A document information extraction and processing method and device based on power word structure

Abstract

The present invention discloses a document information extraction and processing method and device based on electric power word structure, which relates to the technical field of document information processing. The method includes: for target electric power documents, identifying extraction tasks and redefining task paradigms, determining word-unit nodes and constructing word-unit directed graphs; constructing a large language question-answering model based on the disambiguation principle, traversing word-unit directed graphs, and determining word-unit relationships; building an information extraction engine with a document layout encoder; identifying word-unit directed graphs and word-unit relationships, intervening in information extraction engines to perform engine parallel retrieval and information extraction on target electric power documents, and determining a multi-dimensional electric power information system; and complementarily fusing multi-dimensional electric power information systems as information extraction results for target electric power documents. The method solves the technical problems of insufficient accuracy and low completeness of existing electric power document information retrieval, and achieves the effect of improving the accuracy and completeness of electric power document information extraction through multi-dimensional parallel retrieval and information fusion based on task paradigms.

Description

Technical Field

[0001] The present invention belongs to the technical field of document information processing, and in particular relates to a method and device for extracting and processing document information based on an electric power word unit structure. Background Art

[0002] In the power industry, document information extraction and processing technology faces growing demands and challenges. With the rapid development of informatization and big data technology, power companies have accumulated a large number of different types of document data, including power operation rules, operation and maintenance records, dispatch records, electricity consumption and electricity fee marketing data, customer service data, and fault repair records. These documents carry key information elements of the power system, such as time, place, people, events, results, and implementation methods. They are rich in content and huge in information. However, document information in the power industry often has problems of low standardization and inconsistent formats. Most documents lack effective structured information, which makes it difficult to directly apply these data to scientific decision-making, intelligent operation and maintenance, or refined services.

[0003] At present, the existing technology mainly uses the method of term retrieval, builds a bag-of-words library related to electricity, and configures the extraction logic in combination with the algorithm model to perform automated retrieval and extraction. And related technological innovations are mainly based on this method, cutting in from aspects such as extraction logic optimization. However, due to the richness of terms in the power field, the heterogeneity of data, and the existence of term ambiguity, the logical complexity of the extraction process is too high, and it is difficult to ensure the completeness of the extraction and the degree of fit with the extraction requirements, resulting in low retrieval efficiency and insufficient precision. Therefore, it is of great practical significance to develop an efficient document information extraction technology that can extract key, multi-dimensional information from massive power documents to improve the information management level and intelligent decision-making capabilities of the power industry. Summary of the invention

[0004] In order to solve the technical problems of insufficient accuracy and low completeness of existing power document information retrieval, the purpose of the present invention is to provide a document information extraction and processing method and device based on power word structure, which improves the accuracy and completeness of power document information extraction through multi-parallel retrieval and information fusion based on task paradigm.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A document information extraction and processing method based on power word structure, characterized in that the method comprises the following steps:

[0007] 1) For the target power document, identify the extraction task and redefine the task paradigm, determine the word-element nodes and construct a word-element directed graph, and the power direction of the task paradigm is different;

[0008] 2) constructing a large language question-answering model based on the disambiguation principle, traversing the word-unit directed graph, and determining word-unit relationships, wherein the large language question-answering model is a bidirectional logic;

[0009] 3) Building an information extraction engine with the document layout encoder, wherein the information extraction engine is a parallel engine matrix;

[0010] 4) Identify the word-unit directed graph and the word-unit relationship, intervene in the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determine a multi-element power information system, wherein the multi-element power information system is related to the word-unit directed graph. Figure 1 One to one correspondence;

[0011] 5) Complementarily integrating the multiple power information systems as the information extraction result of the target power document.

[0012] Furthermore, in step 2), the large language question answering model is constructed based on the disambiguation principle, including:

[0013] Based on the question and answer samples, perform sample supervised pre-training and determine the pre-training model;

[0014] Determine downstream task categories for documents in the power field;

[0015] Introduce the disambiguation principle and construct a word corpus;

[0016] The pre-trained model is fine-tuned in combination with the word-meta corpus and the downstream task class to determine the large language question-answering model.

[0017] The determining of word-unit relationships includes:

[0018] Traversing the word-unit directed graph, performing forward question-answering logic recursion, and determining the first word-unit relationship;

[0019] Perform reverse question-answer logic verification to determine the second word structure;

[0020] Based on the second word-gram structure, verify the first word-gram relationship, and if the verification is consistent, determine the word-gram relationship;

[0021] If the verification does not match, the ambiguous point is located to compensate for the first word-gram relationship, the word-gram relationship is determined, and the ambiguous point is added to the word-gram corpus, wherein the word-gram corpus is regularly updated with a model.

[0022] In step 3), the construction of the information extraction engine includes:

[0023] Introduce initialization encoder;

[0024] Performing multi-bit twinning on the initialized encoder, fine-tuning the encoder group based on the task paradigm class, and determining the encoder array;

[0025] The information extraction engine is constructed based on the encoder array in a parallel relationship.

[0026] In step 4), the engine parallel retrieval and information extraction include:

[0027] Traversing the word-element directed graph, identifying entity word-element nodes, scanning the target power document, and determining a binary network label based on the entity;

[0028] Determining a search term path based on the binary network label;

[0029] Setting a word-position joint code, wherein the joint code is used to perform a chaotic word-unit sequence constraint;

[0030] The target power document is retrieved and extracted by combining the search word path with the joint code.

[0031] After determining the search term path, include:

[0032] Based on the search term path, entity nesting is identified and entity nesting features are determined;

[0033] With respect to the entity nesting feature, decoupling is performed based on the word-gram relationship to determine a decoupled word-gram path.

