Aviation data revision system based on artificial intelligence

Abstract

The invention relates to the technical field of text processing, in particular to an aeronautical data revision system based on artificial intelligence, comprising: a text acquisition unit used for acquiring a target aeronautical data text and performing phrasing operation on the target aeronautical data text to obtain a first clause sequence of the target aeronautical data text, and obtaining a standard aviation data text, and performing sentence segmentation operation on the standard aviation data text to obtain a second clause sequence of the target aviation data. According to the method, the part needing to be revised is quickly identified from the target aviation data text through the automatic text processing flow, and comparison and revision are performed in combination with the standard aviation data, so that the manual revision time and cost can be greatly reduced, the revision accuracy is improved, human errors are avoided, and the revision efficiency is improved. And by introducing the pre-trained large language model, the system can better understand semantics in the text.

Description

Technical Field

[0001] The present invention relates to the technical field of text processing, and in particular to an aviation information revision system based on artificial intelligence. Background Art

[0002] Aeronautical data refers to various information and documents related to aviation activities, flight operations, navigation management, aircraft maintenance, aviation safety, etc. These data are crucial for airlines, pilots, air traffic controllers, aircraft maintenance personnel, etc. to ensure flight safety, compliance and efficient aviation operations.

[0003] Traditional systems usually rely on manual inspection, comparison and revision of aviation data, which is a cumbersome and time-consuming process. Each update requires personnel to check the original data and the revised content one by one, which may lead to a long revision cycle, especially when a large amount of data needs to be updated. During the manual revision process, personnel often need to manually perform many repetitive tasks, such as finding, copying, pasting and updating multiple data points, which increases unnecessary workload; and traditional systems usually rely on manual operation, so they are easily affected by fatigue, negligence and misjudgment. For example, missing certain updated content, misreading standard data, incorrectly correcting data, etc., may affect the accuracy of aviation data; and most traditional systems do not make full use of automation technology, and require manual search and revision of information in multiple materials; and when revising texts in traditional systems, manual search and separate update of images and text content are usually required, resulting in the lack of unity and integration of graphic and text materials. The updated data may lack consistency in visual and information expression, and the text and image coordination is not accurate enough. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an aviation information revision system based on artificial intelligence.

[0005] The technical solution adopted to solve the above technical problems is: an aeronautical information revision system based on artificial intelligence, an aeronautical information revision system based on artificial intelligence, characterized in that it includes: a text acquisition unit, the text acquisition unit being used to acquire a target aviation information text, perform a sentence operation on the target aviation information text to obtain a first clause sequence of the target aviation information text, acquire a standard aviation information text, perform a sentence operation on the standard aviation information text to obtain a second clause sequence of the target aviation information; a sentence matching unit, the sentence matching unit being used to perform semantic matching on each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause, that is, to obtain a second clause set corresponding to the first clause sequence; A sentence interpretation unit, the sentence interpretation unit is used to perform clause interpretation on each second clause in the second clause set according to a pre-trained large language model to obtain a clause interpretation corresponding to the second clause, that is, to obtain a clause interpretation set corresponding to the second clause set; A data revision unit, the data revision unit is used to perform context-aware generation on the clause interpretation set to obtain a data revision text corresponding to the clause interpretation set; A data generation unit, the data generation unit is used to match the second clause set with the standard aerial image set to obtain the standard aerial image corresponding to the second clause, and add the standard aerial image to the corresponding position in the data revision text to obtain the aerial image and text revision data.

[0006] Preferably, semantically matching each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause includes: Performing a word segmentation operation on the first clause and the second clause to obtain a first word sequence corresponding to the first clause and a second word sequence corresponding to the second clause; Performing a character segmentation operation on the first clause and the second clause to obtain a first character sequence corresponding to the first clause and a second character sequence corresponding to the second clause; Mapping the first word and the first character corresponding to the first word into a first word vector according to the weight of the pre-trained Word2vec model; mapping the second word and the second character corresponding to the second word into a second word vector according to the weight of the pre-trained Word2vec model; Regularizing the first word vector and the second word vector to obtain a first standard word vector corresponding to the first word vector and a second standard word vector corresponding to the second word vector.

[0007] Preferably, semantically matching each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause includes: Performing difference encoding on the first standard word vector and the second standard word vector to obtain a difference vector between the first standard word vector and the second standard word vector; Interactively encoding the first standard word vector and the second standard word vector to obtain an interactive vector between the first standard word vector and the second standard word vector; The difference vector and the interaction vector are fused to obtain a fused vector, and the fused vector is classified to obtain a matching degree between the first clause and the second clause.

