Archive key element identification and classification method based on improved LSTM semantic modeling
Through the improved LSTM semantic modeling method, the accuracy and efficiency of the identification and classification of key elements in archive digital processing are solved, and efficient identification and classification of complex and diverse data are achieved, which improves the overall efficiency of archive management.
Patent Information
- Application Number
- CN202411515203.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-29
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the current digital processing of archives, the accuracy and inefficiency of key elements identification and classification are insufficient and inefficient, especially when processing complex and diverse unstructured data, it is difficult to accurately capture semantic relationships and implicit information.
Using improved LSTM semantic modeling methods, through streaming data cleaning, word segmentation, word vectorization and multi-level LSTM models, combined with adaptive learning mechanisms and context-aware mechanisms, we can capture semantic features and long-distance dependencies in the text, and carry out key tasks. Feature identification and classification.
It significantly improves the accuracy and efficiency of archive processing, can efficiently identify complex semantic structures, reduce noise interference, improve classification accuracy and system stability, and is suitable for diversified and large-scale data processing.
Smart Images

Figure CN119026601A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to a method for identifying and classifying key elements of archives based on improved LSTM semantic modeling. Background Art
[0002] Archives digitization is an important part of modern information management. Its significance lies in converting a large number of paper archives into digital data forms that are easy to store, retrieve and analyze. This can not only greatly improve the management efficiency of archives, but also effectively extend the preservation life of archives and reduce information loss caused by paper aging or accidental damage. The process of archive digitization usually includes scanning paper archives, optical character recognition (OCR), content structuring, key element identification and classification, etc. Among these links, the identification and classification of key elements of archives is particularly important because it can extract core information from massive text data and scientifically classify and manage archives. This process not only helps to improve the retrieval efficiency of archives, but also lays the foundation for subsequent data analysis and utilization. Therefore, key element identification and classification is a key link in archives digitization, which is of great significance to the efficient management of archives.
[0003] Although digital processing of archives has made significant progress in improving management efficiency and information security, the current identification and classification of key elements of archives still face some challenges. Traditional methods often rely on rules and template matching, which are difficult to cope with the complexity and diversity of archive content, especially when dealing with large amounts of unstructured data, and their accuracy and efficiency are insufficient. In addition, the semantic relationships in archives are complex, and traditional methods are difficult to capture the implicit information, resulting in inaccurate classification results.
[0004] In the field of artificial intelligence, semantic modeling based on long short-term memory networks (LSTM) provides an effective path to solve these problems. LSTM can capture long-distance dependencies in text through its memory ability, thereby more accurately understanding the semantic structure and key elements in archives. This method not only improves the accuracy of element recognition, but also continuously optimizes classification rules and improves classification efficiency through the adaptive ability of deep learning models. Therefore, the semantic modeling method based on improved LSTM shows obvious advantages in the identification and classification of key elements of archives, bringing new breakthroughs to the digital processing of archives. Summary of the invention
[0005] 1. Technical issues to be resolved The technical problem to be solved by the present invention is how to provide a method for identifying and classifying key elements of archives based on improved LSTM semantic modeling, so as to solve the problems of insufficient classification accuracy and low efficiency of element extraction in the current digital processing of archives.
[0006] (II) Technical solution In order to solve the above technical problems, the present invention proposes a method for identifying and classifying key elements of archives based on improved LSTM semantic modeling, which comprises the following steps: S1, the streaming archive data cleaning module performs preliminary cleaning on the archive text; S2. Perform word segmentation on the cleaned text to divide the text into independent words or phrases; S3, remove stop words after word segmentation; S4, performing word vectorization processing on the processed words, converting the text into a numerical vector representation; S5, input the text data after word vectorization into the LSTM model, and capture the semantic features and long-distance dependencies in the text through long short-term memory units; Among them, the LSTM model was improved for the archive classification task, including: introducing a multi-level structure to capture local and long-distance semantic features, enhancing the model's ability to understand complex semantic structures; in addition, an adaptive learning mechanism was introduced to adjust the weight of the memory unit according to the archive type, and through the context-aware memory and forgetting mechanism, key elements were accurately identified; S6, LSTM model identifies the key elements in the text based on the extracted semantic features and generates corresponding element labels; S7. Further semantic verification and context analysis are performed on the identified key elements to ensure the accuracy of the recognition results; S8. Based on the verified key elements, the archive text is automatically assigned to the preset categories to complete the automatic classification processing of the archive.
