Multi-label automatic classification method and system for literature data

Abstract

The invention relates to a multi-label automatic classification method and system for literature data, and the method employs two different models: one is a long-term short-term memory network which is good at capturing text sequence information, and the other is a convolutional neural network which can effectively extract the local features of a text, and the two models carry out the feature extraction of the text independently. Then, the extracted features are integrated through a specific fusion strategy, further feature compression and abstraction are carried out through MLP, and finally a text classification result is obtained. By fusing the two characteristics together, the method aims to ensure that the finally extracted characteristics have richness in time and space at the same time, so that the text classification performance is remarkably improved, especially for complex texts with a type crossing phenomenon. Compared with a traditional manual classification method and a current method of using a large model to apply prompt word engineering classification, the method provided by the invention is faster.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of document classification, and particularly relates to a document data multi-label automatic classification method and system. BACKGROUND

[0002] As a key link of document management and knowledge mining, the research process of scientific literature classification has evolved from traditional manual classification, classification based on statistical machine learning, to current deep learning methods. In the early stage, scientific literature classification mainly relied on manual analysis by field experts, which was time-consuming and difficult to apply on a large scale. Subsequently, statistical machine learning methods such as Naive Bayes and Support Vector Machine were introduced, which improved the classification efficiency, but were still limited by the sufficiency of feature extraction and classification accuracy. In recent years, pre-trained language models have shown significant advantages in scientific literature classification due to their powerful feature extraction and pattern recognition capabilities. These models effectively capture the semantic information and context relationships of scientific literature through pre-training on large-scale text data, significantly improving the classification performance.

[0003] After utilizing pre-trained language models for semantic representation, the current widely used deep learning techniques are mostly limited to using a single neural network stack or certain neural network stacks, i.e., usually a multi-layer perceptron (MLP) based neural network stack. However, due to the performance bottleneck of single models, their accuracy and efficiency in handling complex text classification tasks are often unsatisfactory.

[0004] In addition, in the field of text classification, the current mainstream methods mostly focus on solving single-type classification problems, i.e., assuming that each document belongs to only one category. However, this assumption often ignores the universal phenomenon of literature type intersection, i.e., in the real world, a document may cover and apply to multiple types or topics at the same time. The existing methods cannot accurately reflect this complexity, and it is necessary to explore methods that can handle multi-label classification problems to cope with the situation where a document may belong to multiple types at the same time, thereby improving the accuracy of document classification.

[0005] In addition, under the background of the deep integration of artificial intelligence and knowledge management, the field of Chinese scientific and technological literature is facing a double dilemma in the construction of annotated corpus: on the one hand, due to the characteristics of high complexity of professional terms and strong semantic correlation, the cost of obtaining high-quality manually annotated corpus increases exponentially; on the other hand, the traditional annotation system based on rules or fixed classification framework has been difficult to adapt to the multi-dimensional and dynamic needs of modern scientific research management. In particular, the classification standards used by different institutions differ significantly in concept granularity, hierarchical structure and business logic, resulting in a serious semantic gap when cross-platform literature resources are circulated. Under this background, the pseudo-annotation technology based on large language models has become an important way to solve the shortage of corpus due to its high efficiency and automation, but the existing methods generally face the problem of sensitivity to the length of literature text - when dealing with scientific and technological literature containing titles, abstracts, keywords and other information, the existing large language models perform significantly worse in context modeling, fine-grained semantic analysis and computational efficiency. SUMMARY

[0006] In order to solve the problem of literature classification, and considering the reality that one literature may belong to multiple types (i.e. type intersection phenomenon) in text classification, the present application provides a literature data multi-label automatic classification method and system, which uses a deep learning multi-model fusion method to solve the problem of text multi-classification; and also solves the problems of lack of high-quality annotated corpus of Chinese scientific and technological literature and low efficiency and uneven quality of manual annotation.