[0008] Preferably, performing difference encoding on the first standard word vector and the second standard word vector to obtain a difference vector between the first standard word vector and the second standard word vector includes: Performing context encoding on the first standard word vector and the second standard word vector according to a bidirectional GRU to obtain a first context vector corresponding to the first standard word vector and a second context vector corresponding to the second standard word vector; The distance between the first context vector and the second context vector is determined according to the Manhattan distance, and the first context vector and the second context vector are differentially encoded according to the distance to obtain a difference vector between the first standard word vector and the second standard word vector.

[0009] Preferably, interactively encoding the first standard word vector and the second standard word vector to obtain an interactive vector between the first standard word vector and the second standard word vector includes: The first standard word vector and the second standard word vector are interactively encoded according to the multi-head attention mechanism to obtain an interaction vector between the first standard word vector and the second standard word vector, wherein the expression of the interaction vector is as follows: ; in, represents the interaction vector, , and Indicates the query, key, and value projection matrices of the header, and represents the first standard word vector and the second standard word vector, represents the multi-head attention mechanism, Represents a concatenation operation.

[0010] Preferably, context-aware generation is performed on the clause interpretation set to obtain a document revision text corresponding to the clause interpretation set, including: Dependency syntactic analysis is performed on the clause interpretation set to generate a weighted directed graph, wherein the expression of the weighted directed graph is as follows: ; in, represents a weighted directed graph, Represents a collection of nodes. represents the set of dependency edges, represents a syntactic weight set, wherein the nodes in the node set correspond to the lexical units in the clause interpretation in the clause interpretation set, the dependency edges in the dependency edge set correspond to the dependency relationship types between the lexical units, and the syntactic weights in the syntactic weight set correspond to the syntactic association strength; The clause interpretation set is subjected to semantic role labeling to obtain a semantic role labeling result, and a hyperedge set is constructed based on the semantic role labeling result, wherein the expression of the hyperedge in the hyperedge set is as follows: ; in, represents a hyperedge, It represents the 0th predicate-argument structure in the clause interpretation; A hypergraph corresponding to the weighted directed graph is constructed based on the hyperedge set, wherein the expression of the hypergraph is as follows: ; in, represents a hypergraph, Represents a set of hyperedges.

[0011] Preferably, context-aware generation is performed on the clause interpretation set to obtain a document revision text corresponding to the clause interpretation set, further comprising: Performing feature encoding on the hypergraph based on a graph attention network to obtain a context-aware feature vector corresponding to the hypergraph; A prototype vector is constructed based on the context-aware feature vector, wherein the expression of the prototype vector is as follows: ; in, represents the prototype vector, represents the average pooling operation, represents the context-aware feature vector; The semantic deviation of each clause interpretation in the clause interpretation set is calculated based on the prototype vector, wherein the calculation formula of the semantic deviation is as follows: ; in, Represents prototype vector and candidate clause interpretation The semantic deviation between Indicates the interpretation of candidate clauses The output vector of The candidate clause interpretation is dynamically modulated based on a gating mechanism, wherein the expression of the dynamically modulated candidate interpretation is as follows: ; in, represents the dynamic modulation candidate interpretation representation, represents the activation function, represents the bias term, Represents the initialization of hyperedge connections in the hypergraph; When the semantic deviation is less than a preset deviation threshold, the candidate interpretation corresponding to the semantic deviation is used as a clause of the data revision text.

[0012] Preferably, performing image-text matching on the second clause set and the standard aerial image set to obtain the standard aerial image corresponding to the second clause includes: Extracting features from the standard aerial image according to bottom-up attention to generate a regional feature vector, dividing the standard aerial image into a plurality of blocks to obtain a position feature vector, and extracting features from the second clause according to Bi-GRU to obtain a text feature vector; The position feature vector is transformed into a query vector through a linear layer, and the region feature vector is used as a key and value vector for cross attention to obtain a visual feature vector; The text feature vector and the visual feature vector are sampled for similarity to obtain an overall similarity. If the overall similarity is higher than a preset similarity threshold, the second clause matches the standard aerial image.