[0034] In step 5), the complementary integration of the multi-element power information system includes:

[0035] Determine the heterogeneous characteristics of information based on multi-modality and multi-granularity;

[0036] Based on the information heterogeneity characteristics, determining a bidirectional heterogeneous relationship, wherein the de-heterogeneity method is any characteristic within the information heterogeneity characteristics;

[0037] Based on the bidirectional heterogeneous relationship, the de-heterogeneity processing before the complementary fusion of the multi-element power information system and the display heterogeneity processing after the complementary fusion are performed.

[0038] A document information extraction and processing device based on electric power word unit structure, characterized by comprising:

[0039] Word-element directed graph construction module: for the target power document, identify the extraction task and redefine the task paradigm, determine the word-element nodes and construct the word-element directed graph, and the power direction of the task paradigm is different;

[0040] A word-unit relationship determination module: constructs a large language question-answering model based on the disambiguation principle, traverses the word-unit directed graph, and determines the word-unit relationship, wherein the large language question-answering model is a bidirectional logic;

[0041] Information extraction engine construction module: in combination with the document layout encoder, an information extraction engine is constructed, wherein the information extraction engine is a parallel engine matrix;

[0042] A multi-element power information system determination module: identifies the word-element directed graph and the word-element relationship, intervenes in the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determines a multi-element power information system, wherein the multi-element power information system and the word-element directed graph are related. Figure 1 One to one correspondence;

[0043] Information extraction result acquisition module: complementarily integrates the multiple power information systems as the information extraction result of the target power document.

[0044] The beneficial effects of the present invention are:

[0045] For the target power document, identify the extraction task and redefine the task paradigm, determine the word-unit node and construct a word-unit directed graph, the power direction of the task paradigm is different; build a large language question-answering model based on the disambiguation principle, traverse the word-unit directed graph, and determine the word-unit relationship, wherein the large language question-answering model is a bidirectional logic; with the document layout encoder, build an information extraction engine, wherein the information extraction engine is a parallel engine matrix; identify the word-unit directed graph and the word-unit relationship, intervene in the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determine a multi-element power information system, wherein the multi-element power information system and the word-unit directed graph Figure 1 One-to-one correspondence; complementary fusion of the multiple power information systems as the information extraction result of the target power document. This solves the technical problems of insufficient accuracy and low completeness of existing power document information retrieval, and achieves the effect of improving the accuracy and completeness of power document information extraction through multiple parallel retrieval and information fusion based on the task paradigm. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic flow chart of a method for extracting and processing document information based on an electric power word-element structure provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of the structure of a document information extraction and processing device based on an electric power word-gram structure provided in an embodiment of the present invention.

[0048] Explanation of the accompanying drawings: word-unit directed graph construction module 1, word-unit relationship determination module 2, information extraction engine construction module 3, multi-element power information system determination module 4, information extraction result acquisition module 5. DETAILED DESCRIPTION

[0049] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below.

[0050] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0051] A document information extraction and processing method based on power word structure, such as Figure 1 As shown, the method includes:

[0052] For the target power document, the extraction task is identified and the task paradigm is redefined, the word-unit nodes are determined and a word-unit directed graph is constructed, and the power direction of the task paradigm is different.

[0053] In an embodiment of the present invention, when processing a target power document, the system terminal first analyzes the content in the target power document and identifies the specific information extraction task to be performed. This step is not only to find keywords from the document, but also to have an in-depth understanding of the extraction task based on the specific needs of the power industry. In order to ensure that the extraction process can adapt to the needs of different power businesses, it is necessary to redefine the task paradigm. The task paradigm refers to the description method or task model of the task, which can reflect different power fields, such as equipment maintenance, fault analysis, scheduling management or electricity bill settlement. Each task paradigm is biased towards a specific power direction to ensure that the extracted content is closely related to the business needs. For example, when processing documents about equipment failures, the task paradigm may be biased towards the cause of the failure, the scope of impact and the repair method, while when processing scheduling records, it pays more attention to time, resource allocation and execution results. Subsequently, through semantic analysis, the word element nodes in the document are determined. The word element is the smallest semantic unit in the document, including nouns (such as power equipment, time, place), verbs (such as occurrence, repair) and attributives or other modifiers (such as serious). These word elements are usually associated with specific events or concepts in the document. Once the word-element nodes are determined, the system terminal will construct a word-element directed graph based on these word-element. This word-element directed graph is a data structure used to represent the relationship between word-element. Each word-element is used as a node in the graph, and the relationship between word-element (such as the action relationship between verbs and nouns, or the description relationship between modifiers and nouns) is used as a directed edge in the graph to show the logical or semantic relationship between word-element. For example, as a word-element node, power equipment will have a directed edge with the occurrence of a fault, indicating the action relationship between the two. The purpose of constructing this word-element directed graph is to help the system terminal extract information more accurately in subsequent steps by clearly showing the interaction between word-element. For example, when it is necessary to extract information about "equipment failure", the word-element related to the equipment failure, such as the type of failure, repair time, etc., can be quickly located through the directed graph. Through this step, a clear semantic and logical structure foundation can be provided for subsequent information extraction, ensuring that the system terminal can efficiently and accurately extract key information related to business needs, and improve the accuracy and adaptability of document information extraction.

[0054] A large language question-answering model is constructed based on the disambiguation principle, the word-unit directed graph is traversed, and word-unit relationships are determined, wherein the large language question-answering model is a bidirectional logic.

