Automatic construction method and system of electromagnetic spectrum knowledge graph

Abstract

The invention provides an automatic construction method and system for an electromagnetic spectrum knowledge graph, and relates to the technical field of knowledge graph construction, and the method comprises the steps: obtaining a predetermined electromagnetic field text data set, and carrying out the data labeling to obtain an electromagnetic entity labeling set; carrying out transfer learning on the ERNIE model based on the predetermined electromagnetic field text data set and the electromagnetic entity labeling set to obtain a general information extraction model; extracting an electromagnetic spectrum entity from the input text sequence according to a predetermined text extraction structure by using a general information extraction model; performing knowledge fusion processing on the electromagnetic spectrum entity to obtain electromagnetic fusion knowledge; and constructing an electromagnetic spectrum knowledge graph. According to the data integration method and device, the technical problem that the data integration efficiency is poor due to the fact that formats and structures of multi-source data are different is solved, and the data integration efficiency is improved by constructing the general information extraction model, automatically extracting information related to the electromagnetic spectrum from the text and constructing the electromagnetic spectrum knowledge graph.

Description

Technical Field

[0001] The present application relates to the technical field of knowledge graph construction, and in particular to a method and system for automatically constructing an electromagnetic spectrum knowledge graph. Background Art

[0002] Knowledge graphs are a technology that organizes and represents information in a structured manner, storing and displaying real-world entities, concepts, and their relationships in a graph format. Electromagnetic information extraction involves complex event logic and diverse representations, including electromagnetic emergency and recovery strategies, and involves multiple specialized entities. Rule-based approaches are difficult to effectively identify entity elements in electromagnetic information extraction. Therefore, establishing a general information extraction model is an effective way to extract event elements and entities from electromagnetic information extraction. Given that electromagnetic information extraction text often includes components such as frequency bands, orbits, radars, and satellites, constructing an electromagnetic spectrum knowledge graph requires collecting specialized electromagnetic spectrum data, such as spectrum allocation reports, equipment specifications, and policy documents. Entities and relationships, such as spectrum bands, equipment types, geographic locations, and spectrum usage rules, are extracted from these data. This information is integrated into a structured knowledge graph to support applications such as spectrum management, spectrum sharing, and spectrum optimization. While constructing an electromagnetic spectrum knowledge graph is crucial for spectrum management, it faces numerous challenges in practice. The differences in the formats and structures of multi-source data are a major technical issue. Electromagnetic spectrum data are usually scattered in different data sources, such as technical reports, academic papers, policy documents, etc. The formats and structures of these data sources vary, which brings difficulties to data integration.

[0003] In summary, the existing technology has a technical problem of poor data integration efficiency due to differences in the formats and structures of multi-source data. Summary of the Invention

[0004] The purpose of this application is to provide an automated construction method and system for an electromagnetic spectrum knowledge graph, in order to solve the technical problem in the prior art of poor data integration efficiency due to differences in the formats and structures of multi-source data.

[0005] In view of the above problems, this application provides a method and system for automatically constructing an electromagnetic spectrum knowledge graph.

[0006] In the first aspect, the present application provides an automated construction method for an electromagnetic spectrum knowledge graph, which is implemented by an automated construction system for an electromagnetic spectrum knowledge graph, wherein the automated construction method for an electromagnetic spectrum knowledge graph includes: obtaining a predetermined electromagnetic field text data set, performing data annotation on the predetermined electromagnetic field text data set, and obtaining an electromagnetic entity annotation set; calling an ERNIE model, and performing transfer learning on the ERNIE model based on the predetermined electromagnetic field text data set and the electromagnetic entity annotation set to obtain a general information extraction model; using the general information extraction model, extracting electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure; performing knowledge fusion processing on the electromagnetic spectrum entities to obtain electromagnetic fusion knowledge; and constructing an electromagnetic spectrum knowledge graph based on the electromagnetic fusion knowledge.

[0007] Optionally, the data annotation includes an operation subject, an operation verb, an operation object, an operation object complement, an operation preposition adverbial, an operation postposition adverbial, an operation preposition independent word, an operation postposition independent word, and an independent noun.

[0008] Optionally, the predetermined electromagnetic field text data set is subjected to text preprocessing to obtain a standard text data set; the electromagnetic entity annotation set is subjected to annotation normalization processing to obtain a standard entity annotation set; the standard text data set and the standard entity annotation set are divided into a training set and a validation set according to a preset ratio; the training set is input into the ERNIE model for back propagation training to obtain an initial information extraction model; predefined evaluation indicators are used, and the initial information extraction model is evaluated using the validation set, and the parameters of the initial information extraction model are adjusted according to the evaluation results to obtain the general information extraction model, wherein the predefined evaluation indicators include accuracy, recall rate and score value.

[0009] Optionally, entity disambiguation processing is performed on the electromagnetic spectrum entity to obtain a first electromagnetic entity; coreference resolution processing is performed on the first electromagnetic entity to obtain a second electromagnetic entity; and entity alignment processing is performed on the second electromagnetic entity to obtain the electromagnetic fusion knowledge.

[0010] Optionally, based on preset entity rules, a first entity is extracted from the electromagnetic spectrum entity; based on the predetermined electromagnetic field text data set, multiple candidate entities corresponding to the first entity are generated; based on a multi-level attention mechanism, first neighboring information of the first entity is matched in the electromagnetic spectrum entity; the semantic similarity between the first neighboring information and the multiple candidate entities is calculated to obtain multiple similarity values; the maximum similarity value among the multiple similarity values ​​is selected as the first candidate entity; based on the first candidate entity, the electromagnetic spectrum entity is associated mapped to obtain the first electromagnetic entity.

