A method, device, and medium for processing NOTAM text based on semantic enhancement

Abstract

This invention relates to the field of text processing technology, and in particular to a method, device, and medium for processing navigational notice text based on semantic enhancement. The method includes: first, acquiring a content carrier to be processed; then, acquiring the semantic vector and glyph feature vector of the content carrier; concatenating the two types of vectors to form an enhanced text representation; extracting temporal features from the enhanced text representation to obtain temporal features containing forward and backward logical relationships within the text; acquiring the weights of words and sentences in the temporal features and performing weighting to obtain weighted word representations and weighted sentence representations; performing correction processing on the weighted representations to generate corrected text; and finally, validating the corrected text and outputting the target text. This invention can improve the accuracy and efficiency of content carrier processing.

Description

Technical Field

[0001] This invention relates to the field of text processing technology, and in particular to a method, apparatus and medium for processing navigational notice text based on semantic enhancement. Background Technology

[0002] Notices to Airmen (NOTAMs) serve as the core information carrier for aviation safety operations. They bear the important responsibility of disseminating key information such as airspace dynamics, navigation facility status, and flight restrictions in real time. The accuracy of their content directly determines the rationality of flight path planning, the effectiveness of aviation meteorological warnings, and the reliability of air traffic control decisions. They are a crucial link in ensuring aviation operational safety.

[0003] However, NOTAM texts have significant domain-specific characteristics: on the one hand, the texts are dense with technical terms and contain mixed expressions in multiple languages; on the other hand, their semantic structure is complex, containing a large number of causal logic chains (such as "construction → runway closure") and conditional constraints (such as "validity period must be later than the start time"), which makes it difficult for manual verification mode to overcome bottlenecks. Not only is the processing efficiency low and unable to meet the real-time requirements of aviation information, but there is also a high rate of missed detections, and key errors are easily missed due to human error.

[0004] Current mainstream NOTAM verification technologies can be divided into two categories, but both have obvious shortcomings:

[0005] Logical validation techniques based on predefined keywords and regular expressions: These techniques can only achieve simple format matching and fixed logical judgments, and cannot handle semantic ambiguity problems (such as being unable to distinguish the semantic difference between "runway closed" and "runway partially closed"), and are even more difficult to capture the contextual relationships in the text. Their validation accuracy is extremely low in complex semantic scenarios.

[0006] Traditional machine learning model-based text classification and verification techniques, such as Support Vector Machine (SVM), Text Convolutional Neural Network (TextCNN), and Bidirectional Long Short-Term Memory (BiLSTM), rely heavily on manual feature engineering and can only extract local text features, failing to effectively capture long-distance semantic relationships. Furthermore, due to the lack of deep embedding of domain knowledge, the models have insufficient semantic representation capabilities for civil aviation terminology, resulting in scattered attention to key information and making it difficult to meet the requirements for high-precision verification.

[0007] In summary, existing technologies cannot simultaneously address the accuracy, efficiency, and domain adaptability of NOTAM verification. There is an urgent need for an intelligent verification solution that can deeply understand text semantics, dynamically focus on key information, and integrate domain knowledge to improve the automation level of NOTAM processing and reduce labor costs and aviation safety risks. Summary of the Invention

[0008] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0009] According to a first aspect of the present invention, a method for processing navigational notice text based on semantic enhancement is provided, the method comprising the following steps:

[0010] S100, Obtain the content carrier to be processed, wherein the content carrier to be processed is a navigation notice text.

[0011] S200: Obtain the semantic vector and glyph feature vector of the content carrier to be processed.

[0012] S300 generates an enhanced content carrier representation based on semantic vectors and glyph feature vectors.

[0013] S400 performs a temporal feature extraction operation on the enhanced content carrier representation to obtain the corresponding temporal features.

[0014] S500: Obtain the target weight of each word and sentence in the time-series features, and combine the target weights to weight the words and sentences in the time-series features to obtain weighted word representations and weighted sentence representations.

[0015] S600 performs correction processing on the weighted word representation and weighted sentence representation to generate the corrected content carrier.

[0016] S700, verify the modified content carrier to obtain the target content carrier.

[0017] According to a second aspect of the present invention, an electronic device is provided, including a processor and a memory; the processor executes the steps of the method described in the first aspect of the present invention by invoking a program or instructions stored in the memory.

[0018] According to a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a program or instructions that cause a computer to perform the steps of the method described in the first aspect of the present invention. The present invention has at least the following beneficial effects:

[0019] (1) Improve the accuracy of semantic understanding and adapt to the professional attributes of civil aviation.

[0020] This method employs a pre-trained language model adapted to the civil aviation field, combined with glyph feature vectors obtained from a glyph feature extraction model, and concatenates them to form an enhanced text representation. This design breaks through the limitations of traditional single semantic models, accurately recognizing civil aviation professional terms such as "runway number (e.g., 09 / 27) and time format (e.g., DDHHMMZ)," and solving the problem of "misjudging similar-looking characters (e.g., miswriting 'altitude 900' as 'altitude 90o')" through glyph features. This makes the semantic understanding accuracy suitable for the core attribute of "dense professional terminology" in NOTAM texts, and significantly improves the accuracy of semantically related error recognition compared to traditional general text processing models.

[0021] (2) Capture textual logical relationships to reduce the rate of missed logical errors.

[0022] By employing a temporal modeling network to perform contextual modeling on the enhanced text representation, it can effectively capture the "forward logical chains (such as the causal relationship of 'construction → runway closure')" and "backward constraint relationships (such as the conditional constraint of 'expiration date → start time')" in the NOTAM text. This design addresses the pain point that existing regular expression validation and traditional machine learning models cannot handle long-distance semantic relationships, avoiding safety hazards caused by the failure to detect logical errors such as "start time is later than expiration date" and "contradictions between facility status and operational description." The false negative rate of logical errors is lower than that of traditional technologies, significantly improving the reliability of text processing.

[0023] (3) Focus on verifying key information to ensure the accuracy of core elements

[0024] By weighting words and sentences in the context-modeled text using an attention mechanism, the system can dynamically focus on core safety elements such as "time, coordinates, and altitude" in NOTAMs, avoiding the problem of scattered attention to key information caused by the "equal treatment" of all fields in traditional models. Combined with a validation module that integrates domain rules, it can prioritize the validation of core elements while ensuring that the text complies with civil aviation regulations (such as the requirement that runway numbers be on the airport entity list), thereby improving the error detection rate of key fields and further reducing aviation safety risks.

[0025] (4) Adapt to real-time processing scenarios and improve business response efficiency

[0026] The integrated domain rule verification module implicitly possesses dynamic update capabilities, adapting to business scenarios where NOTAMs (Notifications to Airmen) require real-time airspace updates and rapid response to specification updates. Compared to traditional technologies where rule updates necessitate system restarts and interruption handling, this method enables real-time iteration of domain rules, ensuring that text verification always aligns with the latest civil aviation regulations, improving business response efficiency, and reducing verification failures caused by rule lag.

[0027] (5) Reduce labor costs and improve the level of automation.

[0028] By automating the entire process from feature extraction to context modeling, attention weighting, error detection and correction, and rule validation, this approach replaces the traditional manual verification method, which is characterized by low efficiency and high false negative rates. This significantly reduces the time and labor costs associated with manual review. Testing has shown that in processing multiple batches of NOTAM texts, the processing latency for a single text is controllable, and the error detection recall rate and format false positive rate are both superior to existing technologies. This significantly improves the automation level of NOTAM text processing and provides technical support for the high efficiency of aviation information processing.