[0055] In one embodiment, in order to solve the problem of semantic ambiguity in documents. In documents of the power industry, some words (such as nouns or verbs) may have multiple meanings. For example, phase can refer to the phase angle of alternating current in the power system, or it can be used to indicate the working state of power equipment. In order to ensure the accuracy of the extracted information, these semantic ambiguities need to be eliminated. Therefore, based on the disambiguation principle, the system terminal performs sample supervision pre-training and fine-tuning through question and answer samples to build a large language question and answer model. The role of the large language question and answer model is to determine the specific meaning of each word in the document by asking questions and answering. During the model training process, the system terminal will use a large number of question and answer samples for pre-training so that it can recognize common power industry vocabulary and context. When the model is applied to the actual document, it can traverse the previously constructed word-unit directed graph and ask questions for each word-unit node, and determine the correct meaning of the word-unit and its relationship with other words by analyzing the contextual relationship and the association between words. In addition, the large language question and answer model adopts bidirectional logic. This means that the model will not only analyze the relationship between tokens from the forward direction (for example, from the beginning to the end of the sentence), but also check from the backward direction (for example, from the end of the sentence to the forward direction) to ensure logical consistency. This two-way logic helps to understand the context of tokens more comprehensively, reduce ambiguity or errors that may occur in one-way analysis, improve the accuracy of the information extraction process, and avoid misjudgments due to unclear semantics.

[0056] Furthermore, the present invention provides the large language question answering model constructed based on the disambiguation principle, including:

[0057] Based on question and answer samples, sample supervised pre-training is performed to determine the pre-training model; for documents in the power field, downstream task classes are determined; the disambiguation principle is introduced to construct a word meta-corpus; the pre-training model is fine-tuned in combination with the word meta-corpus and the downstream task class to determine the large language question and answer model.

[0058] Preferably, the system terminal collects a large number of training samples containing question-answer pairs, which are mainly used to solve various information extraction tasks in the power field. These samples not only cover common terms, concepts and events in the power industry, but also include standard answers to different questions. For example, for the question of what is the cause of the transformer failure, the annotated answer should include a specific description of the cause of the failure, such as overload operation causing winding overheating. After the data collection and annotation are completed, the system terminal preprocesses the collected data, including removing noise data in the document, such as irrelevant symbols, image annotations, etc., and word segmentation of the text in the document, and converting each word into a token for subsequent model training. Tokens can be representations at the word, phrase or even character level. The question-answer pair samples are then formatted into an input form acceptable to the model. Each question-answer pair is converted into two main parts, namely an input part and an output part, and the input part includes questions and context. Subsequently, all the converted input parts and output parts are divided to form a training set and a validation set, and then the training set is used for supervised learning. In supervised learning, the system terminal uses the training set to train the pre-trained model. The pre-trained model architecture usually adopts the Transformer model in deep learning or language models such as BERT (Bidirectional Encoder Representations from Transformers). Taking BERT as an example, the system terminal uses BERT to build a pre-trained model, including the input layer, BERT encoding layer, fully connected layer, and output layer, and then inputs the training set into the pre-trained model for forward propagation, and gradually performs semantic encoding through the BERT encoding layer. The pre-trained model applies the attention mechanism to each word unit, processes the bidirectional context information, and generates a high-dimensional vector representation. The output layer is responsible for predicting the starting and ending positions of the answer. During the forward propagation process, the system terminal passes the input data layer by layer, and calculates the prediction results of the starting and ending positions corresponding to each word unit in the power document through the BERT encoding layer, fully connected layer, and output layer. After that, the cross entropy loss function is used to calculate the loss value between the model's prediction result and the actual answer, and the difference between the predicted starting and ending positions and the actual positions in the input data is compared. Then, the gradient of the loss value to the weights of each layer of the model is calculated layer by layer through the back propagation algorithm, and the parameters in the BERT model are updated. During the back propagation process, the AdamW optimizer is used to optimize the model parameters and adjust the weights to minimize the value of the cross entropy loss function. This process is repeated many times. Through multiple rounds of iterative training, the model is gradually optimized to improve its performance in the power document information extraction task. The training process continues until the set maximum number of iterations is reached or the model loss value converges.After training, the model is tested using the validation set to evaluate the accuracy of the model in the question-answering task in the power document, including the model's precision, recall rate, and F1 value when extracting information. If the accuracy of the model on the validation set meets expectations, the currently trained pre-trained model is output. If the validation result is not ideal, hyperparameters such as learning rate and batch size are adjusted to further improve the performance of the model. In addition, the system terminal determines multiple downstream task classes based on specific power field needs. Downstream task classes refer to application scenarios that are further refined from the pre-trained model, such as fault analysis, equipment status monitoring, and scheduling optimization. These task classes correspond to different power business needs, and the model needs to be specifically adjusted for these different tasks to ensure that the model performs best in specific tasks. For example, for fault analysis tasks, the model may focus more on identifying equipment names, fault types, and causes, while in scheduling optimization tasks, the model needs to identify information such as time, location, and resource allocation. When processing power documents, the problem of polysemy of terms is often encountered. For example, words such as phase or capacity have multiple meanings. Therefore, the disambiguation principle is introduced to solve the semantic problems of these polysemous words. The system terminal will build a large word-meta corpus based on common word-meta in power documents. This corpus not only contains the terms in the power industry, but also records the specific meaning and usage of each word-meta in different contexts. By analyzing the corpus, the information ambiguity caused by word-meta polysemy in the document can be effectively solved. Then, the word-meta corpus is combined with the downstream task class to fine-tune the pre-trained model. During the fine-tuning process, the system terminal will input the training set into the pre-trained model again. The model will tokenize the input data and convert these word-meta into vector representation through the BERT encoding layer. In this process, the word-meta corpus will play a role. Proprietary words and terms in the power industry, such as phase and capacity, have been added to the word-meta corpus. The model uses these professional words and combines disambiguation information when generating word-meta vectors, so that the model can correctly understand and process polysemous words in power documents. For example, when the model encounters polysemous words such as phase, it will select the appropriate explanation based on the prompts of the corpus to ensure the accuracy of semantics. The pre-trained model will extract the relationship between the question and the context from the perspective of two-way semantic analysis based on the input context and question. In this process, downstream task classes play a key role. The model will process the input content in a targeted manner according to the needs of different downstream task classes (such as equipment failure analysis, operation and maintenance record extraction, etc.). For example, for equipment failure analysis tasks, the model will pay special attention to words related to equipment names and failure types, and associate this information with the semantics of the question, and then generate predictions of the starting and ending positions of the answer through the output layer of the pre-trained model.At this point, the downstream task class can also guide the output processing. For example, in the scheduling optimization task, the model will be more inclined to extract information related to time and place, while in the fault analysis task, the model will pay more attention to the extraction of equipment status and fault causes. Finally, the pre-trained model is fine-tuned through loss calculation and back propagation and optimization, and this process is repeated until the preset number of fine-tuning times. After fine-tuning, the pre-trained model can not only recognize and process complex power documents, but also effectively deal with terminology ambiguity, thereby finally determining a large language question-answering model suitable for power document information extraction. The entire process combines the disambiguation principle with the pre-training of question-answering samples and fine-tuning for tasks in the power field to ensure that the model can accurately understand and process polysemous words in power documents and effectively extract the required key information.

