A Multi-Label Text Classification Method and System Based on Multi-Information Filtering Encoding

Abstract

The present invention discloses a multi-label text classification method and system based on multi-information filtering coding, the method comprising: obtaining a text training set containing labels; obtaining word vectors of text and labels based on the text training set; filtering and feature extraction of text information and label information through multi-information filtering coding to obtain filtered text features and label features; obtaining a final text vector representation through interaction of filtered text features and filtered label features, combined with filtered text features, to train a multi-label text classification model; and performing multi-label text classification on a text data set to be classified using the trained multi-label text classification model. The present invention extracts semantic features in the filtered information in the text space and the label space respectively by filtering noise and redundant information of the text information and label information, and captures more information related to classification through interaction of filtered semantic features, thereby improving the accuracy of multi-label text classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to a multi-label text classification method and system based on multi-information filtering coding. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Multi-label text classification (MLTC) is a key task in natural language processing, which is widely used in fields such as sentiment recognition, question answering, and web page tagging. It aims to learn a model that can assign appropriate multiple labels to unknown text. However, the number of documents and labels and the complex relationship between labels pose challenges to multi-label text classification.

[0004] Traditional multi-label text classification usually adopts methods to learn semantically enhanced representations of text, such as the deep learning method for extreme multi-label classification (XML-CNN), which combines improved convolutional neural networks and dynamic maximum pooling technology to extract high-level features to learn compact document representations; the attention-aware deep model based on label trees (AttentionXML), which uses recursive neural networks to encode text; and the Transformer-based model captures the relevance of words in a sequence to provide contextual representations. Although the above methods fully analyze the text content, they ignore information outside the text, resulting in poor accuracy in the final text classification.

[0005] To this end, the existing technology integrates text and other information for multi-label text classification. On the one hand, considering that in multi-label text classification, since labels share the same document subset, there is a semantic connection between documents and labels, so there is a dependency relationship between labels. Therefore, in the existing technology, the multi-label text classification problem is solved by learning the label structure, label content meaning and label co-occurrence pattern. For example, the Sparse Local Embedding Method for Extreme Multi-Label Classification (SLEEC) divides the dataset into multiple clusters and detects the embedding vector in each cluster by capturing nonlinear label correlations; the Label Specific Dual Graph Neural Network (LDGN) learns label-specific components from documents with the help of category information, and uses a dual graph convolutional network (GCN) to model the interaction between these components by jointly reconstructing the graph dynamically and counting label co-occurrences.

[0006] On the other hand, multi-label data usually has a long-tail distribution problem, that is, the class imbalance problem. When the number of labels is too large, a small subset of labels (called head labels) has a large number of documents, and a large subset of labels (called tail labels) has only a few documents. For commonly used prediction performance indicators, the impact of tail labels is much smaller than that of head labels. Therefore, existing technologies capture features by analyzing information other than text to alleviate the long-tail distribution problem, such as knowledge transfer from head labels to tail labels, tail label or head label instance sampling strategies, etc.

[0007] However, in the prior art, when analyzing text and information outside the text (such as label information) and performing multi-label text classification, it should be noted that there are some noise or redundant content in the document and label information. The prior art ignores some complex semantic information hidden in the noise or redundant content, which will affect the acquisition of semantic features and connections between labels and documents, resulting in a decrease in the accuracy of the final multi-label text classification. Summary of the invention

[0008] To address the deficiencies of the above-mentioned prior art, the present invention provides a multi-label text classification method and system based on multi-information filtering coding, which filters the noise and redundant information of text information and label information in multi-label text classification data, and extracts semantic features in the filtered information in the text space and the label space respectively. Through the interaction of filtered semantic features, more classification-related information is captured, the influence of filtering noise and redundant information on semantic feature extraction is reduced, the long-tail distribution problem is solved, and the accuracy and precision of multi-label text classification are improved.

[0009] In a first aspect, the present disclosure provides a multi-label text classification method based on multi-information filtering coding.