[0007] In order to solve the above technical problems, the present application provides a literature data multi-label automatic classification method, characterized in that it comprises the following steps:

[0008] Step one, original data collection: collect literature data from multiple sources of websites and journals through crawler technology and store it as a structured data table;

[0009] Step two, text preprocessing: clean the collected literature data, including block extraction of title, abstract, keyword and free word, filling of missing values and removal of irrelevant content;

[0010] Step three, building a classification model:

[0011] a, input layer: mark the separation of different input parts through special marks in the BERT model for the input of title, keyword, abstract and free word alone or in combination;

[0012] b, encoding layer: use the pre-trained BERT model to map the text into a vector representation;

[0013] c. Feature extraction layer: The Bi-LSTM network and CNN network are used in parallel to extract the sequence-dependent features and local features of the text respectively. The sequence features output by the Bi-LSTM are combined with the local features output by the CNN to form a comprehensive feature.

[0014] d. Classification layer: compresses and abstracts the fused features through a multi-layer perceptron, and outputs multi-label classification probabilities through a Sigmoid activation function;

[0015] Step 4: Train the classification model;

[0016] Step 5: Input the literature data preprocessed in step 2 into the trained classification model and output the multi-label classification results.

[0017] Furthermore, the text generated in step 2 is processed by a pseudo-annotation method to generate a supervised dataset, and the classification model in step 3 is trained. The pseudo-annotation method steps are as follows:

[0018] S1. Construct a document type feature set, segment the preprocessed text, and remove stop words and special characters;

[0019] S2, based on character similarity and semantic similarity strategies, preliminary classification of document types;

[0020] S3. Integrate the classification results of different information sources, determine the final document type label through weight distribution, and form a pseudo-annotated dataset.

[0021] Furthermore, the Bi-LSTM network is a bidirectional long short-term memory network that captures contextual dependencies through forward and reverse sequence processing, and uses the last layer of hidden state as the overall feature representation of the sequence.

[0022] Furthermore, the LSTM network effectively captures long dependencies in the sequence by introducing input gates, forget gates, output gates, and cell state mechanisms. The calculation process of each part is as follows:

[0023] The input gate determines whether new information is added to the cell state. The calculation process is as follows:

[0024] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0025] Among them, W i is the learnable weight, b i is the bias vector;

[0026] The forget gate determines the information retained in the cell state at the previous time, and the calculation process is as follows:

[0027] f t = σ(W f · [h t-1 , x t ] + b f )

[0028] Where W f is a learnable weight, b f is a bias vector, h t-1 is the hidden state at the previous time, and x t is the input at the current time;

[0029] The output gate is used to determine the information carried by the hidden state o t at the current time, and the calculation process is as follows:

[0030] o t = σ(W o · [h t-1 , x t ] + b o )

[0031] Where W o is a learnable weight, and b0 is a bias vector;

[0032] The candidate memory cell calculation process is as follows:

[0033]

[0034] Where W C is a learnable weight, and b C is a bias vector;

[0035] Memory cell update, calculation process as follows:

[0036]

[0037] C t-1 is the memory cell state at the previous time, and ⊙ represents element multiplication, C t is the memory cell state at the current time;

[0038] The hidden state calculation process is as follows:

[0039] h t = o t ☉ tanh(C t )

[0040] For the forward LSTM, the hidden state of each time step is calculated in turn according to the time step order For the reverse LSTM, the hidden state of each time step is calculated in reverse order of time steps

[0041] At each time step t, the hidden states of the forward and reverse LSTMs are concatenated to obtain a vector representation of each time step

[0042]

[0043] The vector representation of the entire sequence is:

[0044]

[0045] Further, the CNN network uses multi-size convolution kernels to extract local features of the text, and generates a comprehensive feature vector through a max-pooling operation.

[0046] A document data multi-label automatic classification system, characterized in that it comprises the following modules:

[0047] Data acquisition module: collect multi-source literature data through crawler technology;

[0048] Preprocessing module: clean and block process literature data, extract title, abstract, keyword and free word;

[0049] Pseudo-labeling module: generate pseudo-labeling dataset based on character similarity, semantic similarity and weight allocation;

[0050] Classification module: including classification model, used for multi-label classification of preprocessed text;

[0051] Visualization module: display classification results, support online index and query of literature.