[0013] Preferably, similarity sampling is performed on the text feature vector and the visual feature vector to obtain overall similarity, including: Calculating the similarity between the text feature vector and the visual feature vector, wherein the similarity uses cosine similarity; Determining a semantic relationship between the text feature vector and the visual feature vector according to the similarity; Determining a weighted similarity of the block in the standard aerial image according to a semantic relationship between the text feature vector and the visual feature vector; The overall similarity between the text feature vector and the visual feature vector is determined according to the weighted similarity of the blocks in the standard aerial image.

[0014] Preferably, the calculation formula of the semantic relationship is as follows: ; in, represents the semantic relationship between the second clause corresponding to the text feature vector and the standard aerial image corresponding to the visual feature vector, represents a mask that is equal to the input when it is positive, and 0 otherwise, represents the similarity between the text feature vector and the visual feature vector, Indicates the preset parameter threshold, represents the number of blocks in a standard aerial image, represents the activation function; The calculation formula of the weighted similarity is as follows: ; in, represents the weighted similarity of the block in the standard aerial image, A visual feature vector representing a block; The calculation formula of the overall similarity is as follows: ; in, represents the overall similarity between the text feature vector and the visual feature vector, Represents a text feature vector.

[0015] The beneficial effects of the present invention are as follows: (1) The present invention can quickly identify the parts that need to be revised from the target aviation data text through an automated text processing process (such as text sentence segmentation, semantic matching, clause interpretation, etc.), and compare and revise them in combination with standard aviation data. This can greatly reduce the time and cost of manual revision, improve the accuracy of revision, and avoid human errors. By introducing a pre-trained large language model, the system can better understand the semantics in the text. Especially in complex aviation data, the use of a large language model can ensure an in-depth understanding of the aviation data text and provide accurate interpretations based on the context, ensuring that the revised text is more semantically accurate and professional. (2) The present invention automatically selects and inserts standard aviation images related to the revised text by matching the text with a set of standard aviation images. This automated text-to-text matching function avoids the tedious process of manually searching and inserting images, ensuring that the revised aviation data is both accurate and professional. The text revision can correspond to the corresponding image information, which improves the integrity and practicality of the data, and can perform context-aware generation, that is, to generate appropriate revision content based on the context of the entire text, rather than just revising individual sentences one by one. This can ensure that the revised data is more coherent and natural, without logical jumps or information gaps; (3) The present invention ensures that all revised data follow unified standards through the reference of standard aviation data and standard aviation images, avoiding inconsistencies or non-standardization in the content of the data. This is particularly important in the aviation field because accurate and standardized aviation data is crucial to flight safety and operations. In addition, the generated aviation graphic revision data not only provides text revisions, but also improves the visualization of the data through the combination of graphics and text. This enables aviation personnel to understand the aviation data more clearly and intuitively, especially the specific images and data that may be involved in flight operations, which increases the ease of use and practical application value of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the system architecture of an overall system in an embodiment of the present invention.

[0017] Figure numerals: 1. text acquisition unit; 2. sentence matching unit; 3. sentence interpretation unit; 4. data revision unit; 5. data generation unit. DETAILED DESCRIPTION

[0018] Embodiment 1, as Figure 1 As shown, the present invention proposes an artificial intelligence-based aviation information revision system, comprising: The text acquisition unit 1 is used to acquire a target aviation data text, perform a sentence operation on the target aviation data text to obtain a first clause sequence of the target aviation data text, acquire a standard aviation data text, perform a sentence operation on the standard aviation data text to obtain a second clause sequence of the target aviation data; Sentence matching unit 2, sentence matching unit 2 is used to semantically match each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause, that is, to obtain a second clause set corresponding to the first clause sequence; The sentence interpretation unit 3 is used to perform a clause interpretation on each second clause in the second clause set according to the pre-trained large language model to obtain a clause interpretation corresponding to the second clause, that is, to obtain a clause interpretation set corresponding to the second clause set; The data revision unit 4 is used to generate a context-aware sub-clause interpretation set to obtain a data revision text corresponding to the sub-clause interpretation set; The data generation unit 5 is used to match the second clause set with the standard aerial image set to obtain the standard aerial image corresponding to the second clause, and add the standard aerial image to the corresponding position in the data revision text to obtain the aerial image revision data.