[0010] A multi-label text classification method based on multi-information filtering coding, comprising:

[0011] Obtain the text dataset to be classified and the text training set containing multiple labels;

[0012] Vectorize the text in the text training set and the labels in the label set corresponding to the text training set respectively to obtain the word vectors of the text and the word vectors of the label;

[0013] Based on the vector representation of text and label, the text information and label information are filtered and feature extracted through multi-information filtering coding to obtain filtered text features and filtered label features;

[0014] By interacting the filtered text features with the filtered label features, we get the text-specific label representation, and combine it with the filtered text features to get the final text vector representation. Based on the final text vector representation, we train a multi-label text classification model.

[0015] Use the trained multi-label text classification model to perform multi-label text classification on the text dataset to be classified.

[0016] A further technical solution also includes: performing data preprocessing on the acquired text data set to be classified and the text training set containing labels; the preprocessing includes cleaning non-text data, removing stop words and restoring morphology.

[0017] A further technical solution is to filter and extract features of text information and tag information based on vector representation of text and tags through multi-information filtering coding to obtain filtered text features and filtered tag features, specifically:

[0018] Based on the word vector of the text, the text information is filtered and feature extracted through text information filtering coding to obtain filtered text features;

[0019] Based on the word vector of the tag, the tag information is filtered and the feature is extracted through the tag information filtering encoding to obtain the filtered tag feature.

[0020] A further technical solution is to filter and extract features of text information based on the word vector of the text through text information filtering coding to obtain filtered text features. The specific steps are:

[0021] Based on the text word vector, the word vector is converted to the frequency domain through fast Fourier transform;

[0022] Modulate the spectrum in the frequency domain, filter the word vectors of the text, and obtain the filtered text information;

[0023] The filtered text information is converted to the time domain by inverse fast Fourier transform and its representation is updated;

[0024] Based on the word vector of the text and the filtered and updated text information, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered text vector;

[0025] Based on the filtered text vector, the initial filtered text features are obtained through multi-layer perception processing;

[0026] Based on the filtered text vector and the initially filtered text features, residual connection, Dropout and layer normalization operations are performed in sequence to obtain the filtered text features.

[0027] A further technical solution is to filter and extract the label information based on the label word vector through label information filtering coding to obtain the filtered label features. The specific steps are as follows:

[0028] Based on the label word vector, the word vector is converted to the frequency domain through fast Fourier transform;

[0029] Modulate the spectrum in the frequency domain, filter the word vector of the label, and obtain the filtered label information;

[0030] The filtered label information is converted to the time domain through inverse fast Fourier transform and its representation is updated;

[0031] Based on the word vector of the label and the filtered and updated label information, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered label vector;

[0032] Based on the filtered label vector, the initial filtered label features are obtained through multi-layer perception processing;

[0033] Based on the filtered label vector and the initially filtered label features, residual connection, Dropout and layer normalization operations are performed in sequence to obtain the filtered label features.

[0034] A further technical solution is to obtain a text-specific label representation by interacting the filtered text features with the filtered label features. The specific steps are as follows:

[0035] Based on the filtered text features and the filtered label features, an attention value between the filtered text features and the filtered label features is obtained through an attention mechanism guided by filtering information;

[0036] The filtered tag features are weighted based on the attention values ​​to obtain text-specific tag representations.

[0037] A further technical solution combines the filtered text features to obtain the final text vector representation, and trains a multi-label text classification model based on the final text vector representation. The specific steps are:

[0038] Concatenate the filtered text features and the text-specific tag representation to obtain concatenated features;

[0039] The concatenated features are weighted to obtain weighted features, and the weighted features are subjected to averaging and maximum pooling operations to obtain averaging and maximum pooling features.

[0040] Based on the averaged features and the maximum pooled features, the features are merged to obtain the final text vector representation.

