Knowledge organization system construction method and device, equipment, medium and program

Abstract

The present application relates to the field of computer, and provide a kind of knowledge organization system construction method, device, equipment, medium and program, method includes: from the target field data set of prior acquisition retrieves text paragraph containing target field term, and corresponding retrieval result is input into term generation model, obtains term generation result;For each term in term generation result, respectively input term into basic information prediction model and semantic recognition model, obtain the basic information and semantic recognition result of corresponding term;According to term and the basic information of term, utilize classification model, obtain corresponding classification result;According to the term basic information, attribute and relationship of each term and term classification result, generate the entry of corresponding term, and based on all entries generated, construct knowledge organization system.The present application improves the representation ability of knowledge organization system to complex knowledge by comprehensively considering multi-dimensional knowledge types, so that it can better present various types of knowledge.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, device, medium, and program for constructing a knowledge organization system. Background Technology

[0002] With the advent of the digital age, massive amounts of data and text information are exploding. Faced with this vast amount of complex, diverse, and structurally varied data, the introduction of a Knowledge Organization System (KOS) is crucial for integrating and organizing this information. KOS categorizes and indexes this information according to certain logic and rules, transforming the originally chaotic data into an ordered whole, making it easier for users to quickly locate and find the information they need. A Knowledge Organization System is a general term for semantic tools used to express and organize human knowledge structures. These include taxonomy, subject tables, thesaurus, and other new knowledge organization models that have emerged with the development of Semantic Web technology, such as ontology, relational data, visualized topic maps, knowledge maps, and knowledge graphs.

[0003] Currently, the main methods for building knowledge organization systems involve collecting data, extracting knowledge from the collected data and building relationships, and reviewing and filtering the updated data when the knowledge organization system needs to be updated to determine whether the system should be updated.

[0004] However, the above methods focus on constructing entity attributes and relationships between entities, neglecting basic entity information, definitions, classifications, and broader conceptual frameworks. This results in an incomplete understanding of the essential characteristics of entities and their position within the knowledge system, leading to incomplete knowledge representation and impacting the system's comprehensiveness and accuracy. Furthermore, when collecting large amounts of data in a particular domain, these methods require manual processes such as word extraction, classification, and association. As information expands, significant manpower is needed to continuously update the knowledge organization system. Summary of the Invention

[0005] This invention provides a method, apparatus, device, medium, and program for constructing a knowledge organization system, which addresses the shortcomings of existing technologies that only focus on entity attributes and relationships between entities, resulting in incomplete knowledge representation. By comprehensively considering multi-dimensional knowledge types, it improves the ability of the knowledge organization system to represent complex knowledge, enabling it to better process and present various types of knowledge.

[0006] This invention provides a method for constructing a knowledge organization system, comprising: acquiring a target domain dataset, the target domain dataset including multiple text paragraphs obtained based on target domain files; retrieving text paragraphs containing target domain terms from the target domain dataset, and inputting the corresponding retrieval results into a term generation model to obtain term generation results output by the term generation model, the term generation results including at least one term; wherein, the term generation model is obtained through unsupervised training based on unsupervised term training text; for each term in the term generation results, the term is input into a basic information prediction model to obtain basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained through unsupervised training based on unsupervised term training text; For each term in the terminology generation results, the term is input into a semantic recognition model to obtain the semantic recognition result of the corresponding term, which includes attributes and relationships. The semantic recognition model is fine-tuned based on the terminology training data and the corresponding attribute and relationship labels. For each term in the terminology generation results, a classification model is used to obtain the corresponding classification result based on the terminology training data and the corresponding classification labels. Based on the basic information, attributes, relationships, and classification results of each term, a term entry for the corresponding term is generated, and a knowledge organization system is constructed based on all generated terms.

[0007] According to a knowledge organization system construction method provided by the present invention, text paragraphs containing target domain terms are retrieved from a target domain dataset, and the corresponding retrieval results are input into a terminology generation model to obtain terminology generation results output by the terminology generation model. The method includes: using keywords pre-set based on target domain terms to perform keyword retrieval or similarity matching on the target domain dataset to obtain text paragraphs containing target domain terms as retrieval results; and, based on the retrieval results and combined with preset prompt words, calling the terminology generation model to generate terminology generation results corresponding to the retrieval results based on the preset prompt words; wherein, the preset prompt words are set based on the contextual environment required for the terminology in the target domain.

[0008] According to a knowledge organization system construction method provided by the present invention, after obtaining the term generation results output by the term generation model, the method includes: filtering the terms in the term generation results to remove duplicate terms; and / or, verifying the terms in the term generation results based on preset verification rules to remove terms that do not conform to the preset verification rules; wherein the preset verification rules include rules for limiting terms to conform to the target domain.

[0009] According to a knowledge organization system construction method provided by the present invention, before inputting terms into a semantic recognition model, the method includes: acquiring term training data and corresponding attribute labels and relation labels; converting the term training data and corresponding attribute labels and relation labels according to a preset format to obtain format-converted instruction fine-tuning data; wherein, the preset format is used to limit the arrangement format of the attributes and relations corresponding to the terms; and fine-tuning the semantic recognition model according to the instruction fine-tuning data and using a preset fine-tuning method.