[0019] In the present invention, the target aviation information text may contain specific information in the aviation field, while the standard aviation information text is a standard version used for comparison and normalization; semantic matching refers to judging the similarity or equivalence of two sentences in meaning through natural language processing technology, rather than just through literal similarity; the pre-trained large language model interprets the matched second clause (the clause from the standard aviation information text), that is, generates an easy-to-understand explanation or conversion, which helps to ensure that the extracted content is more clearly and accurately conveyed to users; the large language model refers to a language processing model with strong semantic understanding and generation capabilities trained through deep learning technology (such as GPT, BERT, etc.); context-aware generation refers to adjusting and improving the interpretation content according to context information to ensure that the interpretation is more in line with actual application scenarios; image-text matching refers to combining text content with related image content to make the information more visual and easy to understand, which is particularly important in aviation information because visual information can usually help users better understand highly technical content.

[0020] Embodiment 2, an artificial intelligence-based aviation information revision system proposed by the present invention, compared with embodiment 1, this embodiment further includes: semantically matching each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause, including: Perform word segmentation on the first clause and the second clause to obtain a first word sequence corresponding to the first clause and a second word sequence corresponding to the second clause; Perform character segmentation on the first clause and the second clause to obtain a first character sequence corresponding to the first clause and a second character sequence corresponding to the second clause; Map the first word and the first character corresponding to the first word to a first word vector according to the weights of the pre-trained Word2Vec model; map the second word and the second character corresponding to the second word to a second word vector according to the weights of the pre-trained Word2Vec model; Regularize the first word vector and the second word vector to obtain a first standard word vector corresponding to the first word vector and a second standard word vector corresponding to the second word vector.

[0021] In this embodiment, word segmentation is to split a continuous text into meaningful basic units (such as words or phrases). In Chinese text processing, since there are no spaces to separate words, it is necessary to use a word segmentation algorithm to split the text into individual words. In the above operations, the first clause and the second clause are respectively split into their own word sequences for subsequent processing; character segmentation is to split the text by characters, and each word is disassembled into individual characters. For example, "computer" is split into characters such as "ji", "suan", "ji", etc. This step is usually used for further text processing, especially for some models that require character-level processing; Word2Vec is a word vectorization model that maps each word to a vector (a point in a high-dimensional space), and this vector can express the semantic relationship of the word. The Word2Vec model is trained based on a large amount of text data to generate the vector representation of each word. Through this vector representation, the similarity between different words can be quantified and compared; regularization is a commonly used mathematical and statistical method used to adjust the model output or data to make it conform to a certain specification or standard. Regularization operations usually refer to adjusting the word vector so that it is within a unified scale or range. Common regularization methods include L2 regularization, that is, adjusting the vector to a unit length.

[0022] In an optional embodiment, semantic matching is performed between each first clause in the first clause sequence and the second clause sequence to obtain a second clause that matches the first clause, including: Perform differential encoding on the first standard word vector and the second standard word vector to obtain a difference vector between the first standard word vector and the second standard word vector; Perform interaction encoding on the first standard word vector and the second standard word vector to obtain an interaction vector between the first standard word vector and the second standard word vector; The difference vector and the interaction vector are fused to obtain a fused vector, and the fused vector is classified to obtain a matching degree between the first clause and the second clause.

[0023] In an optional embodiment, performing difference encoding on the first standard word vector and the second standard word vector to obtain a difference vector between the first standard word vector and the second standard word vector includes: Performing context encoding on the first standard word vector and the second standard word vector according to the bidirectional GRU to obtain a first context vector corresponding to the first standard word vector and a second context vector corresponding to the second standard word vector; The distance between the first context vector and the second context vector is determined according to the Manhattan distance, and the first context vector and the second context vector are differentially encoded according to the distance to obtain a differential vector between the first standard word vector and the second standard word vector.

[0024] It should be noted that the bidirectional GRU is an extension of the GRU. It encodes the sequence data in two directions (from front to back and from back to front) respectively, so that the model can not only use the context information of the current word, but also combine the future context information to more comprehensively understand the input data. Context encoding refers to the processing of words or characters through a certain model (such as GRU, LSTM, Transformer, etc.) to capture their semantic information in the context. The goal of context encoding is to convert the representation of each word or symbol (such as word vector) into a new representation containing its contextual information. Manhattan distance, also known as city block distance, is a method for calculating the distance between two points. In an n-dimensional space, Manhattan distance is calculated as the sum of the absolute values ​​of the differences between two points in each dimension. Difference encoding is the encoding process of the difference between two vectors. Its core purpose is to calculate the difference between two vectors and convert it into a new vector representation. The difference vector is a new vector obtained by calculating the difference between two vectors (here, the first context vector and the second context vector). This difference vector represents the semantic difference between two words (or their contexts) and is usually used to analyze the semantic relationship or difference between them.

