Navigation notice text processing method and device based on semantic enhancement and medium

Abstract

The invention relates to the technical field of text processing, in particular to a navigation announcement text processing method and device based on semantic enhancement and a medium, and the method comprises the steps: firstly obtaining a to-be-processed content carrier; then obtaining a semantic vector and a font feature vector of a to-be-processed content carrier; splicing the two types of vectors to form enhanced text representation; performing time sequence feature extraction on the enhanced text representation to obtain a time sequence feature containing a forward and backward logic relationship of the text; obtaining weights of words and sentences in the time sequence features and completing weighting to obtain weighted word representation and weighted sentence representation; performing correction processing on the weighted representation to generate a corrected text; and finally, checking the corrected text, and outputting a target text. According to the invention, the processing accuracy and efficiency of the content carrier can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of text processing, in particular to a navigation notice text processing method, device and medium based on semantic enhancement. BACKGROUND

[0002] As the core information carrier of aviation safety operation, the navigation notice (NOTAM) undertakes the important responsibility of publishing real-time airspace dynamics, navigation facility status, flight restrictions and other key information. The accuracy of its content directly determines the rationality of flight path planning, the effectiveness of aviation weather warning and the reliability of control decision, and is a key link to ensure the safety of aviation operation.

[0003] However, the navigation notice text has significant field-specificity: on the one hand, the text is dense in professional terms, and there are mixed language expressions; on the other hand, its semantic structure is complex, containing a large number of causal logic chains (such as "construction → runway closed") and conditional constraint relationships (such as "valid period to need later than the starting time"), which makes the manual checking mode face insurmountable bottlenecks, not only low in processing efficiency, but also unable to meet the real-time demand of aviation information, and has a high omission rate, which is easy to miss key errors due to human error.

[0004] The current mainstream navigation notice checking technology can be divided into two categories, but both have obvious defects: Logical checking technology based on pre-defined keywords and regular expressions: this kind of technology can only realize simple format matching and fixed logic judgment, cannot handle semantic ambiguity problems (for example, it cannot distinguish the semantic difference between "runway closed" and "runway partially closed"), and is more difficult to capture the context association relationship in the text, so the checking accuracy is very low in complex semantic scenarios.

[0005] Text classification checking technology based on traditional machine learning models: such as support vector machine (SVM), text convolutional neural network (TextCNN), pure bidirectional long short-term memory network (BiLSTM), etc. This kind of technology highly depends on manual feature engineering, can only extract local text features, and cannot effectively capture long-distance semantic association; at the same time, due to the lack of deep embedding of domain knowledge, the model has insufficient semantic representation ability for civil aviation professional terms, which leads to the dispersion of key information attention, and it is difficult to meet the high-precision checking demand.

[0006] In summary, the existing technology cannot balance the accuracy, efficiency and field adaptability of navigation notice checking, and an intelligent checking scheme that can deeply understand the text semantics, dynamically focus on key information and integrate domain knowledge is urgently needed to improve the automation level of navigation notice processing and reduce the labor cost and aviation safety risk. SUMMARY

[0007] In view of the above technical problems, the technical scheme adopted by the present application is: According to the first aspect of the present application, a semantic enhancement-based navigation notice text processing method is provided, which comprises the following steps: S100, obtaining a to-be-processed content carrier, the to-be-processed content carrier being a navigation notice text.

[0008] S200, obtaining a semantic vector and a character shape feature vector of the to-be-processed content carrier.

[0009] S300, generating an enhanced content carrier representation based on the semantic vector and the character shape feature vector.

[0010] S400, performing a timing feature extraction operation on the enhanced content carrier representation to obtain a corresponding timing feature.

[0011] S500, obtaining a target weight of each word and each sentence in the timing feature, and weighting the words and sentences in the timing feature in combination with the target weight to obtain a weighted word representation and a weighted sentence representation.

[0012] S600, performing a correction process on the weighted word representation and the weighted sentence representation to generate a corrected content carrier.

[0013] S700, verifying the corrected content carrier to obtain a target content carrier.

[0014] According to the second aspect of the present application, an electronic device is provided, comprising a processor and a memory; the processor is configured to execute the steps of the method according to the first aspect of the present application by calling programs or instructions stored in the memory.

