Electric power scientific research information semantic recognition method and system

Abstract

The invention discloses an electric power scientific research information semantic recognition method and system, and relates to the technical field of semantic recognition, and the method comprises the steps: building an electric power scientific research literature database, carrying out the multi-level text recognition, constructing field word vectors based on the multi-level text recognition result, and configuring the text semantic association between the field word vectors; after the target recognition phrase is obtained, search clustering under the minimum segmentation granularity is carried out on the target recognition phrase, and a search clustering result is established; executing similar matching of the field word vectors based on the search clustering result, and establishing a similar matching result; and returning a similar matching result to the target recognition phrase, executing clustering trust verification under text semantic association, and generating a semantic recognition result according to a clustering trust verification result. The technical problem that electric power scientific research information semantic understanding in the prior art depends on a rule template and cannot adapt to complex semantic variants is solved, and the technical effect of improving semantic recognition accuracy and cross-phrase semantic matching capacity is achieved.

Description

Technical Field

[0001] The present invention relates to the field of semantic recognition technology, and in particular to a method and system for semantic recognition of electric power scientific research information. Background Art

[0002] Power research information is typically presented in a variety of formats, including literature, charts, and experimental data. It features complex content and diverse semantic expressions. In the semantic recognition process, rule-based or template-based approaches are often used to extract and understand this information. While this approach is effective for structured or semantically fixed content, it often struggles to accurately capture the underlying meaning of scientific research passages characterized by frequent terminology updates, variable sentence structures, and complex contextual relationships, leading to inaccurate semantic recognition results. Summary of the Invention

[0003] The present application provides a method and system for semantic recognition of electric power scientific research information, which is used to solve the technical problem in the prior art that the semantic understanding of electric power scientific research information relies on rule templates and cannot adapt to complex semantic variants.

[0004] In view of the above problems, the present application provides a method and system for semantic recognition of electric power scientific research information.

[0005] A first aspect of the present application provides a method for semantic recognition of electric power scientific research information, the method comprising: After capturing electric power research literature through the interface, an electric power research literature database is established, and the data in the electric power research literature database includes text data, chart data, and experimental data; multi-level text recognition is performed on the electric power research literature database, and domain word vectors are constructed based on the multi-level text recognition results, and text semantic associations between domain word vectors are configured; after obtaining the target recognition segment, the target recognition segment is search clustered at the minimum segmentation granularity to establish a search clustering result; similarity matching of domain word vectors is performed based on the search clustering result to establish a similarity matching result; the similarity matching result is returned to the target recognition segment, clustering trust verification under text semantic association is performed, and a semantic recognition result is generated according to the clustering trust verification result.

[0006] A second aspect of the present application provides a semantic recognition system for electric power scientific research information, the system comprising: A database establishment module is used to establish an electric power scientific research literature database after capturing electric power scientific research literature through an interface. The data in the electric power scientific research literature database includes text data, chart data, and experimental data; a text recognition module is used to perform multi-level text recognition on the electric power scientific research literature database, construct domain word vectors based on the multi-level text recognition results, and configure text semantic associations between domain word vectors; a search clustering module is used to perform search clustering on the target recognition segment at the minimum segmentation granularity after obtaining the target recognition segment, and establish a search clustering result; a similarity matching module is used to perform similarity matching of domain word vectors based on the search clustering result, and establish a similarity matching result; a semantic recognition result generation module is used to return the similarity matching result to the target recognition segment, perform clustering trust verification under text semantic association, and generate a semantic recognition result based on the clustering trust verification result.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: After capturing electric power research literature through an interface, this application establishes an electric power research literature database, wherein the data in the electric power research literature database includes text data, chart data, and experimental data; multi-level text recognition is performed on the electric power research literature database, domain word vectors are constructed based on the multi-level text recognition results, and text semantic associations between domain word vectors are configured; after obtaining the target recognition segment, search clustering is performed on the target recognition segment at the minimum segmentation granularity to establish a search clustering result; similarity matching of domain word vectors is performed based on the search clustering result to establish a similarity matching result; the similarity matching result is returned to the target recognition segment, clustering trust verification is performed under text semantic association, and semantic recognition results are generated according to the clustering trust verification result. The present invention solves the technical problem in the prior art that the semantic understanding of electric power research information relies on rule templates and cannot adapt to complex semantic variations. By constructing multi-level text recognition and a search clustering mechanism based on semantic vectors, the technical effect of improving semantic recognition accuracy and cross-segment semantic matching capabilities is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 A flowchart of a method for semantic recognition of electric power scientific research information provided in an embodiment of the present application; Figure 2 A schematic diagram of the structure of a semantic recognition system for electric power scientific research information provided in an embodiment of the present application.