[0041] In a second aspect, the present disclosure provides a multi-label text classification system based on multi-information filtering coding, comprising:

[0042] A data acquisition module is used to obtain a text data set to be classified and a text training set containing multiple labels;

[0043] A vector conversion module is used to vectorize the text in the text training set and the labels in the label set corresponding to the text training set, respectively, to obtain the word vectors of the text and the word vectors of the label;

[0044] The feature acquisition module is used for filtering and extracting features of text information and label information based on word vectors of text and labels through multi-information filtering coding to obtain filtered text features and filtered label features;

[0045] A multi-label text classification model training module is used to obtain a text-specific label representation through the interaction of filtered text features and filtered label features, obtain a final text vector representation by combining the filtered text features, and train a multi-label text classification model based on the final text vector representation;

[0046] The classification module is used to perform multi-label text classification on the text dataset to be classified using the trained multi-label text classification model.

[0047] In a third aspect, the present disclosure further provides an electronic device, comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method described in the first aspect are completed.

[0048] In a fourth aspect, the present disclosure further provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the method described in the first aspect.

[0049] One or more of the above technical solutions have the following beneficial effects:

[0050] 1. The present invention proposes a multi-label text classification method and system based on multi-information filtering coding. Through learnable text information filtering coding and learnable label information filtering coding, useless information in text representation and label representation is filtered out, and semantic features are extracted in text space and label space respectively, thereby reducing useless information in text representation and label representation and extracting more complex semantic features.

[0051] 2. The present invention proposes a multi-label text classification method and system based on multi-information filtering coding, which guides the attention process by utilizing filtering features, generates different text-specific label representations for each text, and highlights the important features in the filtered text features and text-specific label representations through maximum pooling and averaging operations to obtain the final document representation. The final document representation is used to train a multi-label text classification model to make the classification results of multi-label text more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0053] Figure 1 This is an overall framework diagram of the method described in Example 1 of the present invention;

[0054] Figure 2 The figure is an overall flow chart of the method described in the first embodiment of the present invention. DETAILED DESCRIPTION

[0055] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0056] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0057] Embodiment 1

[0058] This embodiment provides a multi-label text classification method based on multi-information filtering coding, such as Figure 1 As shown, the following steps are included:

[0059] Step 1: Obtain the text data set to be classified and the text training set containing multiple labels;

[0060] Step 2: vectorize the text in the text training set and the label in the label set corresponding to the text training set respectively to obtain the word vector of the text and the word vector of the label;

[0061] Step 3: Based on the vector representation of text and label, the text information and label information are filtered and feature extracted by multi-information filtering coding to obtain filtered text features and filtered label features;

[0062] Step 4: obtain text-specific label representation through interaction between filtered text features and filtered label features, obtain the final text vector representation by combining the filtered text features, obtain the predicted label of the text based on the final text vector representation, train the multi-label text classification model multiple times according to the difference between the predicted label of the text and the true label of the text, update the model parameters, and thus obtain the trained multi-label text classification model;

[0063] Step 5: Use the trained multi-label text classification model to perform multi-label text classification on the text dataset to be classified.

[0064] In step 1 of the method described in this embodiment, a text data set to be classified and a text training set containing labels are obtained. The text data set to be classified includes multiple texts to be classified, and the text training set containing labels includes multiple texts with artificial labels. It should be noted that a text may be considered to be a topic related to religion, politics, finance or education at the same time, that is, the same text can have multiple labels.

[0065] Preferably, the acquired text data set to be classified and the text training set containing labels are subjected to data preprocessing, and the preprocessing includes cleaning non-text data, removing stop words and restoring morphology, specifically including:

[0066] Step 1.1: Clean up non-text data. Text data may contain non-text content such as HTML tags, URL addresses, punctuation marks, etc. These contents are of no help to classification and may even have a negative impact. Therefore, the non-text data in the text data should be removed;

[0067] Step 1.2: Remove stop words. Auxiliary verbs and function words exist in large quantities in all texts, but they cannot reflect the meaning of the text. Generally speaking, removing these stop words from the text can make the model better fit the semantic features, thereby improving the generalization ability of the model;

[0068] Step 1.3: Word form restoration. For example, although the words am, is, and are have the same meaning, they are completely different contents for computers. Word form restoration can unify them to facilitate subsequent processing and analysis.