[0015] According to the third aspect of the present application, a computer readable storage medium is provided, which stores programs or instructions, and the programs or instructions are configured to make a computer execute the steps of the method according to the first aspect of the present application. (1) Improve the accuracy of semantic understanding and adapt to the professional attributes of civil aviation The present method uses a pre-trained language model adapted to the civil aviation field, combines the character shape feature vector obtained by the character shape feature extraction model, and forms an enhanced text representation by splicing. This design breaks through the limitations of traditional single semantic models, can accurately identify civil aviation professional terms such as "runway number (e.g. 09 / 27) and time format (e.g. DDHHMMZ)", and can solve the problem of "similar character misjudgment (e.g. 'height 900' miswritten as 'height 90o') through character shape features. The semantic understanding precision is adapted to the core attribute of the navigation notice text "dense professional terms", and compared with the traditional general text processing model, the accuracy of semantic related error recognition is significantly improved.

[0016] (2) Capture the logical association of the text, reduce the logical error missing rate The temporal modeling network is used to model the context of the enhanced text representation, which can effectively capture the "forward logical chain (such as 'construction → runway closed' causal association)" and "backward constraint relationship (such as 'valid period to → start time' conditional restriction)" in the navigation notice text. This design solves the pain points of existing regular check and traditional machine learning models that cannot handle long-distance semantic association, avoids safety hazards caused by logical errors such as "start time later than valid period to" and "facility state and operation description contradiction" not being detected, reduces the logical error missing rate compared to traditional technology, and significantly improves the reliability of text processing.

[0017] (3) Focus on key information verification to ensure the accuracy of core elements By weighting the words and sentences in the context modeling text through the attention mechanism, the "time, coordinates, height" and other core safety elements in the navigation notice can be dynamically focused on, avoiding the problem of key information attention dispersion caused by traditional models treating all fields "equally". Combined with the verification module integrated with domain rules, core elements can be verified first, and the text can also be ensured to comply with civil aviation domain specifications (such as runway number must be in the airport entity list), the key field error detection rate is improved, further reducing the risk of aviation safety.

[0018] (4) Adapt to real-time processing scenarios to improve business response efficiency The integrated domain rule verification module has the ability to dynamically update, adapting to the business scenario of real-time publishing of airspace dynamics and quick response to specification updates. Compared to traditional technology which requires system restart and processing interruption for rule updates, this method can realize real-time iteration of domain rules, ensuring that text verification always complies with the latest civil aviation specifications, improving business response efficiency and reducing the problem of verification failure caused by rule lag.

[0019] (5) Reduce the cost of human dependence and improve the level of automation Through the whole process of automatic processing of "feature extraction-context modeling-attention weighting-error detection and correction-rule verification", the traditional manual verification mode of "low efficiency and high missing rate" is replaced, greatly reducing the manual review time and labor cost. According to the test, in the processing of multiple batches of navigation notice texts, the processing time delay of a single text is controllable, and the error detection recall rate and format false positive rate are better than existing technologies, significantly improving the automation level of navigation notice text processing and providing technical support for the efficiency of aviation information processing.

[0020] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present application, nor to limit the scope of the present application. Other features of the present application will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description only represent some of the embodiments of the present application, and all other drawings obtained by those skilled in the art without creative effort based on these drawings also belong to the protection scope of the present application.

[0022] Figure 1 A flow chart of a semantic enhancement-based navigation notice text processing method provided by the embodiment of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort also belong to the protection scope of the present application.

[0024] 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 application belongs. The terminology used in the description of the application herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used in this description, the singular forms "a", "an" and "the" include plural references unless the context clearly dictates otherwise. The term "and / or" includes any and all combinations of one or more of the associated listed items.

[0025] It should be noted that some of the example embodiments are described as processes that are depicted as flow diagrams. Although the flow diagrams can describe the processes as a sequential process, many of the steps can be performed in parallel, concurrently or simultaneously. In addition, the order of the steps can be re-arranged. A process can be terminated when its operations are completed, but could also occur under some other condition or event occurred, some or all of the steps could also be skipped, the name of the process is changed, or the process could be modified in some manner before, during, or after execution. A process described as a process, function, procedure, subroutine, algorithm, or the like can be implemented using software, firmware, hardware, or any combination of them.

[0026] The present application aims to overcome the problems in the prior art, and provide a navigation notice text verification method supporting semantic depth analysis and dynamic focusing on key fields, which is used for detecting and correcting semantic errors, syntax errors and format non-standard problems in navigation notice texts.

[0027] Further, the embodiment of the present application provides a semantic enhancement-based navigation notice text processing method, as shown in Figure 1 The method can include the following steps: S100, obtaining a content carrier to be processed.

[0028] In the embodiment of the application, the specific form of the content carrier is text, and the content carrier to be processed is a navigation notice text. For the convenience of description of the subsequent technical solutions, the content carrier and the text can be used interchangeably in the following description, that is, the content carrier can be collectively referred to as text in expression, and the text also corresponds to the aforementioned content carrier in technical connotation, both of which refer to the literal carrier carrying information in the field of navigation notice, without changing the core technical attributes and scope.

