A document tag generation method, system and readable storage medium
The document tags are generated through the article topic model and named entity extraction algorithm, which solves the problems of non-text document retrieval difficulties and confidential and private content security, and achieves fast and accurate document retrieval and security management.
Patent Information
- Application Number
- CN202110173106.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-02-08
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-02-08
AI Technical Summary
It is difficult for the prior art to quickly and accurately retrieve document content in non-text forms, especially confidential or private data when it is disclosed.
The document tag generation method based on the article topic model and named entity extraction algorithm is adopted to generate accurate correlation tags through automated tag library generation, tag classification, tag feedback mechanism and tag library maintenance to ensure the rapid retrieval and security of content.
It realizes automatic label initialization of a large number of documents and automatic label assignment of new documents, which improves the accuracy and security of retrieval, simplifies document management, and improves retrieval efficiency.
Smart Images

Figure CN113220872B_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to the field of information processing technology, and in particular to a document tag generation method, system and readable storage medium. [Background technology]
[0002] With the development of Internet technology and the rapid popularization of various network technologies, a large number of documents will be stored in media such as cloud disks. However, for these data that are not stored in text form, it is difficult for people to find the content they need in the article through direct retrieval or find the article they need through retrieval. This problem is especially serious when the content of the document is confidential or private and is not convenient to be made public. Therefore, we designed a document label generation method based on article topic model and text classification algorithm, which can meet the needs of fast and accurate retrieval of document content, and can also ensure the security of confidential and private content through content permission control.
[0003] Therefore, it is necessary to study a document tag generation method, system and readable storage medium to address the shortcomings of the prior art and to solve or alleviate one or more of the above problems. [Summary of the invention]
[0004] In view of this, the present invention provides a document tag generation method, system and readable storage medium, which can obtain accurate associated tags by processing document content.
[0005] In one aspect, the present invention provides a method for generating a document tag, the method comprising the following steps:
[0006] S1 Automatic tag library generation: Through the article topic model and named entity extraction algorithm, keyword tags and noun entity tags are obtained from a limited number of documents to generate an automatic tag library;
[0007] S2 label classification algorithm: Use the text multi-classification model to process the original documents and new documents in the system, match the classification results of the text multi-classification model with the automatic label library generated by S1, and generate corresponding labels;
[0008] S3 tag feedback mechanism: for documents that have not been tagged, the method in S1 is used again to obtain keyword tags and noun entity tags and add them to the automatic tag library;
[0009] Post-maintenance of the S4 tag library: Filter and / or deduplicate manually entered tags and add them to the automated tag library, and regularly train the text multi-classification model.
[0010] According to the above aspects and any possible implementation, an implementation is further provided, wherein S1 specifically includes:
[0011] S11: clustering a limited number of documents using the article topic model, extracting unique keywords in each category, obtaining keyword tags and storing them in a backup tag library;
[0012] S12: extract named entities from the documents in S11 using a named entity extraction algorithm, obtain noun entity tags, filter and store them in a backup tag library;
[0013] S13: performing deduplication and synonym merging processing on noun entity tags and keyword tags in the backup tag library to reduce the number of tags in the backup library;
[0014] S14: Process the labels according to preset conditions to generate an automated label library.
[0015] According to the above aspects and any possible implementation, an implementation is further provided, wherein the preset conditions in S14 include but are not limited to merging similar tags, and the merging similar tags specifically refers to expressing the same product with different names.
[0016] According to the above aspects and any possible implementation manner, an implementation manner is further provided, wherein S11 specifically includes:
[0017] S111: corpus preprocessing to obtain a candidate keyword set;
[0018] S112: Keywords in the document are ranked from high to low using tf-idf and textrank scoring methods, and stop words and parts of speech that do not show the topic meaning are eliminated. The resulting tokens file is used as the input file for LDA;
[0019] S113: Generate an LDA model based on the corpus, using the LDAmodel provided by gensim;
[0020] S114: Apply the LDA model in S113 to a certain document to obtain the topic probability distribution and keywords of the document. If the words obtained after the document segmentation are among the candidate keywords, they are put into the reserve library as keyword tags, otherwise they are included in the stop word library.