[0052] Further, in the classification module, the output features of Bi-LSTM and CNN are flattened and concatenated before being input into MLP for dimension reduction.

[0053] Further, the pseudo-labeling module integrates an artificial review function for correcting the quality of pseudo-labeling labels.

[0054] A computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to implement a document data multi-label automatic classification method.

[0055] An electronic device comprising a memory, a processor, and a computer program stored on the memory, characterized in that the processor executes the computer program to implement a document data multi-label automatic classification method.

[0056] Advantages:

[0057] The present application aims at the problems of high labeling cost, large classification system difference and low processing efficiency of long text in the field of Chinese scientific literature labeling, and proposes a literature type pseudo-labeling method, which can significantly improve the efficiency of supervised dataset construction.

[0058] Meanwhile, unlike the previous text classification methods that rely solely on a single deep learning model, the present method combines two different models: one is the Long Short-Term Memory Network (LSTM) which is good at capturing text sequence information, and the other is the Convolutional Neural Network (CNN) which can effectively extract local features of text such as words and phrases. These two models independently extract features from the text. Then, the extracted features are integrated through a specific fusion strategy, and then passed through MLP for further feature compression and abstraction, and finally the classification result of the text is obtained. The LSTM network can extract the time sequence characteristics contained in the text due to its long-term memory ability, which is crucial for understanding the context information of the text. While the CNN network is known for its strong spatial feature extraction ability, which can efficiently capture the local structure and pattern in the text. By combining these two characteristics together, the present method aims to ensure that the final extracted features are rich in both time and space, thereby significantly improving the performance of text classification, especially for complex texts that have type crossing phenomena.

[0059] The scientific literature classification method and system proposed by the present application can quickly and accurately identify the subject or research field of a document by receiving its detailed information as input, and then assign it to the appropriate categories. Compared with traditional manual classification and current classification methods using large models with prompt word engineering, the present method is faster. This capability greatly promotes the automation of scientific literature classification and archiving process, providing researchers with more reliable and efficient literature management and retrieval tools, significantly improving their work efficiency and accuracy in scientific research activities. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 Schematic diagram of the classification system module of the present application;

[0061] Figure 2 Flow chart of the classification process of the present application;

[0062] Figure 3 Structure diagram of the classification model. DETAILED DESCRIPTION

[0063] To make the purpose, content and advantages of the present application clearer, the specific embodiments of the present application are described in further detail below.

[0064] The literature data multi-label automatic classification method proposed by the present application includes the following steps:

[0065] Step one, original data collection:

[0066] The data used in the application is collected from different websites and journals. In this part, the crawler technology is used to complete the aggregation of the original data, which is in the form of a csv data table.

[0067] Step two, text preprocessing:

[0068] Since the directly collected data is multi-source and the data quality is difficult to guarantee, the collected original data needs to be cleaned. Specifically, the content is blocked according to the structure of the literature, including the title, abstract, keywords, and free words. The default value is filled to complete the missing values, and the irrelevant content generated in the collection process is removed, including HTML tags, headers and footers.

[0069] Step three, pseudo-labeling the preprocessed text as a supervised data set for classification model training. Preferably, the pseudo-labeled text can be manually reviewed as a supervised data set;

[0070] The specific steps are as follows:

[0071] S1, a plurality of information sources are used to construct a literature type feature set, and data preprocessing is performed, the literature type feature set includes the title, abstract, keywords, free words and classification number of the literature;

[0072] S2, classify the preprocessed literature type and determine the literature type, the classification method includes character similarity and semantic similarity strategy;

[0073] S3, comprehensive integration, according to the weight, comprehensive decision of type information from different information sources, determine the final literature type.

[0074] S4, expert review and judgment, according to the result of S3, review the literature classification result and determine whether it is correct.

[0075] In S1, the data preprocessing includes removing punctuation, stop words, numbers and special characters, and using a dictionary-based word segmentation method combined with new word discovery technology to segment Chinese text.

[0076] In S2, for short text information of keywords and free words, a character similarity calculation method is used, a Chinese keyword dictionary and a synonym dictionary are constructed to expand the label set, and the similarity between the literature text and the type label is calculated.