[0059] Furthermore, the present invention provides the determining word-unit relationship, comprising:

[0060] The word-gram directed graph is traversed, and forward question-answering logic recursion is performed to determine the first word-gram relationship; reverse question-answering logic verification is performed to determine the second word-gram structure; based on the second word-gram structure, the first word-gram relationship is verified, and if the verification is consistent, the word-gram relationship is determined; if the verification is inconsistent, the ambiguous point is located to compensate for the first word-gram relationship, the word-gram relationship is determined, and the ambiguous point is added to the word-gram corpus, wherein the word-gram corpus is regularly updated with a model.

[0061] Optionally, the system terminal starts traversing from the constructed word-element directed graph, starting from the head of the document, and analyzes each word-element one by one. The nodes in the word-element directed graph represent words, and the edges represent the semantic or logical relationship between them. Through the forward question-answering logic, recursive analysis is performed to gradually identify the relationship between word-element. For example, in an analysis scenario of power equipment failure, starting from the question, the logical relationship between word-element such as equipment name, fault phenomenon, and influencing factor is identified in turn, and a preliminary first word-element relationship structure is determined. After the forward analysis is completed, the system terminal verifies through reverse logic. The reverse verification process starts from the end of the word-element directed graph, reversely verifies the word-element relationship obtained by forward recursion, and ensures that the association between word-element conforms to the context logic. Specifically, the system terminal first determines the terminal word-element in the forward reasoning process. This terminal word-element is usually the last important information point in the target information. For example, in the document of power failure analysis, the terminal word-element may be the cause of the failure or the solution. This terminal word-element represents the final conclusion in the entire reasoning chain. Starting from the end word, the word directed graph is traversed in reverse, and the relationship between the words is gradually derived and verified. For example, in the power failure analysis scenario, the cause of the fault is traced back to the relevant equipment status and environmental factors, etc., to verify whether these preceding words can be logically and consistently led to the end word. In each step, the system terminal will determine whether there is a logical path, connect each preceding word to the end word, and verify each node in the forward reasoning chain and the relationship between them layer by layer. In the reverse reasoning process, the system terminal will pay special attention to those words that may cause ambiguity. For these words, the context information and word corpus will be combined to ensure that the correct word meaning is used in the reverse reasoning process. Through this disambiguation process, ambiguous problems that may not be identified in the forward reasoning can be further corrected. After completing the reverse reasoning and verification, the second word structure will be generated. This structure is a word relationship graph derived based on reverse reasoning, corresponding to the first word relationship generated in the forward reasoning process. The generation process of the second word structure ensures that the information derived from the document not only conforms to the forward logic, but is also fully verified in the reverse reasoning. The second word-gram structure generated by the reverse check is used to verify the first word-gram relationship. If the second word-gram structure is logically consistent with the first word-gram relationship, then the relationship between the word-grams can be determined to be correct, and the first word-gram relationship is output as the final word-gram relationship. This step is to ensure the consistency between the forward reasoning and the reverse check to prevent misunderstandings or deviations caused by one-way reasoning. If the reverse check finds that the first word-gram relationship is inconsistent with the second word-gram structure, the system terminal will analyze the differences and identify possible ambiguous points between the word-grams. An ambiguous point refers to a word-gram having different meanings in different contexts, or an unclear relationship between it and other word-grams. The system terminal will compensate for these ambiguous points through context analysis or by using previously accumulated disambiguation experience.For example, the specific meaning of a word is determined by the surrounding word information, and the relationship between word units is corrected. After processing the ambiguous point, the system terminal updates the first word unit relationship to ensure that it is consistent with the context. After the compensation is completed, the relationship between word units is finally determined. For each located ambiguous point, the system terminal will also record it in the word unit corpus for reference when processing similar problems later. The word unit corpus is continuously improved through these records to provide more accurate disambiguation information for future question-answering tasks. In addition, the word unit corpus will also be regularly updated. As the system terminal processes more and more documents and word data, the word unit meanings, relationships, and disambiguation information in the corpus will be regularly optimized and expanded. The model update process is based on the latest word unit relationships and processed ambiguous points, so that the system terminal can better understand the complex relationships of word units when processing future documents, and improve the accuracy and efficiency of information extraction.

[0062] An information extraction engine is constructed in conjunction with a document layout encoder, wherein the information extraction engine is a parallel engine matrix.