[0007] (III) Beneficial effects The present invention proposes a method for identifying and classifying archive key elements based on improved LSTM semantic modeling. The present invention proposes a method for identifying and classifying archive key elements based on improved LSTM semantic modeling. The method adopts artificial intelligence technology and improves the application of LSTM model for archive classification tasks, thereby significantly improving the accuracy and efficiency of archive processing. First, the multi-level structure of the LSTM model can capture local and long-distance semantic features in the text, making the understanding of complex archive content more comprehensive. Secondly, the adaptive learning mechanism of the model can dynamically adjust the weights of memory units according to different archive types to ensure that the accuracy of key element identification remains at a high level when facing different semantic structures. In addition, combined with the context-aware memory and forgetting mechanism, the model can reduce noise interference when identifying key elements, further improving the accuracy of classification. Due to these innovative improvements, the present invention can not only cope with diversified and complex archive content, but also maintain efficient operation when processing large amounts of data. The application of this method greatly improves the quality and efficiency of archive digital processing, and provides strong technical support for the intelligence of archive management systems. Through the present invention, more accurate archive classification and more effective key element extraction can be achieved, ultimately improving the overall efficiency of archive management. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 It is a flow chart of the method for identifying and classifying key elements of archives based on improved LSTM semantic modeling of the present invention; Figure 2 This is a diagram of an improved LSTM model for the file classification task of the present invention. DETAILED DESCRIPTION
[0009] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.
[0010] The technical problem to be solved by the present invention is how to provide a method for identifying and classifying key elements of archives based on improved LSTM semantic modeling, so as to solve the problems of insufficient classification accuracy and low efficiency of element extraction in the current digital processing of archives.
[0011] In order to solve the above technical problems, the present invention proposes a method for identifying and classifying key elements of archives based on improved LSTM semantic modeling, which comprises the following steps: S1, the streaming archive data cleaning module performs preliminary cleaning on the archive text; S2. Perform word segmentation on the cleaned text to divide the text into independent words or phrases; S3, remove stop words after word segmentation to reduce the interference of noise data; S4, performing word vectorization processing on the processed words, converting the text into a numerical vector representation; S5, input the text data after word vectorization into the LSTM model, and capture the semantic features and long-distance dependencies in the text through long short-term memory units; Here, the LSTM model is improved for the archive classification task, including the introduction of a multi-level structure to capture local and long-range semantic features and enhance the model's ability to understand complex semantic structures. In addition, the introduction of an adaptive learning mechanism can adjust the weights of memory units according to the archive type, improve the accuracy of feature extraction, and accurately identify key elements and reduce noise interference through a context-aware memory and forgetting mechanism.
[0012] S6, LSTM model identifies the key elements in the text based on the extracted semantic features and generates corresponding element labels; S7. Further semantic verification and context analysis are performed on the identified key elements to ensure the accuracy of the recognition results; S8. Based on the verified key elements, the archive text is automatically assigned to the preset categories to complete the automatic classification processing of the archive. Example
[0013] The technical problem to be solved by the present invention is to identify and classify key elements in the process of digital processing of archives, and to study how to use the targeted improved LSTM semantic modeling technology to improve the accuracy of element extraction and classification when processing complex and diverse archive contents, so as to ensure efficient and intelligent archive management in large-scale data processing.
[0014] The present invention is aimed at digital processing of archives and proposes a method for identifying and classifying key elements of archives based on improved LSTM semantic modeling. The method effectively processes complex and diverse archive contents through an improved LSTM model, improves the accuracy of element extraction and classification, and ensures efficient and intelligent management in large-scale archive data processing. This technology not only improves the classification accuracy of archives, but also ensures the stability and reliability of the system when facing complex semantics.
[0015] Figure 1 It is the main working flow diagram of the technical solution of the present invention, and the whole working flow is as follows: S1. The streaming archive data cleaning module performs preliminary cleaning of the archive text.
[0016] The streaming file data cleaning module is responsible for the preliminary cleaning of file texts to ensure the accuracy and effectiveness of subsequent processing. This module automatically extracts the valid information from the file texts and removes irrelevant content such as special characters, HTML tags, and other noisy data. This step is crucial for file digitization processing and directly affects the subsequent effects of word segmentation, semantic analysis, and element recognition. The streaming file data cleaning module operates in an efficient and real-time manner to ensure that the file data reaches the expected cleanliness and standardization before entering the next processing link.