[0077] In S3, the weight allocation method is to allocate weights to character similarity and semantic similarity respectively.

[0078] Step four, design a classification model for classifying the preprocessed text:

[0079] The classification model comprises an input layer, an encoding layer, a feature extraction layer, and a classification layer:

[0080] The input layer: takes the title (T), keywords (K), abstract (A), and free words (F) as input, alone or in combination, and uses special tokens [CLS] and [SEP] in the BERT model to separate different input parts;

[0081] Specifically, in order to maintain the consistency of the input, special tokens [CLS] and [SEP] in the BERT model are introduced, where [CLS] is used to represent the beginning of the sentence, and [SEP] is used to separate different input parts. For example, when T and K are input as a combination, it can be represented as X = [CLS] + T + [SEP] + K + [SEP].

[0082] The encoding layer: uses a pre-trained BERT model to map text into a vector representation; that is, using a pre-trained BERT model, the characters are mapped to a vector representation space to obtain an n*d dimensional matrix, where n represents the length of the character sequence, and d is the vector dimension, which is generally 768 by default for the BERT model;

[0083] The feature extraction layer: includes LSTM network and CNN network arranged side by side, the feature vector obtained by the encoding layer is input into the LSTM network and the CNN network at the same time, the LSTM network is used to capture the sequence dependence features in the text sequence, and the CNN is used to extract the local features of the text through the sliding of the convolution kernel. After the features extracted by the two networks are fused, they are put into a multi-layer perceptron (MLP) for further feature fusion and extraction. The fused features have both long dependence features and local features.

[0084] The classification layer: outputs multi-label classification probabilities through a Sigmoid activation function.

[0085] The classification model processing flow is as follows: first, use a pre-trained BERT model to map characters to a vector representation space, X = {x1, x2,..., xn}, where n is the length of the character sequence, and d is the vector dimension, which is generally 768 by default for the BERT model; n After that, the feature vector is simultaneously passed through the LSTM network and the CNN network, the LSTM network is used to capture the long-term dependence in the text sequence, and the CNN is used to learn the local features of the text through the sliding of the convolution kernel. After the features extracted by the two networks are fused, they are put into a multi-layer perceptron (MLP) for further feature fusion and extraction. The fused features have both long dependence features and local features. The specific steps are as follows:

[0086] a. Semantic representation vectorization: When processing the text sequence X using the BERT model, the text sequence is segmented by a Chinese tokenizer (such as BertTokenizer) to obtain a set of text sequences X = {x1, x2,..., xn}, and then the BERT module is used, which simultaneously considers word embedding, paragraph embedding and position embedding, to map the text sequence to a d-dimensional vector space, obtaining a matrix of n*d dimensions, where n represents the length of the character sequence. n} and then using the BERT module, which simultaneously considers word embedding, paragraph embedding and position embedding, to map the text sequence to a d-dimensional vector space, obtaining a matrix of n*d dimensions, where n represents the length of the character sequence.

[0087] x i = PLM(x i )

[0088] where PLM is the BERT model and x i is the language representation vector.

[0089] b. LSTM feature extraction: The Bi-LSTM model is composed of two independent LSTM networks, which process the input sequence in forward and reverse order respectively. Based on the semantic representation obtained by the BERT model, the vector representation of each character is taken as the input of the Bi-LSTM. At each time step, the forward LSTM and the reverse LSTM output a hidden state vector respectively, and the vector is spliced to obtain the spliced hidden state vector at this time step. In this method, the last layer hidden state of the Bi-LSTM network is taken as the feature representation of the entire sequence. The LSTM network can effectively capture long dependencies in the sequence by introducing input gate, forget gate, output gate and cell state mechanism. Each part of the calculation process is as follows:

[0090] The input gate determines whether new information is added to the cell state, and the calculation process is as follows:

[0091] i t = σ(W i ·[h t-1 ,x t ]+b i )

[0092] where W i is a learnable weight and b i is a bias vector.

[0093] The forget gate determines the information retained in the cell state at the previous time step, and the calculation process is as follows:

[0094] f t = σ(W f ·[h t-1 , x t ]+b f )

[0095] where W f is a learnable weight and b fis the bias vector, h t-1 is the hidden state of the previous time step, x t is the input of the current time step.