[0063] In one embodiment, the document layout encoder is a tool for understanding the structure of a document. Electric power documents usually include information in various forms such as text, tables, charts, and pictures. The document layout encoder can effectively parse the document according to the layout of the document (such as paragraph structure, title, header and footer, table layout, etc.). The document layout encoder not only recognizes the text content, but also parses the data in the table, the annotations in the picture, etc. Through this preprocessing, the system terminal can accurately standardize different types of information to ensure that subsequent information extraction is more targeted. For example, the document layout encoder can recognize that the title part is a device fault report, and the body part is the fault phenomenon and cause, and encode this information into a specific format. The system terminal introduces an initial document layout encoder and fine-tunes the initial document layout encoder through the task paradigm class to build a parallel processing information extraction engine matrix. Each information extraction engine in the information extraction engine matrix is ​​responsible for a specific task in the document, such as extracting the device name, fault cause, operating parameters, etc. Different engines run simultaneously, and multiple parts of the document can be processed in parallel instead of processing each part in sequence. This enables the system terminal to process a large number of electric power documents in a short time and quickly extract key information. The structure of the engine matrix indicates that the information extraction engines are organized and collaborated according to certain rules, and each engine can share intermediate processing results, thereby improving the overall processing effect. For example, after one engine extracts a transformer fault, it can pass the information to another engine, which further extracts specific details related to the transformer fault, such as coil damage or overload operation. By building a parallel engine matrix, the system terminal can efficiently and accurately extract the required information from the power documents, especially when processing complex power documents, it can quickly obtain key information to support subsequent decision-making and analysis.

[0064] Furthermore, the present invention provides the construction information extraction engine, comprising:

[0065] An initialization encoder is introduced; multi-bit twinning is performed on the initialization encoder, and the encoder group is fine-tuned based on the task paradigm class to determine the encoder array; and the information extraction engine is constructed based on the encoder array in a parallel relationship.

[0066] Preferably, the system terminal introduces multiple initialized document layout encoders as initialization encoders, which are used to perform initial processing on the input power documents and convert the original document data into a machine-processable form. The power document may contain multiple types of data (such as text, tables, pictures, etc.), and the initialization encoder needs to preprocess these data and uniformly encode different types of information. Subsequently, the initialization encoder is copied and set in parallel so that multiple identical initialization encoders can process different inputs at the same time. Each twin encoder can be assigned to different data types or tasks to ensure that the system terminal can process different information parts in the power document in parallel in multiple dimensions. These twin encoders share the same infrastructure and weights, but will focus on different input tasks. For example, one twin encoder can process power equipment information, another twin encoder can process fault data, and another can process operation time or parameter information in the document. Each twin initialization encoder needs to be fine-tuned according to a specific task. Task paradigm class is a rule or pattern set in the system terminal for different power business tasks (such as equipment fault analysis, scheduling optimization, etc.). The fine-tuning process is to train each twin encoder specifically according to these task paradigms to better adapt it to specific task requirements. In this process, the system terminal will select the appropriate task paradigm class according to the document type and business requirements. For example, if it is a fault analysis task, select the fault analysis task paradigm. Then, according to the requirements of the task paradigm class, select the data set related to the task and use these data sets for fine-tuning. For example, for the equipment fault analysis task, the training data set will contain key information such as the device name, fault phenomenon, and cause. The encoder group will be trained and parameter adjusted according to the task requirements to ensure that the encoder can extract key information related to the task from the document. Among them, the training and parameter adjustment process is carried out through steps such as forward propagation, loss calculation, back propagation, gradient calculation, and parameter update, and the execution method is the same as above. After fine-tuning, multiple encoders will be optimized for different task paradigm classes, and each encoder has the ability to process a certain type of information. At this time, the system terminal will combine these fine-tuned encoders into an encoder array. The encoder array is a structure composed of multiple fine-tuned encoders working in parallel. Each encoder is responsible for extracting different parts of information related to a specific task from the document. The parallel structure of the array enables the system to process multiple blocks of information in a document simultaneously, improving processing speed and accuracy.

[0067] Identify the word-unit directed graph and the word-unit relationship, intervene in the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determine a multi-element power information system, wherein the multi-element power information system is related to the word-unit directed graph. Figure 1 One to one correspondence.

[0068] In one embodiment, the system terminal identifies each word in the document and the semantic or logical relationship between them according to the word-unit directed graph constructed by the target power document, and gradually identifies the role of each word in the document and the connection between them according to these relationships. Once the word-unit directed graph and word-unit relationship are identified, the system terminal starts the information extraction engine and intervenes in the information extraction process. In this process, the information extraction engine determines the search word-unit path according to the relationship in the word-unit directed graph, and locates the key information in the document in combination with the text-position joint encoding, and extracts it one by one. Through the parallel processing of these engines, different information can be extracted from multiple parts of the document at the same time. When the information extraction engine completes the retrieval and extraction, the system terminal will classify and organize all the extracted power information to form a complete multi-element power information system. This information system is composed of multiple different power information modules, such as equipment information, operating status, fault phenomenon, fault cause, operation record, etc. The system terminal will ensure that the extracted multi-element power information system is consistent with the original word-unit directed graph. The information of each module is extracted based on the relationship between the word-unit and the word-unit directed graph in the document, ensuring the comprehensiveness and consistency of the information. The nodes and relationships in the word-element directed graph reflect the information structure in the document, and the multivariate power information system is a concrete expression of these words and their relationships. Therefore, the multivariate power information system corresponds one-to-one with the word-element directed graph, ensuring the accuracy and completeness of information extraction.

[0069] Furthermore, the present invention provides the engine parallel retrieval and information extraction, comprising:

[0070] The word-meta directed graph is traversed to identify entity word-meta nodes, the target power document is scanned, and a binary network label based on the entity is determined; based on the binary network label, a search word-meta path is determined; a text-position joint encoding is set, wherein the joint encoding is used to perform chaotic word-meta sequence constraints; and the target power document is retrieved and extracted by combining the search word-meta path with the joint encoding.