[0017] S2. Perform word segmentation on the cleaned text, splitting the text into independent words or phrases.
[0018] Perform word segmentation on the cleaned text. The word segmentation processing module is responsible for splitting the text into independent words or phrases, which is the basis for semantic analysis of file content. The word segmentation processing module uses a fine-grained word segmentation algorithm to decompose the coherent text content into the smallest recognizable semantic units, ensuring that each word or phrase can accurately express its original meaning.
[0019] After automatic word segmentation is completed, it also checks for polysemous words or word segmentation errors. If an error is detected, the system will first perform a round of automatic word segmentation correction. If the automatic correction cannot completely solve the problem, the system will prompt the user to perform manual correction. After the manual correction is completed, the system will check the word segmentation result again to ensure its accuracy before proceeding to the next step. This combined automatic and manual word segmentation optimization mechanism effectively improves the accuracy of word segmentation and the efficiency of text processing.
[0020] During the word segmentation process, the module automatically processes polysemous words, synonyms, and optimizes the relevance between words, providing high-quality input data for subsequent semantic feature extraction. This step is crucial for improving the accuracy of element recognition.
[0021] S3. Remove stop words after word segmentation to reduce the interference of noisy data.
[0022] After the word segmentation process is completed, the stop word removal module filters the stop words in the text. Stop words are usually some words that do not have substantial meaning in semantic analysis, such as "de", "shi", "zai", etc. They appear frequently in the text but contribute little to semantic parsing. By removing these stop words, the module can effectively reduce the interference of noisy data, reduce information redundancy, and improve the accuracy and efficiency of the model in subsequent element recognition and classification tasks. This step ensures the refinement of the input data and the purity of the semantics.
[0023] S4. Perform word vectorization on the processed words, converting the text into a numerical vector representation.
[0024] The processed words are vectorized to convert each word in the text into a numerical vector representation, which is a key step in inputting the text into the LSTM model. Through word vectorization, the words in the text will be converted into vectors with fixed dimensions, so that the LSTM model can understand and process the semantic features of the text. The word vectorization module uses a pre-trained word embedding model or generates word vectors through self-training to ensure that the vector representation of each word can fully reflect its semantic relationship in the text, providing high-quality input data for subsequent semantic modeling.
[0025] S5. Input the text data after word vectorization into the LSTM model, and capture the semantic features and long-distance dependencies in the text through long short-term memory units.
[0026] In this step, the text data after word vectorization is input into the LSTM model, which uses its long short-term memory unit to capture the semantic features and long-distance dependencies in the text. The LSTM model can effectively retain important information in the context and ignore irrelevant data through its unique memory and forgetting mechanism, thereby understanding the complex semantic structure in the text. The core of this step is to accurately extract key information from the text through the deep learning ability of LSTM, laying a solid foundation for subsequent feature recognition and classification.
[0027] S6, LSTM model identifies the key elements in the text based on the extracted semantic features and generates corresponding element labels; In this step, the LSTM model identifies the key elements in the text based on the semantic features extracted in the previous step and generates corresponding element labels. Through the LSTM model's deep understanding of semantics, it can accurately locate and annotate key information in the text, such as core elements such as people, places, and time. These element labels not only provide a basis for the classification and retrieval of archives, but also help the system further understand and organize the content of archives and improve the intelligent level of archive management. This step is a key link in achieving accurate archive classification.
[0028] S7. Further semantic verification and context analysis are performed on the identified key elements to ensure the accuracy of the recognition results; In this step, the key elements identified by the LSTM model are further semantically verified and contextually analyzed to ensure the accuracy of the recognition results. Through in-depth analysis of the text context, it is verified whether the elements extracted by the model are consistent with the overall semantics to avoid recognition errors caused by insufficient local information. The semantic verification module uses cross-validation for analysis. This module is implemented through the following steps: S71. First, the identified key elements are compared with the context of the paragraph in which they are located to check whether they are consistent with the global semantics.
[0029] S72. Then, based on predefined semantic rules and similar archive contents in the database, the semantics of the key elements are matched and verified to ensure that the elements can remain consistent in different contexts.
[0030] S73. Then, using a semantic similarity calculation algorithm, cross-compare the element with related elements in other text fragments to further verify its correctness in different contexts.