[0069] After the data preprocessing is completed, step 2 is performed, that is, the texts in the text training set and the labels in the label set corresponding to the text training set are vectorized respectively to obtain the word vectors of the texts in the text training set and the word vectors of the labels in the label set.

[0070] For the text training set, based on the preprocessed text in the training set, each word in the text in the training set is converted into its corresponding vector representation. At the same time, based on the label set of the training set, the label is randomly initialized into the form of a vector.

[0071] After the texts in the text training set and the labels in the label set corresponding to the text training set are vectorized, a multi-label text classification model based on multi-information filtering coding is constructed, namely, the following steps 3 and 4.

[0072] like Figure 2As shown, in step 3, based on the word vectors of the text and the label, the text information and the label information are filtered and feature extracted through multi-information filtering coding to obtain filtered text features and filtered label features. Specifically, it includes:

[0073] Based on the word vector of the text, the text information is filtered and feature extracted through text information filtering coding to obtain filtered text features;

[0074] Based on the word vector of the tag, the tag information is filtered and the feature is extracted through the tag information filtering encoding to obtain the filtered tag feature.

[0075] First, the specific steps of the above text information filtering and encoding include:

[0076] Step 3.1.1: Based on the word vector of the text, the word vector is converted to the frequency domain through fast Fourier transform to obtain the text word vector represented in the frequency domain. The conversion process is:

[0077]

[0078] Where f(.) represents one-dimensional fast Fourier transform, F word word vectors representing the input text, Where R represents a real number, C represents a complex number, v represents the number of word vectors of the text, and d x Represents the dimension of the word vector, X word is a complex tensor.

[0079] Step 3.1.2: Modulate the spectrum in the frequency domain through the learnable text filter to filter the text word vector represented in the frequency domain, specifically:

[0080]

[0081] in, represents a learnable text filter, which is adaptively optimized during model training. ⊙ represents a dot product operation. Represents the filtered text information.

[0082] Step 3.1.3: Convert the filtered text information to the time domain and update its representation through inverse fast Fourier transform, specifically:

[0083]

[0084] Among them, f -1 (.) represents inverse fast Fourier transform, Represents the converted text information, that is, the text information after filtering and updating.

[0085] Step 3.1.4: Based on the word vector of the text and the filtered and updated text information, perform residual connection, Dropout and layer standardization operations in sequence to obtain the filtered text vector Specifically:

[0086]

[0087] Among them, LayerNorm means layer normalization, and Dropout means randomly discarding a part of neurons.

[0088] Step 3.1.5: Based on the filtered text vector, in the forward propagation layer, after being processed by a multi-layer perceptron with Gelu as the activation function, the initial filtered text features are obtained. Specifically:

[0089]

[0090] Among them, Gelu is the activation function, W1, b1, W2, and b2 are learnable parameters.

[0091] Step 3.1.6: Based on the filtered text vector and the initially filtered text features, residual connection, dropout and layer standardization operations are performed in sequence to obtain the filtered text features H, which are specifically:

[0092]

[0093] Secondly, the specific steps of the above tag information filtering and encoding include:

[0094] Step 3.2.1: Based on the word vector of the label, the word vector is converted to the frequency domain through fast Fourier transform to obtain the label word vector represented in the frequency domain. The conversion process is:

[0095]

[0096] Where f(.) represents one-dimensional fast Fourier transform, F label Represents the word vector of the input label, R represents a real number, Where R represents a real number, C represents a complex number, L represents the number of word vectors of the labels in the label set corresponding to the text training set, and d x Represents the dimension of the word vector, X label is a complex tensor.

[0097] Step 3.2.2: Modulate the spectrum in the frequency domain through the learnable label filter to filter the label words represented in the frequency domain. Specifically:

[0098]

[0099] in, represents a learnable label filter, which is adaptively optimized during model training. ⊙ represents a dot product operation. Indicates the filtered tag information.

[0100] Step 3.2.3: Convert the filtered label information to the time domain and update its representation through inverse fast Fourier transform, specifically:

[0101]

[0102] Among them, f -1 (.) represents inverse fast Fourier transform, Indicates the converted label information, that is, the filtered and updated label information.