[0010] Description of the accompanying drawings: database establishment module 11, text recognition module 12, search clustering module 13, similarity matching module 14, semantic recognition result generation module 15. DETAILED DESCRIPTION

[0011] This application provides a method and system for semantic recognition of electric power scientific research information, aiming to solve the technical problem in the existing technology that the semantic understanding of electric power scientific research information relies on rule templates and cannot adapt to complex semantic variations. By constructing a multi-level text recognition and a search clustering mechanism based on semantic vectors, the application achieves the technical effect of improving the accuracy of semantic recognition and the ability of cross-segment semantic matching.

[0012] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0013] It should be noted that any variations of the terms "include" and "have" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0014] Example 1, as Figure 1 As shown, the present application provides a method for semantic recognition of electric power scientific research information, the method comprising: Step S100: After capturing electric power research literature through the interface, an electric power research literature database is established, wherein the data in the electric power research literature database includes text data, chart data, and experimental data.

[0015] In this embodiment, a public document data interface, such as a REST API, is first used to send a search request to a public document repository such as CNKI or Wanfang Data, retrieving electric power research literature based on predefined keywords and timeframes. Text extraction methods are then used to process the literature, extracting text data such as titles, abstracts, and main text. Image recognition is also used to extract chart data from the literature. Table parsing methods are then used to identify experimental data, including experimental conditions and test results, from the literature.

[0016] Finally, the extracted text data, chart data and experimental data are classified, organized and stored according to the preset template to complete the construction of the electric power scientific research literature database.

[0017] Furthermore, in the method provided in the application embodiment, the establishment of the electric power research literature database further includes: A crawling cycle for electric power scientific research literature is set, and update crawling of electric power scientific research literature is performed at a crawling cycle node; and the electric power scientific research literature database is dynamically updated according to the update crawling result.

[0018] In this embodiment, a technical expert first sets a crawling cycle for electric power research literature, for example, to update and crawl electric power research literature on a daily, weekly, or monthly basis. At each set crawling cycle, the document data interface is called, starting with the crawling time at the end of the previous cycle and crawling new electric power research literature data after that time point. This process completes the update crawling of electric power research literature.

[0019] Subsequently, the updated crawling results are compared with the data in the existing electric power scientific research literature database. By adopting the data comparison method, the new documents or updated content of the existing documents are identified, and the identified changes are written into the electric power scientific research literature database to complete the dynamic update of the electric power scientific research literature database.

[0020] Step S200: performing multi-level text recognition on the electric power scientific research literature database, constructing domain word vectors based on the multi-level text recognition results, and configuring text semantic associations between the domain word vectors.

[0021] In this embodiment of the present application, when performing multi-level text recognition on the electric power research literature database, the document structure of the documents in the electric power research literature database is first read to complete the identification and division of paragraphs. Then, the method of named entity recognition and keyword extraction is applied to extract the domain vocabulary of the documents in the electric power research literature database. Next, the extracted vocabulary is located based on the paragraph position, and a self-attention mechanism is used to perform multi-level semantic analysis to generate a multi-level text recognition result.

[0022] Next, domain word vectors are constructed based on the multi-level text recognition results. This process first extracts domain vocabulary from the multi-level text recognition results. Each domain word is converted into a corresponding vector representation using the Word2Vec method, resulting in a domain word vector containing semantic features. The cosine similarity method is then used to calculate the semantic similarity between each domain word vector. The similarity ranges from 0 to 1, with a higher value indicating a greater semantic proximity between the two words. To ensure the accuracy of semantic association, a similarity threshold, such as 0.75, is set as the judgment criterion. Specifically, when the similarity between two domain word vectors is greater than or equal to 0.75, they are considered semantically related and a corresponding textual semantic association is established. For example, the similarity between the domain word vectors representing short-circuit current and fault current is 0.82, exceeding the threshold, and therefore configured as having a semantic association. However, the similarity between the domain word vectors corresponding to power flow analysis and line coding is 0.48, below the threshold, and therefore no association is established. This completes the configuration of textual semantic associations between domain word vectors.