[0010] According to a knowledge organization system construction method provided by the present invention, the classification model includes a first-level classifier, a second-level classifier, a first- and third-level classifier, and a second- and third-level classifier. The first-level classifier is obtained by training on terminology classification training samples and the first-level classifiers corresponding to the terminology classification training samples. The second-level classifier is obtained by training on terminology classification training samples and the second-level classifiers corresponding to the terminology classification training samples. The first- and third-level classifiers are obtained by training the first-level classifier using the corresponding third-level classifiers and their corresponding terminology classification training samples. The second- and third-level classifiers are obtained by training the second-level classifier using the corresponding third-level classifiers and their corresponding terminology classification training samples. The first-level, second-level, and third-level classifiers are obtained by dividing the terminology classification training samples according to a preset hierarchical division rule from high to low. For each term in the terminology generation result, based on the term and its basic information, the classification model is used to obtain the corresponding classification result, including: for each term in the terminology generation result, classifying each term with... The basic information corresponding to the terms is concatenated to obtain the text to be classified. This text is then input into a primary classifier and a secondary classifier to obtain the primary predicted classification result output by the primary classifier and the secondary predicted classification result output by the secondary classifier. The primary predicted classification result includes the primary predicted classification number of the text, and the secondary predicted classification result includes the secondary predicted classification number of the text. It is then determined whether the secondary predicted classification number of the same text belongs to the corresponding primary predicted classification number. If so, the text is classified using the secondary and tertiary classifiers to obtain the tertiary predicted classification result. Otherwise, the text is classified using the primary and tertiary classifiers to obtain the tertiary predicted classification result. Based on the primary, secondary, and tertiary predicted classification results of the text, the classification result for the corresponding term is obtained.

[0011] According to a knowledge organization system construction method provided by the present invention, a three-level classifier includes a fusion classifier and a base classifier. The three-level classifier is used to classify the text to be classified to obtain a three-level predicted classification number. The method includes: inputting the text to be classified into each base classifier to obtain the probability value output by each base classifier; and outputting the probability value output by each base classifier to the fusion classifier for classification decision to obtain the three-level predicted classification result output by the fusion classifier.

[0012] This invention also provides a knowledge organization system construction apparatus, comprising: a data acquisition module for acquiring a target domain dataset, the target domain dataset including multiple text paragraphs obtained based on target domain files; a terminology generation module for retrieving text paragraphs containing target domain terms from the target domain dataset and inputting the corresponding retrieval results into a terminology generation model to obtain terminology generation results output by the terminology generation model, the terminology generation results including at least one term; wherein, the terminology generation model is obtained through unsupervised training based on unsupervised terminology training text; an information prediction module for each term in the terminology generation results, inputting the term into a basic information prediction model to obtain basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained through unsupervised training based on unsupervised terminology training text; and semantic recognition. The terminology module takes each term in the terminology generation results and inputs it into a semantic recognition model to obtain the semantic recognition result of the corresponding term. The semantic recognition result includes attributes and relationships. The semantic recognition model is fine-tuned based on the terminology training data and the corresponding attribute and relationship labels of the terminology training data. The terminology classification module takes each term in the terminology generation results and, based on the terminology and its basic information, uses a classification model to obtain the corresponding classification result. The classification model is fine-tuned based on the terminology training data and the corresponding classification labels of the terminology training data. The system construction module generates entries for the corresponding terms based on the basic information, attributes, relationships, and classification results of each term, and builds a knowledge organization system based on all the generated entries.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the knowledge organization system construction method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the knowledge organization system construction method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the knowledge organization system construction method as described above.

[0016] The knowledge organization system construction method, apparatus, device, medium, and program provided by this invention generate corresponding terms from a pre-acquired target domain dataset using a terminology generation model. Further, multiple models, including a basic information prediction model, a semantic recognition model, and a classification model, are used to process the terms from different perspectives. Through clear division of labor and mutual cooperation among these models, the basic information, attributes, relationships, and classification results of the terms are comprehensively and deeply mined. This improves the quality and efficiency of terminology processing while comprehensively considering multi-dimensional knowledge types, allowing different types of knowledge to be integrated into a single term. This helps break down barriers between knowledge and avoids the limitations of understanding only a single dimension of information. It facilitates a deeper understanding and accurate application of terms. Furthermore, when new knowledge or information appears, since the term is constructed based on information from various dimensions, only the information in the corresponding dimensions needs to be updated. This enables the knowledge organization system to more realistically reflect the knowledge structure and relationships in the objective world, enhancing its ability to represent complex knowledge and better process and present various types of knowledge. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is one of the flowcharts illustrating the knowledge organization system construction method provided by the present invention; Figure 2 This is the second flowchart illustrating the knowledge organization system construction method provided by this invention; Figure 3 This is a schematic diagram of the structure of the knowledge organization system construction device provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Figure 1 This is a flowchart illustrating the knowledge organization system construction method provided by the present invention, such as... Figure 1 As shown, the method includes: S11, Obtain the target domain dataset, which includes multiple text paragraphs obtained based on the target domain file; S12, retrieve text paragraphs containing target domain terms from the target domain dataset, and input the retrieval results into the term generation model to obtain the term generation results output by the term generation model. The term generation results include at least one term. The term generation model is obtained by unsupervised training based on unsupervised term training text. S13, For each term in the term generation result, the term is input into the basic information prediction model to obtain the basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained by unsupervised training based on unsupervised term training text; S14, for each term in the term generation result, the term is input into the semantic recognition model to obtain the semantic recognition result of the corresponding term output by the semantic recognition model. The semantic recognition result includes attributes and relations. The semantic recognition model is obtained by fine-tuning the pre-trained model based on the term training data and the attribute labels and relation labels corresponding to the term training data. S15, for each term in the term generation result, based on the term and its basic information, use a classification model to obtain the corresponding classification result; wherein, the classification model is trained based on the term training data and the classification labels corresponding to the term training data. S16. Based on the basic information, attributes, and relationships of each term and the term classification results, generate entries for the corresponding terms, and construct a knowledge organization system based on all generated entries.

[0021] It should be noted that the step numbers "S1N" in this manual do not represent the chronological order of the knowledge organization system construction methods. The following details will explain further. Figure 2 The present invention describes a method for constructing a knowledge organization system.

[0022] Step S11: Obtain the target domain dataset, which includes multiple text paragraphs obtained based on the target domain file.

[0023] In this embodiment, obtaining the target domain dataset includes: collecting files in the target domain; extracting text from the collected files to obtain the text content corresponding to each file; performing noise reduction and segmentation on each text content to obtain corresponding text paragraphs; and removing duplicates from all obtained text paragraphs to obtain the target domain dataset.