[0021] According to the above aspects and any possible implementation, an implementation is further provided, wherein the score of the candidate keyword in S114 = the probability of each topic * the probability of each word belonging to the topic.
[0022] According to the above aspects and any possible implementation manner, an implementation manner is further provided, wherein S12 specifically includes:
[0023] S121: corpus preprocessing to obtain clean corpus;
[0024] S122: Use the Mingming entity extraction model generated by the BERT+(LSTM)+CRF network to annotate the named entities appearing in the corpus;
[0025] S123: Calculate the frequency of named entities appearing in the document, and remove some named entities using preset rules;
[0026] S124: Put the noun entities within the preset frequency into the reserve library as noun entity tags.
[0027] According to the aspects described above and any possible implementation method, an implementation method is further provided, in which the training method of the text multi-classification model in S2 is: using a multi-level attention network to train the text multi-classification model.
[0028] According to the above aspects and any possible implementation, an implementation is further provided, wherein the text multi-classification model in S2 specifically includes:
[0029] Word encoding: Each word in the sentence is embedded into a word vector in the distribution vector space, and the sentence is encoded using a bidirectional GRU. By connecting the two-way vectors, the sentence representation is obtained while capturing contextual information.
[0030] Word attention mechanism: Use the multiplicative attention mechanism, use a layer of MLP to obtain the implicit representation of the sentence, then use the word-level context vector to measure the importance of the word, and use the softmax function to get the normalized importance;
[0031] Sentence encoding: Use bidirectional GRU to encode the sentence vector;
[0032] Sentence Attention Mechanism: Use the multiplicative attention mechanism, use a layer of MLP to obtain the implicit representation of the sentence, and then use the sentence vector to measure the importance of the word.
[0033] According to the above aspects and any possible implementation, a document tag generation system based on article topic model and named entity extraction is further provided, the system comprising:
[0034] The automatic tag library generation module obtains keyword tags and noun entity tags from a limited number of documents through the article topic model and named entity extraction algorithm to generate an automatic tag library;
[0035] The label classification algorithm module processes the original documents and new documents in the system through the text multi-classification model to generate corresponding labels;
[0036] The tag feedback mechanism module re-enters the automatic tag library generation module for documents that have not obtained tags, obtains keyword tags and noun entity tags, and adds them to the automatic tag library;
[0037] The tag library post-maintenance module filters and / or removes duplicate tags entered manually and adds them to the automated tag library, while regularly training the text multi-classification model.
[0038] According to the above aspects and any possible implementation, a readable storage medium is further provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the document tag generation methods are implemented.
[0039] Compared with the prior art, the present invention can achieve the following technical effects:
[0040] The present invention constructs a document tag generation method based on article topic model and named entity extraction. For batch document files stored in a fixed space by users, a tag library can be automatically initialized, and the tag corresponding to each document can be calculated. In subsequent use, tags can be automatically assigned to newly added documents, which is convenient for users to store and manage documents and accurately retrieve them.
[0041] Two model technologies are used in the process of initializing the extraction of tags. The article body model can help cluster documents and extract keywords with topic marks under similar topics; while the named entity recognition technology helps us extract the description body of a single article to ensure the comprehensive coverage of tags and improve the accuracy and comprehensiveness of the initialization tag library.
[0042] In the subsequent multi-label classification model, a multi-level attention network is used to classify documents, ensuring the effectiveness and speed of the model output.
[0043] Of course, any product implementing the present invention does not necessarily need to achieve all of the above-mentioned technical effects at the same time.
Brief Description of the Drawings
[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0045] Figure 1 is a flow chart of a document tag generation method provided by an embodiment of the present invention;
[0046] Figure 2It is an article topic model diagram provided by an embodiment of the present invention;
[0047] Figure 3 is a named entity extraction graph provided by an embodiment of the present invention;
[0048] Figure 4 It is a multi-label classification algorithm diagram provided by an embodiment of the present invention. [Specific implementation method]
[0049] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0050] It should be clear that the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0051] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.