[0029] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart illustrating a semantically enhanced NOTAM text processing method provided in this embodiment of the invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0034] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0035] The purpose of this invention is to overcome the problems of the prior art and provide a method for verifying NOTAM text that supports semantic depth analysis and dynamically focuses on key fields, for detecting and correcting semantic errors, grammatical errors and non-standard formatting issues in NOTAM text.

[0036] Furthermore, embodiments of the present invention provide a semantically enhanced method for processing NOTAM text, such as... Figure 1 As shown, the steps may include the following:

[0037] S100: Obtain the carrier of the content to be processed.

[0038] In this embodiment of the invention, the specific form of the content carrier is text, and the content carrier to be processed is a NOTAM (Notice to Airmen) text. For the convenience of the subsequent description of the technical solution, the terms "content carrier" and "text" may be used interchangeably in the following description. That is, the content carrier can be collectively referred to as "text" in terms of expression, and the text also corresponds to the aforementioned content carrier in terms of technical connotation. Both refer to textual carriers that carry information in the field of NOTAMs, without changing their core technical attributes and scope.

[0039] NOTAM texts are the core information carriers for aviation safety operations. They are used to release key information in real time, such as airspace dynamics, navigation facility status (e.g., runway and navigation beacon operation status), and flight restrictions (e.g., temporary no-fly zones and altitude restrictions). The text content includes time (e.g., effective / expiration time in "DDHHMMZ" format, such as "120600Z"), coordinates (e.g., airspace geographic coordinates, airport runway coordinates), altitude (e.g., flight restriction altitude, minimum safe altitude), and technical terms (e.g., "RWY" for runway, "CLSD" for closure, and "MAINT" for maintenance). It also has the characteristics of complex semantic structure (including a forward causal chain of "construction → facility closure" and a backward constraint relationship of "expiration must be later than the start time"), which is adapted to the technical requirements of subsequent semantic feature extraction and contextual logic modeling.

[0040] In this embodiment of the invention, the method of obtaining the text to be processed includes retrieving it from the database of the aviation management system and receiving text data pushed in real time by relevant civil aviation departments. After obtaining the text, it needs to be preprocessed: removing redundant spaces, unifying the text encoding format, and truncating or padding the text length to the maximum length of the sequence that can be adapted to the model input (e.g., 512 tokens) to ensure that the text to be processed conforms to the subsequent model input specifications and provides a high-quality data source for the accurate extraction of semantic vectors and glyph feature vectors.

[0041] In one illustrative embodiment, the obtained text to be processed may be in the format: "RWY09 / 27CLSDDUETOMAINTFROM120600ZTO121200Z".

[0042] S200: Obtain the semantic vector and glyph feature vector of the content carrier to be processed.

[0043] In this embodiment of the invention, a semantic vector of the content carrier to be processed is obtained based on a pre-trained language model, and a glyph feature vector of the content carrier to be processed is obtained based on a glyph feature extraction model. The pre-trained language model is fine-tuned using a civil aviation corpus, and its embedding layer integrates a civil aviation terminology dictionary to output semantic vectors representing the semantic information of the text. In one illustrative embodiment, the pre-trained language model can be an ERNIE series, a BERT series, the China Aviation Information Technology Qianrang model, or Aviation GPT, etc. The following common process is followed when adapting different models:

[0044] Step 1: Construct a corpus for the civil aviation field

[0045] (1) Obtaining original data: The sources include civil aviation regulations and documents (airspace management specifications, navigation facility operation standards, rules for writing NOTAMs, etc.), historical NOTAM texts (including airspace dynamics, navigation facility status, and valid flight restriction notices, which must include key elements such as complete time, coordinates, and altitude, as well as professional terms such as "RWY", "CLSD", and "MAINT"), civil aviation professional auxiliary texts (related descriptions of notices in airport operation manuals and related content of navigation facility maintenance records), and supplementing the diversity of data scenarios.

[0046] (2) Corpus cleaning and preprocessing: Delete incomplete (key elements missing), disordered (no standard time marker "Z") and redundant corpus; unify synonymous terms with different expressions according to civil aviation regulations (e.g., "runway closed" is unified as "RWYCLSD"); convert all corpus into a unified encoding format, and split the "element description section" and "restriction description section" according to the NOTAM structure to ensure consistent input format.

[0047] (3) Corpus annotation and division: The time, coordinates, height and professional terms in the corpus are annotated and divided into training set (model fine-tuning), validation set (parameter adjustment) and test set (adaptation effect evaluation) in a 7:2:1 ratio.

[0048] Step 2: Fine-tuning the general pre-trained language model

[0049] (1) Model initialization before fine-tuning: Load the general model weights and expand the embedding layer. That is, add the extra word embedding matrix corresponding to the civil aviation professional terminology dictionary to the embedding layer. After the terminology is converted into word vectors, it is connected. According to the average length of the flight notification text, the maximum length of the input sequence is set to 512 tokens (matching the processing capacity of the time series modeling network). The text with excessive length is truncated according to the principle of "prioritizing the retention of key elements". The insufficient length is filled with domain-specific placeholders.

[0050] (2) Fine-tuning task and loss function settings: The dual-task fine-tuning of "masked language modeling + next sentence prediction" is adopted. The hybrid loss function is "0.7×masked language modeling loss + 0.3×next sentence prediction loss". Both losses adopt the cross-entropy loss function. Among them, the masked language modeling task accounts for 80%. The specific operation is to randomly mask 15% of the civil aviation professional terms in the training set. There are three masking methods: 80% of the mask positions are replaced with the [MASK] symbol, 10% of the mask positions are replaced with random civil aviation terms, and 10% of the mask positions retain the original terms. The model is trained through this task to predict the correct terms of the mask positions according to the context. The next sentence prediction task accounts for 20%. The specific operation is to split the corpus in the training set into the form of "previous sentence-next sentence" pairs, and at the same time randomly generate 50% of "non-continuous sentence pairs" (i.e. the previous sentence and the next sentence have no actual contextual relationship). The model is trained through this task to determine whether the input sentence pairs are continuous context.

[0051] (3) Fine-tuning process parameter control: The AdamW optimizer was used with an initial learning rate of 2e-5, which was linearly decayed by 10% every 3 rounds; 10 rounds of training were set, with 32 corpora input in each batch. In each round, the accuracy of term prediction and sentence pair judgment was evaluated using the validation set. If there was no improvement in performance for 2 consecutive rounds, the training was stopped and the optimal weights were saved; a dropout layer (probability 0.2) was added to the fully connected layer to reduce overfitting.

[0052] (4) Model validation after fine-tuning: Calculate the cosine similarity between the semantic vector and the key element annotation using the test set (≥0.85) to verify the effectiveness of the vector representation; concatenate the semantic vector input with the glyph feature vector to verify the effect of the enhanced text representation after fusion on the accuracy of capturing logical relationships in the context modeling module (the logical error recognition rate needs to be improved by ≥15%).

[0053] In one illustrative embodiment of the present invention, the ERNIE series pre-trained language model is selected, and the output semantic vector dimension is 768.