[0011] Optionally, multiple entity words are extracted from the first electromagnetic entity, and a second entity word is selected from the multiple entity words; second neighboring information of the second entity word is matched in the first electromagnetic entity; based on preset linguistic rules, multiple second pronouns are extracted from the second neighboring information; probability evaluation is performed on the multiple second pronouns through a coreference evaluation function to obtain multiple pronoun probability values; the multiple second pronouns whose pronoun probability values ​​exceed a preset threshold are extracted to construct a second pronoun chain; the second pronoun with the highest occurrence frequency in the second pronoun chain is selected as the second standard entity, and the multiple second pronouns in the second pronoun chain are replaced to obtain a second updated pronoun chain; based on the multiple entity words, multiple updated pronoun chains are obtained, and the first electromagnetic entity is associated mapped to obtain the second electromagnetic entity.

[0012] Optionally, a first text sequence is obtained, an input electromagnetic entity is extracted from it, and entity attribute information of the input electromagnetic entity is obtained at the same time; based on the entity attribute information, a match is performed in the electromagnetic spectrum knowledge graph to determine whether there are electromagnetic entities with the same attributes; if the input electromagnetic entity contains an electromagnetic entity with the same attributes as the electromagnetic spectrum knowledge graph, a prompt is issued to the user terminal; if the feedback information of the user terminal is confirmation, the electromagnetic spectrum knowledge graph is updated.

[0013] In the second aspect, the present application also provides an automated construction system for an electromagnetic spectrum knowledge graph, which is used to execute the automated construction method for the electromagnetic spectrum knowledge graph as described in the first aspect, wherein the automated construction system for the electromagnetic spectrum knowledge graph includes: a data annotation module, which is used to obtain a predetermined electromagnetic field text data set, perform data annotation on the predetermined electromagnetic field text data set, and obtain an electromagnetic entity annotation set; a transfer learning module, which is used to call the ERNIE model, and perform transfer learning on the ERNIE model based on the predetermined electromagnetic field text data set and the electromagnetic entity annotation set to obtain a general information extraction model; an entity extraction module, which is used to use the general information extraction model to extract electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure; a knowledge fusion processing module, which is used to perform knowledge fusion processing on the electromagnetic spectrum entities to obtain electromagnetic fusion knowledge; and a knowledge graph construction module, which is used to construct an electromagnetic spectrum knowledge graph based on the electromagnetic fusion knowledge.

[0014] One or more technical solutions provided in this application have at least the following beneficial effects:

[0015] The method involves obtaining a predetermined electromagnetic field text dataset, annotating the predetermined electromagnetic field text dataset, and obtaining an electromagnetic entity annotation set. The method then calls the ERNIE model and performs transfer learning on the ERNIE model based on the predetermined electromagnetic field text dataset and the electromagnetic entity annotation set to obtain a general information extraction model. The general information extraction model is then used to extract electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure. Knowledge fusion processing is then performed on the electromagnetic spectrum entities to obtain electromagnetic fusion knowledge. Based on the electromagnetic fusion knowledge, an electromagnetic spectrum knowledge graph is constructed. In other words, by performing transfer learning on the ERNIE pre-trained language model, a general information extraction model is constructed to automatically extract key information related to the electromagnetic spectrum from unstructured text. The extracted information is then subjected to knowledge fusion processing to construct an electromagnetic spectrum knowledge graph, significantly improving the efficiency of data integration.

[0016] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, which can be implemented in accordance with the contents of the description, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically listed below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in this application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and a person of ordinary skill in the art can obtain other drawings based on the provided drawings without creative work.

[0018] Figure 1 This is a flowchart of the method for automatically constructing the electromagnetic spectrum knowledge graph in this application;

[0019] Figure 2 This is a schematic diagram of the structure of the automatic construction system of the electromagnetic spectrum knowledge graph of this application.

[0020] Explanation of the accompanying drawings: data labeling module 11, transfer learning module 12, entity extraction module 13, knowledge fusion processing module 14, knowledge graph construction module 15. DETAILED DESCRIPTION

[0021] This application addresses the existing technical problem of poor data integration efficiency due to differences in the format and structure of multi-source data by providing a method and system for automatically constructing an electromagnetic spectrum knowledge graph. By performing transfer learning on the ERNIE pre-trained language model, a general information extraction model is constructed to automatically extract key information related to the electromagnetic spectrum from unstructured text. This extracted information is then subjected to knowledge fusion processing to construct an electromagnetic spectrum knowledge graph, significantly improving data integration efficiency.

[0022] Below, the technical solutions in this application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this application, rather than all the embodiments of this application. It should be understood that this application is not limited to the example embodiments described herein. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should also be noted that, for the convenience of description, only the parts related to this application, rather than all of them, are shown in the accompanying drawings.

[0023] For example, see the attached Figure 1 The present application provides a method for automatically constructing an electromagnetic spectrum knowledge graph, wherein the method is executed by an automated construction system for an electromagnetic spectrum knowledge graph, and the method specifically includes the following steps:

[0024] S100: Acquire a predetermined electromagnetic field text dataset, and perform data annotation on the predetermined electromagnetic field text dataset to obtain an electromagnetic entity annotation set.

[0025] Furthermore, the present application S100 includes:

[0026] The data annotation includes an operation subject, an operation verb, an operation object, an operation object complement, an operation preposition adverbial, an operation postposition adverbial, an operation preposition independent word, an operation postposition independent word, and an independent noun.