[0052] like Figure 1 As shown, the present invention provides a document tag generation method, the method comprising the following steps:
[0053] S1 Automatic tag library generation: Through the article topic model and named entity extraction algorithm, keyword tags and noun entity tags are obtained from a limited number of documents to generate an automatic tag library;
[0054] S2 label classification algorithm: Use the text multi-classification model to process the original documents and new documents in the system, match the classification results of the text multi-classification model with the automatic label library generated by S1, and generate corresponding labels;
[0055] S3 tag feedback mechanism: for documents that have not been tagged, the method in S1 is used again to obtain keyword tags and noun entity tags and add them to the automatic tag library;
[0056] Post-maintenance of the S4 tag library: Filter and / or deduplicate manually entered tags and add them to the automated tag library, and regularly train the text multi-classification model.
[0057] The S1 specifically includes:
[0058] S11: clustering a limited number of documents using the article topic model, extracting unique keywords in each category, obtaining keyword tags and storing them in a backup tag library;
[0059] S12: extract named entities from the documents in S11 using a named entity extraction algorithm, obtain noun entity tags, filter and store them in a backup tag library;
[0060] S13: performing deduplication and synonym merging processing on noun entity tags and keyword tags in the backup tag library to reduce the number of tags in the backup library;
[0061] S14: Process the labels according to preset conditions to generate an automated label library.
[0062] The preset conditions in S14 include but are not limited to merging similar tags, and the merging similar tags specifically refers to: expressing the same product with different names.
[0063] The S11 specifically includes:
[0064] S111: corpus preprocessing to obtain a candidate keyword set;
[0065] S112: Keywords in the document are ranked from high to low using tf-idf and textrank scoring methods, and stop words and parts of speech that do not show the topic meaning are eliminated. The resulting tokens file is used as the input file for LDA;
[0066] S113: Generate an LDA model based on the corpus, using the LDAmodel provided by gensim;
[0067] S114: Apply the LDA model in S113 to a certain document to obtain the topic probability distribution and keywords of the document. If the words obtained after the document segmentation are among the candidate keywords, they are put into the reserve library as keyword tags, otherwise they are included in the stop word library.
[0068] The score of the candidate keyword in S114 = the probability of each topic * the probability of each word belonging to the topic.
[0069] The S12 specifically includes:
[0070] S121: corpus preprocessing to obtain clean corpus;
[0071] S122: Use the Mingming entity extraction model generated by the BERT+(LSTM)+CRF network to annotate the named entities appearing in the corpus;
[0072] S123: Calculate the frequency of named entities appearing in the document, and remove some named entities using preset rules;
[0073] S124: Put the noun entities within the preset frequency into the reserve library as noun entity tags.
[0074] The text multi-classification model in S2 is trained by using a multi-level attention network. The text multi-classification model in S2 specifically includes:
[0075] Word encoding: Each word in the sentence is embedded into a word vector in the distribution vector space, and the sentence is encoded using a bidirectional GRU. By connecting the two-way vectors, the sentence representation is obtained while capturing contextual information.
[0076] Word attention mechanism: Use the multiplicative attention mechanism, use a layer of MLP to obtain the implicit representation of the sentence, then use the word-level context vector to measure the importance of the word, and use the softmax function to get the normalized importance;
[0077] Sentence encoding: Use bidirectional GRU to encode the sentence vector;
[0078] Sentence Attention Mechanism: Use the multiplicative attention mechanism, use a layer of MLP to obtain the implicit representation of the sentence, and then use the sentence vector to measure the importance of the word.
[0079] The present invention also provides a document tag generation system based on article topic model and named entity extraction, the system comprising:
[0080] The automatic tag library generation module obtains keyword tags and noun entity tags from a limited number of documents through the article topic model and named entity extraction algorithm to generate an automatic tag library;
[0081] The label classification algorithm module processes the original documents and new documents in the system through the text multi-classification model to generate corresponding labels;
[0082] The tag feedback mechanism module re-enters the automatic tag library generation module for documents that have not obtained tags, obtains keyword tags and noun entity tags, and adds them to the automatic tag library;
[0083] The tag library post-maintenance module filters and / or removes duplicate tags entered manually and adds them to the automated tag library, while regularly training the text multi-classification model.
[0084] The present invention also provides a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the document tag generating methods are implemented.