[0031] Through multi-level semantic comparison and cross-validation, possible misjudgments are corrected to ensure that the final output feature labels are accurate and reliable.
[0032] S8. Based on the verified key elements, the archive text is automatically assigned to the preset categories to complete the automatic classification processing of the archive.
[0033] In this step, according to the key element information verified in the previous step, the archive text is automatically assigned to the preset category to complete the automatic classification of the archive. For multiple key elements after verification, the system uses a weighted classification algorithm to analyze the importance of each key element in the archive text and comprehensively determine the category that best suits the archive content. Specifically, S81. First, according to the preset classification criteria, all identified key elements are assigned weights. The weights are determined comprehensively by multiple dimensions such as the frequency of occurrence of the elements in the text, semantic relevance, and contextual importance.
[0034] S82. Then, the classification value of each key element is weighted and compared with different classification labels. By calculating the matching degree between the classification label and the key element, the possibility of the archive text belonging to each classification is obtained.
[0035] S83. Finally, using the weighted matching degree, select the classification with the highest matching degree as the final classification result.
[0036] In order to further improve the accuracy of classification, the present invention also adopts a multi-level classification structure. In this structure, a rough classification is first performed to divide the archive text into major categories; then, based on the refined classification criteria, the deep semantic information of the key elements is used to perform secondary fine classification to ensure that the archive text is accurately assigned to the most appropriate subcategory.
[0037] Through the combination of this multi-factor weighted classification algorithm and multi-level classification structure, the system can automatically and accurately complete the classification task of archives while processing a large number of key factors, thereby effectively improving the intelligence level and classification accuracy of archive management. By matching the identified elements with predefined classification criteria, the system can efficiently and accurately classify the archives into the most suitable categories. This classification process not only improves the efficiency of archive management, but also ensures the accuracy of the classification results, facilitating subsequent query and utilization. This automated classification step significantly reduces manual intervention and improves the level of intelligence in archive processing.
[0038] Among them, it is mentioned in S5 that the text data after word vectorization is input into the LSTM model improved for the task of extracting and classifying key elements of archives. This part is described in detail below.
[0039] like Figure 2 , demonstrated the LSTM model improvement method for archive key element extraction and classification tasks, focusing on solving the following key problems: the diversity and complexity of archive content, the accurate extraction of key elements, and the efficiency and robustness of classification tasks.
[0040] The LSTM model includes: an input layer, multiple LSTM units, a multi-task learning module, and a global context module; Figure 2 The left side of is the input layer, which means that the archive text is used as the input of the model after preprocessing (word segmentation, stop word removal, word vectorization, etc.). After the input text is processed by word vectorization, it enters the LSTM unit in the form of a numerical vector.
[0041] The LSTM unit is the core part of the model, which can be used to capture the long-term and short-term dependencies in the text through mechanisms such as memory gate, forget gate, and input gate. Figure 2 In the example, you can see the following key components: Input gate, used to control whether new information enters the memory unit; The forget gate is used to decide whether to forget the previous memory; The output gate is used to decide which parts of the memory are passed as output to the next time step; Candidate memory unit, used to store candidate new memory content; Memory cells are used to store long-term dependency information.
[0042] These gating mechanisms help the model extract key elements from complex archival content by incorporating contextual semantic information and retaining or forgetting specific information.
[0043] In the operation process of LSTM unit combined with attention mechanism, first, the input data of each time step, that is, the archival data after word vectorization, is sent to the LSTM unit. Before entering the LSTM, these data will undergo two attention weighting processes to highlight important information. The first attention weighting process is performed by a fully connected neural network, which is based on the current input data and the hidden state of the previous time step. , calculate the first attention weight value. This weight value indicates the importance of the current input data. Then, the first attention weight value is calculated with the current input data dot product to obtain the first weighted input sequence.
[0044] Next, the first weighted input sequence enters the second fully connected neural network and combines the memory cell state of the previous time step , the second fully connected neural network calculates the second weight value, and the second weight value will be dot-producted with the first weighted input sequence to obtain the final input sequence that completes the attention weighting.
[0045] The input sequence after attention weighting, together with the hidden state of the previous time step , and are input into the LSTM unit together.
[0046] Inside the LSTM unit, a series of gating mechanisms are used to determine which information should be retained, which should be forgotten, and which new information should be stored. This includes the forget gate, input gate, memory unit update, and output gate calculation.