[0025] In an optional embodiment, interactively encoding the first standard word vector and the second standard word vector to obtain an interactive vector between the first standard word vector and the second standard word vector includes: The first standard word vector and the second standard word vector are interactively encoded according to the multi-head attention mechanism to obtain an interaction vector between the first standard word vector and the second standard word vector, where the expression of the interaction vector is as follows: ; in, represents the interaction vector, , and Indicates the query, key, and value projection matrices of the header, and represents the first standard word vector and the second standard word vector, represents the multi-head attention mechanism, Represents a concatenation operation.

[0026] It should be noted that the attention mechanism is a technology used in neural networks to allow the model to focus on different parts of the input (such as different words in a sequence). It determines which parts should be paid more attention to by calculating the weights of different parts of the input (i.e., attention weights).

[0027] In an optional embodiment, context-aware generation is performed on the clause interpretation set to obtain a document revision text corresponding to the clause interpretation set, including: The dependency syntactic analysis is performed on the clause interpretation set to generate a weighted directed graph, where the expression of the weighted directed graph is as follows: ; in, represents a weighted directed graph, Represents a collection of nodes. represents the set of dependency edges, represents a syntactic weight set, wherein the nodes in the node set correspond to the lexical units in the clause interpretation in the clause interpretation set, the dependency edges in the dependency edge set correspond to the dependency relationship types between the lexical units, and the syntactic weights in the syntactic weight set correspond to the syntactic association strength; The clause interpretation set is semantically labeled to obtain the semantic role labeling result, and a hyperedge set is constructed based on the semantic role labeling result. The expression of the hyperedge in the hyperedge set is as follows: ; in, represents a hyperedge, It represents the 0th predicate-argument structure in the clause interpretation; The hypergraph corresponding to the weighted directed graph is constructed based on the hyperedge set, where the expression of the hypergraph is as follows: ; in, represents a hypergraph, Represents a set of hyperedges.

[0028] It should be noted that dependency parsing is a syntactic analysis method that regards lexical units in a sentence as nodes and the relationships between nodes (such as subject-predicate relationship, verb-object relationship, etc.) as directed edges, and analyzes the dependency relationships between lexical units. The result it generates is usually a directed graph, in which nodes represent words or words, and edges represent the syntactic relationships between them; a weighted directed graph is a concept in graph theory, in which each edge corresponds to a weight value, and the weight represents a certain characteristic or importance of the edge. In syntactic analysis, the weight of an edge usually represents the strength, degree of association, or some other characteristics of the dependency relationship; the dependency edge set is the set of all edges in a weighted directed graph, each edge represents a dependency relationship, usually indicating the syntactic association between words (such as subject-predicate relationship, verb-object relationship, etc.), and each edge connects two lexical units, indicating the grammatical relationship between them; semantic role labeling is to label the lexical units in a sentence, and label In order to identify the semantic relationship and function of each component in a sentence, a hyperedge set is an edge set in a hypergraph. A hypergraph is a graph theory structure in which edges (i.e., hyperedges) can connect multiple nodes, not just two. Each hyperedge can represent a more complex semantic structure (such as a predicate-argument structure). When constructing a weighted directed graph, a hyperedge represents a higher-level semantic structure (such as a predicate and its related arguments). A hyperedge is an edge in a hypergraph. It not only connects two nodes, but usually connects multiple nodes. In semantic analysis, a hyperedge can represent some complex semantic relationship in a sentence, such as the relationship between a predicate and multiple arguments. A predicate-argument structure refers to the semantic structure of a predicate and its related arguments. A hypergraph is an extension of graph theory in which edges (hyperedges) can connect multiple nodes, not just two nodes.

[0029] In an optional embodiment, context-aware generation is performed on the clause interpretation set to obtain a document revision text corresponding to the clause interpretation set, further comprising: The hypergraph is feature encoded based on the graph attention network to obtain the context-aware feature vector corresponding to the hypergraph; The prototype vector is constructed based on the context-aware feature vector, where the expression of the prototype vector is as follows: ; in, represents the prototype vector, represents the average pooling operation, represents the context-aware feature vector; The semantic deviation of each clause interpretation in the clause interpretation set is calculated based on the prototype vector, where the calculation formula of the semantic deviation is as follows: ; in, Represents prototype vector and candidate clause interpretation The semantic deviation between Indicates the interpretation of candidate clauses The output vector of Based on the gating mechanism, the candidate clause interpretation is dynamically modulated to express the candidate interpretation, wherein the expression of the dynamically modulated candidate interpretation representation is as follows: ; in, represents the dynamic modulation candidate interpretation representation, represents the activation function, represents the bias term, Represents the initialization of hyperedge connections in the hypergraph; When the semantic deviation is less than a preset deviation threshold, the candidate interpretation corresponding to the semantic deviation is used as a clause of the data revision text.