[0054] In this embodiment of the invention, the character shape feature extraction model is a model capable of extracting structural features of Chinese characters, used to output character shape feature vectors representing text character shape information. Preferably, it is a convolutional neural network model (containing at least two convolutional layers). In a specific embodiment of the invention, a three-layer convolutional neural network model is used, with convolutional sizes of 3×3, 5×5, and 3×3 respectively. This size design is suitable for the characteristics of short Chinese characters and small structural differences in navigation notice texts: the 3×3 small-size convolutional kernel is used to capture the local stroke features of Chinese characters, and the 5×5 large-size convolutional kernel is used to capture the overall structural features of Chinese characters. Specifically, the first 3×3 convolutional layer outputs 64-dimensional features, the second 5×5 convolutional layer outputs 128-dimensional features, and the third 3×3 convolutional layer fuses and reduces the dimensionality of the features output from the first two layers, ultimately outputting a character shape feature vector with a dimension of 50. This vector can effectively distinguish the character shape differences between similar-looking characters such as "900" and "90o", and "height" and "high change".

[0055] The technical advantages of the S200 are:

[0056] (1) Improve the accuracy of semantic understanding and model compatibility in the civil aviation field: Through the design of "multi-model selection + common fine-tuning process", it is ensured that pre-trained language models with different architectures can accurately identify civil aviation professional terms (such as "RWY" and "CLSD") and semantic associations unique to fields such as "construction → closure". After testing, the term recognition accuracy is ≥20% higher than that of general pre-trained models, and the semantic association misjudgment rate is ≥18% lower. At the same time, the dimensional design of the 768-dimensional semantic vector can be seamlessly spliced ​​with the subsequent glyph feature vectors, providing an accurate semantic basis for context modeling. Moreover, the multi-model selection expands the protection scope of the technical solution and avoids the technical avoidance risk caused by the limitation of a single model.

[0057] (2) Accurately solve the problem of misjudging similar characters: In response to the common errors of similar characters in navigation notices, the design of three-layer differential convolution size can capture the local strokes and overall structural features of Chinese characters. According to the test, the 50-dimensional character feature vector has an accuracy of ≥95% in distinguishing similar characters such as “900 / 90o” and “height / high change”. This makes up for the shortcomings of traditional single semantic models that cannot identify visual similarity errors, significantly reduces the risk of verification omissions caused by similar character errors, and further ensures the accuracy of navigation notice text processing.

[0058] (3) Provide highly adaptable feature input for subsequent modules: The dimension matching design of semantic vector (768-dimensional) and glyph feature vector (50-dimensional) ensures that the 818-dimensional enhanced text representation after splicing can be directly input into the temporal modeling network without additional dimension conversion; the dual-dimensional features can simultaneously meet the needs of the context modeling module for capturing semantic associations and the attention module for focusing on key information, avoiding insufficient subsequent processing accuracy due to single feature dimensions. According to the test, the efficiency of the entire navigation notice text processing process is ≥12% higher than that of the traditional solution.

[0059] S300 generates an enhanced content carrier representation based on semantic vectors and glyph feature vectors.

[0060] This step uses the vector output by S200 as a base to concatenate the semantic vector and the glyph feature vector, forming an enhanced text representation that integrates semantic and glyph information. The semantic vector (such as the 768-dimensional vector output by the ERNIE series models) and the glyph feature vector (such as the 50-dimensional vector output by the convolutional neural network) must maintain alignment at the text unit level. That is, for each word (or sub-word) in the text to be processed, its corresponding semantic vector (768-dimensional) and glyph feature vector (50-dimensional) are used as a set of associated vectors to ensure that the vector concatenation objects are feature representations of the same text unit, avoiding feature fusion failure due to text unit misalignment.

[0061] The concatenation process is performed using a dimension stacking method: for each group of aligned semantic vectors and glyph feature vectors, the dimensions are concatenated in a fixed order of "semantic vector first, glyph feature vector second". That is, the concatenated vector dimension of a single text unit is "semantic vector dimension + glyph feature vector dimension" (e.g., 768 + 50 = 818 dimensions). If the text to be processed contains n words (or sub-words), the final enhanced text representation is an "n×818" dimension matrix (n is the number of text units, matching the limit of 512 tokens for the maximum length of the input sequence of the pre-trained language model in S200, i.e., n≤512). Each row in the matrix corresponds to the fusion feature of a text unit, and each column corresponds to a feature dimension (the first 768 columns are semantic feature dimensions, and the last 50 columns are glyph feature dimensions).

[0062] After concatenation, feature consistency verification is required: the dimensional integrity of the fused vector of each text unit in the concatenation matrix is ​​calculated (to ensure no missing or redundant dimensions) and the matching of the number of text units with the length of the input text in S200 is calculated (to ensure no missing or additional text units are added). If the verification passes, the enhanced text representation is output. If the verification finds a dimensional mismatch (such as some text units containing only semantic vectors without glyph feature vectors) or a misalignment of the number of text units, the process automatically backtracks to S200, re-extracts the missing vectors of the corresponding text units, until an enhanced text representation that meets the requirements is generated.

[0063] In this embodiment of the invention, the core value of enhanced text representation lies in: retaining the semantic vector's ability to accurately represent civil aviation professional terms (such as "RWY" and "CLSD") and textual logical connections (such as "construction → closure"), while incorporating the glyph feature vector's ability to distinguish differences in Chinese character structure (such as the stroke differences between "900" and "90o", and the structural differences between "height" and "high change"), thus achieving dual information fusion of "semantic understanding + visual feature recognition". This provides a more comprehensive feature foundation for the subsequent context modeling module to capture textual logical relationships and for the attention module to focus on key information, avoiding the limitations of single semantic features or single glyph features in text representation.

[0064] S400 performs a temporal feature extraction operation on the enhanced content carrier representation to obtain the corresponding temporal features.

[0065] In this embodiment of the invention, a temporal modeling network is used to perform contextual modeling on the enhanced text representation, resulting in context-modeled text containing forward and backward logical relationships of the text, i.e., temporal features.

[0066] In this embodiment of the invention, the temporal modeling network is preferably a bidirectional long short-term memory (Bi-LSTM) network. Its core design revolves around the temporal correlation and logical dependency of NOTAM texts. The chain-like relationship of time-event-restriction in NOTAMs (e.g., RWYMAINT → RWYCLSD → ATCINSTRUCTION) needs to simultaneously capture both the forward temporal flow and the reverse dependency relationship. The bidirectional structure can simultaneously learn the forward logic of the text from left to right (e.g., the order of events) and the reverse correlation from right to left (e.g., tracing the cause from the result). Specific network parameter settings are as follows:

[0067] The hidden dimension is set to 256. This dimension design is based on the 818-dimensional feature scale of the enhanced text representation. The high-dimensional features are compressed to a dimension suitable for capturing logical relationships through non-linear transformation (256 dimensions can balance feature retention and computational efficiency, and avoid overfitting caused by excessive dimensionality).

[0068] The network has two layers. The first layer is used to initially extract local temporal correlations (such as the collocation relationship between adjacent terms, the combination of "RWY09 / 27" and "CLSD"). The second layer is used to integrate global logical chains (such as the complete event sequence of "maintenance → shutdown → recovery").

[0069] Each layer is set with dropout regularization (probability 0.2) to perform dropout operation on the sequence dimension, which reduces the model's over-reliance on features of specific time positions and enhances its generalization ability to announcement texts of different formats;

[0070] The output layer uses a linear transformation to concatenate the forward hidden state and the backward hidden state of the bidirectional LSTM (256×2=512 dimensions) to form a context vector containing bidirectional logical information.