[0024] It should be noted that the files can include literature, reports, and professional books within the target field, obtained through web scraping, Application Programming Interfaces (APIs), or other data collection tools. The specific files collected can be tailored to the actual design requirements. Noise reduction methods can be set based on the noise that may be introduced during text extraction, such as removing redundant HyperText Markup Language (HTML) tags and using Uniform Resource Locators (URLs). Further limitations are not specified here. Additionally, the collected files are unstructured data.

[0025] Furthermore, each text content is segmented to obtain corresponding text paragraphs. This includes: reading the text content and identifying paragraphs based on punctuation and formatting information, or segmenting based on content theme or semantics. It should be noted that segmentation based on content theme or semantics can be achieved using existing Natural Language Processing (NLP) tools, which will not be elaborated upon here.

[0026] In an optional embodiment, after obtaining the target domain dataset, the method further includes: monitoring the source of the corresponding file so as to obtain the corresponding updated file after the file is updated, and after processing it into a dataset in the manner described above, obtaining the corresponding updated terms in steps S12-S15 below, so as to update the knowledge organization system using the updated terms. This will not be repeated here.

[0027] It should be noted that updating a previously created knowledge organization system using updated terms includes: retrieving terms within the knowledge organization system based on the terms corresponding to the updated terms, obtaining system terms within the knowledge organization system that match the terms of the updated terms; comparing the updated terms with the system terms; if they match, then the knowledge organization system is not updated; otherwise, the updated terms replace the system terms within the knowledge organization system to update the knowledge organization system.

[0028] Alternatively, the target domain dataset can also be a dataset within a previously disclosed target domain that contains terms from that domain; no further limitations are imposed here.

[0029] Step S12: Retrieve text paragraphs containing target domain terms from the target domain dataset, and input the corresponding retrieval results into the terminology generation model to obtain the terminology generation results output by the terminology generation model. The terminology generation results include at least one term. The terminology generation model is obtained through unsupervised training based on unsupervised terminology training text. It should be noted that the terminology generation results can be in list format, document format, or other representation formats, depending on the output of the selected and trained terminology generation model; no further limitations are made here.

[0030] In this embodiment, text paragraphs containing target domain terms are retrieved from the target domain dataset, and the corresponding retrieval results are input into the terminology generation model to obtain the terminology generation results output by the terminology generation model. This includes: using keywords previously set based on target domain terms to perform keyword retrieval or similarity matching on the target domain dataset to obtain text paragraphs containing target domain terms as retrieval results; based on the retrieval results and combined with preset prompt words, the terminology generation model is invoked to generate terminology generation results corresponding to the retrieval results based on the preset prompt words; wherein, the preset prompt words are set based on the contextual environment required for the terminology in the target domain.

[0031] It is worth noting that if keyword retrieval of the target domain dataset is performed using keywords pre-defined based on the target domain, a retrieval system should be built on the target domain dataset using preset retrieval techniques beforehand. This facilitates subsequent keyword-based retrieval of relevant terminology text within the target domain dataset. It should be added that the preset retrieval techniques can employ existing methods such as inverted indexes, and the specific choice depends on the actual retrieval requirements; no further limitations are imposed here. Furthermore, when performing similarity matching on the target domain dataset using keywords pre-defined based on the target domain, similarity matching algorithms such as cosine similarity algorithms can be used, and the specific choice depends on the actual design requirements; no further limitations are imposed here.

[0032] In one alternative embodiment, the terminology generation model can be selected based on actual design requirements or prior experience, such as the Tongyi Qianwen Big Language Model. The preset prompt words can be designed based on actual design requirements or prior experience, such as using prompt words to build the COSTAR framework, etc., without further limitations here.

[0033] It's important to note that when writing preset prompts using the COSTAR framework, they must be written according to the context required for the terminology in the target domain. This context includes the domain background and related concepts. Based on this context, prompts can be written to guide the model in generating relevant terms. For example, a preset prompt could be: "[Background Description] Please generate a list of terms related to the XX domain based on the above background, using a scholarly and objective tone. The target audience is knowledge engineers; please use the lightweight data exchange format JSON."

[0034] In one optional embodiment, after obtaining the term generation result output by the term generation model, the method includes: filtering the terms in the term generation result to remove duplicate terms; and / or, verifying the terms in the term generation result based on preset verification rules to remove terms that do not conform to the preset verification rules; wherein the preset verification rules include rules for limiting terms to conform to the target domain.

[0035] Specifically, based on preset verification rules, the terms in the terminology generation results are verified to remove terms that do not conform to the preset verification rules. This includes: matching each term in the terminology generation results with terms in the preset terminology database based on the preset verification rules, and determining that the corresponding term in the terminology generation results passes the verification based on the matching results; otherwise, the term fails the verification and is removed from the terminology generation results. The preset terminology database is constructed using terminology in the target domain based on expert experience or prior experience.

[0036] Furthermore, after determining that the corresponding term in the terminology generation result has failed verification, the method further includes: pushing the terminology that has failed verification to the corresponding user terminal for review to determine whether it conforms to the corresponding target domain, and receiving the review result returned by the user terminal; if the review result is that the term conforms to the target domain, then retain the corresponding terminology and use the term to update the corresponding preset terminology library; if the review result is that the term does not conform to the target domain, then remove the corresponding term from the terminology generation result.

[0037] Step S13: For each term in the term generation result, input the term into the basic information prediction model to obtain the basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained by unsupervised training based on unsupervised term training text.