[0096] The output gate is used to determine the hidden state of the current time step o t The information to be carried is calculated as follows:

[0097] o t = σ(W o · [h t-1 , x t ] + b o )

[0098] where W o is the learnable weight, and b0 is the bias vector.

[0099] The candidate memory cell calculation process is as follows:

[0100]

[0101] where W C is the learnable weight, and b C is the bias vector.

[0102] Memory cell update, the calculation process is as follows:

[0103]

[0104] C t-1 is the memory cell state of the previous time step, and ⊙ represents element multiplication. C t is the memory cell state of the current time step

[0105] The hidden state calculation process is as follows:

[0106] h t = o t ⊙ tanh(C t )

[0107] For the forward LSTM, the hidden state of each time step is calculated in sequence according to the time step For the backward LSTM, the hidden state of each time step is calculated in sequence according to the time step in reverse order

[0108] In each time step t, the hidden states of the forward and backward LSTMs are concatenated to obtain the vector representation of each time step

[0109]

[0110] The vector representation of the entire sequence is:

[0111]

[0112] c. CNN feature extraction: Add a channel dimension to the output vector generated by the BERT model, convert the three-dimensional tensor into a four-dimensional tensor, use a convolution kernel of size 3 to capture local features (the appropriate convolution kernel size can be set according to the application scenario), and then apply the maximum pooling operation to the feature map generated by each convolution kernel to extract the most important features in each convolution kernel. The features extracted by different convolution kernels are spliced ​​together to form a comprehensive feature vector as the feature representation of the sequence.

[0113] The calculation process is as follows,

[0114] x′ i =x i unsqueeze (-1)

[0115] Where x′ i It is the feature representation after the number of channels is increased.

[0116] The convolution operation is expressed as follows:

[0117] C = Conv2D(x′ i )

[0118] The maximum pooling calculation process is as follows:

[0119] P=MaxPool2d(C)

[0120] Assume there are C convolution kernels, and after maximum pooling, we get C features c1, c2, ..., c c . These features are stitched together to form a comprehensive feature vector Z C :

[0121] Z C =[z1,z2,…,z c ]

[0122] d. Feature fusion: The output features of the CNN path and the LSTM path are flattened, expanded, concatenated, and fused to obtain feature representations for text classification.

[0123] The calculation process is as follows:

[0124] F=[H bi ; Z C ]

[0125] F is the fused vector representation.

[0126] c. Enter the classification layer to realize the text type classification: After the feature extraction layer, the classification layer is introduced to realize the mapping of the high-dimensional feature vector to the category label space. Specifically, the vector dimension is changed by alternately stacking the full connection layer and the dropout layer, and then the Sigmoid activation function is used to convert the elements of each output vector into a value between 0 and 1 to obtain the probability that the literature belongs to each category, and the type classification is realized.

[0127] The linear layer calculation process is as follows:

[0128] L = W h F + b h

[0129] Wherein, W h and b h are the weight and bias of the Linear layer, and L is the output of the Linear layer.

[0130] The Sigmoid function is used, and the probability calculation process is as follows:

[0131]

[0132] Wherein, p i is the probability that the literature belongs to the i-th type.

[0133] Step five, train the classification model according to the data set obtained in step three.

[0134] The existing data can also be directly trained, but the pseudo-labeling method in step three is used to automatically generate pseudo-labels through pseudo-labeling data to form a pseudo-labeling data set, which can effectively reduce the dependence on manual labeling. At the same time, through the pseudo-label reasoning mechanism, the classification standard difference of different institutions can be dynamically mapped, the pseudo-labeling data set is used for model semi-supervised learning training, and the samples with a training confidence higher than a certain value are iteratively put into the training set (generally, the confidence value will decrease with the increase of the iteration number, and the initial confidence value is set to 80%), and the classification performance of the model is enhanced. This method can be applied to open scene classification tasks.