[0023] Furthermore, in the method provided in the embodiment of the application, the multi-level text recognition on the electric power scientific research literature database further includes: The document structure of the documents in the electric power research literature database is read, and paragraphs are marked according to the document structure; domain vocabulary extraction of the documents in the electric power research literature database is performed based on named entity recognition and keyword extraction; based on the domain vocabulary extraction results and domain vocabulary positioning, self-attention multi-level semantic analysis under paragraph marking is performed to establish a multi-level text recognition result.

[0024] In an embodiment of the present application, the document structure of the documents in the electric power scientific research document database is first read, and by adopting a paragraph format recognition method, the format information such as title number, paragraph indentation, line break, etc. in the document content is parsed to determine the boundaries of each chapter and paragraph, complete the identification of the document structure, and mark the paragraphs according to the document structure.

[0025] The content of the documents in the electric power research literature database is then processed based on named entity recognition and keyword extraction, and domain vocabulary extraction is performed on the documents within the electric power research literature database. Named entity recognition uses dictionary matching and template recognition to identify power-related terms such as short-circuit current. Keyword extraction uses the TF-IDF method to identify important domain-specific terms within the paragraph, such as power flow analysis and relay protection. The extracted content constitutes the domain vocabulary, and its specific location within the paragraph is recorded to form domain vocabulary location information.

[0026] Finally, based on the domain vocabulary extraction and domain vocabulary location results, self-attention multi-level semantic analysis is performed under paragraph identification. This process first extracts the context of each domain vocabulary based on the domain vocabulary extraction and location results, combined with paragraph identification, and sets local-global analysis weights based on the paragraph position. A calibrated multi-level search space is then configured, and the context is hierarchically divided and spatially adjusted. Finally, a multi-level attention semantic analysis is performed using the set local-global analysis weights, resulting in a multi-level text recognition result.

[0027] Furthermore, in the method provided in the embodiment of the application, the self-attention multi-level semantic analysis based on paragraph identification is performed based on the domain vocabulary extraction results and domain vocabulary positioning to establish a multi-level text recognition result, which also includes: The domain vocabulary context is extracted according to the domain vocabulary positioning and the paragraph identification, and the local-global analysis weight of the domain vocabulary context is configured according to the paragraph identification; a calibrated multi-level search space is configured, and after the calibrated multi-level search space is used to perform adaptive hierarchical space configuration of the domain vocabulary context, a multi-level attention semantic analysis is performed using the local-global analysis weight to establish a multi-level text recognition result.

[0028] In an embodiment of the present application, when extracting the domain vocabulary context based on domain vocabulary positioning and paragraph identification, a fixed window sliding method is adopted, with the position of each domain vocabulary in the paragraph as the center, and 5 words are extracted forward and backward to form the domain vocabulary context. Next, the local-global analysis weight of the domain vocabulary context is configured according to the paragraph identification. In this process, the local-global analysis weight is configured according to the paragraph identification and the frequency of occurrence of the domain vocabulary. Specifically, the word frequency normalization and paragraph priority method are adopted. When a vocabulary appears for the first time in the current paragraph, it is given a local weight of 1.0. If it appears repeatedly in other paragraphs, it is given a global weight of 0.6. When a certain vocabulary co-occurs across more than three paragraphs, it indicates that it has a global semantic effect, and its global weight is increased to 0.8. Through this process, the configuration of the local-global analysis weight of the domain vocabulary context is completed.

[0029] Next, we configure a calibrated multi-level search space. This process divides the calibrated multi-level search space into three layers: L1, L2, and L3. L1 identifies the local space within the current paragraph, L2 represents the space of adjacent paragraphs semantically related to the current paragraph, and L3 represents the global space of the entire document containing related or common domain vocabulary. When configuring the calibrated multi-level search space, we assign boundary parameters such as space labels and context start and end positions to each layer.

[0030] The calibrated multi-level search space is then used to adaptively configure the hierarchical space of domain vocabulary context. This configuration dynamically adjusts the context window size based on the distribution density and semantic coverage of domain vocabulary in each layer. A default window of ±5 words is used to extract context at the L1 layer. If a domain vocabulary appears more than twice in the L2 layer, the window is expanded to ±8 words. In the L3 layer, if the word co-occurs repeatedly across multiple paragraphs, the window is expanded to ±10 words. This process completes the configuration of the adaptive hierarchical space.