[0038] It is worth noting that when using the basic information prediction model to predict the basic information of corresponding terms, it is also necessary to define the content of the basic information. For example, the basic information includes the Chinese word form, pinyin, and English translation of the term. A term containing the above information should be selected from professional books in the target field as an example for model learning. Corresponding prompt samples should be configured, and the obtained prompt samples should be input into the basic information prediction model so that the model can generate the basic information of the corresponding term based on the patterns learned from the prompt samples. For example, when the target field is the new energy vehicle field, the basic information related to the term "biofuel" in the 2012 edition of the Chinese Scientific and Technological Terminology System (New Energy Vehicle Volume) can be selected as an example for model learning.

[0039] In one alternative embodiment, a one-shot learning technique can be used to generate basic information for the corresponding terms.

[0040] Step S14: For each term in the term generation result, input the term into the semantic recognition model to obtain the semantic recognition result of the corresponding term output by the semantic recognition model. The semantic recognition result includes attributes and relationships. The semantic recognition model is obtained by fine-tuning the pre-trained model based on the term training data and the attribute labels and relationship labels corresponding to the term training data.

[0041] In one alternative embodiment, before inputting terms into the semantic recognition model, the process includes: downloading the original model file of the semantic recognition model and deploying it to a local server. It should be noted that the semantic recognition model can be selected based on actual recognition needs or prior experience, such as choosing the open-source large language model ChatGLM; no further limitations are imposed here.

[0042] In an optional embodiment, before inputting terms into the semantic recognition model, the method further includes: obtaining term training data and corresponding attribute labels and relation labels; converting the term training data and corresponding attribute labels and relation labels according to a preset format to obtain instruction fine-tuning data; wherein, the preset format is used to limit the arrangement format of the attributes and relations corresponding to the terms; and fine-tuning the semantic recognition model according to the instruction fine-tuning data using a preset fine-tuning method.

[0043] It should be added that the preset fine-tuning method can be selected according to actual design needs. For example, the prompt fine-tuning (P-Tuning) method can be selected to fine-tune the original model file of the semantic recognition model deployed to the local server to obtain the semantic recognition model. In addition, the semantic recognition model can be fine-tuned using preset fine-tuning methods based on instruction fine-tuning data. This includes: replacing traditional discrete text prompts with learnable continuous embeddings to optimize only prompt-related parameters (such as embedding layers and encoders), freezing the main parameters of the pre-trained model, and then using instruction fine-tuning data to fine-tune the model to significantly reduce the training load; and introducing an encoder composed of a Long Short-Term Memory (LSTM) network or a Multilayer Perceptron (MLP) during the fine-tuning process to extract features from virtual tags, thereby enhancing the correlation of features and accelerating training.

[0044] It's worth noting that by using trainable LSTMs or MLPs to dynamically generate virtual labeled embeddings, different embeddings can be generated based on different input data. This allows for a better understanding and adaptation to complex, sequentially dependent tasks, providing greater flexibility and adaptability. Furthermore, the recurrent structure of LSTMs better handles and understands long-term dependencies and complex contextual information, making it suitable for tasks requiring fine-grained control and understanding of complex contexts. Additionally, in P-Tuning, the parameters of the LSTM model can be shared across multiple tasks, thereby improving the generalization ability of the fine-tuned semantic recognition model and reducing the training requirements for each individual task.

[0045] Furthermore, the acquired terminology training data and its corresponding attributes and relationships can be obtained from relevant professional books based on the actual target field. For example, if the target field is the field of new energy vehicles, 510 entries containing both attributes and relationships can be selected from the Chinese Scientific and Technological Terminology System (New Energy Vehicle Volume) published in 2012 to obtain terminology training data and the corresponding attribute and relationship labels. No further limitations are imposed here.

[0046] Additionally, attributes are used to characterize the features corresponding to a term, used to describe, distinguish, or classify it. Relationships are used to characterize the connections, dependencies, or interactions between a corresponding term and other terms. Relationships include hierarchical relationships and compositional relationships. Hierarchical relationships characterize the position of a term in a classification, such as the class or group to which the term belongs. Compositional relationships characterize which larger entities or sets the term, as a component, belongs to. The preset format can be configured according to the actual arrangement requirements of the attributes and relationships involved. For example, it can be: { Instruction: "You are an expert in the field of new energy vehicles and have a deep understanding of related terminology. Based on the given term, please provide its attributes and relationships. Please maintain the complete format. Term: Methane Related Attributes and Relationships:", "input": "", Output:

Entry Attributes

Characteristics

Key Indicators

Disadvantages

Characteristics

Characteristics

Physical Properties

Scale

Boiling Point

Density

Ignition Point

Melting Point

Flash Point

Form

Function

Uses

Uses

Entry Relationship

Hierarchical Relationship

Classification

Classification

Classification

Classification

Classification

Classification

Composition Relationship

Belongs to

Belongs to

Belongs to

Belongs to

Belongs to

Belongs to

Belongs to

Belongs to

Input / Output

Materials-Finished Product

Materials-Finished Product

Materials-Finished Product

[0047] It should be added that the runtime environment for the P-Tuning fine-tuning task can be configured based on considerations such as computational requirements, resource availability, model complexity, and the size of the training dataset, or prior experience. For example, a runtime environment with a single GPU V100 32GB card can be selected. No further restrictions are imposed here.

[0048] Step S15: For each term in the term generation result, based on the term and its basic information, a classification model is used to obtain the corresponding classification result; wherein, the classification model is trained based on the term training data and the classification labels corresponding to the term training data.

[0049] In this embodiment, the classification model includes a first-level classifier, a second-level classifier, a first- and third-level classifier, and a second- and third-level classifier. The first-level classifier is trained based on term classification training samples and the first-level classifiers corresponding to the term classification training samples. The second-level classifier is trained based on term classification training samples and the second-level classifiers corresponding to the term classification training samples. The first- and third-level classifiers are trained on the first-level classifier using the corresponding third-level classifiers and their corresponding term classification training samples. The second- and third-level classifiers are trained on the second-level classifier using the corresponding third-level classifiers and their corresponding term classification training samples. The first-level, second-level, and third-level classifiers are obtained by dividing the term classification training samples according to a preset hierarchical division rule from high to low. The term classification training samples include terms and their corresponding basic information.