[0027] Specifically, we obtained a predefined electromagnetic text dataset from sources such as academic journals and technical reports in the field. We then annotated this dataset using the data annotation platform doccano. On the doccano platform, we created a sequence annotation project. Based on the defined entity label categories, we identified and labeled nine entity types: action subjects, action verbs, action objects, action object complements, action preposition adverbials, action postposition adverbials, action preposition independent words, action postposition independent words, and independent nouns.

[0028] For example, for the following text: "The R-105D radio transceiver is a VHF portable radio transceiver used by the Soviet military. It uses a frequency of 36.0-46.1 MHz, a total radio weight of 21 kg, and a communication range of 10 kilometers when using a short whip antenna and 30-40 kilometers when using a more complex antenna." The data annotation process is as follows: Operation subject (sub): R-105D radio transceiver, Operation verb (act): is, for, Operation object (obj): VHF portable radio transceiver used by the Soviet military, 10 kilometers, 30-40 kilometers, Operation object complement (com): none, Operation preposition adverbial (preadv): operating frequency, total radio weight, communication range, Operation postposition adverbial (endadv): none, Operation preposition independent word (preind): when using a short whip antenna, when using a more complex antenna, Operation postposition independent word (endind): none, Independent noun (ind): 36.0-46.1 MHz, 21 kg.

[0029] After data annotation of a predetermined electromagnetic text dataset, an electromagnetic entity annotation set is obtained. This electromagnetic entity annotation set is a set of annotated data obtained after data annotation of the electromagnetic text dataset. It is used to identify and label electromagnetic entities in the text, including components such as frequency bands, orbits, radars, and satellites. Through data annotation, raw text data is converted into structured data for subsequent processing and analysis.

[0030] S200: Retrieve an ERNIE model, and perform transfer learning on the ERNIE model based on the predetermined electromagnetic field text dataset and the electromagnetic entity annotation set to obtain a general information extraction model.

[0031] Furthermore, the present application S200 includes:

[0032] The predetermined electromagnetic field text dataset is subjected to text preprocessing to obtain a standard text dataset; the electromagnetic entity annotation set is subjected to annotation normalization processing to obtain a standard entity annotation set; the standard text dataset and the standard entity annotation set are divided into a training set and a validation set according to a preset ratio; the training set is input into the ERNIE model for back propagation training to obtain an initial information extraction model; predefined evaluation indicators are used, and the initial information extraction model is evaluated using the validation set, and the parameters of the initial information extraction model are adjusted according to the evaluation results to obtain the general information extraction model, wherein the predefined evaluation indicators include accuracy, recall rate and score value.

[0033] Specifically, text preprocessing is performed on a predetermined electromagnetic text dataset. This involves cleaning and formatting the raw text dataset to produce content that can be recognized by the ERNIE model, known as a standard text dataset. This process may include removing excess spaces and punctuation, unifying capitalization, and removing stop words. Annotation normalization involves unifying and standardizing the electromagnetic entity annotation set to produce content that can be recognized by the ERNIE model, known as a standard entity annotation set.

[0034] ERNIE is an existing large-scale knowledge-enhanced pre-training model for language understanding and generation. Its core feature is the integration of autoregressive networks and autoencoder networks, and the introduction of large-scale knowledge graph data, which significantly improves the model's performance in natural language processing tasks. The ERNIE series of models is characterized by the integration of world knowledge to enhance language representation. Through pre-training and fine-tuning, it can be used for a variety of NLP tasks such as information extraction, text classification, named entity recognition, etc. ERNIE adopts a multi-paradigm unified pre-training framework that combines autoregressive networks (Transformer-XL structure) and autoencoder networks. The autoregressive network supports long text modeling, while the autoencoder network incrementally constructs pre-training tasks through multi-task learning, and continuously performs semantic understanding learning.

[0035] The General Information Extraction Model (GEM) leverages the linguistic expressiveness of the ERNIE model, encoding the extraction structures of different tasks into a unified representation through the Structured Extraction Language (SEL). Developed based on the ERNIE knowledge-enhanced pre-trained model, the GIEM is a general information extraction framework that enables unified modeling for tasks such as entity extraction, relationship extraction, event extraction, and sentiment analysis, while also exhibiting excellent transfer and generalization capabilities.

[0036] The pre-trained parameters of ERNIE are used as initial weights and loaded into the framework of a general information extraction model. The model is fine-tuned using a standard text dataset and a standard entity annotation set to adapt to the needs of specific tasks.

[0037] The process of dividing a standard text dataset and a standard entity annotation set into a training set and a validation set according to a preset ratio. Typically, the training set is used for model training, and the validation set is used for model evaluation. For example, 80% of the data is used as the training set, and the remaining 20% ​​is used as the validation set. Model training involves inputting the training set into the ERNIE model for backpropagation training to obtain an initial information extraction model. ERNIE is a Transformer-based pre-trained language model that has achieved excellent results in natural language processing. For example, the ERNIE model is trained on text data in the training set to learn the representation and relationships of electromagnetic entities in text.

[0038] The process of evaluating the initial information extraction model using predefined evaluation indicators. The predefined evaluation indicators include accuracy, recall, and score. Accuracy evaluates the proportion of entities whose true results and predicted results are consistent, while recall evaluates the proportion of entities whose true results and predicted results are consistent. The score is the harmonic mean of accuracy and recall, and the higher the score, the better it reflects the recognition effect of the model. The calculation expression of accuracy is: The calculation expression of recall rate is: The calculation expression of the score value is: Among them, T p To predict the number of correct entities, F p is the number of entities that are not of this type but are mistakenly predicted to be of this type, F n The number of entities in this category that were mistakenly predicted as entities of other categories, P is the precision, R is the recall, and F is the score. We use both precision and recall as metrics, and stipulate that the information extraction model must achieve a minimum accuracy of 75% and a minimum recall of 75%.