[0103] Step 3.2.4: Based on the word vector of the label and the filtered and updated label information, perform residual connection, Dropout and layer normalization operations in sequence to obtain the filtered label vector. Specifically:

[0104]

[0105] Among them, LayerNorm means layer normalization, and Dropout means randomly discarding a part of neurons.

[0106] Step 3.2.5: Based on the filtered label vector, in the forward propagation layer, after being processed by a multi-layer perceptron with Gelu as the activation function, the initial filtered label feature is obtained. Specifically:

[0107]

[0108] Among them, Gelu is the activation function, W3, b3, W4, and b4 are learnable parameters.

[0109] Step 3.2.6: Based on the filtered label vector and the initially filtered label feature, residual connection, Dropout and layer normalization operations are performed in sequence to obtain the filtered label feature G, which is specifically:

[0110]

[0111] In step 4, the text-specific label representation is obtained by interacting the filtered text features with the filtered label features, and the final text vector representation is obtained by combining the filtered text features. Based on the final text vector representation, a multi-label text classification model is trained.

[0112] In order to generate text representation by associating text information with tag information, filtered features are used to guide the attention process and generate different text-specific tag representations to enhance document representation. Specifically, based on the filtered text features and the filtered tag features, the attention value between the filtered text features and the filtered tag features is obtained through an attention mechanism guided by the filtered information, and the filtered tag features are weighted based on the attention value to obtain the text-specific tag representation.

[0113] In this embodiment, further, based on the filtered label features and the filtered text features, the attention score between the filtered label features and the filtered text features is calculated using the attention mechanism guided by the filtered features. The specific steps are:

[0114] Step 4.1: Use the attention mechanism to obtain the attention score α of the filtered feature, specifically:

[0115]

[0116] Among them, H j Represents the jth word vector in the filtered text feature, G i Represents the i-th label vector in the filtered label feature, () T represents the transpose operation, and exp() is the exponential function.

[0117] Step 4.2: Use the attention value to weight the filtered label features to obtain the text-specific label representation, specifically:

[0118]

[0119] Then, the filtered text features and text-specific label representations are concatenated to obtain concatenated features; the concatenated features are weighted to obtain weighted features, and the weighted features are respectively subjected to averaging and maximum pooling operations to obtain averaging features and maximum pooling features; based on the averaging features and maximum pooling features, the features are combined to obtain the final text vector representation, further highlighting the important role of key information. The specific steps are as follows:

[0120] Step 4.3: Concatenate the filtered text features and the text-specific label representation to obtain the concatenated features. The formula is:

[0121] P i =[H i ; U i ] (13)

[0122] Among them, [x; y] means that two types of components are spliced ​​along a certain dimension.

[0123] Step 4.4: Weight the concatenated features to obtain weighted features. The formula is:

[0124] A=(W5P+b5) (14)

[0125] Among them, W5 and b5 are trainable parameters.

[0126] Step 4.5: Perform mean and maximum pooling operations on the weighted features along a certain dimension to obtain mean features and maximum pooling features. The formula is:

[0127]

[0128] P m =Maxpooling(A1...A v ) (16)

[0129] Among them, P s is the feature representation obtained by averaging along the 0th dimension of the feature, P m It is the feature representation obtained by performing maximum pooling along the 0th dimension of the feature.

[0130] Step 4.6: Merge features to get the final text vector representation P o , the formula is:

[0131] P o =P s +P m (17)

[0132] The final text vector representation is obtained through the above scheme, and a multi-label text classification model is trained based on the final text vector representation. Specifically, based on the final text vector representation P o , the predicted label y of the text is obtained through the Sigmoid layer prediction pred ,y pred It represents the probability value of the multi-label text classification model predicting that the text belongs to each category. The Sigmoid layer is actually a Sigmoid activation function, and the formula is:

[0133]

[0134] Based on the predicted labels of the text obtained through the Sigmoid layer, the multi-label text classification model is trained multiple times using back propagation, the difference between the predicted text labels and the true text labels is compared, and the parameters of the multi-label text classification model are updated to obtain a trained multi-label text classification model.