[0050] It is important to note that the number of tertiary classifiers should be configured as closely as possible to the types of secondary classification numbers involved, so as to ensure that regardless of whether the secondary predicted classification number of the same term belongs to the corresponding primary predicted classification number, there is a corresponding tertiary classifier to classify the term.

[0051] In addition, the preset hierarchical division rules can be configured according to the Chinese Library Classification (CLC) or other classification needs. The preset hierarchical division rules include at least three division levels and classification rules within each level. Taking three levels as an example, there are three levels from high to low: the first level, the second level, and the third level. The first level divides the terms in the target field according to the subject area to obtain the first classification number. The second level can further divide the terms belonging to the same first classification number based on the first level, according to the professional field or specific technical field, to obtain the second classification number. The third level can further divide the terms belonging to the same second classification number based on the second classification number, according to the specific research theme, technical point, or application, to obtain the third classification number.

[0052] Accordingly, for each term in the terminology generation results, based on the term and its basic information, a classification model is used to obtain the corresponding classification result, including: for each term in the terminology generation results, concatenating each term with its corresponding basic information to obtain the text to be classified; inputting the text to be classified into a first-level classifier and a second-level classifier respectively, to obtain the first-level predicted classification result of the text to be classified output by the first-level classifier and the second-level predicted classification result of the text to be classified output by the second-level classifier. The first-level predicted classification result includes the first-level predicted classification number of the text to be classified, and the second-level predicted classification result includes the first-level predicted classification number of the text to be classified. Secondary predicted classification number; determine whether the secondary predicted classification number of the same text to be classified belongs to the corresponding primary predicted classification number. If the secondary predicted classification number of the same text to be classified belongs to the corresponding primary predicted classification number, then the secondary and tertiary classification number classifiers are used to classify the text to be classified, and the tertiary predicted classification result of the text to be classified is obtained; otherwise, the primary and tertiary classification number classifiers are used to classify the text to be classified, and the tertiary predicted classification result of the text to be classified is obtained. Based on the primary predicted classification result, the secondary predicted classification result, and the tertiary predicted classification result of the text to be classified, the classification result of the corresponding term is obtained.

[0053] It should be added that, for each text to be classified, when inputting the same text to the primary classifier and the secondary classifier respectively, the individual text to be classified can be input into the primary classifier and the secondary classifier for classification respectively; or, the corresponding text to be classified obtained from at least two terms in the term generation result can be input into the primary classifier, and the text to be classified input into the primary classifier can be input into the secondary classifier. Accordingly, the primary prediction classification result includes the primary classification number of each text to be classified corresponding to the input, and the secondary prediction classification result includes the secondary classification number of each text to be classified corresponding to the input.

[0054] Specifically, the three-level classifier includes a fusion classifier and base classifiers. The three-level classifier is used to classify the text to be classified and obtain the three-level predicted classification number. This includes: inputting the text to be classified into each base classifier to obtain the probability value output by each base classifier; and outputting the probability value output by each base classifier to the fusion classifier for classification decision to obtain the three-level predicted classification result output by the fusion classifier.

[0055] It should be added that the base classifier can be set according to actual design requirements. For example, the three-level classifier can use ensemble learning stacking to integrate 10 base binary classifiers and 1 fusion classifier for each low-level category. No further restrictions are made here.

[0056] In addition, if there is only one text to be classified in the primary classifier and the secondary classifier, the primary classifier will perform vector transformation on the corresponding input text to be classified and classify the vector features obtained by transformation. The secondary classifier is the same as the primary classifier, and no further restrictions are made here.

[0057] If the text to be classified input into the primary and secondary classifiers includes at least two elements, the primary and secondary classifiers can be implemented using the FastText fast text classification algorithm. Specifically, the primary classifier includes a word embedding layer, a pooling layer, and a classification layer. The text to be classified is input into the primary classifier to obtain the primary predicted classification number. This process involves: inputting at least two texts to be classified into the word embedding layer for vector transformation to obtain corresponding vector features; inputting the corresponding vector features into the pooling layer for summation to obtain text vector features; and inputting the text vector features into the classification layer for classification mapping to obtain the corresponding primary predicted classification number. Similarly, the method for classifying the text using the secondary classifier can refer to the method described above for classification using the primary classifier, and will not be repeated here.

[0058] It should be noted that in steps S13-S15, reference Figure 2 For each term in the term generation result, the term is input into the basic information prediction model, the term is input into the semantic recognition model, and the corresponding classification result is obtained by using the classification model based on the term and the basic information of the term. This can be done by: traversing each term in the term generation result and inputting the traversed terms into the basic information prediction model and the semantic recognition model respectively; and after obtaining the basic information output by the basic information prediction model, concatenating the basic information with the corresponding traversed terms and inputting it into the classification model for classification.

[0059] Accordingly, after generating the corresponding term entries, the process also includes: determining whether the term generation results have been traversed completely; if so, ending the traversal; otherwise, continuing to traverse the term generation results to generate entries as described above. It should be noted that the knowledge organization system can be constructed after the initial acquisition of the corresponding term, and updated using subsequently generated terms, or it can be created for all terms after their generation. For details on term generation and knowledge organization system creation, please refer to the following description, which will not be repeated here.

[0060] Step S16: Generate entries for each term based on its basic information, attributes, relationships, and classification results, and construct a knowledge organization system based on all generated entries.

[0061] In one optional embodiment, before generating entries for corresponding terms based on the basic information, attributes, relationships, and classification results of each term, the process includes: for each term in the term generation results, automatically accessing web pages and extracting the term's definition and annotations by writing scripts or using web crawlers; and / or, querying a pre-built electronic dictionary or database for each term in the term generation results, the dictionary may include professional dictionaries such as WordNet and CNKI (China National Knowledge Infrastructure) based on cognitive linguistics; and / or, sending a query request to a pre-selected retrieval website via API or web interface, and extracting the term's definition and annotations from the returned results, the retrieval website may be an existing retrieval website, without further limitation here.