[0085] The system of the present invention is designed to process a large amount of data stored in a document format. Through the early processing of the document topic model of multiple documents, a large number of corresponding screening tags are extracted, and then after a manual screening process, an initial document tag library is generated; then, the multi-classification algorithm of the text is used to judge the content / source of each article, generate corresponding tags, and assist the subsequent precise retrieval and personal recommendation functions.
[0086] The present invention uses a variety of leading natural language processing methods to perform article topic modeling and text classification on document data, and helps users quickly and accurately find the required content and files through classification technology, thus realizing a fast, accurate and personalized document tag generation method. The specific content of the method includes the following steps:
[0087] S1 automatic tag library generation
[0088] S1.1 As Figure 2 As shown in the figure, multiple documents are clustered using the article topic model, and unique keywords in each category are extracted and stored in the backup tag library; the process of generating documents by the LDA article topic algorithm is as follows:
[0089] (1) Corpus preprocessing to obtain a set of candidate keywords (words obtained after document segmentation);
[0090] (2) The keywords of each article are ranked from high to low using tf-idf and textrank scoring methods, and stop words and parts of speech that do not show the topic meaning are eliminated. The resulting tokens file is used as the input file of LDA;
[0091] (3) Generate an LDA model based on the corpus, using the LDAmodel provided by gensim;
[0092] (4) Apply the model to a certain text to obtain the topic probability distribution and keywords of the text. If the word obtained after document segmentation is among the candidate keywords, then extract it as a keyword. For each candidate keyword, its score = the probability of each topic * the probability of each word belonging to the topic;
[0093] (5) Some words that cannot reflect the topic meaning are included in the stop word library (manual screening).
[0094] S1.2 As Figure 3 As shown in the figure, the named entity extraction algorithm is used to extract the named entities in the processed documents, filter them and store them in the backup tag library; the process of BERT+(LSTM)+CRF using the named entity extraction model is as follows:
[0095] (1) Corpus preprocessing to obtain clean corpus;
[0096] (2) The Mingming entity extraction model generated by the BERT+(LSTM)+CRF network is used to annotate the expected named entities. The parameterized form of CRF is shown below;
[0097]
[0098] Where Z(x)=∑ y exp[∑ i,k v k t k (y i-1 ,y i ,x,i)+∑ i,l u l s l (y i ,x,i)],
[0099] t k and l is the characteristic function, v k and u l is the corresponding weight, t k is the state transition function, v k is the corresponding weight; s l is the emission function, u l is the corresponding weight.
[0100] (3) Calculate the frequency of named entities appearing in the text and use rules to remove some named entities;
[0101] (4) The top-ranked noun entities are taken as our tag words and put into the tag library.
[0102] S1.3 performs simple deduplication and synonym merging processing on the tags in the backup tag library to reduce the number of tags in the backup library;
[0103] S1.4 According to the needs of business personnel, merge similar tags, such as different product name expressions, etc., to generate the final tag library
[0104] S2 label classification algorithm, such as Figure 4 Shown:
[0105] S2.1 Organize existing text and corresponding labels;
[0106] S2.2 performs model training and uses a multi-level attention network to train a text multi-classification model. Our document model multi-classification model mainly consists of four parts:
[0107] (1) Word encoding: embed each word in the sentence into a word vector in the distribution vector space, and use a bidirectional GRU to encode the sentence. By connecting the two-way vectors, the representation of the sentence is obtained, while capturing contextual information;
[0108] x it =W e w it , t∈[1,T]
[0109]
[0110]
[0111] (2) Word attention mechanism: Using the multiplicative attention mechanism, we first use a layer of MLP to obtain the implicit representation of the sentence, then use the word-level context vector to measure the importance of the word, and then use the softmax function to obtain the normalized importance.
[0112] u it =tanh(w w h it +b w ),
[0113]
[0114]
[0115] (3) Sentence encoding: Similar to word encoding, sentence vectors are encoded using bidirectional GRU;
[0116] u it =tanh(w w h it +b w ),
[0117]
[0118]
[0119] (4) Sentence attention mechanism. Similar to the attention mechanism used for words, sentence vectors are used to measure the importance between sentences.