[0030] It should be noted that the graph attention network is a deep learning model specifically used for learning graph structured data. It introduces an attention mechanism to perform weighted aggregation on each node in the graph, thereby automatically assigning different importance weights to neighboring nodes during the learning process. This method can more effectively capture the relationship between different nodes in the graph structure, and is especially suitable for heterogeneous graphs and nodes with multiple features. The context-aware feature vector refers to the feature representation obtained when encoding the hypergraph through the graph attention network, which can perceive the contextual information between nodes. The prototype vector is a vector used to represent a specific category or semantics, which is usually obtained through pooling operations (such as average pooling) based on multiple vectors. The semantic deviation is used to measure the distance between two vectors (such as prototypes). The smaller the semantic deviation, the closer the semantics between the prototype vector and the candidate clause interpretation, while the larger the deviation, the greater the semantic difference between the two. Semantic deviation is often used in natural language processing tasks to evaluate the quality of candidate answers or explanations. Dynamic modulation of candidate interpretation representation refers to the use of a gating mechanism to dynamically adjust the candidate clause interpretation so that the representation of the candidate interpretation is more suitable for the current context or task requirements. By adjusting these representations, the model can capture semantic information more accurately and improve the final prediction accuracy. Hyperedge connection initialization refers to the initial connection relationship of each hyperedge in the hypergraph. These connection relationships define which nodes are related by hyperedges.

[0031] In an optional embodiment, performing image-text matching on the second clause set and the standard aerial image set to obtain the standard aerial image corresponding to the second clause includes: The standard aerial image is feature extracted according to bottom-up attention to generate a regional feature vector, the standard aerial image is divided into several blocks to obtain a position feature vector, and the second clause is feature extracted according to Bi-GRU to obtain a text feature vector; The position feature vector is transformed into the query vector through a linear layer, and the regional feature vector is used as the key and value vector for cross attention to obtain the visual feature vector; The text feature vector and the visual feature vector are sampled for similarity to obtain the overall similarity. If the overall similarity is higher than a preset similarity threshold, the second clause matches the standard aerial image.

[0032] It should be noted that bottom-up attention is an attention mechanism based on image features. It pays attention layer by layer from low-level features of the image (such as edges, colors, textures, etc.) to higher-level semantic information. In image processing, it is common to extract features through convolutional neural networks (CNNs) and use these features to perform weighted calculations in the image to determine which areas the model should focus on. Regional feature vectors are feature representations extracted from images to describe a specific area (such as an image block). These features are usually extracted through convolutional neural networks (CNNs) or similar deep learning models, reflecting the visual information of the area, such as color, shape, texture, etc.; Bi-GRU is a bidirectional gated recurrent unit (GRU), a variant of RNN (recurrent neural network). When processing sequence data, it considers both the forward and backward information of the input sequence to better capture contextual relationships. Compared with unidirectional RNNs, Bi-GRU can provide more contextual information, so it is often used to process text data in NLP tasks; query vectors, key vectors, and value vectors are important components of attention mechanisms (such as self-attention, cross-attention, etc.).

[0033] In an optional embodiment, similarity sampling is performed on the text feature vector and the visual feature vector to obtain the overall similarity, including: Calculate the similarity between the text feature vector and the visual feature vector, where the similarity uses cosine similarity; Determine the semantic relationship between the text feature vector and the visual feature vector based on the similarity; Determine the weighted similarity of the blocks in the standard aerial image according to the semantic relationship between the text feature vector and the visual feature vector; The overall similarity between the text feature vector and the visual feature vector is determined based on the weighted similarity of the blocks in the standard aerial image.