[0062] In one optional embodiment, the formal concept description includes terms, definitions and annotations, attributes and relationships. Before generating the corresponding term entries based on the basic information, attributes and relationships and classification results of each term, the method further includes: generating the formal concept description of the corresponding term by taking the basic information, attributes and relationships, classification results, definitions and annotations of the term in a preset structured form.

[0063] Furthermore, the basic information, attributes and relationships, classification results, definitions and annotations corresponding to the terms are used to generate formal concept descriptions of the corresponding terms in a preset structured format. This includes: creating a template based on the terms, definitions and annotations, attributes and relationships in a preset structured format; and filling the template with the basic information, attributes and relationships, classification results, definitions and annotations corresponding to the terms to obtain the corresponding formal concept descriptions.

[0064] It should be added that formal concept description is used to characterize the form in which the corresponding terms are represented in a predefined structured form. Furthermore, formal concept description also includes examples, which represent one or more instances used in actual contexts. Therefore, before generating entries for the corresponding terms based on their basic information, attributes, relationships, and classification results, the process includes: automatically accessing web pages and extracting text containing the target terms by writing scripts or using web crawlers; using regular expressions or other text processing techniques to identify sentences containing the target terms; filtering sentences containing the target terms; and / or, querying a pre-built electronic dictionary or database for each term in the term generation results; and / or, sending query requests to pre-selected retrieval websites via API or web interface and extracting examples of the terms from the returned results.

[0065] Accordingly, a formal concept description for the corresponding term is generated, including: taking the basic information, attributes and relationships, classification results, definitions and annotations, and examples of the term and generating a formal concept description for the corresponding term in a preset structured form. For specific implementation methods, please refer to the method described above for generating formal concept descriptions by creating templates. When creating templates, it is also necessary to combine them with examples, which will not be repeated here.

[0066] In one optional embodiment, after generating the corresponding term, the method includes: verifying the accuracy of the term and removing terms that fail the verification.

[0067] Specifically, the accuracy of the entries is verified, including: using natural language processing (NLP) tools to check whether the grammar and spelling of the entries are correct, and verifying them based on correctness; and / or, using text similarity algorithms based on a previously constructed reference dataset to evaluate the consistency between the entries and the corresponding entries in the reference dataset, and verifying them based on consistency; wherein, the reference dataset is a previously acquired professional dictionary, professional knowledge base, or a dataset of entries constructed based on prior experience or expert experience; and / or, using NLP techniques to analyze examples to evaluate the correctness of the examples, or to check whether the source of the examples is reliable.

[0068] It should be added that natural language processing tools can be such as the text proofreading tool LanguageTool or the grammarly API, and text similarity algorithms can be such as cosine similarity, edit distance, etc. The specific choice can be made according to the actual verification needs, and no further restrictions are made here.

[0069] In this embodiment, a knowledge organization system is constructed based on all generated terms, including: extracting terms from the terms and determining synonyms, near-synonyms, and related vocabulary of the terms; determining the hierarchical relationship between terms according to the classification results; organizing the extracted terms, synonyms, near-synonyms, related vocabulary, and hierarchical relationship between terms into a structured thesaurus; creating metadata for each term by using preset elements to formulate the format and preset encoding rules, including basic information, attributes and relationships, classification results, definitions and annotations, and formal concept descriptions; and obtaining the knowledge organization system based on the thesaurus and metadata of the terms.

[0070] In summary, this invention utilizes a terminology generation model to generate corresponding terms from a pre-acquired target domain dataset. Furthermore, it employs multiple models, including a basic information prediction model, a semantic recognition model, and a classification model, to process the terms from different perspectives. Through clear division of labor and mutual collaboration among these models, the basic information, attributes, relationships, and classification results of the terms are comprehensively and deeply mined. This improves the quality and efficiency of terminology processing. By comprehensively considering multi-dimensional knowledge types, different types of knowledge can be integrated into a single term, helping to break down barriers between knowledge and avoiding the limitations of understanding only a single dimension. This facilitates a deeper understanding and accurate application of terms. Moreover, when new knowledge or information emerges, since the term is constructed based on information from various dimensions, only the information in the corresponding dimensions needs to be updated. This allows the knowledge organization system to more realistically reflect the knowledge structure and relationships in the objective world, enhancing its ability to represent complex knowledge and enabling it to better process and present various types of knowledge.

[0071] The knowledge organization system construction apparatus provided by the present invention is described below. The knowledge organization system construction apparatus described below and the knowledge organization system construction method described above can be referred to in correspondence.

[0072] Figure 3 A schematic diagram of a knowledge organization system construction device is shown, the device comprising: Data acquisition module 31 acquires the target domain dataset, which includes multiple text paragraphs obtained based on the target domain file; The terminology generation module 32 retrieves text paragraphs containing target domain terms from the target domain dataset and inputs the corresponding retrieval results into the terminology generation model to obtain the terminology generation results output by the terminology generation model. The terminology generation results include at least one term. The terminology generation model is obtained by unsupervised training based on unsupervised terminology training text. The information prediction module 33 inputs each term in the term generation result into the basic information prediction model to obtain the basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained by unsupervised training based on unsupervised term training text; The semantic recognition module 34 inputs each term in the term generation result into the semantic recognition model to obtain the semantic recognition result of the corresponding term output by the semantic recognition model. The semantic recognition result includes attributes and relationships. The semantic recognition model is obtained by fine-tuning a pre-trained model based on the term training data and the attribute labels and relationship labels corresponding to the term training data. The terminology classification module 35, for each term in the terminology generation results, uses a classification model based on the term and its basic information to obtain the corresponding classification result; wherein, the classification model is trained based on the terminology training data and the classification labels corresponding to the terminology training data; System construction module 36 generates entries for each term based on its basic information, attributes, relationships, and classification results, and builds a knowledge organization system based on all generated entries.