[0031] Afterwards, multi-level attention semantic analysis is performed using the local-global analysis weights. This process adopts a weighted semantic similarity scoring method combined with a context vector superposition method. First, the semantic similarity between the domain vocabulary and each word in its context is calculated in the three semantic levels L1, L2, and L3 (for example, using the cosine similarity of the word vector); then the similarity value is multiplied by the local or global analysis weight of the fragment to which the context belongs to obtain the weighted semantic contribution value of the context word; then the weighted contribution values of all context words in each layer are accumulated to form the semantic score of the layer; finally, the scores of the three levels L1, L2, and L3 are normalized according to the preset fusion strategy to form a multi-level semantic expression vector of the domain vocabulary in the full text.

[0032] Finally, based on the semantic analysis results, a multi-level text recognition result is established. The multi-level text recognition result includes the contextual extraction content of the domain vocabulary, the configured local and global analysis weights, the level identification and boundary information of the calibrated multi-level search space, the weighted semantic contribution value of the context in each semantic layer and the fused semantic score, and the multi-level semantic expression vector of the domain vocabulary in the full text.

[0033] Step S300: After obtaining the target recognition segment, search clustering is performed on the target recognition segment at the minimum segmentation granularity to establish a search clustering result.

[0034] In the embodiment of the present application, a preset target recognition segment is first obtained, where the target recognition segment is a text segment that needs to be processed for semantic recognition.

[0035] Next, the target recognition segment is searched and clustered at the minimum granularity. This process begins by segmenting the target segment at the minimum granularity. The search space is configured based on the number of minimum granularities, and a mapping is established between the search space and the corresponding minimum granularity. Subsequently, within the search space, search clustering operations are performed at increasing granularity, with the corresponding granularity as the search center. After clustering the entire search space, the granularity clustering results from each space are integrated to obtain the search clustering result.

[0036] Furthermore, in the method provided in the embodiment of the application, the search clustering of the target recognition segment at the minimum segmentation granularity to establish the search clustering result further includes: After the target recognition segment is segmented into the minimum granularity, the search space is configured according to the number of minimum segmentation granularities, and a mapping between the search space and the corresponding minimum segmentation granularity is established; within the search space, the corresponding minimum segmentation granularity is used as the search center, and incremental granularity search clustering is performed; after traversing the cluster search for each search space, a search clustering result is established based on the granularity search clustering results of all search spaces.

[0037] In an embodiment of the present application, the target recognition segment is first segmented into the minimum granularity, that is, a word segmentation method based on character segmentation is adopted to divide the segment into character units, and each Chinese character is used as a minimum segmentation granularity. Next, according to the number of minimum segmentation granularities, an equal number of search spaces are configured, that is, each character corresponds to an independent search space. For example, 8 search spaces are established for 8 characters, and each search space uses its corresponding single character as the search center. At the same time, the association between each search space and the corresponding character is recorded through the index mapping relationship to complete the mapping of the search space and the corresponding minimum segmentation granularity.

[0038] Next, within the search space, the search center is the corresponding minimum segmentation granularity, and an increasing granularity search clustering operation is performed. Specifically, starting from the search center, the character combination is expanded forward and backward in sequence, gradually increasing the combination length, for example, from a 1-character combination to a 2-character, 3-character combination until the set maximum search range is reached. At each combination stage, the semantic similarity between the combination and the search center is calculated (for example, using a cosine similarity calculation method based on word vectors) to determine whether the combination constitutes a semantic cluster. Even if a combination is judged to be clustered successfully, the search process continues to expand to a larger granularity to ensure that all possible combinations in the search space are traversed.

[0039] After traversing all search spaces, the clustering results obtained in each search space are summarized and integrated. The integration process uses a semantic overlap determination mechanism and cluster range merging rules to merge possible semantic overlap or boundary overlap combinations between multiple search spaces, ultimately establishing the search clustering results.

[0040] Step S400: performing similarity matching of domain word vectors based on the search clustering results to establish similarity matching results.

[0041] In an embodiment of the present application, when performing similarity matching of domain word vectors based on the search clustering results, the user's input data is first obtained, and the input trust identifier set therein is extracted. Next, the search constraints are configured according to the input data and the trust identifier, and the range of semantic deviation allowed for similar matching is clarified. Subsequently, under the configured constraints, for each text fragment extracted from the search clustering results, a word vector similarity calculation method (such as cosine similarity) is used to compare one by one with the preset standard domain vocabulary vector to obtain the similarity score between each fragment and the standard vocabulary. Based on the score, the pre-set threshold is combined to determine whether a valid match is constituted, and the semantically similar domain vocabulary is screened out to form the final similarity matching result.