[0135] That is to say, in this embodiment, the multi-information filtering encoding of the multi-label text classification model is used to filter and extract features of text information and label information, and the attention mechanism is used to interact with text and label features to obtain a vector representation of the text, thereby training the multi-label text classification model and obtaining a trained multi-label text classification model.

[0136] In step 5, the trained multi-label text classification model is used to perform multi-label text classification on the text dataset to be classified.

[0137] The multi-label text classification method in this embodiment filters out useless information in text representation and label representation through learnable text information filtering coding and learnable label information filtering coding, and extracts semantic features in text space and label space respectively, thereby reducing useless information in text representation and label representation and extracting more complex semantic features. At the same time, by using the filtered features to guide the attention process, different text-specific label representations are generated for each text, and the important features in the filtered text features and text-specific label representations are highlighted through maximum pooling and averaging operations to obtain the final document representation. The multi-label text classification model is trained using the final document representation to make the classification results of the multi-label text more accurate.

[0138] Embodiment 2

[0139] This embodiment provides a multi-label text classification system based on multi-information filtering coding, including:

[0140] A data acquisition module is used to obtain a text data set to be classified and a text training set containing multiple labels;

[0141] A vector conversion module is used to vectorize the text in the text training set and the labels in the label set corresponding to the text training set, respectively, to obtain the word vectors of the text and the word vectors of the label;

[0142] The feature acquisition module is used for filtering and extracting features of text information and label information based on word vectors of text and labels through multi-information filtering coding to obtain filtered text features and filtered label features;

[0143] A multi-label text classification model training module is used to obtain a text-specific label representation through the interaction of filtered text features and filtered label features, obtain a final text vector representation by combining the filtered text features, and train a multi-label text classification model based on the final text vector representation;

[0144] The classification module is used to perform multi-label text classification on the text dataset to be classified using the trained multi-label text classification model.

[0145] Embodiment 3

[0146] This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the multi-label text classification method based on multi-information filtering coding as described above are completed.

[0147] Embodiment 4

[0148] This embodiment also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps in the multi-label text classification method based on multi-information filtering coding as described above are completed.

[0149] The steps involved in the above embodiments 2 to 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0150] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0152] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A multi-label text classification method based on multi-information filtering coding, characterized in that: include: Obtain the text dataset to be classified and the text training set containing multiple labels; Vectorize the text in the text training set and the labels in the label set corresponding to the text training set respectively to obtain the word vectors of the text and the word vectors of the label; Based on the vector representation of text and labels, the text information and label information are filtered and feature extracted through multi-information filtering coding to obtain filtered text features and filtered label features, including: Based on the word vector of the text, the text information is filtered and feature extracted through text information filtering coding to obtain the filtered text features. The specific steps are: Based on the text word vector, the word vector is converted to the frequency domain through fast Fourier transform; Modulate the spectrum in the frequency domain, filter the word vectors of the text, and obtain the filtered text information; The filtered text information is converted to the time domain and its representation is updated by inverse fast Fourier transform; Based on the word vector of the text and the filtered and updated text information, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered text vector; Based on the filtered text vector, the initial filtered text features are obtained through multi-layer perception processing; Based on the filtered text vector and the initially filtered text features, residual connection, dropout and layer normalization operations are performed in sequence to obtain filtered text features; Based on the word vector of the tag, the tag information is filtered and feature extracted through tag information filtering encoding to obtain the filtered tag features. The specific steps are: Based on the word vector of the tag, the word vector is converted to the frequency domain through fast Fourier transform; Modulate the spectrum in the frequency domain, filter the word vector of the label, and obtain the filtered label information; The filtered label information is converted to the time domain through inverse fast Fourier transform and its representation is updated; Based on the word vector of the label and the filtered and updated label information, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered label vector; Based on the filtered label vector, the initial filtered label features are obtained through multi-layer perception processing; Based on the filtered label vector and the initially filtered label features, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered label features; By interacting the filtered text features with the filtered label features, we get the text-specific label representation, and combine it with the filtered text features to get the final text vector representation. Based on the final text vector representation, we train a multi-label text classification model. Use the trained multi-label text classification model to perform multi-label text classification on the text dataset to be classified.