[0073] It should be noted that the relevant functional modules in the embodiments of the present invention can be implemented by a hardware processor. The specific principle is the same as that of the method described in the above embodiments. More detailed explanations will not be repeated here.

[0074] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a knowledge organization system construction method. This method includes: acquiring a target domain dataset, which includes multiple text paragraphs obtained based on target domain files; retrieving text paragraphs containing target domain terms from the target domain dataset and inputting the corresponding retrieval results into a terminology generation model to obtain terminology generation results output by the terminology generation model, wherein the terminology generation model includes at least one term; wherein the terminology generation model is obtained through unsupervised training based on unsupervised terminology training text; and for each term in the terminology generation results, inputting the term into a basic information prediction model to obtain basic information of the corresponding term output by the basic information prediction model; wherein the basic information prediction model is based on unsupervised terminology training text. The process involves unsupervised training. For each term in the term generation results, the term is input into a semantic recognition model to obtain the semantic recognition result of the corresponding term, which includes attributes and relationships. The semantic recognition model is fine-tuned based on the term training data and the corresponding attribute and relationship labels. For each term in the term generation results, a classification model is used to obtain the corresponding classification result based on the term training data and the corresponding classification labels. Based on the basic information, attributes, relationships, and classification results of each term, a term entry is generated, and a knowledge organization system is constructed based on all generated terms.

[0075] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the knowledge organization system construction method provided by the above methods. The method includes: acquiring a target domain dataset, the target domain dataset including multiple text paragraphs obtained based on target domain files; retrieving text paragraphs containing target domain terms from the target domain dataset, and inputting the corresponding retrieval results into a term generation model to obtain a term generation result output by the term generation model, the term generation result including at least one term; wherein, the term generation model is obtained by unsupervised training based on unsupervised term training text; for each term in the term generation result, inputting the term into a basic information prediction model to obtain the output of the basic information prediction model. The process involves several steps: First, generating basic information for each term. The basic information prediction model is trained unsupervised on unsupervised term training text. For each term generated, it is input into a semantic recognition model to obtain the semantic recognition result, which includes attributes and relationships. This model is fine-tuned based on the term training data and its corresponding attribute and relationship labels. Second, for each term generated, a classification model is used to obtain the corresponding classification result based on the term and its basic information. This model is trained on the term training data and its corresponding classification labels. Finally, based on the basic information, attributes, relationships, and classification results of each term, a term entry is generated, and a knowledge organization system is constructed based on all generated terms.

[0077] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for constructing a knowledge organization system provided by the methods described above. This method includes: acquiring a target domain dataset, the target domain dataset comprising multiple text paragraphs obtained based on target domain files; retrieving text paragraphs containing target domain terms from the target domain dataset, and inputting the corresponding retrieval results into a term generation model to obtain term generation results output by the term generation model, the term generation results including at least one term; wherein the term generation model is obtained through unsupervised training based on unsupervised term training text; for each term in the term generation results, inputting the term into a basic information prediction model to obtain basic information of the corresponding term output by the basic information prediction model; wherein the basic information prediction model outputs basic information of the term. This information prediction model is obtained through unsupervised training of unsupervised terminology training text. For each term in the terminology generation result, the term is input into a semantic recognition model to obtain the semantic recognition result of the corresponding term, which includes attributes and relationships. The semantic recognition model is fine-tuned based on the terminology training data and the corresponding attribute and relationship labels. For each term in the terminology generation result, a classification model is used to obtain the corresponding classification result based on the terminology training data and the corresponding classification labels. Based on the basic information, attributes, relationships, and classification results of each term, terms for the corresponding terms are generated, and a knowledge organization system is constructed based on all generated terms.

[0078] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a knowledge organization system, characterized in that, include: Obtain a target domain dataset, which includes multiple text paragraphs obtained based on target domain files; The text paragraphs containing target domain terms are retrieved from the target domain dataset, and the corresponding retrieval results are input into the terminology generation model to obtain the terminology generation result output by the terminology generation model. The terminology generation result includes at least one term. The terminology generation model is obtained by unsupervised training based on unsupervised terminology training text. For each term in the terminology generation result, the term is input into the basic information prediction model to obtain the basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained by unsupervised training based on unsupervised terminology training text; For each term in the term generation result, the term is input into a semantic recognition model to obtain the semantic recognition result of the corresponding term output by the semantic recognition model. The semantic recognition result includes attributes and relationships. The semantic recognition model is obtained by fine-tuning a pre-trained model based on the term training data and the attribute labels and relationship labels corresponding to the term training data. For each term in the terminology generation result, a corresponding classification result is obtained using a classification model based on the term and its basic information; wherein, the classification model is trained based on the terminology training data and the classification labels corresponding to the terminology training data; Based on the basic information, attributes, relationships, and classification results of each term, generate entries for the corresponding terms, and construct a knowledge organization system based on all generated entries; Retrieve text paragraphs containing target domain terms from the target domain dataset, and input the corresponding retrieval results into the terminology generation model to obtain the terminology generation results output by the terminology generation model, including: Using keywords previously set based on target domain terms, keyword retrieval or similarity matching is performed on the target domain dataset to obtain text paragraphs containing the target domain terms as retrieval results; Based on the search results and combined with preset prompts, the terminology generation model is invoked to generate terminology results corresponding to the search results, based on the preset prompts; wherein, the preset prompts are set based on the contextual environment required for the terminology in the target domain, and the contextual environment includes the domain background and related concepts involved in the domain; The process of inputting the term into the semantic recognition model includes: Based on the instruction-based fine-tuning data, a preset fine-tuning method is used to replace traditional discrete text prompts with learnable continuous embeddings to optimize prompt-related parameters. The main parameters of the pre-trained model are frozen, and the model is fine-tuned using the instruction-based fine-tuning data. These prompt-related parameters include embedding layer and encoder parameters. During the fine-tuning process, an encoder composed of a long short-term memory network or a multilayer perceptron is introduced to dynamically generate virtual tag embeddings.