[0039] The initial information extraction model is evaluated using a validation set, and the model's precision, recall, and score on the validation set are calculated. Based on the evaluation results, model parameters are adjusted to improve performance. For example, parameters such as the learning rate and batch size are adjusted to achieve better precision and recall. Through transfer learning, a general information extraction model is developed that automatically extracts electromagnetic spectrum-related information from text. This model is then applied to construct an electromagnetic spectrum knowledge graph, improving data integration efficiency.

[0040] S300: Using the general information extraction model, extracting electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure.

[0041] Specifically, based on the characteristics of the electromagnetic spectrum field, a predefined text extraction structure is used, that is, the rules or patterns of the content that need to be extracted from the input text sequence. The input text sequence and the pre-defined text extraction structure are used as input to the general information extraction model to generate electromagnetic spectrum entities. The general information extraction model adopts a discriminative extraction model and uses the ERNIE pre-trained model as an encoder to learn the lexical, syntactic and semantic information of the language by setting up a variety of pre-training tasks. The ERNIE encoding layer consists of two layers, the lower layer is a general representation module that can capture the underlying basic features of natural language; the upper layer is a specific task representation module that can perform feature encoding according to the adapted task. When fine-tuning the model, the lower-layer general representation module is fixed, and only the upper-layer specific task representation module is fine-tuned to improve training efficiency.

[0042] For the general representation module, Transformer-XL is used as the backbone network, and a memory loop mechanism is used to model longer text sequence dependencies, enhancing the learning of the relationship between the contextual content in long text sequences. By setting the number of layers and parameters, the ability to capture the desired lexical and syntactic low-level language features is enhanced. Although the task-specific representation module also uses Transformer-XL as the backbone network, it will learn high-level semantic features based on different task modes. Transformer-XL has been improved on the basis of Transformer, solving its problems of context fragmentation and slow inference speed. Based on the fragment recursion mechanism, it improves the ability to capture long-term dependencies. During training, Transformer-XL takes input in the form of fixed-length fragments, can cache the state of the previous fragment, and reuse the hidden layer state of the previous time slice when calculating the current segment, enabling Transformer-XL to model longer-term dependencies.

[0043] A general information extraction model (GEEM) is used as a methodological tool. The GEEM is a deep learning-based model that automatically extracts entities, attributes, and relationships from text. The GEEM utilizes an ERNIE encoder to encode text sequences and extract rich semantic information. The model then predicts the location and type of each electromagnetic spectrum entity in the text based on the defined structural extraction pattern and the encoded text information. The decoded entities are output in a structured format generated by a structured coding language to obtain electromagnetic spectrum entities. By using the GEEM to extract electromagnetic spectrum entities from input text sequences, GEEM-related information can be automatically extracted from large amounts of text, improving information extraction efficiency.

[0044] S400: Performing knowledge fusion processing on the electromagnetic spectrum entity to obtain electromagnetic fusion knowledge.

[0045] Furthermore, the present application S400 includes:

[0046] Perform entity disambiguation processing on the electromagnetic spectrum entity to obtain a first electromagnetic entity; perform coreference resolution processing on the first electromagnetic entity to obtain a second electromagnetic entity; perform entity alignment processing on the second electromagnetic entity to obtain the electromagnetic fusion knowledge.

[0047] Furthermore, the present application further comprises the following steps:

[0048] Based on preset entity rules, a first entity is extracted from the electromagnetic spectrum entity; based on the predetermined electromagnetic field text data set, multiple candidate entities corresponding to the first entity are generated; based on a multi-level attention mechanism, first neighboring information of the first entity is matched in the electromagnetic spectrum entity; semantic similarity between the first neighboring information and the multiple candidate entities is calculated to obtain multiple similarity values; the maximum similarity value among the multiple similarity values ​​is selected as the first candidate entity; based on the first candidate entity, association mapping is performed on the electromagnetic spectrum entity to obtain the first electromagnetic entity.

[0049] Specifically, knowledge fusion is the process of integrating information from multiple extractions into a unified and consistent knowledge. The main goal is to solve the problems of heterogeneity and redundancy between different data sources in order to provide a more comprehensive and accurate knowledge representation. In the process of knowledge fusion, several key technologies are particularly important, including entity disambiguation, coreference resolution, entity alignment, and conflict detection. The task of entity disambiguation is to solve the polysemy problem of a named entity referent corresponding to multiple entity concepts based on the contextual information in the text. The coreference resolution problem can be described as finding all pronouns with coreference relationships and linking an entity word to each pronoun. The role of entity alignment in the construction of knowledge graphs is to find that two entity representations have the same meaning and point to the same object in the real world.

[0050] Entity disambiguation addresses polysemy, where a named entity designation corresponds to multiple entity concepts. It also addresses semantic ambiguity in electromagnetic field texts involving polysemy or homonymy. In text, the same electromagnetic spectrum entity may be referred to by different names or descriptions, such as L-band, microwave, and 1-10 GHz. Based on pre-set entity rules, the first entity is extracted from the electromagnetic spectrum entities. These rules can be applied to the electromagnetic spectrum entities extracted by the general information extraction model based on entity attributes, such as frequency range, purpose, or importance, to identify the first entity.