[0120] u i =tanh(w s h i +b s ),
[0121]
[0122]
[0123] S2.3 processes the newly entered documents in real time and obtains corresponding labels.
[0124] S3 tag feedback mechanism
[0125] S3.1 Filter out documents without label results;
[0126] S3.2 puts the screened documents into the S1 automatic tag library generation module in batches for tag extraction again, and puts the newly generated tags into the tag library.
[0127] S4 tag library post-maintenance
[0128] S4.1 filters / de-duplicates the labels manually entered by system users, selectively adds them to the label library, and uses text distance for preliminary matching;
[0129] S4.2 Regularly retrain the text multi-classification model to obtain a model that can handle new labels.
[0130] The present invention constructs a document tag generation method based on article topic model and named entity extraction. For batch document files stored by users in a fixed space, the tag library can be automatically initialized, and the tag corresponding to each document can be calculated. In subsequent use, the newly added documents can be automatically assigned tags, which is convenient for users to store and manage documents and accurately retrieve them. In the process of initializing the extraction of tags, we used two model technologies. The article body model can help cluster documents and extract keywords with topic marks under similar topics; and the named entity recognition technology helps us extract the descriptive body of a single article, ensuring comprehensive coverage of tags and improving the accuracy and comprehensiveness of the initialized tag library. In the subsequent multi-label classification model, we used a multi-level attention network to classify documents, ensuring the effect and speed of the model output. The innovative points of the present invention specifically include:
[0131] (1) First, we used two NLP technologies, topic modeling and named entity extraction, to automatically generate a tag library, predict the tags we need, and facilitate business personnel to organize and classify them. We improved the problems of inconsistent standards and large numbers of automatic labeling products on the market, and facilitated the introduction of relevant knowledge by business personnel.
[0132] (2) The traditional manual labeling process is introduced into an automated labeling mechanism using a multi-label classification model. Combined with the sorted label library mentioned in (1), it is possible to obtain document labeling results that are standardized, meet the needs of business personnel, and are highly accurate. In addition, the speed is greatly improved compared to manual labeling.
[0133] (3) We introduced a post-maintenance mechanism for the tag library, which allows business personnel to add, delete, and modify tags in the tag library on a regular basis, add new products and focus points, and delete useless focus points, so that the tags can truly help business personnel improve retrieval efficiency and accuracy. In addition, our model will be iteratively trained based on the adjustment of tags to optimize the model effect.
[0134] Table 1 below compares the three tag generation methods: manual tagging, automatic generation of subject tags, and document tag generation method based on article topic model and named entity extraction (tag library + tag):
[0135]
[0136] The above is a detailed introduction to a document tag generation method, system and readable storage medium provided by the embodiments of the present application. The description of the above embodiments is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
[0137] For example, certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that hardware manufacturers may use different nouns to refer to the same component. This specification and claims do not use differences in names as a way to distinguish components, but use differences in the functions of components as the criteria for distinction. As mentioned throughout the specification and claims, "including" and "comprising" are open-ended terms, so they should be interpreted as "including / including but not limited to". "Approximately" means that within an acceptable error range, those skilled in the art can solve the technical problem within a certain error range and basically achieve the technical effect. The subsequent description of the specification is a preferred embodiment of the present application, but the description is for the purpose of illustrating the general principles of the present application, and is not used to limit the scope of the present application. The scope of protection of the present application shall be determined by the definition of the attached claims.
[0138] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a product or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such a product or system. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the product or system including the elements.
[0139] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.
[0140] The above description shows and describes several preferred embodiments of the present application, but as mentioned above, it should be understood that the present application is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the application concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present application, and should be within the scope of protection of the claims attached to the present application.