[0071] The modeling process is carried out in three steps:

[0072] Input adaptation: The n×818-dimensional enhanced text representation (n≤512) output by S300 is input into Bi-LSTM in the order of text units, and each time step corresponds to the fusion feature of a text unit;

[0073] Bidirectional encoding: The forward LSTM encodes sequentially from the first text cell to the last text cell, capturing the positive logic of cause → effect (e.g., navigation facility failure → runway closure); the backward LSTM encodes in reverse order from the last text cell to the first text cell, capturing the reverse correlation of result → cause (e.g., runway closure ← navigation facility failure).

[0074] Feature fusion: Perform an element-wise addition fusion operation on the forward hidden state and the backward hidden state at each time step to generate context modeling text (dimension n×256) containing bidirectional logical information. The output vector of each text unit contains both its semantic role in the forward time series and its logical position in the reverse dependency.

[0075] The technical advantages of the S400 are:

[0076] (1) In view of the characteristics of precise time coordinates and mandatory causal relationship in the navigation notice, the gating mechanism of Bi-LSTM can effectively solve the long-distance dependency forgetting problem of traditional recurrent neural networks and accurately capture the logical relationship across paragraphs, such as the implicit relationship between the construction time at the beginning of the notice and the restricted height at the end.

[0077] (2) The 256-dimensional hidden layer is adapted to the dimensions of the enhanced text representation, ensuring that key semantic and glyph information are preserved while compressing features, providing high-quality input for the subsequent attention module to focus on core elements such as time, coordinates, and constraints;

[0078] (3) The bidirectional modeling capability can identify logical contradictions in the text (such as RWYCLSD but not specifying the closing time), providing logical-level abnormal features for subsequent error detection and improving the comprehensiveness of the notification text verification.

[0079] S500: Obtain the target weight of each word and sentence in the time-series features, and combine the target weights to weight the words and sentences in the time-series features to obtain weighted word representations and weighted sentence representations.

[0080] In this embodiment of the invention, an attention mechanism is used to obtain the target weight of each word and each sentence in the temporal features. The words and sentences in the temporal features are weighted by combining the target weights to obtain weighted word representations and weighted sentence representations.

[0081] In this embodiment of the invention, the attention mechanism is a rule-guided hierarchical attention mechanism. Unlike traditional general attention mechanisms (such as Transformer attention) that rely solely on data-driven weight allocation, this mechanism employs a two-stage design of initial weights learned from data and weights corrected according to domain rules. This accurately focuses on core safety elements such as time, coordinates, and altitude in the navigation notification text, solving the problem of traditional mechanisms' scattered attention to key domain information. The hierarchical attention mechanism consists of two layers: word-level attention and sentence-level attention, corresponding to word units and sentence units representing temporal features, respectively.

[0082] In this embodiment of the invention, an attention mechanism is used to obtain the weights of each word and each sentence in the temporal features, specifically including:

[0083] S501, based on the semantic relevance of temporal features, generates initial word weights and initial sentence weights by pre-setting nonlinear transformation components and activation function components, respectively, to ensure that the weights reflect the semantic importance of the text itself.

[0084] S501 may specifically include:

[0085] For each word unit in the temporal features, the hidden state vector corresponding to each word unit is obtained. After the hidden state vector is input into the word-level nonlinear transformation component for feature mapping, it is connected to the first activation function component to output the initial weight of each word unit.

[0086] In this embodiment of the invention, the word-level nonlinear transformation component includes a trainable feature transformation matrix and a bias term. Specifically, the initial weights of each word unit satisfy the following condition: α i initial =softmax((W q ·h i +b q ) / (d k 1 / 2 )).

[0087] Where, α i initial h represents the initial weight of the i-th word unit in the time-series feature, where i ranges from 1 to n, and n is the number of word units in the time-series feature. i W is the latent state vector of the i-th word unit in the temporal features, with a dimension of 256, representing the bidirectional contextual information of the i-th word; q and b q For trainable word-level feature transformation parameters, specifically, Wq This is a word-level feature transformation matrix with dimensions 256×256, b q This is a word-level bias term with a dimension of 256. k 1 / 2 d is the scaling factor. k =256, to avoid the softmax output becoming too extreme (some weights approaching 1, others approaching 0) due to excessively large values ​​after feature transformation, and to ensure that the initial weights can reasonably distinguish the semantic importance of different words (e.g., the initial weights of "CLSD" and "120600Z" are higher than those of function words such as "THE" and "A"). · represents dot product.

[0088] For the sentence units in the temporal features, a pooling operation is performed on the word latent state vectors contained in each sentence unit to obtain a sentence vector. The sentence vectors are then input into the sentence-level nonlinear transformation component for feature mapping and then connected to the second activation function component to output the initial sentence weights of each sentence unit.

[0089] In this embodiment of the invention, the sentence-level nonlinear transformation component includes a trainable feature transformation matrix and a bias term. Specifically, the initial weights of each sentence satisfy the following condition: β j initial =sigmoid(W s ·s j +b s ).

[0090] Where, β j initia Let s be the initial sentence weight of the j-th sentence unit in the time-series features. j W is the sentence vector of the j-th sentence unit in the temporal feature, obtained by average pooling of the latent state vectors of all words contained in the j-th sentence. It has a dimension of 256 and represents the global semantics of the j-th sentence; j ranges from 1 to m, where m is the number of sentences in the temporal feature, and W... s and b s For trainable sentence-level feature transformation parameters, specifically, W s This is a sentence-level feature transformation matrix with dimensions 256×128, b s This is a sentence-level bias term with a dimension of 128. In this embodiment of the invention, the trainable parameters are initialized using a Xavier normal distribution to ensure gradient stability during training.

[0091] In this embodiment of the invention, sigmoid is used instead of softmax because there is no constraint that the total weight of sentences is 1 (such as the requirement that "event description sentences" and "restriction description sentences" in announcements both have high weights), which can simultaneously increase the attention share of key sentences.

[0092] S502, based on preset domain rules, the initial weights of words and sentences are adjusted through a dynamically configurable rule storage component to obtain the target weights of each word and each sentence in the temporal features.

[0093] In this embodiment of the invention, the preset domain rules are core field weight adjustment rules formulated based on domain normative documents. The core fields include at least time-related fields, coordinate-related fields, altitude-related fields, and facility status-related fields that characterize key domain attributes. Different types of core fields correspond to preset weighting ratios. Specifically, the weighting ratios are divided according to the degree of influence of the elements on aviation safety:

[0094] First-level bonus (+30%): Time fields (e.g., "120600Z" "121200Z"), coordinate fields (e.g., "N30°24′E120°18′"), and height fields (e.g., "FL350" "1500FT"). Errors in these types of elements directly cause security risks and require the highest priority attention.

[0095] Secondary bonus (+15%): Facility status fields (such as “RWYCLSD”, “NAVAIDOUT”) and operation instruction fields (such as “ATCCLEARANCE”, “MAINTAIN”). These elements are associated with operational decisions and require secondary priority attention.

[0096] Level 3 Bonus (+5%): Airport / airspace identification fields (such as "PEK" and "ZBAA"). These elements ensure the accuracy of information positioning and should be given basic priority.

[0097] In this embodiment of the invention, the rule storage component adopts an editable structured data format, supporting the addition, deletion, and modification of the type classification and corresponding weighting ratio of the core fields, enabling rule iteration without adjusting the main architecture of the attention mechanism. The structured data format includes, but is not limited to, JSON configuration files, XML configuration files, and YAML configuration files. In one specific embodiment, the rule storage component is a JSON configuration file.