2. The knowledge organization system construction method according to claim 1, characterized in that, After obtaining the term generation result output by the term generation model, the following is included: The terms in the terminology generation results are filtered to remove duplicate terms; and / or, Based on preset verification rules, the terms in the terminology generation results are verified to remove terms that do not conform to the preset verification rules; wherein, the preset verification rules include rules for limiting terms to conform to the target domain.

3. The knowledge organization system construction method according to claim 1, characterized in that, The instruction fine-tuning data is obtained through the following methods: Obtain terminology training data and the corresponding attribute and relation labels for the terminology training data; The terminology training data and the corresponding attribute and relation labels are converted according to a preset format to obtain format-converted instruction fine-tuning data; wherein, the preset format is used to limit the arrangement format of the attributes and relations corresponding to the terms.

4. The knowledge organization system construction method according to claim 1, characterized in that, The classification model includes a primary classifier, a secondary classifier, a first tertiary classifier, and a second tertiary classifier. The primary classifier is trained based on terminology classification training samples and the primary classifiers corresponding to those training samples. The secondary classifier is trained based on terminology classification training samples and the secondary classifiers corresponding to those training samples. The first tertiary classifier is trained on the primary classifier using the corresponding tertiary classifiers and their corresponding terminology classification training samples. The second tertiary classifier is trained on the secondary classifier using the corresponding tertiary classifiers and their corresponding terminology classification training samples. The primary, secondary, and tertiary classifiers are obtained by dividing the terminology classification training samples based on a preset hierarchical division rule from high to low. For each term in the terminology generation result, based on the term and its basic information, a classification model is used to obtain the corresponding classification result, including: For each term in the terminology generation result, the term and its corresponding basic information are concatenated to obtain the text to be classified; The text to be classified is input into the first-level classifier and the second-level classifier respectively, to obtain the first-level predicted classification result of the text to be classified output by the first-level classifier and the second-level predicted classification result of the text to be classified output by the second-level classifier. The first-level predicted classification result includes the first-level predicted classification number of the text to be classified, and the second-level predicted classification result includes the second-level predicted classification number of the text to be classified. Determine whether the secondary predicted classification number of the same text to be classified belongs to the corresponding primary predicted classification number. If the secondary predicted classification number of the same text to be classified belongs to the corresponding primary predicted classification number, then the text to be classified is classified using the second and third-level classifiers to obtain the tertiary predicted classification result of the text to be classified; otherwise, the text to be classified is classified using the first and third-level classifiers to obtain the tertiary predicted classification result of the text to be classified. Based on the first-level predicted classification result, the second-level predicted classification result, and the third-level predicted classification result of the corresponding text to be classified, the classification result of the corresponding term is obtained.

5. The knowledge organization system construction method according to claim 4, characterized in that, The three-level classifier includes a fusion classifier and a base classifier. It is used to classify the text to be classified, obtaining a three-level predicted classifier, including: The text to be classified is input into each of the base classifiers to obtain the probability values ​​output by each base classifier; The probability values ​​output by each base classifier are fed into the fusion classifier for classification decision-making, resulting in a three-level predicted classification result output by the fusion classifier.

6. A knowledge organization system construction device, characterized in that, include: The data acquisition module acquires a target domain dataset, which includes multiple text paragraphs obtained based on target domain files; The terminology generation module retrieves text paragraphs containing target domain terms from the target domain dataset and inputs the corresponding retrieval results into the terminology generation model to obtain the terminology generation result output by the terminology generation model. The terminology generation result includes at least one term. The terminology generation model is obtained by unsupervised training based on unsupervised terminology training text. The information prediction module, for each term in the terminology generation result, inputs the term into the basic information prediction model to obtain the basic information of the corresponding term output by the basic information prediction model; wherein, the basic information prediction model is obtained by unsupervised training based on unsupervised terminology training text; The semantic recognition module, for each term in the term generation result, inputs the term into the semantic recognition model to obtain the semantic recognition result of the corresponding term output by the semantic recognition model. The semantic recognition result includes attributes and relationships. The semantic recognition model is obtained by fine-tuning a pre-trained model based on the term training data and the attribute labels and relationship labels corresponding to the term training data. The terminology classification module, for each term in the terminology generation result, uses a classification model based on the term and its basic information to obtain the corresponding classification result; wherein, the classification model is obtained by fine-tuning a pre-trained model based on the terminology training data and the classification labels corresponding to the terminology training data; The system construction module generates entries for each term based on its basic information, attributes, relationships, and classification results, and builds a knowledge organization system based on all generated entries. The terminology generation module is specifically used for: Using keywords previously set based on target domain terms, keyword retrieval or similarity matching is performed on the target domain dataset to obtain text paragraphs containing the target domain terms as retrieval results; Based on the search results and combined with preset prompts, the terminology generation model is invoked to generate terminology results corresponding to the search results, based on the preset prompts; wherein, the preset prompts are set based on the contextual environment required for the terminology in the target domain, and the contextual environment includes the domain background and related concepts involved in the domain; The process of inputting the term into the semantic recognition model includes: Based on the instruction-based fine-tuning data, a preset fine-tuning method is used to replace traditional discrete text prompts with learnable continuous embeddings to optimize prompt-related parameters. The main parameters of the pre-trained model are frozen, and the model is fine-tuned using the instruction-based fine-tuning data. These prompt-related parameters include embedding layer and encoder parameters. During the fine-tuning process, an encoder composed of a long short-term memory network or a multilayer perceptron is introduced to dynamically generate virtual tag embeddings.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the knowledge organization system construction method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the knowledge organization system construction method as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the knowledge organization system construction method as described in any one of claims 1 to 5.

Images