[0071] Preferably, the system terminal starts from the head of the document, traverses each node of the word-element directed graph in turn, and identifies the entity word-element node therein. The entity word-element node refers to the word-element in the document that actually represents a specific object or concept such as power equipment, parameters, time, etc. For example, transformers, voltages, time, etc. in the word-element directed graph are all entity word-element nodes. These nodes are usually the most important part of the information extraction process because they represent the core data in the power document. After identifying the entity word-element node, the information extraction engine is started to perform a comprehensive scan of the target power document, find all content matching the entity word-element, gradually determine the position of each entity word-element node in the document, and generate a binary network label. The binary network label is a binary representation used to indicate whether a certain entity word-element appears at a specific document position. Each entity word-element node will correspond to a binary label, 1 indicates that the entity word-element exists at the position, and 0 indicates that it does not exist. For example, if the transformer is mentioned three times in the document, the corresponding three positions will be marked as 1, and the other positions where the transformer is not mentioned will be marked as 0. Subsequently, based on the generated binary network labels, the system terminal analyzes all entity word-element nodes marked as 1 in the binary network labels to find the semantic associations between them. Based on these associations, the word-element retrieval path is determined so that the content related to these word-element can be extracted in sequence in the subsequent steps. The retrieval word-element path refers to the path for information extraction planned by the system terminal based on the distribution of word-element in the document. This path helps the system terminal find the association between different word-element nodes in the document to ensure that relevant information can be extracted coherently. For example, there is a direct association between transformer and running time in the document, which will connect these word-element to form a word-element retrieval path. Since the word-element order may be chaotic or irregular in the document, the system terminal generates a joint encoding for each word-element, and the encoding content contains the semantic information of the word-element itself and the specific position of the word-element in the document. The role of the joint encoding is to encode the word-element together with its position to ensure that the required information can be accurately extracted even if the word-element order in the document changes. This encoding method helps the system terminal filter out situations that do not conform to the word-element path or are in a chaotic order, and improves the accuracy of information extraction. After that, the determined search word path is combined with the joint coding to perform information extraction. The system terminal will search the entity word nodes in the document one by one along the search path, and use the joint coding to correct the order and position of the word. Through this process, all relevant information can be extracted from the document to form an information extraction result, ensuring that the extracted result is complete and accurate.

[0072] Furthermore, the present invention provides that after determining the search word path, the method includes:

[0073] Based on the search word-meta path, entity nesting is identified to determine entity nesting features; with respect to the entity nesting features, decoupling is performed based on the word-meta relationship to determine a decoupled word-meta path.

[0074] Optionally, after planning the search word path, the system terminal will scan the entity word node in the target document through the information extraction engine to identify the existence of entity nesting. Entity nesting refers to the complex nesting relationship between multiple entity words in a sentence or paragraph. For example, some long sentences may contain multiple device names, fault types or operation steps, which makes the logical relationship between word words complicated and easily causes recognition errors, such as device A fails and causes device B to overheat. Once the sentences or paragraphs with entity nesting are identified, the system terminal will extract the nesting features of these sentences. The nesting features include the complex structure, logical relationship and association pattern between word words in the sentence. These nesting features help the system terminal understand the logical relationship of word words more clearly and avoid information confusion or erroneous extraction. For long sentence segmentation, the long sentence is decomposed by punctuation, conjunctions, etc., and the entity word words in each clause are separately identified. This can avoid a complex long sentence from being misunderstood as a whole. For example, the failure of device A causing the overheating of device B will be divided into two parts: the failure of device A and the overheating of device B. For logical structure analysis, for sentences involving causal and conditional relationships, logical connectives (such as cause, if, etc.) will be identified, and the entities in the sentence will be decomposed accordingly. For example, in sentences such as "If the current is too large, device X may fail", the conditional association between current and device X will be identified. For the extracted entity nesting features, these nested relationships will be decoupled. The purpose of decoupling is to decompose complex word-element relationships so that the system terminal can process each entity word-element one by one to avoid extraction errors caused by nesting. Through decoupling, a separate word-element path can be regenerated for each nested entity to ensure the accuracy of extraction. For nested entities with causal relationships, the cause part will be extracted first, and then the result part. For example, transformer damage, which causes the cooling system to fail, will be decomposed into two word-element paths, one to extract the relationship between transformer and damage, and the other to extract the relationship between cooling system and failure. For nested entities with conditional relationships, the entity word-element of the conditional part will be extracted first, and then the entity of the main sentence part will be extracted. For example, if the current is too large, device X may fail and will be decomposed into two word-element paths, one to extract the current and the excessive condition, and the other to extract the entity relationship between device X and the failure. For multiple entities in a long sentence, they will be separated into their own independent word-element paths. For example, if devices A and B overheat, they will be decomposed into two paths, one to deal with device A and overheating, and the other to deal with device B and overheating. After the decoupling is completed, the system terminal will reallocate each decoupled entity word-element to a separate retrieval path. In this way, the system terminal can search along these decoupled word-element paths one by one during the information extraction process to avoid ambiguity or extraction errors caused by nesting.

[0075] The multiple power information systems are complementarily integrated as the information extraction result of the target power document.

[0076] In one embodiment, the system terminal determines the heterogeneous characteristics of the information based on multimodality and multi-granularity by analyzing different information modules in the extracted multi-element power information system. Each information module may come from different data forms or dimensions (such as text, table, image, etc.), so there is heterogeneity. Subsequently, according to the determined heterogeneous characteristics, the bidirectional heterogeneous relationship between the information modules is identified, and a suitable de-heterogeneous method is selected. These heterogeneous characteristics will help the system terminal to integrate information before fusion and eliminate the structural differences between the information. Afterwards, based on the bidirectional heterogeneous relationship, the multi-element power information system is de-heterogeneous to ensure that the information remains consistent during the fusion process, and then these de-heterogeneous information modules are complementary fused. Then, the complementary fused multi-element power information system is used as the information extraction result of the target power file. This result is obtained after the system terminal deeply processes the target power file, and contains all important power information. The fused extraction result can provide a reliable basis for subsequent decision-making, analysis or report generation.