[0098] During the weight adjustment process, the core fields contained in the time-series features are first identified using the field matching component to determine the word units and sentence units whose weights need to be adjusted. Then, the initial weights of the corresponding words and sentences are corrected based on the preset bonus ratios in the rule storage component. After correction, the weight values ​​are subject to boundary constraint processing to ensure that the target weights are within the preset effective range. The specific adjustment process may include:

[0099] Field matching: Identify the core fields mentioned above in the time series features using regular expressions to determine the word units and sentence units whose weights need to be adjusted (e.g., if "120600Z" is identified, locate its corresponding word unit and sentence unit).

[0100] Weighting: For each matched word unit, its initial weight is multiplied by (1 + the weighting ratio). For example, if the weight after the first-level weighting is α... i initial ×1.3; For sentence units containing core fields, multiply their initial weight by (1 + the highest-level bonus percentage for that sentence). For example, if a sentence contains time and height fields, adjust the bonus to β based on the first-level bonus. j initial ×1.3;

[0101] Boundary constraints: If the corrected weight exceeds 1, it is truncated to 1 (to avoid weight overflow that could lead to distortion in subsequent weighted calculations); if it is below 0.1, it is increased to 0.1 (to ensure that non-critical elements still retain basic attention).

[0102] Furthermore, the word hidden state vector h i With the corresponding corrected weight h i Element-wise multiplication yields a weighted word representation, highlighting the semantic features of keywords. The word weighting formula is h. i w =α i ·h i h i w Let be the weighted word representation of the i-th word.

[0103] Furthermore, the sentence vector s j With the corresponding correction weight β j Element-wise multiplication is performed to obtain a weighted sentence representation, which strengthens the global semantics of key sentences. The sentence weighting formula is: s j w =β j ·s j s j w The weighted sentence representation for the j-th sentence.

[0104] The final output of weighted word representations and weighted sentence representations not only preserves the semantic connections of the text itself, but also strengthens the features of key security elements through domain rules, providing key feature inputs for subsequent error detection and correction.

[0105] S600 performs correction processing on the weighted word representation and weighted sentence representation to generate the corrected content carrier.

[0106] Furthermore, the S600 specifically includes:

[0107] S601, perform error detection operation on the weighted word representation and the weighted sentence representation, and output error location labels based on the BIO annotation system. The error location labels indicate at least one of the following error types: semantic error, syntax error, and formatting error.

[0108] S601 specifically includes:

[0109] S6011, Establish the association between word representation and sentence representation according to the text logical order: concatenate the weighted sentence representation of each sentence unit with the weighted word representation of all word units contained in the sentence to form a sentence-word hierarchical feature matrix. The sentence-word hierarchical feature matrix is ​​used to enable error detection to simultaneously cover local word-level anomalies and global sentence-level logical contradictions.

[0110] S6012 inputs the sentence-word level feature matrix into the error detection classifier and outputs the misalignment position label.

[0111] The error detection classifier consists of a 3-layer fully connected network, with the following configuration for each layer:

[0112] The first layer has a hidden dimension of 512 and uses the ReLU activation function to perform a non-linear transformation on the sentence-word level feature matrix to extract common error features.

[0113] The second layer, with a hidden dimension of 256, embeds a civil aviation error feature extraction submodule to capture error patterns specific to areas such as facility status conflicts and time format anomalies.

[0114] The third layer has an output dimension of 4, corresponding to the BIO label space and error type encoding. The BIO label space includes B - error start, I - error continuation, O - no error, and E - error end. The error type encoding includes semantic error = 1, syntax error = 2, and format error = 3. The label probability distribution of each word unit is output through the softmax function.

[0115] The error detection classifier identifies error types using the following logic:

[0116] To address semantic errors: Based on the semantic association of civil aviation terms in the weighted features (such as the contradiction between "RWYCLSD (runway closed)" and "AVBL (available)"), identify problems such as "facility status conflict" and "time logic contradiction", and label them as "B-semantic error", "I-semantic error" and "E-semantic error".

[0117] Regarding grammatical errors: In accordance with the standard that "the description of the elements of the Notice to Airmen must contain a complete subject-verb-object structure", identify problems such as incomplete components and improper collocation (such as "CLSD" missing the subject "RWYXX"), and mark them as B-format error, I-format error, and E-format error;

[0118] For format errors: Refer to the standards for civil aviation time (DDHHMMZ format), coordinate format (NXX°XX′EXXX°XX′ format), and altitude format (XXXFT / FLXXX format) to identify issues such as missing symbols and misplaced characters (e.g., "120600" is missing "Z"), and mark them as B-format error, I-format error, or E-format error;

[0119] Word units without errors are uniformly marked with "O".

[0120] S6013 integrates the BIO tags and error type codes of all word units to generate a labeled sequence containing text location-error type, providing a precise location basis for error correction in S602.

[0121] S602, use a sequence-to-sequence model to perform a correction operation on the error content indicated by the error location label and generate the corresponding corrected text.

[0122] In this embodiment of the invention, the sequence-to-sequence model adopts an architecture of bidirectional GRU encoder-unidirectional GRU decoder + domain attention mechanism, which is adapted to the characteristics of short error segments and strong correction rules in Notices to Airmen text, wherein:

[0123] Encoder: A bidirectional GRU network with 256 hidden dimensions is used. The input is a concatenation feature of the error location label sequence and weighted word representation. It encodes the semantic relationship between the context of the error location (such as the keywords "FROM" and "TO" before and after the error time "120600").

[0124] Decoder: Employs a unidirectional GRU network with 256 hidden dimensions. It receives the context vector output by the encoder and combines it with a "domain attention mechanism" (prioritizing error locations and related core civil aviation fields) to generate corrected candidate content word by word.

[0125] Output constraint layer: Integrates with the civil aviation professional terminology dictionary (including standardized expressions such as time, coordinates, facility codes, etc.) and the softmax function to ensure that the generated corrections conform to the domain format and semantic requirements, and avoid new standard conflicts.

[0126] S602 specifically includes:

[0127] S6021, the error location label sequence, weighted word representation, and weighted sentence representation output by S601 are concatenated according to a preset dimension to form the input features of the sequence-to-sequence model; the input features include error location information, word-level semantic features, and sentence-level logical features, which are used to provide error context association information for the model.

[0128] S6022, based on the error type indicated by the error location label, invokes preset domain rule resources to perform targeted correction:

[0129] For semantic errors: the domain knowledge graph in the integrated domain rule verification module is invoked. Based on the civil aviation entity-standard value mapping relationship stored in the knowledge graph, for facility status conflict errors, the status description that conforms to the latest timestamp is retained; for logical missing errors, the associated required elements are supplemented (such as supplementing the closing time interval for "RWYCLSD").

[0130] For grammatical errors: call the standard sentence template library for Notices to Airmen (NOASA) and, for errors involving incomplete components, complete the missing subject, predicate, or object (e.g., add the subject "RWY09" for "CLSD"); for errors involving disordered word order, adjust the word order according to the template of "facility code + status description + restrictions".

[0131] For formatting errors: The regular expression engine's standard format library is called. For time field errors, the time zone identifier "Z" is added and the format "DDHHMMZ" is verified. For coordinate field errors, the angle symbols "°" and "′" are added and the format "NXX°XX′EXXX°XX′" is verified. For height field errors, the unit "FT" or "FL" is added and the numerical range is verified.