[0034] In an optional embodiment, the calculation formula of the semantic relationship is as follows: ; in, represents the semantic relationship between the second clause corresponding to the text feature vector and the standard aerial image corresponding to the visual feature vector, represents a mask that is equal to the input when it is positive, and 0 otherwise, represents the similarity between the text feature vector and the visual feature vector, Indicates the preset parameter threshold, represents the number of blocks in a standard aerial image, represents the activation function; The calculation formula of weighted similarity is as follows: ; in, represents the weighted similarity of the block in the standard aerial image, A visual feature vector representing a block; The calculation formula for overall similarity is as follows: ; in, represents the overall similarity between the text feature vector and the visual feature vector, Represents a text feature vector.

[0035] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. An artificial intelligence-based aviation information revision system, characterized in that: include: A text acquisition unit (1), the text acquisition unit (1) is used to acquire a target aviation information text, perform a sentence splitting operation on the target aviation information text to obtain a first clause sequence of the target aviation information text, acquire a standard aviation information text, perform a sentence splitting operation on the standard aviation information text to obtain a second clause sequence of the target aviation information; A sentence matching unit (2), the sentence matching unit (2) being used to perform semantic matching on each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause, that is, to obtain a second clause set corresponding to the first clause sequence; A sentence interpretation unit (3), the sentence interpretation unit (3) being used to perform a clause interpretation on each second clause in the second clause set according to a pre-trained large language model, so as to obtain a clause interpretation corresponding to the second clause, that is, to obtain a clause interpretation set corresponding to the second clause set; A data revision unit (4), the data revision unit (4) being used to perform context-aware generation on the clause interpretation set to obtain a data revision text corresponding to the clause interpretation set; A data generation unit (5), the data generation unit (5) is used to perform image-text matching on the second clause set and the standard aerial image set to obtain the standard aerial image corresponding to the second clause, and add the standard aerial image to the corresponding position in the data revision text to obtain the aerial image-text revision data.

2. The aeronautical information revision system based on artificial intelligence according to claim 1, characterized in that: Semantically matching each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause includes: Performing a word segmentation operation on the first clause and the second clause to obtain a first word sequence corresponding to the first clause and a second word sequence corresponding to the second clause; Performing a character segmentation operation on the first clause and the second clause to obtain a first character sequence corresponding to the first clause and a second character sequence corresponding to the second clause; Mapping the first word and the first character corresponding to the first word into a first word vector according to the weight of the pre-trained Word2vec model; mapping the second word and the second character corresponding to the second word into a second word vector according to the weight of the pre-trained Word2vec model; Regularizing the first word vector and the second word vector to obtain a first standard word vector corresponding to the first word vector and a second standard word vector corresponding to the second word vector.

3. The aeronautical information revision system based on artificial intelligence according to claim 2, characterized in that: Semantically matching each first clause in the first clause sequence with the second clause sequence to obtain a second clause matching the first clause includes: Performing difference encoding on the first standard word vector and the second standard word vector to obtain a difference vector between the first standard word vector and the second standard word vector; Interactively encoding the first standard word vector and the second standard word vector to obtain an interactive vector between the first standard word vector and the second standard word vector; The difference vector and the interaction vector are fused to obtain a fused vector, and the fused vector is classified to obtain a matching degree between the first clause and the second clause.

4. The aeronautical information revision system based on artificial intelligence according to claim 3, characterized in that: Performing difference encoding on the first standard word vector and the second standard word vector to obtain a difference vector between the first standard word vector and the second standard word vector includes: Performing context encoding on the first standard word vector and the second standard word vector according to a bidirectional GRU to obtain a first context vector corresponding to the first standard word vector and a second context vector corresponding to the second standard word vector; The distance between the first context vector and the second context vector is determined according to the Manhattan distance, and the first context vector and the second context vector are differentially encoded according to the distance to obtain a difference vector between the first standard word vector and the second standard word vector.

5. The aeronautical information revision system based on artificial intelligence according to claim 4, characterized in that: Interactively encoding the first standard word vector and the second standard word vector to obtain an interactive vector between the first standard word vector and the second standard word vector includes: The first standard word vector and the second standard word vector are interactively encoded according to the multi-head attention mechanism to obtain an interaction vector between the first standard word vector and the second standard word vector, wherein the expression of the interaction vector is as follows: ； in, represents the interaction vector, , and Indicates the query, key, and value projection matrices of the header, and represents the first standard word vector and the second standard word vector, represents the multi-head attention mechanism, Represents a concatenation operation.