[0051] Based on a predetermined electromagnetic field text dataset, multiple candidate entities corresponding to a first entity are generated. These candidate entities share similar attributes with the first entity, typically representing different meanings of a single word. A multi-level attention mechanism is used to extract neighboring information about the first entity from the text, namely, its contextual information, including other words related to the entity in the sentence or paragraph. For example, if the context mentions radar equipment operating in the 10-20 GHz frequency band, it is inferred that the term "radar" here is more likely to refer to radar equipment than radar waves. This multi-level attention mechanism not only uses contextual information to determine the meaning of entities but also establishes more fine-grained associations across different levels (e.g., word, sentence, and document levels). This multi-level attention mechanism is used to extract neighboring information about the first entity from the text, namely, its contextual information, including other words related to the entity in the sentence or paragraph. For example, if the context mentions radar equipment operating in the 10-20 GHz frequency band, it is inferred that the term "radar" here is more likely to refer to radar equipment than radar waves. This multi-level attention mechanism not only uses contextual information to determine the meaning of entities but also establishes more fine-grained associations across different levels (e.g., word, sentence, and document levels).

[0052] Using semantic similarity calculation methods, such as word embedding similarity or sentence embedding similarity, the similarity between the first neighboring information and multiple candidate entities is calculated. Multiple similarity values ​​are obtained, each corresponding to a candidate entity. A simple comparison algorithm is used to select the maximum value from the multiple similarity values ​​calculated. The first candidate entity is associated and mapped with the electromagnetic spectrum entity to obtain a first electromagnetic entity, which contains information about the first entity and its related candidate entities. In other words, this most consistent meaning is embedded in the electromagnetic spectrum entity to achieve entity disambiguation. The resulting first electromagnetic entity is semantically unambiguous and unique. Entity disambiguation resolves the polysemy problem, where a named entity referent corresponds to multiple entity concepts, eliminates semantic ambiguity in different contexts, and ensures entity uniqueness.

[0053] Furthermore, the present application further comprises the following steps:

[0054] Extract multiple entity words from the first electromagnetic entity and select a second entity word from the multiple entity words; match second neighboring information of the second entity word in the first electromagnetic entity; extract multiple second pronouns from the second neighboring information based on preset linguistic rules; perform probability evaluation on the multiple second pronouns using a coreference evaluation function to obtain multiple pronoun probability values; extract the multiple second pronouns whose pronoun probability values ​​exceed a preset threshold and construct a second pronoun chain; select the second pronoun with the highest occurrence frequency in the second pronoun chain as the second standard entity, replace the multiple second pronouns in the second pronoun chain to obtain a second updated pronoun chain; obtain multiple updated pronoun chains based on the multiple entity words, perform association mapping on the first electromagnetic entity, and obtain the second electromagnetic entity.

[0055] Specifically, multiple entity words are extracted from the first electromagnetic entity, including device names and technical parameters, such as L-band, microwave, and radar. A second entity word is randomly selected from the multiple entity words as the second entity word. Secondary pronouns closely related to the second entity word are matched within the first electromagnetic entity. Based on pre-set linguistic rules, such as pronouns pointing to preceding noun phrases and demonstrative pronouns pointing to previously mentioned entities, multiple secondary pronouns are extracted from the secondary pronouns. For example, for the text: "The United States has newly developed an ANS11 radar that uses the S-band; this radar is equipped on the Nimitz aircraft carrier," "ANS11 radar" is the second entity word, while "its" and "the radar" are second pronouns. The task of coreference resolution is to link the useful information described by "its" and "the radar" to the "ANS11 radar" entity.

[0056] The co-reference evaluation function is used to calculate the probability of multiple second-referencing pronouns and second-entity words co-referring to each other, and obtain multiple reference probability values. The co-reference evaluation function is a function used to evaluate the co-reference relationship between pronouns or noun phrases and other noun phrases in a text. Co-reference relationship refers to the situation where two or more noun phrases refer to the same entity in a text. Co-reference evaluation functions are commonly used in natural language processing tasks such as text summarization, machine translation, and question-answering systems to determine the co-reference relationship between pronouns or noun phrases and other noun phrases in a text. The co-reference evaluation function can be used to solve the problem of reference resolution in text, that is, to determine the co-reference relationship between pronouns or noun phrases and other noun phrases in a text. For example, in a text summarization task, the co-reference evaluation function can help determine whether a pronoun or noun phrase in a text co-refers with a noun phrase in the summary, thereby generating a more accurate and coherent summary.

[0057] If the probability exceeds a preset threshold, the secondary pronoun is considered to have a coreference relationship with the second entity word and is extracted to construct a second coreference chain. All secondary pronouns in the second coreference chain have a coreference relationship with the second entity word. The preset threshold is typically set based on specific application scenario requirements and expert experience. By extracting secondary pronouns that exceed the preset threshold, coreference relationships in the text can be identified, supporting subsequent coreference resolution and entity alignment.

[0058] The occurrence frequencies of all second reference terms in the second reference chain are counted, the second reference term with the highest occurrence frequency in the second reference chain is selected as the second standard entity, and the positions of multiple second reference terms in the second reference chain in the first electromagnetic entity are replaced to obtain a second updated reference chain, where the second updated reference chain is a word in which words with a coreference relationship are replaced.

[0059] Repeat the above process for multiple entity words to obtain an updated referent chain corresponding to each entity word. Based on these updated referent chains, replace all coreference words in the first electromagnetic entity to make the text clearer and replace all different expressions in the text with the same entity.

[0060] The role of entity alignment in knowledge graph construction is to discover that two entities have the same meaning and point to the same object in the real world. Traditional entity alignment methods can only be applied to simple prediction tasks, that is, they can only determine the relationship between entities that have appeared in the training phase. For entities that have not appeared in the training phase, since the corresponding representation vectors cannot be obtained, it is impossible to determine whether there is a relationship between them. The entity vectors generated by the pre-trained model can overcome this problem. Therefore, this project uses an algorithm that combines the general information extraction model of the deep learning model with the TransE algorithm for entity alignment. TransE is a commonly used entity alignment algorithm. The main idea is to transform the problem of measuring the rationality of triples in the vectorized knowledge graph into the problem of measuring the distance between the head entity and the tail entity. Due to its simplicity and efficiency, this algorithm has attracted widespread attention and a large number of applications in the field.