[0042] Furthermore, in the method provided in the embodiment of the application, performing similarity matching of domain word vectors based on the search clustering results further includes: Obtaining user input data, wherein the input data is provided with an input trust identifier; configuring search constraints according to the input data and the trust identifier, performing similarity calculation of domain word vectors on the search clustering results; and completing similarity matching using the similarity calculation results.

[0043] In an embodiment of the present application, the user's input data is first obtained. The input data is the text to be analyzed actively input by the user. The input data is set with an input trust level identified by the user. The identification is expressed in numerical form. For example, 1.0 represents high trust, 0.7 represents medium trust, and 0.4 represents low trust.

[0044] Next, based on the input data and its corresponding trust identifier, the search constraints are configured using a threshold setting method. Specifically, different semantic similarity threshold intervals are set according to different trust levels. When the input trust level is 1.0, similarity matching must meet high similarity conditions (for example, similarity ≥ 0.85); when the trust level is 0.6, the matching threshold is relaxed (for example, similarity ≥ 0.6). After completing the configuration of the search constraints, the similarity calculation of the domain word vector is performed on the search clustering results. In this process, the Word2Vec method is first used to vectorize the user input text and the text fragments in the search clustering results. Then, the cosine similarity calculation method is used to compare the semantic similarity of the word vector input by the user with the word vector in the clustering results one by one to obtain the similarity calculation result.

[0045] Finally, based on the similarity calculation results and combined with the matching threshold corresponding to the input trust, all matching items with similarities higher than the threshold are screened out to construct the final similarity matching results.

[0046] Furthermore, in the method provided in the embodiment of the application, performing similarity matching of domain word vectors based on the search clustering results to establish similarity matching results further includes: After data extraction of cross-domain documents, a cross-domain database is established; in the process of performing similarity matching of domain word vectors, it is determined whether the similarity matching result fails to meet a preset matching threshold; if the similarity matching result fails to meet the preset matching threshold, the cross-domain database is activated and additional similarity matching is performed; and the similarity matching result is updated according to the additional similarity matching result.

[0047] In the embodiment of the present application, other subject documents related to the target topic are first captured through the interface, and keyword retrieval and content analysis methods are used to extract text data, including titles, abstracts, and texts, and then a cross-domain document database is constructed.

[0048] Next, in the process of performing similarity matching of domain word vectors, the Word2Vec method is first used to convert the user input text and the text fragments in the search clustering results into vector representations, and then the cosine similarity method is used to calculate the matching similarity. Afterwards, based on the input trust identifier configured in the user input, the corresponding preset matching threshold, such as 0.85 or 0.6, is selected to judge the current similarity result. If the similarity matching result cannot meet the preset matching threshold, the cross-domain database is activated and additional similarity matching is performed. In this process, the Word2Vec word vector model and cosine similarity algorithm are also used to compare the user input data with all text fragments in the cross-domain database one by one, and the text items with similarity higher than the matching threshold are screened out to form additional similarity matching results.

[0049] Finally, the original similarity matching results are updated based on the additional similarity matching results. That is, if there is a text item in the additional similarity matching results that meets the matching threshold, this result will be added as a replacement or supplement to the similarity matching results. This ensures that if the user input fails to match successfully in the original domain, it can still be effectively matched through cross-domain semantic compensation.

[0050] Step S500: returning the similarity matching result to the target recognition segment, performing clustering trust verification under text semantic association, and generating a semantic recognition result according to the clustering trust verification result.

[0051] In an embodiment of the present application, the similarity matching results are first returned to the target recognition segment. Combined with the search clustering results obtained above, the target recognition segment is segmented and combined to obtain a segmented combination set. Next, the similarity matching results and the text semantic associations are used to perform global semantic recognition of the segmented combination set, forming a preliminary semantic credibility assessment of each combination. Cluster trust verification is then performed based on this assessment result. Specifically, the semantic confidence of each combination is determined through a confidence screening mechanism, and valid combinations that meet the confidence criteria are screened out. Finally, the semantic combinations that pass the cluster trust verification are output as the final semantic recognition results.