[0077] Furthermore, the present invention provides the complementary integration of the multi-element power information system, including:

[0078] Based on multi-modality and multi-granularity, the information heterogeneous characteristics are determined; based on the information heterogeneous characteristics, a bidirectional heterogeneous relationship is determined, wherein the de-heterogeneous method is any characteristic within the information heterogeneous characteristics; based on the bidirectional heterogeneous relationship, the de-heterogeneous processing before the complementary fusion of the multi-element power information system and the display heterogeneous processing after the complementary fusion are performed.

[0079] Preferably, when fusing the multi-element power information system, the system terminal identifies and determines the heterogeneous characteristics of the information based on multi-modality (such as text, image, table, etc.) and multi-granularity (level of details of information) by analyzing different information sources and data structures. Due to the existence of these heterogeneous information, there are structural differences between the data, which increases the difficulty of fusion. For example, the text description of the equipment and the operation data table belong to different data modes and show heterogeneity. After determining the heterogeneous characteristics, the system terminal defines a bidirectional heterogeneous relationship based on these characteristics. Bidirectional heterogeneous relationship refers to integrating different information modes through unified dimensions so that heterogeneous data can be processed under the same framework. At the same time, this bidirectional heterogeneous relationship can reflect the connection between different information and ensure the logical consistency between heterogeneous information. In this process, any characteristic of the heterogeneous characteristics is selected as the entry point for de-heterogeneity to ensure that the subsequent fusion processing has a basis. The role of the unified dimension is to retain the unique attributes of various types of information while de-heterogeneity, laying the foundation for the subsequent diversified information output. After determining the bidirectional heterogeneous relationship, the system terminal performs de-heterogeneity processing on the multi-element power information system. The goal of de-heterogeneization is to integrate information of different modalities and granularities according to a unified structural framework and eliminate the structural differences between various types of information. For example, the structuring of text information and the parsing of tabular data need to be integrated so that they can be related to each other. After the fusion is completed, the system terminal performs display heterogeneity processing, which means that the fused information will be diversified and converted according to the previous heterogeneous characteristics to adapt it to different display needs. This step ensures that the final output is not only structurally consistent, but also retains the diversity and richness of heterogeneous information. For example, the text description of the power equipment can be displayed together with the operating data in the chart to enrich the expression of the report.

[0080] In the above, refer to Figure 1 A method for extracting document information based on an electric power word structure according to an embodiment of the present invention is described in detail. Figure 2 A document information extraction and processing device based on an electric power word-gram structure according to an embodiment of the present invention is described.

[0081] According to an embodiment of the present invention, a document information extraction and processing device based on an electric power word-element structure is used to solve the technical problems of insufficient precision and low completeness of existing electric power document information retrieval, and achieves the effect of improving the precision and completeness of electric power document information extraction through multi-parallel retrieval and information fusion based on a task paradigm. A document information extraction and processing device based on an electric power word-element structure includes: a word-element directed graph construction module 1, a word-element relationship determination module 2, an information extraction engine construction module 3, a multi-element electric power information system determination module 4, and an information extraction result acquisition module 5.

[0082] Term-unit directed graph construction module 1: The term-unit directed graph construction module 1 is used to identify the extraction task and redefine the task paradigm for the target power document, determine the term-unit node and construct the term-unit directed graph, and the power direction of the task paradigm is different; Term-unit relationship determination module 2: The term-unit relationship determination module 2 is used to construct a large language question-answering model based on the disambiguation principle, traverse the term-unit directed graph, and determine the term-unit relationship, wherein the large language question-answering model is a bidirectional logic; Information extraction engine construction module 3: The information extraction engine construction module 3 is used to match the document layout encoder to construct an information extraction engine, wherein the information extraction engine is a parallel engine matrix; Multi-electromechanical information system determination module 4: The multi-electromechanical information system determination module 4 is used to identify the term-unit directed graph and the term-unit relationship, intervene in the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determine the multi-electromechanical information system, wherein the multi-electromechanical information system and the term-unit directed graph Figure 1 One correspondence; information extraction result acquisition module 5: the information extraction result acquisition module 5 is used to complement and integrate the multiple power information systems as the information extraction result of the target power document.

[0083] Furthermore, the word-unit relationship determination module 2 further includes:

[0084] Based on question and answer samples, sample supervised pre-training is performed to determine the pre-training model; for documents in the power field, downstream task classes are determined; the disambiguation principle is introduced to construct a word meta-corpus; the pre-training model is fine-tuned in combination with the word meta-corpus and the downstream task class to determine the large language question and answer model.

[0085] Furthermore, the word-unit relationship determination module 2 further includes:

[0086] The word-gram directed graph is traversed, and forward question-answering logic recursion is performed to determine the first word-gram relationship; reverse question-answering logic verification is performed to determine the second word-gram structure; based on the second word-gram structure, the first word-gram relationship is verified, and if the verification is consistent, the word-gram relationship is determined; if the verification is inconsistent, the ambiguous point is located to compensate for the first word-gram relationship, the word-gram relationship is determined, and the ambiguous point is added to the word-gram corpus, wherein the word-gram corpus is regularly updated with a model.

[0087] Furthermore, the information extraction engine construction module 3 also includes:

[0088] An initialization encoder is introduced; multi-bit twinning is performed on the initialization encoder, and the encoder group is fine-tuned based on the task paradigm class to determine the encoder array; and the information extraction engine is constructed based on the encoder array in a parallel relationship.