[0047] Finally, the LSTM unit will output the hidden state and memory cell state of the current time step, which will be used for calculation of the next time step.
[0048] Through such a process, the LSTM unit can process archival data more effectively and capture key information, thereby improving the accuracy of identifying and classifying key elements of archives.
[0049] The LSTM unit also includes a context-aware module. The context-aware module located on the right side of the LSTM unit dynamically adjusts the model's attention weight by analyzing the text context. It helps the model better understand the semantic relationship between different time steps, thereby accurately capturing key elements. This module retains valuable information and enhances the global semantic understanding of the text by calculating context dependencies.
[0050] The top of the LSTM model is a multi-task learning module, which is used to optimize two objectives simultaneously: key feature extraction and archive classification. By sharing feature representations, it can improve the performance of the model on both tasks, avoid overfitting, and make use of the complementarity between different tasks.
[0051] The global context module at the bottom of the LSTM model is used to integrate semantic information from the entire text. The local semantic information generated by LSTM at each time step will be combined with the global context to enhance the understanding of the semantic structure in the archival text.
[0052] In order to improve the performance of key element extraction and archive classification at the same time, the method of the present invention adopts a multi-task learning framework. Under this framework, the method of the present invention uses Figure 2 The model structure shown in the figure optimizes two objective functions at the same time, namely the loss function of the key feature extraction task And the loss function of the archive classification task , two tasks share Figure 2 The LSTM model structure shown is trained. The loss function of the entire model is calculated as follows: .in, and is the weight coefficient. By jointly optimizing these two tasks, the model can share useful features between different tasks, thereby improving the overall performance. In order to ensure that the model can converge effectively, the following strategies are adopted: dynamic learning rate adjustment, the learning rate of the model will be dynamically adjusted according to the performance on the validation set. If the model performs poorly for a period of time, the learning rate will be reduced to avoid falling into the local optimal solution and speed up the convergence. At the same time, an early stopping mechanism is adopted. If the performance of the model on the validation set does not improve within a certain round, the training will stop early to prevent overfitting.
[0053] In order to improve the robustness of the model when processing complex archive content, synonym replacement, random insertion and random deletion data augmentation techniques are used. These techniques expand the training data set and improve the generalization ability of the model. The model uses the Adam optimizer for parameter update, combined with the dynamic learning rate mechanism, which helps to accelerate convergence. The input data is input into the model through batch processing, and the gradient is updated after each batch, which can speed up the training speed and stabilize the model performance. Through these optimization strategies and improved model structure, the LSTM model shows good performance and efficient learning ability when processing archive classification and key element extraction tasks.
[0054] Through optimization of model structure, training mechanism and task-specificity, the performance of LSTM model in extracting and classifying key elements of archives has been significantly improved.
[0055] When traditional LSTM models are used to process text tasks, the conventional single-layer LSTM structure may not be sufficient to capture the complex semantic information in archival texts, especially when long-distance dependencies are involved. To this end, the method of the present invention introduces a multi-level LSTM structure, that is, adding more hidden layers to the basic LSTM layer, so that the model can extract deeper semantic features layer by layer. Specifically, assuming the input sequence is , after a multi-layer LSTM structure, the output is ,in For the The output of the layer, for This multi-level structure can fully capture the fine-grained and long-distance dependency information in the text, and improve the model's ability to understand complex archive content.
[0056] In order to further improve the model's ability to process different types of archival texts, the method of the present invention introduces an adaptive memory mechanism. In this mechanism, the model can dynamically adjust the state of the memory unit according to the input archival content, ensuring that the model can flexibly adjust its own memory and forgetting capabilities when processing different content. This mechanism is achieved by introducing a learnable memory gating parameter, the formula is in, is the current time step Memory cell status, For the Gate of Forgetfulness, is the input gate, for The dot product of the attention weights obtained by the second fully connected neural network. and Through dynamic adjustment, the model can adaptively control the state of memory units, thereby showing better generalization ability when processing different types of archival texts. Figure 2The tanh function is mainly used to generate candidate state values. It receives information from the current input and the hidden state of the previous moment and performs a nonlinear transformation on it. The role of the tanh function is to compress the input value into the range of -1 to 1, so that the network can better process and represent the characteristics of the data. The output gate is used in LSTM to control the output information at the current moment. It determines which information should be passed from the current cell state to the output, and the intensity of the output. The output gate receives the input data at the current moment and the hidden state at the previous moment, and after a series of linear transformations and activation function operations, an intermediate result is obtained. This intermediate result is determined by a sigmoid function to determine the output gating value. The sigmoid function compresses the input value into the range of 0 to 1, where 0 means that the output gate is completely closed and 1 means that the output gate is completely opened. The output gate selects the output information based on the gating value and the cell state at the current moment. Specifically, the gating value is multiplied element by element with the cell state, and only when the gating value is close to 1, the corresponding cell state information will be passed to the output. The result after state selection is processed by a linear transformation and activation function (tanh function) to obtain the output value at the current moment.