[0135] Step six, input the data obtained in step two into the trained classification model to realize text classification and visual display: according to the classification result, the literature is distributed to multiple categories, and the literature information is displayed in the literature management system, including title, keyword, author and other information, supporting online index and query of the literature.

[0136] The literature data multi-label automatic classification system provided by the application comprises a data acquisition module, a preprocessing module, a pseudo-labeling module, a classification module and a visual display module.

[0137] Data collection module: for collecting literature data from multi-source websites and journals through crawler technology; including data collection module for collecting literature data;

[0138] Preprocessing module: for cleaning and block processing of literature data, extracting title, abstract, keyword and free word; further configured to: complete missing data by filling default values, and remove HTML tags and header and footer.

[0139] Pseudo-labeling module: based on character similarity, semantic similarity and weight distribution to generate pseudo-labeling dataset;

[0140] Classification module: including the classification model described in the above method, for multi-label classification of preprocessed text;

[0141] In the classification module: the Bi-LSTM network and the CNN network process the vector output by BERT in parallel, and realize space-time feature fusion through feature splicing. The classification layer adopts Sigmoid activation function, outputs the probability of literature belonging to each category, and realizes multi-label classification.

[0142] Visualization module: for displaying literature information in the literature management system according to the classification result, and supporting online index and query.

[0143] The trained classification model is embedded in the literature management system, and the data generated by the data collection module is classified by the classification model to obtain the literature type label, and then the literature is automatically divided into the corresponding theme module according to the theme, and the literature information is displayed in the theme module of the literature management system, including title, keyword, author and other information, supporting online index and query of literature.

[0144] The application also provides a computer readable storage medium, characterized in that it stores a computer program, which is executed by a processor to realize the above-mentioned literature data multi-label automatic classification method.

[0145] The application also provides an electronic device, characterized in that it includes a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the literature data multi-label automatic classification method.

[0146] Experimental result analysis:

[0147] The total number of experimental literature dataset entries is 1385, and the literature comes from the Journal of Explosives and Propellants, Energetic Materials, Journal of Safety and Environment, Chemical Propellants and Polymeric Materials, and Propulsion Technology. According to the research content and field characteristics, the literature is defined as the following nine types: synthesis material technology, high-energy explosive technology, propellant technology, gun propellant technology, engine technology, explosive common basic technology, warhead technology, explosion damage and evaluation technology, and others. The literature data distribution is analyzed, and according to the above type order, if the literature belongs to the type, the corresponding label position is 1, otherwise it is 0, and the statistical results are shown in the following table.

[0148]

[0149] The Chinese pre-training language model BERT-base-Chinese, i.e. the model learned by using Chinese Wikipedia data pre-training, is used to learn the semantic representation information of the literature. In the training process, the training period epochs is set to 10, and the early stopping method is used to avoid overfitting, the batch size batch_size is set to 4, and the learning rate is set to 1e-5. The proposed method is compared with three baseline models, which are BERT, BERT+CNN and BERT+LSTM. The model input is uniform as title+keyword+free word, and the experimental results are as follows:

[0150] Model P R F BERT 0.5742 0.5664 0.5703 BERT+CNN 0.5984 0.6125 0.6054 BERT+LSTM 0.6124 0.6015 0.6069 The method of the invention 0.6268 0.6593 0.6267

[0151] From the above table, it can be clearly seen that in the three key evaluation indicators of accuracy, recall rate and F1 value, the method proposed in this paper is better than the other three baseline models. It shows that the method in this paper can more accurately capture the features and semantic information of the literature, thereby improving the accuracy of literature type prediction.

[0152] The above is only the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and modifications can be made, these improvements and modifications should also be regarded as the protection scope of the present application.