2. A multi-label text classification method based on multi-information filtering coding as claimed in claim 1, characterized in that include: Perform data preprocessing on the acquired text data set to be classified and the text training set containing labels; The preprocessing includes cleaning non-text data, removing stop words and lemmatization.

3. The multi-label text classification method based on multi-information filtering coding as claimed in claim 1 is characterized in that: By interacting the filtered text features with the filtered label features, a text-specific label representation is obtained. The specific steps are as follows: Based on the filtered text features and the filtered label features, an attention value between the filtered text features and the filtered label features is obtained through an attention mechanism guided by filtering information; The filtered tag features are weighted based on the attention values ​​to obtain text-specific tag representations.

4. The multi-label text classification method based on multi-information filtering coding as claimed in claim 1 is characterized in that: Combine the filtered text features to obtain the final text vector representation. Based on the final text vector representation, train a multi-label text classification model. The specific steps are: Concatenate the filtered text features and the text-specific tag representation to obtain concatenated features; The concatenated features are weighted to obtain weighted features, and the weighted features are subjected to averaging and maximum pooling operations to obtain averaging and maximum pooling features. Based on the averaged features and the maximum pooled features, the features are merged to obtain the final text vector representation.

5. A multi-label text classification system based on multi-information filtering coding, characterized in that: include: A data acquisition module is used to obtain a text data set to be classified and a text training set containing multiple labels; A vector conversion module is used to vectorize the text in the text training set and the labels in the label set corresponding to the text training set, respectively, to obtain the word vectors of the text and the word vectors of the label; The feature acquisition module is used for word vectors based on text and labels. It filters and extracts features of text information and label information through multi-information filtering coding to obtain filtered text features and filtered label features, including: Based on the word vector of the text, the text information is filtered and feature extracted through text information filtering coding to obtain the filtered text features. The specific steps are: Based on the text word vector, the word vector is converted to the frequency domain through fast Fourier transform; Modulate the spectrum in the frequency domain, filter the word vectors of the text, and obtain the filtered text information; The filtered text information is converted to the time domain and its representation is updated by inverse fast Fourier transform; Based on the word vector of the text and the filtered and updated text information, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered text vector; Based on the filtered text vector, the initial filtered text features are obtained through multi-layer perception processing; Based on the filtered text vector and the initially filtered text features, residual connection, dropout and layer normalization operations are performed in sequence to obtain filtered text features; Based on the word vector of the tag, the tag information is filtered and feature extracted through tag information filtering encoding to obtain the filtered tag features. The specific steps are: Based on the word vector of the tag, the word vector is converted to the frequency domain through fast Fourier transform; Modulate the spectrum in the frequency domain, filter the word vector of the label, and obtain the filtered label information; The filtered label information is converted to the time domain through inverse fast Fourier transform and its representation is updated; Based on the word vector of the label and the filtered and updated label information, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered label vector; Based on the filtered label vector, the initial filtered label features are obtained through multi-layer perception processing; Based on the filtered label vector and the initially filtered label features, residual connection, dropout and layer normalization operations are performed in sequence to obtain the filtered label features; A multi-label text classification model training module is used to obtain a text-specific label representation through the interaction of filtered text features and filtered label features, obtain a final text vector representation by combining the filtered text features, and train a multi-label text classification model based on the final text vector representation; The classification module is used to perform multi-label text classification on the text dataset to be classified using the trained multi-label text classification model.

6. An electronic device, characterized in that: The invention comprises a memory and a processor and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of a multi-label text classification method based on multi-information filtering coding as claimed in any one of claims 1 to 4 are completed.

7. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of a multi-label text classification method based on multi-information filtering coding as described in any one of claims 1-4.

Images