Claims
1. A document tag generation method based on article topic model and named entity extraction, characterized in that: The method comprises the following steps: S1 Automatic tag library generation: Through the article topic model and named entity extraction algorithm, keyword tags and noun entity tags are obtained from a limited number of documents to generate an automatic tag library; S2 label classification algorithm: Use the text multi-classification model to process the original documents and new documents in the system, match the classification results of the text multi-classification model with the automatic label library generated by S1, and generate corresponding labels; S3 tag feedback mechanism: for documents that have not been tagged, the method in S1 is used again to obtain keyword tags and noun entity tags and add them to the automatic tag library; S4 tag library post-maintenance: filter and / or deduplicate manually entered tags and add them to the automated tag library, and regularly train text multi-classification models; The S1 specifically includes: S11: clustering a limited number of documents using the article topic model, extracting unique keywords in each category, obtaining keyword tags and storing them in a backup tag library; S12: extract named entities from the documents in S11 using a named entity extraction algorithm, obtain noun entity tags, filter and store them in a backup tag library; S13: performing deduplication and synonym merging processing on noun entity tags and keyword tags in the backup tag library to reduce the number of tags in the backup library; S14: Process the labels according to preset conditions to generate an automated label library; The S11 specifically includes: S111: corpus preprocessing to obtain a candidate keyword set; S112: Keywords in the document are ranked from high to low using tf-idf and textrank scoring methods, and stop words and parts of speech that do not show the topic meaning are eliminated. The resulting tokens file is used as the input file for LDA; S113: Generate an LDA model based on the corpus, using the LDAmodel provided by gensim; S114: Apply the LDA model in S113 to a certain document to obtain the topic probability distribution and keywords of the document. If the words obtained after the document segmentation are in the candidate keywords, put them into the reserve library as keyword tags, otherwise put them into the stop word library; The score of the candidate keyword in S114 = the probability of each topic * the probability of each word belonging to the topic; The S12 specifically includes: S121: corpus preprocessing to obtain clean corpus; S122: Use the Mingming entity extraction model generated by the BERT+LSTM+CRF network to annotate the named entities appearing in the corpus; S123: Calculate the frequency of named entities appearing in the document, and remove some named entities using preset rules; S124: Put the noun entities within the preset frequency into the reserve library as noun entity tags.
2. The method according to claim 1, characterized in that The preset conditions in S14 include but are not limited to merging similar tags, and the merging similar tags specifically refers to: expressing the same product with different names.
3. The method according to claim 1, characterized in that The training method of the text multi-classification model in S2 is: using a multi-level attention network to train the text multi-classification model.
4. The method according to claim 3, characterized in that The text multi-classification model in S2 specifically includes: Word encoding: Each word in the sentence is embedded into a word vector in the distribution vector space, and the sentence is encoded using a bidirectional GRU. By connecting the two-way vectors, the sentence representation is obtained while capturing contextual information. Word attention mechanism: Use the multiplicative attention mechanism, use a layer of MLP to obtain the implicit representation of the sentence, then use the word-level context vector to measure the importance of the word, and use the softmax function to get the normalized importance; Sentence encoding: Use bidirectional GRU to encode the sentence vector; Sentence Attention Mechanism: Use the multiplicative attention mechanism, use a layer of MLP to obtain the implicit representation of the sentence, and then use the sentence vector to measure the importance of the word.
5. A document tag generation system based on article topic model and named entity extraction, applying the method of any one of claims 1 to 4 above, characterized in that: The system comprises: The automatic tag library generation module obtains keyword tags and noun entity tags from a limited number of documents through the article topic model and named entity extraction algorithm to generate an automatic tag library; The label classification algorithm module processes the original documents and new documents in the system through the text multi-classification model to generate corresponding labels; The tag feedback mechanism module re-enters the automatic tag library generation module for documents that have not obtained tags, obtains keyword tags and noun entity tags, and adds them to the automatic tag library; The tag library post-maintenance module filters and / or removes duplicate tags entered manually and adds them to the automated tag library, while regularly training the text multi-classification model.
6. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the document tag generating method according to any one of claims 1 to 4 are implemented.
Citation Information
Patent Citations
Dependency tree and attention mechanism-based attribute sentiment classification method
CN108399158A
Method and apparatus for generating document label, electronic device, and storage medium
CN108875059A
Document classification method based on hierarchical multi-attention network
CN109558487A
Text data viewpoint generation method and device and electronic equipment
CN110162594A
Text label generation method and device and computer readable storage medium
CN110580292A
Cited By
Hierarchical Knowledge Extraction and Tag Association Methods for Large Models Targeting Business Text
CN123070288A