[0057] The dynamic adjustment process is as follows: When new archive text is input into the model, LSTM first calculates the values of the forget gate and input gate through the activation function. The values of these gates are dynamically calculated based on the current input and the hidden state of the previous moment; the forget gate controls how much of the memory cell state of the previous moment will be retained, and the input gate determines the proportion of the candidate memory state input at the current moment that enters the memory cell. Through the combined action of these two gates, the memory cell state will be updated to The model continuously optimizes parameters through training data, so that the model can adaptively adjust the values of the forget gate and input gate, thereby flexibly adjusting the retention and update of memory when processing different archival texts. When processing different types of archives, the model can adaptively decide which old information to forget and which new information to retain, ensuring sufficient generalization ability for various types of archive content.
[0058] In the archive text, different words contribute differently to the key elements. In order to highlight the role of key words in element extraction, the method of the present invention introduces an improvement based on the attention mechanism in the LSTM model. The attention mechanism allows the model to dynamically focus on different parts of the input sequence, thereby improving the model's ability to capture important information. By introducing the attention mechanism, the model can dynamically adjust the attention to different words, thereby more accurately identifying the key elements in the archive.
[0059] The extraction of key elements in archive texts often depends on the accurate capture of context information. To this end, the present invention introduces a context-aware mechanism in the LSTM model to enhance the model's ability to identify key elements by capturing context information at each time step. The context-aware mechanism captures the global semantic information of the current time step by introducing a global context vector, as shown in the formula: (The calculation of the output gate and the tanh activation function are included in the LSTM() function), .in, is the hidden state after attention weighting at the i-th time step, where is the attention weight, is the LSTM hidden state. Through weighted summation, the model can combine global context information to further improve the recognition accuracy of key elements. t With h t The calculation method is the same as that of t The output of the current time step is recorded, h t As the input of the next time step, it continues to participate in the next round of calculation) In addition, multi-task learning can effectively prevent model overfitting and improve generalization ability. During the training process, the method of the present invention uses data enhancement technology to expand the training data set and improve the generalization performance of the model. According to the characteristics of archival text, a variety of enhancement methods are designed, such as synonym replacement, random insertion and random deletion.
[0060] In addition, the method of the present invention introduces a dynamic learning rate adjustment strategy, which dynamically adjusts the learning rate according to the performance of the model on the validation set to accelerate the convergence of the model and avoid falling into the local optimum.
[0061] Through the structural optimization and targeted improvement of the LSTM model, the present invention has achieved remarkable results in the task of extracting and classifying key elements of archives. The improved LSTM model can better understand and process complex archive content, improve the recognition accuracy of key elements and the accuracy of classification. These improvements provide strong technical support for intelligent management in archive digital processing, and show good performance and applicability in practical applications.
[0062] The present invention proposes a method for identifying and classifying archive key elements based on improved LSTM semantic modeling. By using artificial intelligence technology and improving the application of LSTM model for archive classification tasks, the accuracy and efficiency of archive processing are significantly improved. First, the multi-level structure of the LSTM model can capture local and long-distance semantic features in the text, making the understanding of complex archive content more comprehensive. Secondly, the adaptive learning mechanism of the model can dynamically adjust the weights of memory units according to different archive types to ensure that the accuracy of key element identification remains at a high level when facing different semantic structures. In addition, combined with the context-aware memory and forgetting mechanism, the model can reduce noise interference when identifying key elements, further improving the accuracy of classification. Due to these innovative improvements, the present invention can not only cope with diversified and complex archive content, but also maintain efficient operation when processing large amounts of data. The application of this method greatly improves the quality and efficiency of archive digital processing, and provides strong technical support for the intelligence of archive management systems. Through the present invention, more accurate archive classification and more effective key element extraction can be achieved, ultimately improving the overall efficiency of archive management.