6. The aeronautical information revision system based on artificial intelligence according to claim 5, characterized in that: Context-aware generation is performed on the clause interpretation set to obtain a document revision text corresponding to the clause interpretation set, including: Dependency syntactic analysis is performed on the clause interpretation set to generate a weighted directed graph, wherein the expression of the weighted directed graph is as follows: ； in, represents a weighted directed graph, Represents a collection of nodes. represents the set of dependency edges, represents a syntactic weight set, wherein the nodes in the node set correspond to the lexical units in the clause interpretation in the clause interpretation set, the dependency edges in the dependency edge set correspond to the dependency relationship types between the lexical units, and the syntactic weights in the syntactic weight set correspond to the syntactic association strength; The clause interpretation set is subjected to semantic role labeling to obtain a semantic role labeling result, and a hyperedge set is constructed based on the semantic role labeling result, wherein the expression of the hyperedge in the hyperedge set is as follows: ； in, represents a hyperedge, It represents the 0th predicate-argument structure in the clause interpretation; A hypergraph corresponding to the weighted directed graph is constructed based on the hyperedge set, wherein the expression of the hypergraph is as follows: ； in, represents a hypergraph, Represents a set of hyperedges.

7. The aeronautical information revision system based on artificial intelligence according to claim 6, characterized in that: Context-aware generation is performed on the clause interpretation set to obtain a document revision text corresponding to the clause interpretation set, further comprising: Performing feature encoding on the hypergraph based on a graph attention network to obtain a context-aware feature vector corresponding to the hypergraph; A prototype vector is constructed based on the context-aware feature vector, wherein the expression of the prototype vector is as follows: ； in, represents the prototype vector, represents the average pooling operation, represents the context-aware feature vector; The semantic deviation of each clause interpretation in the clause interpretation set is calculated based on the prototype vector, wherein the calculation formula of the semantic deviation is as follows: ； in, Represents prototype vector and candidate clause interpretation The semantic deviation between Indicates the interpretation of candidate clauses The output vector of The candidate clause interpretation is dynamically modulated based on a gating mechanism, wherein the expression of the dynamically modulated candidate interpretation is as follows: ； in, represents the dynamic modulation candidate interpretation representation, represents the activation function, represents the bias term, Represents the initialization of hyperedge connections in the hypergraph; When the semantic deviation is less than a preset deviation threshold, the candidate interpretation corresponding to the semantic deviation is used as a clause of the data revision text.

8. The aeronautical information revision system based on artificial intelligence according to claim 1, characterized in that: Performing image-text matching on the second clause set and the standard aerial image set to obtain the standard aerial image corresponding to the second clause includes: Extracting features from the standard aerial image according to bottom-up attention to generate a regional feature vector, dividing the standard aerial image into a plurality of blocks to obtain a position feature vector, and extracting features from the second clause according to Bi-GRU to obtain a text feature vector; The position feature vector is transformed into a query vector through a linear layer, and the region feature vector is used as a key and value vector for cross attention to obtain a visual feature vector; The text feature vector and the visual feature vector are sampled for similarity to obtain an overall similarity. If the overall similarity is higher than a preset similarity threshold, the second clause matches the standard aerial image.

9. The aeronautical information revision system based on artificial intelligence according to claim 8, characterized in that: The text feature vector and the visual feature vector are sampled for similarity to obtain overall similarity, including: Calculating the similarity between the text feature vector and the visual feature vector, wherein the similarity uses cosine similarity; Determining a semantic relationship between the text feature vector and the visual feature vector according to the similarity; Determining a weighted similarity of the block in the standard aerial image according to a semantic relationship between the text feature vector and the visual feature vector; The overall similarity between the text feature vector and the visual feature vector is determined according to the weighted similarity of the block in the standard aerial image.

10. The aeronautical information revision system based on artificial intelligence according to claim 9, characterized in that: The calculation formula of the semantic relationship is as follows: ； in, represents the semantic relationship between the second clause corresponding to the text feature vector and the standard aerial image corresponding to the visual feature vector, represents a mask that is equal to the input when it is positive, and 0 otherwise, represents the similarity between the text feature vector and the visual feature vector, Indicates the preset parameter threshold, represents the number of blocks in a standard aerial image, represents the activation function; The calculation formula of the weighted similarity is as follows: ； in, represents the weighted similarity of the block in the standard aerial image, A visual feature vector representing a block; The calculation formula of the overall similarity is as follows: ； in, represents the overall similarity between the text feature vector and the visual feature vector, Represents a text feature vector.

Images