[0061] Entity alignment combines electromagnetic spectrum entity data from multiple sources, such as literature, databases, or websites, using defined entity alignment rules and selected methods. The matched entities are then fused to construct an electromagnetic spectrum knowledge base with unified identifiers and rich attributes. This integration of information from different data sources improves data integration efficiency.

[0062] S500: Constructing an electromagnetic spectrum knowledge graph based on the electromagnetic fusion knowledge.

[0063] Specifically, key entities, such as equipment, technical parameters, and frequency ranges, are extracted from the fused electromagnetic knowledge. Relationships between entities, such as their attributes and interactions, are also identified. Relationships between entities, such as their attributes (e.g., the frequency range of a radio transceiver) and their interactions (e.g., the connection between a radio transceiver and an antenna), are identified. Relationship extraction algorithms, such as rule-based, machine learning, and deep learning methods, are used to extract relationships between entities from text.

[0064] A knowledge graph is constructed using a graph structure to represent entities and relationships. Each entity is represented as a node, and relationships between entities are represented as edges between nodes. The constructed knowledge graph is stored in a knowledge base. This knowledge graph integrates dispersed knowledge in the electromagnetic spectrum field, improving its accessibility and reusability. By representing electromagnetic spectrum entities, attributes, and relationships in a graph structure, it facilitates information retrieval, knowledge reasoning, and intelligent decision-making, providing important technical support for fields such as electromagnetic spectrum management and radar signal analysis.

[0065] Furthermore, the present application further comprises the following steps:

[0066] A first text sequence is obtained, an input electromagnetic entity is extracted from it, and entity attribute information of the input electromagnetic entity is obtained at the same time; based on the entity attribute information, a match is performed in the electromagnetic spectrum knowledge graph to determine whether there are electromagnetic entities with the same attributes; if the electromagnetic entity with the same attributes as the electromagnetic spectrum knowledge graph exists in the input electromagnetic entity, a prompt is issued to the user terminal; if the feedback information of the user terminal is confirmation, the electromagnetic spectrum knowledge graph is updated.

[0067] Specifically, a first text sequence, i.e., unstructured text containing electromagnetic entity information, is obtained. The first text sequence is preprocessed to conform to the input format of a general information extraction model. The first text sequence is then input into the general information extraction model to extract input electromagnetic entities, such as device names and technical parameters. Attribute extraction techniques are then used to extract attribute information of the electromagnetic entities, such as frequency range and power, from the text.

[0068] The electromagnetic spectrum knowledge graph is searched for electromagnetic entities with identical attributes, performed through graph database queries or similarity matching. If an electromagnetic entity with identical attributes is found, a prompt message is sent to the user, requesting confirmation. Based on the user's feedback, if the confirmation message is correct, the corresponding entity information in the knowledge graph is updated. Conversely, if the user's feedback indicates non-confirmation, it is considered a knowledge conflict, and entry into the electromagnetic spectrum knowledge graph is denied, and the electromagnetic spectrum knowledge graph is not updated.

[0069] In simple terms, if an entity and its attribute information are detected in a text, a query and comparison is performed against the existing electromagnetic spectrum knowledge graph. If the entity is found to be identical to an entity in the electromagnetic spectrum knowledge graph, it is considered a knowledge conflict and, by default, the entity is not allowed to enter the electromagnetic spectrum knowledge graph. However, to facilitate basic knowledge base maintenance and error correction, this conflict will be prompted and the user will be asked to confirm and resolve it.

[0070] The aforementioned conflict detection is performed on all input electromagnetic entities extracted from the first text sequence, and the electromagnetic spectrum knowledge graph is updated based on the conflict detection results. By continuously extracting electromagnetic entities from unstructured text and determining whether they conflict with entities in the electromagnetic spectrum knowledge graph, the electromagnetic spectrum knowledge graph is updated, continuously improving and enriching it. Automatically identifying and integrating new information ensures the dynamic updating and accuracy of the knowledge graph, ensuring the continuous updating and improvement of the electromagnetic spectrum knowledge graph, and providing more accurate and comprehensive information support for spectrum management and decision-making.

[0071] In summary, the automated construction method of the electromagnetic spectrum knowledge graph provided in this application has the following beneficial effects:

[0072] The method involves obtaining a predetermined electromagnetic field text dataset, annotating the predetermined electromagnetic field text dataset, and obtaining an electromagnetic entity annotation set. The method then calls the ERNIE model and performs transfer learning on the ERNIE model based on the predetermined electromagnetic field text dataset and the electromagnetic entity annotation set to obtain a general information extraction model. The general information extraction model is then used to extract electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure. Knowledge fusion processing is then performed on the electromagnetic spectrum entities to obtain electromagnetic fusion knowledge. Based on the electromagnetic fusion knowledge, an electromagnetic spectrum knowledge graph is constructed. In other words, by performing transfer learning on the ERNIE pre-trained language model, a general information extraction model is constructed to automatically extract key information related to the electromagnetic spectrum from unstructured text. The extracted information is then subjected to knowledge fusion processing to construct an electromagnetic spectrum knowledge graph, significantly improving the efficiency of data integration.