[0063] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A method for identifying and classifying key elements of archives based on improved LSTM semantic modeling. It is characterized in that The method comprises the following steps: S1, the streaming archive data cleaning module performs preliminary cleaning on the archive text; S2. Perform word segmentation on the cleaned text to divide the text into independent words or phrases; S3, remove stop words after word segmentation; S4, performing word vectorization processing on the processed words, converting the text into a numerical vector representation; S5, input the text data after word vectorization into the LSTM model, and capture the semantic features and long-distance dependencies in the text through long short-term memory units; Among them, the LSTM model was improved for the archive classification task, including: introducing a multi-level structure to capture local and long-distance semantic features, enhancing the model's ability to understand complex semantic structures; in addition, an adaptive learning mechanism was introduced to adjust the weight of the memory unit according to the archive type, and through the context-aware memory and forgetting mechanism, key elements were accurately identified; S6, LSTM model identifies the key elements in the text based on the extracted semantic features and generates corresponding element labels; S7. Further semantic verification and context analysis are performed on the identified key elements to ensure the accuracy of the recognition results; S8. Based on the verified key elements, the archive text is automatically assigned to the preset categories to complete the automatic classification processing of the archive.
2. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 1, It is characterized in that In S2, after the automatic word segmentation is completed, it is also checked whether there are polysemous words or word segmentation errors; if an error is detected, the system will first perform a round of automatic word segmentation correction; if the automatic correction cannot completely solve the problem, the system will prompt the user to perform manual correction; after the manual correction is completed, the system will check the word segmentation results again to ensure accuracy before proceeding to the next step of processing.
3. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 1, It is characterized in that The S7 includes: S71. First, the identified key elements are compared with the context of the paragraph in which they are located to check whether they are consistent with the global semantics; S72, then, based on predefined semantic rules and similar archive contents in the database, the semantics of the key elements are matched and verified to ensure that the elements can be consistent in different contexts; S73, then, using a semantic similarity calculation algorithm, cross-comparison is performed between the element and related elements in other text segments to further verify its correctness in different contexts; Through multi-level semantic comparison and cross-validation, possible misjudgments are corrected to ensure that the final output feature labels are accurate and reliable.
4. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 3, It is characterized in that The S8 includes: S81. First, according to the preset classification criteria, all identified key elements are assigned weights, and the weights are determined by the frequency of occurrence of the elements in the text, the semantic relevance, and the contextual importance; S82, then, weighted calculation of the classification value of each key element, comparison of the classification value of each key element with different classification labels, and calculation of the matching degree between the classification label and the key element to obtain the possibility that the archive text belongs to each classification; S83. Finally, using the weighted matching degree, select the classification with the highest matching degree as the final classification result.
5. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 4, It is characterized in that This method adopts a multi-level classification structure, in which a coarse classification is first performed to divide the archival texts into major categories; then, based on the refined classification criteria, a secondary fine classification is performed using the deep semantic information of key elements to ensure that the archival texts are accurately assigned to the most appropriate subcategories.
6. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling according to any one of claims 1 to 5, It is characterized in that In S5, the LSTM model includes: an input layer, multiple LSTM units, a multi-task learning module and a global context module; The input layer represents the archival text as the input of the model after preprocessing. After the input text is processed by word vectorization, it enters the LSTM unit in the form of a numerical vector; LSTM units, used to capture long-term and short-term dependencies in text; A multi-task learning module for simultaneously optimizing two objectives: key feature extraction and profile classification; through shared feature representation; The global context module is used to integrate semantic information from the entire text. The local semantic information generated by LSTM at each time step will be combined with the global context to enhance the understanding of the semantic structure in the archival text.
7. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 6, It is characterized in that In the LSTM unit, the following process is performed: First, the input data of each time step, that is, the archival data after word vectorization, is fed into the LSTM unit; before entering the LSTM, these data will undergo two attention weighting processes to highlight important information; the first attention weighting process is performed through a fully connected neural network, which is based on the current input data and the hidden state of the previous time step. , calculate the first attention weight value; Then, the dot product of the first attention weight value and the current input data is calculated to obtain the first weighted input sequence; Next, the first weighted input sequence enters the second fully connected neural network and combines the memory cell state of the previous time step , the second fully connected neural network calculates the second weight value, which will be dot-producted with the first weighted input sequence to obtain the final attention-weighted input sequence; The input sequence after attention weighting, together with the hidden state of the previous time step , and are input into the LSTM unit together; Inside the LSTM unit, a series of gating mechanisms are used to determine which information should be retained, which should be forgotten, and which new information should be stored; this includes the forget gate, input gate, memory unit update, and output gate calculation; Finally, the LSTM unit will output the hidden state and memory cell state of the current time step, which will be used to calculate the next time step; The LSTM unit also includes a context-aware module, which is used to dynamically adjust the model's attention weights by analyzing the text context to help the model better understand the semantic relationships between different time steps, thereby accurately capturing key elements; this module retains valuable information and enhances the global semantic understanding of the text by calculating contextual dependencies.
8. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 7, It is characterized in that This method simultaneously optimizes two objective functions, namely the loss function of the key element extraction task And the loss function of the archive classification task ,The two tasks are trained using the LSTM model structure together; the overall loss function of the model is calculated as follows: ;in, and is the weight coefficient; by jointly optimizing these two tasks, the model can share useful features between different tasks; The following strategies were adopted during training: Dynamic learning rate adjustment: the learning rate of the model will be dynamically adjusted according to the performance on the validation set. If the model performs poorly for a period of time, the learning rate will be reduced to avoid falling into the local optimal solution and speed up the convergence. At the same time, an early stopping mechanism is adopted: if the performance of the model on the validation set does not improve within a certain round, the training will be stopped early to prevent overfitting; The model uses the Adam optimizer to update parameters and combines it with a dynamic learning rate mechanism to accelerate convergence. The input data is input into the model through batch processing, and the gradient is updated after each batch to speed up training and stabilize model performance.
9. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 7, It is characterized in that The LSTM model introduces a multi-level LSTM structure, that is, adding more hidden layers to the basic LSTM layer, so that the model can extract deeper semantic features layer by layer; specifically, assuming the input sequence is , after the multi-layer LSTM structure, the output is ,in For the The output of the layer, for The output of the layer; The LSTM model introduces an adaptive memory mechanism; in this mechanism, the model dynamically adjusts the state of the memory unit according to the input archive content. This mechanism is achieved by introducing a learnable memory gating parameter, which is formulated as ,in, is the current time step Memory cell status, For the Gate of Forgetfulness, is the input gate, for The dot product of the attention weights obtained by the second fully connected neural network; and The model can adaptively control the state of memory units through dynamic adjustment, thus showing better generalization ability when processing different types of archival texts. The dynamic adjustment process is as follows: when new archive text is input into the model, LSTM first calculates the values of the forget gate and input gate through the activation function. The values of these gates are dynamically calculated based on the current input and the hidden state at the previous moment; the forget gate controls how much of the memory cell state at the previous moment will be retained, and the input gate determines the proportion of the candidate memory state input at the current moment that enters the memory cell. Through the joint action of these two gates, the memory cell state will be updated to The model continuously optimizes parameters through training data, so that the model can adaptively adjust the values of the forget gate and input gate, thereby flexibly adjusting the retention and update of memory when processing different archival texts.
10. The method for identifying and classifying key elements of archives based on improved LSTM semantic modeling as claimed in claim 7, It is characterized in that The context-aware mechanism is introduced into the LSTM model to enhance the model's ability to identify key elements by capturing context information at each time step. The context-aware mechanism captures the global semantic information of the current time step by introducing a global context vector. The formula is: in, For the The hidden state after attention weighting at time steps, where is the attention weight, is the hidden state of LSTM; through weighted summation, the model can combine global context information to further improve the recognition accuracy of key elements.
Citation Information
Patent Citations
BERT improved model-based text sentiment analysis method
CN114781392A
Financial archive management method and system, electronic equipment and storage medium
CN116910330A
Electric power production potential safety hazard text clustering method and system
CN117076665A
OCR (Optical Character Recognition) method for production logging-oriented archival data
CN118587719A
Text classification method, apparatus, computer device, and storage medium
WO2022227207A1
Cited By
Structured extraction method for insurance customer service picture text information based on multi-modal large model
CN119559648A
Semantic communication retransmission method and system based on long short-term memory network, terminal and storage medium
CN120050004A
A semantic communication retransmission method, system, terminal and storage medium based on long short-term memory network
CN120050004B
System and method for safe arrangement of archives based on preposition of archive elements
CN120510622A