[0132] S6023, the sequence-to-sequence model decoder is based on the correction logic of S6022, and uses a beam search strategy (beamsize=3) to select the optimal candidate words from the civil aviation professional terminology dictionary, and generates the corrected text sequence word by word; the civil aviation professional terminology dictionary contains standardized terms such as facility codes, time formats, and coordinate expressions that conform to ICAO standards, which are used to constrain the domain compliance of the corrected content.

[0133] S6024, Perform double verification on the generated corrected text:

[0134] Error Residual Validation: Re-enter the corrected text into the error detection logic of S601 to check if the error location label still exists;

[0135] Rule compliance verification: Call the verification module of the integrated domain rules to check whether the corrected text conforms to the entity association rules of the domain knowledge graph and the format specifications of the regular expression engine;

[0136] If the double verification passes, the final corrected text is output; if there are any failed items, the uncorrected locations and corresponding error types are fed back to S6022, and the correction operation is re-executed until a corrected text conforming to civil aviation standards is generated.

[0137] S700, verify the modified content carrier to obtain the target content carrier.

[0138] In this embodiment of the invention, a verification module integrating domain rules is used to verify the corrected text to obtain the target content carrier.

[0139] In this embodiment of the invention, the integrated domain rule verification module includes a domain knowledge graph, a regular expression engine, a knowledge graph incremental update interface, a regular expression rule hot-loading unit, and a conflict resolution unit. Each component is a software logic component, specifically:

[0140] The domain knowledge graph is a software component that stores the mapping relationship between entities and corresponding normative values ​​within the target domain. It pre-constructs a standardized entity-attribute-normative value triple structure to constrain and correct the logic and compliance of entity associations in the content carrier. These triples cover core civil aviation business scenarios, such as: Entity = RWYCLSD (Runway Closed) → Attribute = Related Element → Normative Value = Closure Time Interval (format: DDHHMMZ-DDHHMMZ), Affected Airspace Range (including airport code and coordinate boundaries); Entity = NAVAIDOUT (Navigation Facility Failure) → Attribute = Restriction Conditions → Normative Value = Affected Flight Altitude ≤ FL200, Alternate Airport Identification; Entity = UAVRESTRICT (Unmanned Aerial Vehicle Restricted Area) → Attribute = No-Fly Parameters → Normative Value = No-Fly Time, No-Fly Radius ≤ 5KM. This ensures that the mapping relationship between entities and normative values ​​fully matches the requirements of official documents such as the "Administrative Measures for Civil Aviation Notices" and the "Regulations on Airspace Use Management."

[0141] The regular expression engine is a software component used to perform format compliance checks on specific format fields in the modified content carrier. These specific format fields include at least time fields and coordinate fields, and can be extended to altitude fields and facility code fields. The regular expression engine internally pre-loads a civil aviation-specific regular expression rule library. Each rule in the rule library corresponds to a standard format in the civil aviation field, for example:

[0142] Time-related field rules: ^\d{6}Z(-\d{6}Z)?$ (matches a single timestamp such as 120600Z, or a time range such as 120600Z-121200Z);

[0143] Coordinate field rules: ^N\d{2}°\d{2}′E\d{3}°\d{2}′$ (matches North latitude and East longitude coordinates such as N30°24′E120°18′);

[0144] Altitude field rule: ^(FL\d{3}|\d{1,4}FT)$ (matches flight altitude layers such as FL350, or foot height such as 1500FT);

[0145] Facility code field rule: ^RWY\s+\d{1,2}([ / ]\d{1,2})?$ (matches runway codes such as RWY09, or RWY09 / 27);

[0146] Quickly identify fields with abnormal formats using regular expression matching.

[0147] The incremental update interface for the knowledge graph is a software interface that supports dynamic iteration of domain knowledge. It is used to add, modify, or delete entity-attribute-canonical value mapping relationships in the domain knowledge graph through preset network interface protocols (such as RESTful API, SOAP API). When civil aviation rules are updated (such as adding new navigation facilities or adjusting airspace restriction parameters), users can send a standardization request through this interface (such as a POST request to add "entity=SBASINOP (satellite-based augmentation system failure) → attribute=affected area → canonical value=coverage area code"). After receiving the request, the interface automatically synchronizes to the knowledge graph database, and the update process takes effect in real time without the need to rebuild the overall structure of the knowledge graph.

[0148] The regular expression rule hot-loading unit is a software logic unit that enables dynamic updates of regular expression validation rules. It updates the rule base of the regular expression engine through an editable configuration file (supporting structured formats such as JSON, XML, and YAML), and the update process does not interrupt the operation of the validation system. For example, when a new standard "DDHHMMSSZ" (including second-level precision) is added to the civil aviation time format, the user can edit the JSON configuration file and add the rule ^\d{8}Z(-\d{8}Z)?$ under the "time_rules" field. After saving the file, the regular expression rule hot-loading unit automatically reads the updated content through a file listening mechanism, injects the new rule into the rule base of the regular expression engine, and retains historical rules to ensure compatibility between the old and new formats. The entire process does not affect the validation task of the current modified content carrier.

[0149] The conflict resolution unit is a software logic unit that handles inconsistencies in multi-source verification results. When there is a conflict between the correction result of S602, the error detection result of S601, and the verification result of the integrated domain rule verification module (domain knowledge graph, regular expression engine), the conflict resolution unit determines the final verification result according to a preset priority strategy. In the preset priority strategy, the verification result of the domain knowledge graph has the highest priority (because entity association logic directly affects flight safety), followed by the format verification result of the regular expression engine (format specifications ensure the accuracy of information transmission), and finally the correction result of S602. For example, if S602 corrects "RWY09CLSD" to "RWY09CLSD121200Z-120600Z" (a time interval logic contradiction), the domain knowledge graph verification marks the time interval as incorrect, while the regular expression engine verification marks the format as compliant. In this case, the conflict resolution unit prioritizes the knowledge graph result and determines that the time interval needs to be corrected, rather than accepting the format-compliant result.

[0150] Furthermore, the S700 specifically includes:

[0151] S701, perform structured preprocessing on the modified content carrier, and split the content carrier content according to field type using a preset field parsing component, specifically including:

[0152] Identify and extract entity class fields: Traverse the civil aviation professional terms (such as "RWYCLSD" and "NAVAIDOUT") in the content carrier, associate them with the corresponding attribute information (such as closing time and scope of impact), and form a set of entity class fields for input into the domain knowledge graph;

[0153] Identify and extract formatted fields: Locate time, coordinate, altitude, and facility code fields by keyword matching (such as "Z", "°", "FT", "FL"), remove redundant spaces and special characters (such as meaningless "_" and "-"), and form a set of formatted fields for input into the regular expression engine;

[0154] After preprocessing, a structured list of fields is output to ensure that each field accurately matches the input requirements of the corresponding validation component, avoiding validation omissions due to mixed fields.

[0155] S702, relying on the domain knowledge graph and regular expression engine in the integrated domain rule validation module, performs validation from two dimensions: entity association logic and field format specifications. Specifically, it includes:

[0156] (1) Entity association verification based on domain knowledge graph:

[0157] A. Entity matching: Input the entity class fields extracted by S701 into the domain knowledge graph, and locate the corresponding entity-attribute-canonical value triplet through fuzzy matching of entity names (supports synonym matching, such as "runway closed" matching "RWYCLSD");

[0158] B. Attribute Integrity Validation: Check whether the matched entity contains all required attributes (e.g., "RWYCLSD" must contain "Close Time Interval" and "Affected Spatial Area" attributes). If any are missing, mark "Entity Attribute Missing" error and record the name of the missing attribute.