[0073] Example 2: Based on the same inventive concept as the method for automatically constructing the electromagnetic spectrum knowledge graph in Example 1, this application also provides an automated construction system for the electromagnetic spectrum knowledge graph. Figure 2 , the automatic construction system of the electromagnetic spectrum knowledge graph includes:

[0074] A data annotation module 11 is used to obtain a predetermined electromagnetic field text data set, perform data annotation on the predetermined electromagnetic field text data set, and obtain an electromagnetic entity annotation set; a transfer learning module 12 is used to call the ERNIE model, and perform transfer learning on the ERNIE model based on the predetermined electromagnetic field text data set and the electromagnetic entity annotation set to obtain a general information extraction model; an entity extraction module 13 is used to use the general information extraction model to extract electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure; a knowledge fusion processing module 14 is used to perform knowledge fusion processing on the electromagnetic spectrum entities to obtain electromagnetic fusion knowledge; a knowledge graph construction module 15 is used to construct an electromagnetic spectrum knowledge graph based on the electromagnetic fusion knowledge.

[0075] Furthermore, the data annotation module 11 in the automatic construction system of the electromagnetic spectrum knowledge graph is also used to:

[0076] The data annotation includes an operation subject, an operation verb, an operation object, an operation object complement, an operation preposition adverbial, an operation postposition adverbial, an operation preposition independent word, an operation postposition independent word, and an independent noun.

[0077] Furthermore, the transfer learning module 12 in the automatic construction system of the electromagnetic spectrum knowledge graph is also used to:

[0078] The predetermined electromagnetic field text dataset is subjected to text preprocessing to obtain a standard text dataset; the electromagnetic entity annotation set is subjected to annotation normalization processing to obtain a standard entity annotation set; the standard text dataset and the standard entity annotation set are divided into a training set and a validation set according to a preset ratio; the training set is input into the ERNIE model for back propagation training to obtain an initial information extraction model; predefined evaluation indicators are used, and the initial information extraction model is evaluated using the validation set, and the parameters of the initial information extraction model are adjusted according to the evaluation results to obtain the general information extraction model, wherein the predefined evaluation indicators include accuracy, recall rate and score value.

[0079] Furthermore, the knowledge fusion processing module 14 in the automatic construction system of the electromagnetic spectrum knowledge graph is also used to:

[0080] Perform entity disambiguation processing on the electromagnetic spectrum entity to obtain a first electromagnetic entity; perform coreference resolution processing on the first electromagnetic entity to obtain a second electromagnetic entity; perform entity alignment processing on the second electromagnetic entity to obtain the electromagnetic fusion knowledge.

[0081] Furthermore, the knowledge fusion processing module 14 in the automatic construction system of the electromagnetic spectrum knowledge graph is also used to:

[0082] Based on preset entity rules, a first entity is extracted from the electromagnetic spectrum entity; based on the predetermined electromagnetic field text data set, multiple candidate entities corresponding to the first entity are generated; based on a multi-level attention mechanism, first neighboring information of the first entity is matched in the electromagnetic spectrum entity; semantic similarity between the first neighboring information and the multiple candidate entities is calculated to obtain multiple similarity values; the maximum similarity value among the multiple similarity values ​​is selected as the first candidate entity; based on the first candidate entity, association mapping is performed on the electromagnetic spectrum entity to obtain the first electromagnetic entity.

[0083] Furthermore, the knowledge fusion processing module 14 in the automatic construction system of the electromagnetic spectrum knowledge graph is also used to:

[0084] Extract multiple entity words from the first electromagnetic entity and select a second entity word from the multiple entity words; match second neighboring information of the second entity word in the first electromagnetic entity; extract multiple second pronouns from the second neighboring information based on preset linguistic rules; perform probability evaluation on the multiple second pronouns using a coreference evaluation function to obtain multiple pronoun probability values; extract the multiple second pronouns whose pronoun probability values ​​exceed a preset threshold and construct a second pronoun chain; select the second pronoun with the highest occurrence frequency in the second pronoun chain as the second standard entity, replace the multiple second pronouns in the second pronoun chain to obtain a second updated pronoun chain; obtain multiple updated pronoun chains based on the multiple entity words, perform association mapping on the first electromagnetic entity, and obtain the second electromagnetic entity.

[0085] Furthermore, the automated construction system of the electromagnetic spectrum knowledge graph further includes a graph updating module, which is further configured to:

[0086] A first text sequence is obtained, an input electromagnetic entity is extracted from it, and entity attribute information of the input electromagnetic entity is obtained at the same time; based on the entity attribute information, a match is performed in the electromagnetic spectrum knowledge graph to determine whether there are electromagnetic entities with the same attributes; if the electromagnetic entity with the same attributes as the electromagnetic spectrum knowledge graph exists in the input electromagnetic entity, a prompt is issued to the user terminal; if the feedback information of the user terminal is confirmation, the electromagnetic spectrum knowledge graph is updated.

[0087] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. Figure 1The method and specific examples for automatically constructing the electromagnetic spectrum knowledge graph in Example 1 are also applicable to the automated construction system for the electromagnetic spectrum knowledge graph in this embodiment. Through the detailed description of the automated construction method for the electromagnetic spectrum knowledge graph, those skilled in the art will clearly understand the automated construction system for the electromagnetic spectrum knowledge graph in this embodiment. Therefore, for the sake of brevity, we will not describe it in detail here. As for the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant details, please refer to the method description.

[0088] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0089] Obviously, for those skilled in the art, several improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the scope of protection of the present application.