[0029] The navigation notice text is a core information carrier for safe operation of aviation, which is used to publish key information such as airspace dynamics, navigation facility status (such as runway and navigation station operation), flight restrictions (such as temporary no-fly zones and height restrictions) in real time. The text content includes time (such as effective / inactive time in the format of “DDHHMMZ”, for example, “120600Z”), coordinates (such as airspace geographic coordinates and airport runway coordinates), height (such as flight restriction height and minimum safety height), and professional terms (such as “RWY” for runway, “CLSD” for closed, and “MAINT” for maintenance), and has the characteristics of complex semantic structure (including forward causal chain of “construction→facility closed” and backward constraint relationship of “valid period to need later than starting time”), which adapts to the technical requirements of subsequent semantic feature extraction and context logic modeling.

[0030] In the embodiment of the application, the acquisition method of the text to be processed includes retrieving from an aviation management system database and receiving text data pushed by civil aviation departments in real time. After acquisition, the text needs to be preprocessed: removing redundant spaces, unifying text encoding format, and truncating or completing the text length to the maximum length of the sequence input by the model (such as 512 tokens), to ensure that the text to be processed meets the subsequent model input specifications, and to provide a high-quality data source for accurate extraction of semantic vectors and glyph feature vectors.

[0031] In one illustrative embodiment, the acquired text to be processed can be: “RWY09 / 27CLSD DUETOMAINT FROM 120600Z TO 121200Z”.

[0032] S200, acquiring a semantic vector and a glyph feature vector of the content carrier to be processed.

[0033] In the embodiment of the present application, the semantic vector of the content carrier to be processed is obtained based on a pre-trained language model, and the 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 based on civil aviation field corpus, and the embedding layer of the pre-trained language model is fused with a civil aviation professional term dictionary, which is used to output a semantic vector representing semantic information of the text. In an exemplary embodiment, the pre-trained language model can be selected from ERNIE series, BERT series, Zhonghangxinxi Rongda model, or Aviation GPT, etc. Different models are adapted to follow the following common process: First step: build civil aviation field corpus (1) Obtain original corpus: sources cover civil aviation regulations documents (airspace management specifications, navigation facility operation standards, flight notice writing rules, etc. official documents), historical flight notice texts (including airspace dynamics, navigation facility status, flight restriction type effective notices, which need to contain complete time, coordinates, height, etc. key elements and professional terms such as “RWY”, “CLSD”, “MAINT”), civil aviation professional auxiliary texts (airport operation manual related description, navigation facility maintenance record associated content), and supplement the diversity of corpus scenarios.

[0034] (2) Corpus cleaning and preprocessing: delete incomplete content (missing key elements), format confusion (no standard time identifier “Z”), and redundant corpus; unify the same meaning but different expressions of terms according to civil aviation regulations (such as “runway closed” unified as “RWYCLSD”); convert all corpus to a unified encoding format, split “element description section” and “restriction description section” according to flight notice structure, and ensure consistent input format.

[0035] (3) Corpus annotation and division: annotate time, coordinates, height, and professional terms in the corpus, and divide it into training set (model fine-tuning), validation set (parameter adjustment), and test set (adaptation effect evaluation) according to the ratio of 7:2:1.

[0036] Second step: fine-tuning of general pre-trained language model (1) Model initialization before fine-tuning: load general model weights, extend embedding layer, i.e. add additional word embedding matrix corresponding to civil aviation professional term dictionary to embedding layer, and input term converted to word vector; according to the average length of flight notice text, set the maximum length of input sequence to 512 tokens (match time sequence modeling network processing capacity), and truncate the text longer than the length according to the “key element priority retention” principle, and supplement the length shorter than the length with domain-specific placeholders.

[0037] (2) Fine-tuning task and loss function setting: adopt the double task of "masked language modeling + next sentence prediction" for fine-tuning, and the mixed loss function is "0.7x masked language modeling loss + 0.3x next sentence prediction loss", both of which use cross-entropy loss function. Among them, the proportion of the masked language modeling task is 80%, and the specific operation is to randomly mask 15% of the civil aviation professional terms in the training set, and the masking method includes three kinds: 80% of the masked positions are replaced with [MASK] symbol, 10% of the masked positions are replaced with random civil aviation terms, and 10% of the masked positions remain the original terms. Through this task, the model predicts the correct terms in the masked position according to the context; the proportion of the next sentence prediction task is 20%, and the specific operation is to split the corpus in the training set into "previous sentence-next sentence" sentence pair form, and randomly generate 50% of "non-continuous sentence pairs" (i.e. the previous sentence and the next sentence have no actual context association). Through this task, the model judges whether the input sentence pair is a continuous context.