[0159] C. Compliance Verification of Standard Values: Check whether the value of each attribute meets the requirements of the standard values ​​(e.g., "Closing Time Interval" must meet the requirements of "Start Time < End Time" and "Time Format Consistent with Knowledge Graph"; if "121200Z-120600Z" appears, mark it as "Standard Value Logical Contradiction" error; "Affected Airspace" must contain a valid airport code; if an invalid code "XXX" appears, mark it as "Invalid Standard Value" error).

[0160] D. Output entity validation results: Generate a validation report with the format "Field Name - Error Type - Error Description", such as "RWY09CLSD / Close Time Interval / Logical Contradiction of Standard Value / Start Time 121200Z is later than End Time 120600Z".

[0161] (2) Format validation based on regular expression engine:

[0162] A. Rule matching: The format fields extracted by S701 are classified by type (time, coordinates, altitude, facility code), and then input into the regular expression engine to call the corresponding regular expression rules for matching;

[0163] B. Format Anomaly Identification: If a field does not match any corresponding rule (e.g., time "120600" does not match ^\d{6}Z$), then mark it as a "Format Missing" error (e.g., "Time field / Format Missing / Omitted Time Zone Symbol 'Z'"); if a field partially matches a rule (e.g., coordinate "N3024E12018" does not match ^N\d{2}°\d{2}′E\d{3}°\d{2}′$), then mark it as a "Format Misalignment" error (e.g., "Coordinate field / Format Misalignment / Missing Angle Symbol '°''′'");

[0164] C. Output format verification results: Generate a format verification report with the same format as the entity verification report to ensure that subsequent conflict resolution can be handled uniformly.

[0165] S703 invokes the conflict resolution unit to handle conflicts between the multi-source verification results output by S702 and the previous processing results, specifically including:

[0166] (1) Conflict detection: Compare the correction results of S602, the error detection results of S601, and the entity verification results and format verification results of S702 to identify inconsistencies (e.g., S602 determines that “RWY09CLSD121200Z-120600Z” is corrected, while the domain knowledge graph determines that its “canonical value is logically contradictory”).

[0167] (2) Priority determination: Conflict items are determined according to the priority of "domain knowledge graph results > regular expression engine results > S602 correction results > S601 error detection results";

[0168] (3) Conflict resolution:

[0169] If the domain knowledge graph result conflicts with other results: take the domain knowledge graph result as the standard and generate the instruction "field to be corrected - correction direction" (e.g., "the closing time interval of RWY09CLSD needs to be adjusted to start time < end time, such as 120600Z-121200Z").

[0170] If the result of the regular expression engine conflicts with the S602 correction result: take the regular expression result as the standard and correct the format (e.g., "time field 120600 / time zone identifier 'Z' needs to be added, corrected to 120600Z").

[0171] Output the results after resolution: Integrate the conflict-free validation fields with the correction instructions after conflict handling to form the "Final Validation Report".

[0172] S704, if the final verification report contains correction instructions, the correction instructions are fed back to S602, and S602 re-executes error correction based on the correction instructions (such as correcting the closing time interval according to "120600Z-121200Z") to generate a secondary corrected content carrier; the secondary corrected content carrier is re-entered into S701 to S703 to perform full-process verification. If all fields pass the entity association verification and format verification (without error markings), the content carrier is determined to be the target content carrier; if errors still exist, step four is repeated until a fully compliant target content carrier is generated.

[0173] S705 will organize the target content carrier that has passed the final verification according to the standard format of NOTAM (such as including the paragraphs "notification number-issuance time-entity information-restriction conditions"), output it to the Civil Aviation NOTAM issuance system, and store the "error record-correction instruction-verification result" log during the verification process for subsequent traceability and auditing.

[0174] The present invention provides a semantically enhanced NOTAM text processing method. When tested on a test set containing 50,000 NOTAMs, the semantic error recall rate was 92.7%, the format false positive rate was 3.1%, and the processing latency of a single NOTAM was less than 350ms.

[0175] Based on the same inventive concept, embodiments of the present invention provide a semantically enhanced navigational notice text processing device, including a text acquisition module, a feature extraction module, a feature fusion module, a temporal modeling module, an attention weighting module, an error detection and correction module, and a domain rule verification module. Wherein:

[0176] The text acquisition module is used to acquire the text to be processed, which is a navigation notice text;

[0177] The feature extraction module includes a pre-trained language model unit adapted to the civil aviation field and a character shape feature extraction model unit. The pre-trained language model unit is used to perform semantic encoding on the text to be processed to output a semantic vector, and the character shape feature extraction model unit is used to extract character shape features from the text to be processed to output a character shape feature vector.

[0178] The feature fusion module is used to receive the semantic vector and the glyph feature vector, and perform a concatenation operation on the semantic vector and the glyph feature vector to form an enhanced text representation that integrates semantic and glyph information;

[0179] The temporal modeling module has a built-in temporal modeling network, which is used to receive the enhanced text representation, perform context modeling on the enhanced text representation through the temporal modeling network, and output context modeling text containing the forward and backward logical relationships of the text.

[0180] The attention weighting module has a built-in attention mechanism, which is used to receive the context modeling text, obtain the target word weight of each word unit and the target sentence weight of each sentence unit in the context modeling text through the attention mechanism, and combine the target word weights to weight the word units and sentence units in the context modeling text to output weighted word representations and weighted sentence representations.

[0181] The error detection and correction module is used to receive the weighted word representation and the weighted sentence representation, perform error detection and correction operations on the weighted word representation and the weighted sentence representation, and output corrected text that conforms to the standards of the civil aviation field.

[0182] The domain rule verification module is used to receive the corrected text, perform compliance verification on the corrected text, and output the final target text.

[0183] Furthermore, the temporal modeling network in the temporal modeling module is a bidirectional long short-term memory network (Bi-LSTM). The hidden dimension of the Bi-LSTM network is set to 256, and the Bi-LSTM network includes forward LSTM units and backward LSTM units. The forward LSTM units are used to encode the enhanced text representation in the order of the text, and the backward LSTM units are used to encode the enhanced text representation in the reverse order of the text. The output of the Bi-LSTM network is the context modeling text obtained by fusing the forward encoding result of the forward LSTM unit and the backward encoding result of the backward LSTM unit.

[0184] Furthermore, the attention mechanism in the attention weighting module is a rule-guided hierarchical attention mechanism, including an initial weight calculation unit and a weight correction unit. The initial weight calculation unit is used to model the semantic relevance of the text based on the context, and generate initial word weights and initial sentence weights through a nonlinear transformation component and an activation function component, respectively. The activation function component includes a softmax function sub-component for generating initial word weights and a sigmoid function sub-component for generating initial sentence weights. The weight correction unit has a built-in rule storage sub-unit based on the core field weighting rules predefined in the civil aviation system normative documents and an editable structured data format. It is used to dynamically adjust the initial word weights and initial sentence weights through the rule storage sub-unit to obtain the target word weights and target sentence weights.