Claims

1. An automated construction method for an electromagnetic spectrum knowledge graph, characterized in that: include: Acquire a predetermined electromagnetic field text dataset, and perform data annotation on the predetermined electromagnetic field text dataset to obtain an electromagnetic entity annotation set; Retrieving an ERNIE model, and performing transfer learning on the ERNIE model based on the predetermined electromagnetic field text dataset and the electromagnetic entity annotation set to obtain a general information extraction model; Using the general information extraction model, extracting electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure; Performing knowledge fusion processing on the electromagnetic spectrum entity to obtain electromagnetic fusion knowledge; Based on the electromagnetic fusion knowledge, an electromagnetic spectrum knowledge graph is constructed.

2. The method for automatically constructing an electromagnetic spectrum knowledge graph according to claim 1, characterized in that: The data annotation includes an operation subject, an operation verb, an operation object, an operation object complement, an operation preposition adverbial, an operation postposition adverbial, an operation preposition independent word, an operation postposition independent word, and an independent noun.

3. The method for automatically constructing an electromagnetic spectrum knowledge graph according to claim 1, characterized in that: Based on the predetermined electromagnetic field text dataset and the electromagnetic entity annotation set, transfer learning is performed on the ERNIE model to obtain a general information extraction model, including: Performing text preprocessing on the predetermined electromagnetic field text dataset to obtain a standard text dataset; Performing annotation normalization processing on the electromagnetic entity annotation set to obtain a standard entity annotation set; Dividing the standard text dataset and the standard entity annotation set into a training set and a validation set according to a preset ratio; Inputting the training set into the ERNIE model for back propagation training to obtain an initial information extraction model; Predefined evaluation indicators are used to evaluate the initial information extraction model using the validation set, and parameters of the initial information extraction model are adjusted according to the evaluation results to obtain the general information extraction model, wherein the predefined evaluation indicators include accuracy, recall rate and score value.

4. The method for automatically constructing an electromagnetic spectrum knowledge graph according to claim 1, characterized in that: Performing knowledge fusion processing on the electromagnetic spectrum entity to obtain electromagnetic fusion knowledge includes: performing entity disambiguation processing on the electromagnetic spectrum entity to obtain a first electromagnetic entity; performing coreference resolution processing on the first electromagnetic entity to obtain a second electromagnetic entity; Perform entity alignment processing on the second electromagnetic entity to obtain the electromagnetic fusion knowledge.

5. The method for automatically constructing an electromagnetic spectrum knowledge graph according to claim 4, characterized in that: Performing entity disambiguation processing on the electromagnetic spectrum entity to obtain a first electromagnetic entity includes: extracting a first entity from the electromagnetic spectrum entities based on a preset entity rule; generating a plurality of candidate entities corresponding to the first entity according to the predetermined electromagnetic field text dataset; Based on a multi-level attention mechanism, matching first neighboring information of the first entity in the electromagnetic spectrum entity; Calculating semantic similarities between the first neighboring information and the plurality of candidate entities to obtain a plurality of similarity values; Selecting a maximum similarity value among the multiple similarity values ​​as a first candidate entity; Based on the first candidate entity, association mapping is performed on the electromagnetic spectrum entity to obtain the first electromagnetic entity.

6. The method for automatically constructing an electromagnetic spectrum knowledge graph according to claim 5, characterized in that: Performing coreference resolution on the first electromagnetic entity to obtain a second electromagnetic entity includes: extracting a plurality of entity words from the first electromagnetic entity, and selecting a second entity word from the plurality of entity words; matching second neighboring information of the second entity word in the first electromagnetic entity; extracting a plurality of second pronouns from the second neighboring information based on preset linguistic rules; Performing probability evaluation on the plurality of second pronouns using a coreference evaluation function to obtain a plurality of reference probability values; Extracting the second pronouns whose reference probability values ​​exceed a preset threshold value, and constructing a second pronoun chain; Selecting the second pronoun with the highest frequency of occurrence in the second pronoun chain as the second standard entity, and replacing multiple second pronouns in the second pronoun chain to obtain a second updated pronoun chain; Based on the multiple entity words, multiple updated reference chains are obtained, and association mapping is performed on the first electromagnetic entity to obtain the second electromagnetic entity.

7. The method for automatically constructing an electromagnetic spectrum knowledge graph according to claim 1, characterized in that: Also includes: Acquire a first text sequence, extract an input electromagnetic entity therefrom, and simultaneously acquire entity attribute information of the input electromagnetic entity; Based on the entity attribute information, matching is performed in the electromagnetic spectrum knowledge graph to determine whether there is an electromagnetic entity with the same attribute; If the electromagnetic entity with the same attribute as the electromagnetic spectrum knowledge graph exists in the input electromagnetic entity, a prompt is issued to the user terminal; If the feedback information from the user terminal is confirmation, the electromagnetic spectrum knowledge graph is updated.

8. An automated construction system for electromagnetic spectrum knowledge graphs, characterized by: The steps for implementing the method for automatically constructing an electromagnetic spectrum knowledge graph according to any one of claims 1 to 7, wherein the automated construction system for the electromagnetic spectrum knowledge graph comprises: A data annotation module is used to obtain a predetermined electromagnetic field text data set, perform data annotation on the predetermined electromagnetic field text data set, and obtain an electromagnetic entity annotation set; A transfer learning module is used to call the ERNIE model, and perform transfer learning on the ERNIE model based on the predetermined electromagnetic field text dataset and the electromagnetic entity annotation set to obtain a general information extraction model; an entity extraction module, configured to extract electromagnetic spectrum entities from an input text sequence according to a predetermined text extraction structure using the general information extraction model; A knowledge fusion processing module, configured to perform knowledge fusion processing on the electromagnetic spectrum entity to obtain electromagnetic fusion knowledge; The knowledge graph construction module is used to construct an electromagnetic spectrum knowledge graph based on the electromagnetic fusion knowledge.

Images