[0052] Furthermore, in the method provided in the embodiment of the application, the similarity matching result is returned to the target recognition segment, clustering trust verification under text semantic association is performed, and semantic recognition results are generated according to the clustering trust verification result, which also includes: The target recognition segments are segmented and combined according to the search clustering results to establish a segmentation combination set; the similarity matching results and the text semantic association are used to perform global semantic recognition of the segmentation combination set, and trust screening of the segmentation combination set is performed according to the global semantic recognition results to complete clustering trust verification.

[0053] In an embodiment of the present application, the target recognition segment is first segmented and combined according to the search clustering results. In this process, the target recognition segment is reconstructed by a sliding window combination method. The minimum segmentation granularity is used as the window step size, and sliding starts from the starting position of the target recognition segment. Continuous character combinations with a length of 2 to N (preset maximum length) are extracted at a time to obtain a segmentation combination set.

[0054] Next, we use the similarity matching results and text semantic associations to perform global semantic recognition of the segmentation combination set. In this process, each combination in the segmentation combination set is first vectorized using the Word2Vec word vector model. Then, using a pre-similarity calculation method, we perform semantic similarity calculation on it and the vectors in the similarity matching results, thereby generating a set of global semantic recognition results.

[0055] Afterwards, the trust screening of the segmentation combination set is performed based on the global semantic recognition results. That is, the similarity calculation results in the global semantic recognition results are screened by a pre-set semantic similarity confidence threshold, and only the segmentation combinations whose similarity exceeds the semantic similarity confidence threshold are retained to complete the clustering trust verification.

[0056] Furthermore, in the method provided in the embodiment of the application, generating a semantic recognition result based on the clustering trust verification result further includes: Obtain user correction feedback on semantic recognition results; establish an optimized data set based on the user correction feedback, and perform feedback optimization management of semantic recognition based on the optimized data set.

[0057] In an embodiment of the present application, after obtaining the semantic recognition results, an interactive interface is opened to the user to support the user in correcting the semantic recognition results. The correction content includes replacement text and annotation position of the incorrectly recognized fragment, and the user's correction behavior is recorded as user correction feedback.

[0058] Then, based on the user correction feedback, we used an example alignment method to create an optimized dataset. Specifically, we extracted the original segment, initial recognition results, and corrected text from the user correction feedback. These three were then used to construct semantic recognition example pairs in a one-to-one correspondence. These pairs were then formatted and standardized to form the optimized dataset.

[0059] Finally, based on the optimized dataset, a rule-based induction approach is used to optimize feedback management for semantic recognition. This process involves performing semantic difference analysis on example pairs in the optimized dataset, extracting common features of recognition errors and their correction patterns, and generating new recognition rules or constraints. These are then incorporated into the subsequent recognition process, achieving feedback optimization management for semantic recognition.

[0060] In the embodiments of the present application, in summary, the embodiments of the present application have at least the following technical effects: After capturing electric power research literature through an interface, this application establishes an electric power research literature database, wherein the data in the electric power research literature database includes text data, chart data, and experimental data; multi-level text recognition is performed on the electric power research literature database, domain word vectors are constructed based on the multi-level text recognition results, and text semantic associations between domain word vectors are configured; after obtaining the target recognition segment, search clustering is performed on the target recognition segment at the minimum segmentation granularity to establish a search clustering result; similarity matching of domain word vectors is performed based on the search clustering result to establish a similarity matching result; the similarity matching result is returned to the target recognition segment, clustering trust verification is performed under text semantic association, and semantic recognition results are generated according to the clustering trust verification result. The present invention solves the technical problem in the prior art that the semantic understanding of electric power research information relies on rule templates and cannot adapt to complex semantic variations. By constructing multi-level text recognition and a search clustering mechanism based on semantic vectors, the technical effect of improving semantic recognition accuracy and cross-segment semantic matching capabilities is achieved.

[0061] Example 2, based on the same inventive concept as the method for semantic recognition of electric power scientific research information in the above embodiment, Figure 2 As shown, the present application provides a semantic recognition system for electric power scientific research information. The system and method embodiments in the present application are based on the same inventive concept. The system includes: A database establishment module 11 is used to establish an electric power scientific research literature database after capturing electric power scientific research literature through an interface. The data in the electric power scientific research literature database includes text data, chart data, and experimental data; a text recognition module 12 is used to perform multi-level text recognition on the electric power scientific research literature database, construct domain word vectors based on the multi-level text recognition results, and configure text semantic associations between domain word vectors; a search clustering module 13 is used to perform search clustering on the target recognition segment at the minimum segmentation granularity after obtaining the target recognition segment, and establish a search clustering result; a similarity matching module 14 is used to perform similarity matching of domain word vectors based on the search clustering result, and establish a similarity matching result; a semantic recognition result generation module 15 is used to return the similarity matching result to the target recognition segment, perform clustering trust verification under text semantic association, and generate a semantic recognition result based on the clustering trust verification result.