[0038] (3) Fine-tuning process parameter control: use AdamW optimizer, initial learning rate 2e-5, linear decay 10% every 3 rounds; set 10 rounds of training, input 32 pieces of corpus per batch, use the validation set to evaluate the term prediction accuracy and sentence pair judgment accuracy every round, and stop and save the optimal weight when the performance does not improve for 2 consecutive rounds; add a dropout layer (probability 0.2) to the full connection layer to reduce overfitting.

[0039] (4) Model verification after fine-tuning: calculate the cosine similarity (≥0.85) between the semantic vector and the key element annotation with the test set to verify the effectiveness of the vector representation; input the semantic vector into the word shape feature vector to verify the improvement effect of the fused enhanced text representation on the logical relationship capture accuracy of the context modeling module (the logical error recognition rate needs to be improved by ≥15%).

[0040] In an illustrative embodiment of the application, the ERNIE series pre-training language model is selected, and the output semantic vector dimension is 768.

[0041] In the embodiment of the present application, the character feature extraction model is a model capable of extracting structural features of Chinese characters, and is used to output a character feature vector representing text character shape information, and is preferably a convolutional neural network model (at least containing two convolutional layers). In a specific embodiment of the present application, a convolutional neural network model with three convolutional layers is used, and the convolutional sizes of the three convolutional layers are 3x3, 5x5 and 3x3 in sequence. This size design is suitable for the characteristics of short Chinese character length and small structural differences in the navigation notice text: a 3x3 small size convolution kernel is used to capture the local stroke features of Chinese characters, and a 5x5 large size convolution kernel is used to capture the overall structural features of Chinese characters. Among them, the first 3x3 convolutional layer outputs 64-dimensional features, the second 5x5 convolutional layer outputs 128-dimensional features, and the third 3x3 convolutional layer fuses and reduces the features output by the previous two layers, and finally outputs a 50-dimensional character feature vector. This vector can effectively distinguish the character shape differences between "900" and "90o", "height" and "high variation" and other similar characters.

[0042] The technical effect of S200 is that: (1) Improve the semantic understanding accuracy and model compatibility in the field of civil aviation: Through the design of "multi-model selection + common fine-tuning process", it is ensured that pre-trained language models of different architectures can accurately identify civil aviation professional terms (such as "RWY" "CLSD") and semantic associations specific to the field such as "construction → closed". Through testing, the term recognition accuracy is improved by ≥20% compared with general pre-trained models, and the semantic association misjudgment rate is reduced by ≥18%. At the same time, the dimension design of the 768-dimensional semantic vector can seamlessly splice with the subsequent character feature vector, providing accurate semantic basis for context modeling, and the multi-model selection expands the protection range of the technical solution, avoiding the technical avoidance risk caused by single model limitation.

[0043] (2) Accurately solve the problem of similar character misjudgment: For the common similar character errors in the navigation notice text, the design of three-layer differentiated convolution size can capture the local stroke and overall structure features of Chinese characters. Through testing, the 50-dimensional character feature vector has a discrimination accuracy of ≥95% for "900 / 90o", "height / high variation" and other similar characters, which makes up for the defect that traditional single semantic model cannot identify visual similar errors, significantly reduces the risk of verification omission caused by similar character errors, and further ensures the accuracy of navigation notice text processing.

[0044] (3) Provide high adaptability feature input for subsequent modules: the dimension matching design of semantic vector (768 dimensions) and glyph feature vector (50 dimensions) ensures that the 818-dimensional enhanced text representation after splicing can be directly input into the time series modeling network without additional dimension conversion; the double-dimensional features can meet the needs of context modeling module for semantic association capture and attention module for key information focusing, avoiding insufficient precision in subsequent processing due to single feature dimension, and the efficiency of the entire navigation notice text processing flow is improved by ≥12% compared with traditional solutions.

[0045] S300, based on the semantic vector and the glyph feature vector, generating an enhanced content carrier representation.

[0046] This step is based on the vector output by S200 to splice the semantic vector and the glyph feature vector to form an enhanced text representation that integrates semantic and glyph information. The semantic vector (such as the 768-dimensional vector output by the ERNIE series model) and the glyph feature vector (such as the 50-dimensional vector output by the convolutional neural network) need to be aligned at the text unit level, that is, for each word (or subword) in the text to be processed, its corresponding 1 semantic vector (dimension 768) and 1 glyph feature vector (dimension 50) are taken as a group of associated vectors, ensuring that the vector splicing object is the feature representation of the same text unit, and avoiding feature fusion failure due to text unit misplacement.