[0089] Furthermore, the multi-element power information system determination module 4 further includes:

[0090] The word-meta directed graph is traversed to identify entity word-meta nodes, the target power document is scanned, and a binary network label based on the entity is determined; based on the binary network label, a search word-meta path is determined; a text-position joint encoding is set, wherein the joint encoding is used to perform chaotic word-meta sequence constraints; and the target power document is retrieved and extracted by combining the search word-meta path with the joint encoding.

[0091] Furthermore, the multi-element power information system determination module 4 further includes:

[0092] Based on the search word-meta path, entity nesting is identified to determine entity nesting features; with respect to the entity nesting features, decoupling is performed based on the word-meta relationship to determine a decoupled word-meta path.

[0093] Furthermore, the information extraction result acquisition module 5 also includes:

[0094] Based on multi-modality and multi-granularity, the information heterogeneous characteristics are determined; based on the information heterogeneous characteristics, a bidirectional heterogeneous relationship is determined, wherein the de-heterogeneous method is any characteristic within the information heterogeneous characteristics; based on the bidirectional heterogeneous relationship, the de-heterogeneous processing before the complementary fusion of the multi-element power information system and the display heterogeneous processing after the complementary fusion are performed.

[0095] A document information extraction and processing device based on an electric power word structure provided in an embodiment of the present invention can execute a document information extraction and processing method based on an electric power word structure provided in any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0096] Although the present invention makes various references to certain modules in the device according to the embodiments of the present invention, any number of different modules may be used and run on the user terminal and / or server, and the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0097] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A document information extraction and processing method based on power word structure, characterized in that: The method comprises the following steps: 1) For the target power document, identify the extraction task and redefine the task paradigm, determine the word-element nodes and construct a word-element directed graph, and the power direction of the task paradigm is different; 2) constructing a large language question-answering model based on the disambiguation principle, traversing the word-unit directed graph, and determining word-unit relationships, wherein the large language question-answering model is a bidirectional logic; 3) Building an information extraction engine with the document layout encoder, wherein the information extraction engine is a parallel engine matrix; 4) identifying the word-unit directed graph and the word-unit relationship, intervening the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determining a multi-element power information system, wherein the multi-element power information system corresponds one-to-one to the word-unit directed graph; 5) complementarily integrating the multiple power information systems as the information extraction result of the target power document; In step 4), the engine parallel retrieval and information extraction include: Traversing the word-element directed graph, identifying entity word-element nodes, scanning the target power document, and determining a binary network label based on the entity; Determining a search term path based on the binary network label; Setting a word-position joint code, wherein the joint code is used to perform a chaotic word-unit sequence constraint; Combining the search word path with the joint code, searching and extracting the target power document; After determining the search term path, include: Based on the search term path, entity nesting is identified and entity nesting features are determined; With respect to the entity nesting feature, decoupling is performed based on the word-gram relationship to determine a decoupled word-gram path.

2. The document information extraction and processing method based on the electric power word structure according to claim 1 is characterized in that: In step 2), the large language question answering model is constructed based on the disambiguation principle, including: Based on the question and answer samples, perform sample supervised pre-training and determine the pre-training model; Determine downstream task categories for documents in the power field; Introduce the disambiguation principle and construct a word corpus; The pre-trained model is fine-tuned in combination with the word-meta corpus and the downstream task class to determine the large language question-answering model.

3. The document information extraction and processing method based on the electric power word structure as claimed in claim 2 is characterized in that: The determining of word-unit relationships includes: Traversing the word-unit directed graph, performing forward question-answering logic recursion, and determining the first word-unit relationship; Perform reverse question-answer logic verification to determine the second word structure; Based on the second word-gram structure, verify the first word-gram relationship, and if the verification is consistent, determine the word-gram relationship; If the verification does not match, the ambiguous point is located to compensate for the first word-gram relationship, the word-gram relationship is determined, and the ambiguous point is added to the word-gram corpus, wherein the word-gram corpus is regularly updated with a model.

4. The document information extraction and processing method based on the electric power word structure according to claim 1 is characterized in that: In step 3), the construction of the information extraction engine includes: Introduce initialization encoder; Performing multi-bit twinning on the initialized encoder, fine-tuning the encoder group based on the task paradigm class, and determining the encoder array; The information extraction engine is constructed based on the encoder array in a parallel relationship.

5. The document information extraction and processing method based on the electric power word structure according to claim 1 is characterized in that: In step 5), the complementary integration of the multi-element power information system includes: Determine the heterogeneous characteristics of information based on multi-modality and multi-granularity; Based on the information heterogeneity characteristics, determining a bidirectional heterogeneous relationship, wherein the de-heterogeneity method is any characteristic within the information heterogeneity characteristics; Based on the bidirectional heterogeneous relationship, the de-heterogeneity processing before the complementary fusion of the multi-element power information system and the display heterogeneity processing after the complementary fusion are performed.

6. A document information extraction and processing device based on power word structure, characterized in that: The device is used to implement the document information extraction and processing method based on the power word structure as described in any one of claims 1 to 5, comprising: Word-element directed graph construction module: for the target power document, identify the extraction task and redefine the task paradigm, determine the word-element nodes and construct the word-element directed graph, and the power direction of the task paradigm is different; A word-unit relationship determination module: constructs a large language question-answering model based on the disambiguation principle, traverses the word-unit directed graph, and determines the word-unit relationship, wherein the large language question-answering model is a bidirectional logic; Information extraction engine construction module: in combination with the document layout encoder, an information extraction engine is constructed, wherein the information extraction engine is a parallel engine matrix; A multi-element power information system determination module: identifying the word-element directed graph and the word-element relationship, intervening the information extraction engine to perform engine parallel retrieval and information extraction on the target power document, and determining a multi-element power information system, wherein the multi-element power information system corresponds one-to-one to the word-element directed graph; Information extraction result acquisition module: complementarily integrates the multiple power information systems as the information extraction result of the target power document.

Images