[0062] Furthermore, the system is also used to implement the following functions: The document structure of the documents in the electric power research literature database is read, and paragraphs are marked according to the document structure; domain vocabulary extraction of the documents in the electric power research literature database is performed based on named entity recognition and keyword extraction; based on the domain vocabulary extraction results and domain vocabulary positioning, self-attention multi-level semantic analysis under paragraph marking is performed to establish a multi-level text recognition result.

[0063] Furthermore, the system is also used to implement the following functions: The domain vocabulary context is extracted according to the domain vocabulary positioning and the paragraph identification, and the local-global analysis weight of the domain vocabulary context is configured according to the paragraph identification; a calibrated multi-level search space is configured, and after the calibrated multi-level search space is used to perform adaptive hierarchical space configuration of the domain vocabulary context, a multi-level attention semantic analysis is performed using the local-global analysis weight to establish a multi-level text recognition result.

[0064] Furthermore, the system is also used to implement the following functions: After the target recognition segment is segmented into the minimum granularity, the search space is configured according to the number of minimum segmentation granularities, and a mapping between the search space and the corresponding minimum segmentation granularity is established; within the search space, the corresponding minimum segmentation granularity is used as the search center, and incremental granularity search clustering is performed; after traversing the cluster search for each search space, a search clustering result is established based on the granularity search clustering results of all search spaces.

[0065] Furthermore, the system is also used to implement the following functions: Obtaining user input data, wherein the input data is provided with an input trust identifier; configuring search constraints according to the input data and the trust identifier, performing similarity calculation of domain word vectors on the search clustering results; and completing similarity matching using the similarity calculation results.

[0066] Furthermore, the system is also used to implement the following functions: The target recognition segments are segmented and combined according to the search clustering results to establish a segmentation combination set; the similarity matching results and the text semantic association are used to perform global semantic recognition of the segmentation combination set, and trust screening of the segmentation combination set is performed according to the global semantic recognition results to complete clustering trust verification.

[0067] Furthermore, the system is also used to implement the following functions: After data extraction of cross-domain documents, a cross-domain database is established; in the process of performing similarity matching of domain word vectors, it is determined whether the similarity matching result fails to meet a preset matching threshold; if the similarity matching result fails to meet the preset matching threshold, the cross-domain database is activated and additional similarity matching is performed; and the similarity matching result is updated according to the additional similarity matching result.

[0068] Furthermore, the system is also used to implement the following functions: A crawling cycle for electric power scientific research literature is set, and update crawling of electric power scientific research literature is performed at a crawling cycle node; and the electric power scientific research literature database is dynamically updated according to the update crawling result.

[0069] Furthermore, the system is also used to implement the following functions: Obtain user correction feedback on semantic recognition results; establish an optimized data set based on the user correction feedback, and perform feedback optimization management of semantic recognition based on the optimized data set.

[0070] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0071] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0072] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A semantic recognition method for electric power scientific research information, characterized in that: The method comprises: After capturing electric power research literature through the interface, an electric power research literature database is established, wherein the data in the electric power research literature database includes text data, chart data, and experimental data; Performing multi-level text recognition on the electric power scientific research literature database, constructing domain word vectors based on the multi-level text recognition results, and configuring text semantic associations between the domain word vectors; After obtaining the target recognition segment, performing search clustering on the target recognition segment at the minimum segmentation granularity to establish a search clustering result; Perform similarity matching of domain word vectors based on the search clustering results to establish similarity matching results; The similarity matching result is returned to the target recognition segment, clustering trust verification under text semantic association is performed, and a semantic recognition result is generated according to the clustering trust verification result.

2. The method for semantic recognition of electric power scientific research information according to claim 1, characterized in that: The multi-level text recognition on the electric power scientific research literature database includes: Read the document structure of the document in the electric power scientific research document database and mark the paragraphs according to the document structure; Perform domain vocabulary extraction of documents in the electric power research literature database based on named entity recognition and keyword extraction; According to the domain vocabulary extraction results and domain vocabulary positioning, self-attention multi-level semantic analysis based on paragraph marking is performed to establish multi-level text recognition results.