[0185] Furthermore, the error detection and correction module includes an error detection unit and an error correction unit. The error detection unit incorporates a BIO annotation system and an error detection classifier. The error detection classifier is a 3-layer fully connected network used to perform error detection operations on the weighted word representation and the weighted sentence representation, and outputs error location labels based on the BIO annotation system. The error type indicated by the error location labels includes at least semantic errors, syntax errors, and formatting errors. The error correction unit incorporates a sequence-to-sequence (Seq2Seq) model. The sequence-to-sequence model adopts an architecture of "bidirectional GRU encoder - unidirectional GRU decoder + domain attention mechanism," and is used to perform correction operations on the error content based on the error location labels, and output the corrected text.

[0186] Furthermore, the domain rule verification module includes a domain knowledge graph submodule, a regular expression engine submodule, a knowledge graph incremental update interface submodule, a regular expression rule hot loading submodule, and a conflict resolution submodule, each of which is a software logic component.

[0187] The domain knowledge graph submodule is used to store the mapping relationship between entities and normative values ​​in the form of "entity-attribute-normative value" triples in the civil aviation domain, and to perform entity association compliance verification on the entity class fields in the corrected text;

[0188] The regular expression engine submodule has a built-in civil aviation-specific regular expression rule library, which includes at least time field rules, coordinate field rules, altitude field rules, and facility code field rules, and is used to perform format compliance checks on the format fields in the corrected text.

[0189] The knowledge graph incremental update interface submodule is a software interface that supports a preset network interface protocol and is used to add or modify entity mapping relationships in the domain knowledge graph submodule.

[0190] The regular expression rule hot-loading submodule is used to update the rule base of the regular expression engine submodule through an editable configuration file, and the update process does not require interrupting system operation.

[0191] The conflict resolution submodule is used to handle the conflict between the processing result of the error detection and correction module and the verification result of the domain rule verification module, and to determine the final verification result according to a preset priority strategy, wherein the verification result of the domain knowledge graph submodule has a higher priority than other results.

[0192] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0193] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0194] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0195] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for processing a navigation notice text based on semantic enhancement, characterized in that, The method comprises the following steps: S100, acquiring a to-be-processed content carrier, the to-be-processed content carrier being a navigation notice text; S200, acquiring a semantic vector and a character feature vector of the to-be-processed content carrier; S300, generating an enhanced content carrier representation based on the semantic vector and the character feature vector; S400, performing a time sequence feature extraction operation on the enhanced content carrier representation by using a bidirectional long short-term memory network to obtain corresponding time sequence features; S500, obtaining target weights of each word and each sentence in the time sequence features by using a rule-guided hierarchical attention mechanism, and weighting the words and sentences in the time sequence features according to the target weights to obtain weighted word representations and weighted sentence representations; wherein the rule-guided hierarchical attention mechanism comprises: generating word initial weights and sentence initial weights respectively by a preset nonlinear transformation component and an activation function component based on semantic correlation of the time sequence features; adjusting the word initial weights and the sentence initial weights by a dynamically configurable rule storage component to obtain the target weights of each word and each sentence in the time sequence features based on core field weight adjustment rules formulated according to domain normative files; the core field at least includes a time field, a coordinate field, a height field and a facility state field representing key attributes of the domain; S600, performing correction processing on the weighted word representations and the weighted sentence representations to generate a corrected content carrier; S700, verifying the corrected content carrier by using a verification module integrated with civil aviation domain rules to obtain a target content carrier, the verification module integrated with civil aviation domain rules comprising a domain knowledge graph, a regular expression engine, a knowledge graph incremental update interface and a regular rule hot loading unit; S600 specifically comprises: S601, performing an error detection operation on the weighted word representations and the weighted sentence representations, and outputting error position labels based on a BIO tagging system; S602, performing a correction operation on error content indicated by the error position labels by using a sequence-to-sequence model to generate corresponding corrected content carriers; wherein the sequence-to-sequence model adopts an architecture of a bidirectional GRU encoder-single-direction GRU decoder plus a domain attention mechanism; Wherein, S601 specifically comprises: S6011, establishing an association relationship between word representations and sentence representations in a text logical order: concatenating the weighted sentence representation of each sentence unit and the weighted word representation of all word units contained in the sentence to form a sentence-word hierarchical feature matrix, the sentence-word hierarchical feature matrix being used to cover local word-level abnormalities and global sentence-level logical contradictions simultaneously in error detection; S6012, inputting the sentence-word hierarchical feature matrix into an error detection classifier to output error position labels; S6013, integrating BIO labels and error type encodings of all word units to generate a tagging sequence containing text positions and error types, and providing accurate positioning basis for error correction of S602; S700 specifically comprises: S701, performing structured preprocessing on the correction content carrier, splitting the content carrier content by field type through a preset field parsing component, specifically including: identifying and extracting entity class fields, identifying and extracting format class fields, and outputting a structured field list after preprocessing; S702, relying on the domain knowledge graph and regular expression engine in the integrated field rule verification module to perform verification from two dimensions of entity association logic and field format specification; S703, calling a conflict resolution unit to handle the conflicts between the multi-source verification results output by S702 and the previous processing results, specifically including: conflict detection, priority determination, and conflict handling; S704, if the final verification report contains correction instructions, feed the correction instructions back to S602, and based on the correction instructions, re-execute error correction by S602 to generate a secondary correction content carrier; input the secondary correction content carrier into S701-S703 again to perform full-process verification, if all fields pass the entity association verification and format verification, determine that the content carrier is the target content carrier; if there are still errors, repeat S704 until a completely compliant target content carrier is generated; S705, arrange the target content carrier that passes the final verification according to the standard format of the navigation notice, output to the civil aviation navigation notice publishing system, and store the error record-correction instruction-verification result log in the verification process for subsequent tracing and auditing.

2. The method of claim 1, wherein, Based on the pre-trained language model, the semantic vector of the content carrier to be processed is obtained, and the glyph feature vector of the content carrier to be processed is obtained based on the glyph feature extraction model, the pre-trained language model is fine-tuned based on the civil aviation field corpus, and the embedding layer of the pre-trained language model is fused with the civil aviation professional term dictionary.

3. The method of claim 1, wherein, The semantic association based on the timing feature generates word initial weights and sentence initial weights through a preset nonlinear transformation component and an activation function component, specifically including: For the word units in the timing feature, the hidden state vector corresponding to each word unit is obtained, the hidden state vector is input into the word-level nonlinear transformation component for feature mapping, then input into the first activation function component, and the word initial weight of each word unit is output; For the sentence units in the timing feature, perform a pooling operation on the word hidden state vector contained in each sentence unit to obtain a sentence vector, input the sentence vector into the sentence-level nonlinear transformation component for feature mapping, then input into the second activation function component, and output the sentence initial weight of each sentence unit.

4. The method of claim 1, wherein, The rule storage component adopts an editable structured data format, supports adding, deleting, and modifying the type division of the core field and the corresponding weight addition ratio, and can complete rule iteration without adjusting the attention mechanism main body architecture.

5. The method of claim 1, wherein, During weight adjustment, the core field contained in the timing feature is identified through the field matching component to determine the word units and sentence units whose weights need to be adjusted, then the corresponding word initial weight and the corresponding sentence initial weight are corrected based on the preset addition ratio in the rule storage component, the weight values are boundary constrained after correction, and the corresponding target weight is obtained.

6. An electronic device, comprising: It includes a processor and a memory; The processor is configured to execute the steps of the method according to any one of claims 1 to 5 by invoking the program or instructions stored in the memory.

7. A computer readable storage medium characterized by The computer readable storage medium is configured to store the program or instructions, which cause the computer to execute the steps of the method according to any one of claims 1 to 5.

Images