knowledge graph comprises multi-document similarity identification combined with the double-document similarity identification model and the knowledge graph, and is represented as: where n represents the total number of documents, ω ij represents the weight coefficient of the document d i and d j , and Φ(G, d i , d j ) represents the relationship strength of the entity in the document d i and d j and the knowledge graph.

2. The document review contrast method pre-trained by document understanding of claim 1, wherein: the document identification and the technical content data extraction of the scientific project comprise identifying document information, extracting text, metadata and auxiliary data of the technical content of the scientific project; the text comprises research reports, paper descriptions, paper abstracts, problem-solving of theses and high-frequency words; the metadata comprises author information, publication date and citation data; the auxiliary data comprises charts, formulas and annotations. 3.The document review contrast method pre-trained by document understanding of claim 2, wherein: The pre-processing of the acquired data to extract features includes constructing a document set D = d1, d2,..., d n}, where each document d i contains data collected from a document, constructing a document-term frequency matrix A, the elements a ij in the matrix represent the frequency of term j in document d i , which is represented as: where f ij denotes the number of occurrences of the word j in the document d i , f ik denotes the number of occurrences of the high-frequency word in the document d i , and the TF-IDF transformation applied to the matrix A is denoted by where b ij denotes the elements in the transformed matrix B, |D| denotes the size of the document collection; For each document d i extracting a feature vector is represented as: wherein, denotes a pre-trained word embedding vector for the word j.

4. The document review contrast method pre-trained by document understanding of claim 3, wherein: computing entity relationship strength for the entities in the extracted feature vectors includes identifying entities in the feature vectors and mapping to where m is the embedding dimension, for entities (e1, e2), computing entity relationship strength is represented as: where E(e) represents the high-dimensional space embedding of entity e, e' represents a standard entity, U e represents a set of entities.

5. A system employing the document review contrast method pre-trained by document understanding according to any one of claims 1-4, characterized in that: The method comprises a data processing module, a knowledge graph module and a correlation analysis module. The data processing module is used for identifying documents to extract technical content data of a scientific project, and pre-processing the obtained data to extract features. The knowledge graph module is used for calculating entity relationship strength of entities in the feature vector to construct a knowledge graph model. The correlation analysis module is used for performing deep semantic analysis on the collected documents by using natural language processing and deep learning technology, and performing multi-document correlation evaluation by using the knowledge graph. 6.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-4 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the document review comparison method pre-trained by document understanding according to any one of claims 1 to 4.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the document review comparison method pre-trained by document understanding according to any one of claims 1 to 4.

Images