3. The method for semantic recognition of electric power scientific research information according to claim 2, characterized in that: The self-attention multi-level semantic analysis based on paragraph identification is performed according to the domain vocabulary extraction results and domain vocabulary positioning to establish a multi-level text recognition result, including: Extracting domain vocabulary context according to the domain vocabulary positioning and the paragraph identification, and configuring a local-global analysis weight of the domain vocabulary context according to the paragraph identification; A calibrated multi-level search space is configured, and after using the calibrated multi-level search space to perform adaptive hierarchical space configuration of domain vocabulary context, a multi-level attention semantic analysis is performed using local-global analysis weights to establish a multi-level text recognition result.

4. The method for semantic recognition of electric power scientific research information according to claim 1, characterized in that: The performing search clustering on the target recognition segment at the minimum segmentation granularity to establish the search clustering result includes: After performing minimum granularity segmentation on the target recognition segment, configuring a search space according to the number of minimum segmentation granularities, and establishing a mapping between the search space and the corresponding minimum segmentation granularity; In the search space, the corresponding minimum segmentation granularity is used as the search center to perform increasing granularity search clustering; After traversing cluster search on each search space, the search cluster result is established according to the granularity search cluster results of all search spaces.

5. The method for semantic recognition of electric power scientific research information according to claim 1, characterized in that: The performing similarity matching of domain word vectors based on the search clustering results to establish similarity matching results includes: Acquiring user input data, wherein the input data is provided with an input confidence identifier; After configuring search constraints according to the input data and the trust identifier, performing similarity calculation of domain word vectors on the search clustering results; The similarity calculation results are used to complete similarity matching.

6. A method for semantic recognition of electric power scientific research information according to claim 5, characterized in that: The method of returning the similarity matching result to the target recognition segment, performing clustering trust verification under text semantic association, and generating a semantic recognition result according to the clustering trust verification result includes: Segmenting and combining target recognition segments according to the search clustering results to establish a segmentation and combination set; The similarity matching results and the text semantic association are used to perform global semantic recognition of the segmentation combination set, and trust screening of the segmentation combination set is performed according to the global semantic recognition result to complete clustering trust verification.

7. The method for semantic recognition of electric power scientific research information according to claim 1, characterized in that: The performing similarity matching of domain word vectors based on the search clustering results to establish similarity matching results includes: After extracting data from cross-disciplinary literature, a cross-disciplinary database was established; During the similarity matching process of the domain word vectors, it is determined whether the similarity matching result fails to meet the preset matching threshold; If the similarity matching result fails to meet the preset matching threshold, activating the cross-domain database and performing additional similarity matching; The similar matching results are updated according to the additional similar matching results.

8. The method for semantic recognition of electric power scientific research information according to claim 1, characterized in that: The establishment of the electric power scientific research literature database includes: Set the crawling cycle of electric power scientific research literature and execute the update crawling of electric power scientific research literature at the crawling cycle node; The electric power research literature database is dynamically updated according to the updated crawling results.

9. The method for semantic recognition of electric power scientific research information according to claim 1, characterized in that: Generating a semantic recognition result according to the clustering trust verification result includes: Obtain user correction feedback on semantic recognition results; An optimized data set is established according to the user correction feedback, and feedback optimization management of semantic recognition is performed according to the optimized data set.

10. A semantic recognition system for electric power scientific research information, characterized in that: The system is used to execute the method for semantic recognition of electric power scientific research information according to any one of claims 1 to 9, and the system includes: A database establishment module is used to establish an electric power scientific research literature database after capturing electric power scientific research literature through the interface. The data in the electric power scientific research literature database includes text data, chart data, and experimental data; a text recognition module, configured to perform multi-level text recognition on the electric power research literature database, construct domain word vectors based on the multi-level text recognition results, and configure text semantic associations between the domain word vectors; A search clustering module is used to perform search clustering on the target recognition segment at the minimum segmentation granularity after obtaining the target recognition segment, and establish a search clustering result; A similarity matching module, configured to perform similarity matching of domain word vectors based on the search clustering results and establish similarity matching results; The semantic recognition result generation module is used to return the similarity matching result to the target recognition segment, perform clustering trust verification under text semantic association, and generate a semantic recognition result based on the clustering trust verification result.

Images