Claims

1. A multi-label automatic classification method for document data, characterized in that: The following steps are involved: Step 1: Raw data collection: Use crawler technology to collect literature data from multiple source websites and journals and store them as structured data tables; Step 2: Text preprocessing: Clean the collected literature data, including extracting titles, abstracts, keywords and free words in blocks, filling missing values ​​and removing irrelevant content; Step 3: Build a classification model: a. Input layer: For titles, keywords, abstracts, and free words input individually or in combination, the BERT model uses special markers to separate different input parts; b. Encoding layer: Use the pre-trained BERT model to map text into vector representation; c. Feature extraction layer: The Bi-LSTM network and CNN network are used in parallel to extract the sequence-dependent features and local features of the text respectively. The sequence features output by the Bi-LSTM are combined with the local features output by the CNN to form a comprehensive feature. d. Classification layer: compresses and abstracts the fused features through a multi-layer perceptron, and outputs multi-label classification probabilities through a Sigmoid activation function; Step 4: Train the classification model; Step 5: Input the literature data preprocessed in step 2 into the trained classification model and output the multi-label classification results.

2. The method for automatic multi-label classification of document data according to claim 1, characterized in that: The text generated in step 2 is processed by a pseudo-annotation method to generate a supervised dataset, and the classification model in step 3 is trained. The pseudo-annotation method steps are as follows: S1. Construct a document type feature set, segment the preprocessed text, and remove stop words and special characters; S2, based on character similarity and semantic similarity strategies, preliminary classification of document types; S3. Integrate the classification results of different information sources, determine the final document type label through weight distribution, and form a pseudo-annotated dataset.

3. The multi-label automatic classification method for document data according to claim 1, characterized in that: The Bi-LSTM network is a bidirectional long short-term memory network that captures contextual dependencies through forward and reverse sequence processing and uses the last layer of hidden states as the overall feature representation of the sequence.

4. The method for automatic multi-label classification of document data according to claim 3, characterized in that: The LSTM network effectively captures long dependencies in sequences by introducing input gates, forget gates, output gates, and cell state mechanisms. The calculation process of each part is as follows: The input gate determines whether new information is added to the cell state. The calculation process is as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ) Among them, W i is the learnable weight, b i is the bias vector; The forget gate determines the information retained in the cell state at the previous moment. The calculation process is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) Among them, W f is the learnable weight, b f is the bias vector, h t-1 is the hidden state of the previous moment, x t is the input at the current moment; The output gate is used to determine the current hidden state o t The information to be carried is calculated as follows: the t =σ(W o ·[h t-1 ,x t ]+b o ) Among them, W o is the learnable weight, b0 is the bias vector; The candidate memory unit calculation process is as follows: Among them, W C is the learnable weight, b C is the bias vector; The memory unit is updated and the calculation process is as follows: C t-1 is the memory cell state at the previous moment, ⊙ represents element multiplication, C t is the current state of the memory cell; The hidden state calculation process is as follows: h t =o t ⊙tanh(C t ) For the forward LSTM, the hidden state of each time step is calculated in sequence according to the time step order For reverse LSTM, the hidden state of each time step is calculated in reverse order of the time steps At each time step t, the hidden states of the forward and reverse LSTM are concatenated to obtain the vector representation of each moment The vector representation of the entire sequence is:

5. The method for automatic multi-label classification of document data according to claim 1, characterized in that: The CNN network uses multi-size convolution kernels to extract local features of text and generates a comprehensive feature vector through maximum pooling operation.

6. A multi-label automatic classification system for document data, characterized in that: Includes the following modules: Data collection module: collects multi-source literature data through crawler technology; Preprocessing module: cleans and processes literature data in blocks, extracts titles, abstracts, keywords and free words; Pseudo-annotation module: Generates pseudo-annotation dataset based on character similarity, semantic similarity and weight distribution; Classification module: comprising the classification model according to any one of claims 1 to 5, for performing multi-label classification on the preprocessed text; Visualization module: displays classification results and supports online indexing and query of documents.

7. The multi-label automatic classification system for document data according to claim 6, characterized in that: In the classification module, the output features of Bi-LSTM and CNN are flattened and concatenated and then input into MLP for dimensionality reduction.

8. The multi-label automatic classification system for document data according to claim 6, characterized in that: The pseudo-annotation module integrates a manual review function to correct the quality of the pseudo-annotation labels.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for automatic multi-label classification of document data according to any one of claims 1 to 5 is implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein: When the processor executes the computer program, the method for automatic multi-label classification of document data according to any one of claims 1 to 5 is implemented.

Images