[0047] The splicing process is performed in a dimension stacking manner: for each set of aligned semantic vector and glyph feature vector, the dimensions are spliced in the fixed order of "semantic vector first, glyph feature vector second", that is, the spliced vector dimension of a single text unit is "semantic vector dimension + glyph feature vector dimension" (such as 768+50=818 dimensions); if the text to be processed contains n words (or subwords), the final enhanced text representation is an "n x 818" dimensional matrix (n is the number of text units, matching the limit of the maximum length of 512 tokens input sequence of the pre-trained language model in S200, that is, n≤512), each row of the matrix corresponds to the fusion features of a text unit, and each column corresponds to a type of feature dimension (the first 768 columns are semantic feature dimensions, and the last 50 columns are glyph feature dimensions).

[0048] After splicing, feature consistency verification needs to be performed: by calculating the dimension integrity of each text unit fusion vector in the splicing matrix (to ensure that there is no dimension missing or redundancy), the matching of the number of text units and the length of the input text in S200 (to ensure that there is no text unit missing or additional addition), if the verification is passed, the enhanced text representation is output; if the verification finds that the dimensions do not match (such as some text units only contain semantic vectors without glyph feature vectors) or the number of text units is misaligned, automatically backtrack to S200 to re-extract the missing vectors of the corresponding text units until the enhanced text representation that meets the requirements is generated.

[0049] The core value of enhancing the text representation in the embodiment of the application lies in: retaining the precise representation capability of the semantic vector to the civil aviation field professional terms (such as "RWY" and "CLSD") and the text logical correlation (such as "construction -> closed"), and integrating the distinguishing capability of the character shape feature vector to the structural differences of Chinese characters (such as the stroke difference between "900" and "90o", and the structural difference between "height" and "high variation"), realizing the dual information fusion of "semantic understanding + visual feature recognition", providing a more comprehensive feature basis for the subsequent context modeling module to capture the text logical relationship and the attention module to focus on the key information, and avoiding the limitation of single semantic feature or single character shape feature to the text representation.

[0050] S400, performing a time sequence feature extraction operation on the enhanced content carrier representation to obtain corresponding time sequence features.

[0051] In the embodiment of the application, the context modeling network is used to model the enhanced text representation, and a context modeling text containing the forward and backward logical relationships of the text is obtained, that is, the time sequence features.

[0052] In the embodiment of the application, the context modeling network is preferably a bidirectional long short-term memory network (Bi-LSTM), and the core design is developed around the time sequence correlation and logical dependency of the navigation notice text. The chain relationship of time-event-limit in the navigation notice (such as RWYMAINT (maintenance) -> RWYCLSD (closed) -> ATCINSTRUCTION (control instruction)) needs to capture the forward time sequence flow and the reverse dependency relationship at the same time. The bidirectional structure can simultaneously learn the forward logic (such as the event development sequence) from left to right and the reverse correlation (such as the tracing of the result to the cause) from right to left. The specific network parameter settings are as follows: The hidden dimension is set to 256. The 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 (256 dimensions can balance feature retention and computational efficiency, avoiding overfitting caused by too high dimension); The network layer is 2 layers. The first layer is used to preliminarily extract local time sequence correlation (such as the collocation relationship of adjacent terms, "RWY09 / 27" and "CLSD" combination), and the second layer is used to integrate global logical chain (such as the complete event sequence of "maintenance -> closed -> recovery"); Each layer is provided with dropout regularization (probability 0.2). The dropout operation is performed on the sequence dimension to reduce the excessive dependence of the model on specific time sequence position features and enhance the generalization ability to different format notice texts; The output layer adopts linear transformation to splice the forward hidden state and the backward hidden state of the bidirectional LSTM (256x2=512 dimensions), forming a context vector containing bidirectional logical information.

[0053] The modeling process is executed in three steps: Input adaptation: the n*818-dimensional enhanced text representation output by S300 (n<=512) is input into the Bi-LSTM in the order of text units, and the fusion features of each time step corresponding to a text unit are obtained; Bidirectional encoding: the forward LSTM encodes from the first text unit to the last text unit, capturing the forward logic of cause->effect (such as navigation facility failure->runway closure); the backward LSTM encodes from the last text unit to the first text unit, capturing the reverse association of effect->cause (such as runway closure <- navigation facility failure); Feature fusion: the element-wise addition fusion operation is performed on the forward hidden state and the backward hidden state of each time step to generate context modeling text (dimension n*256) containing bidirectional logical information, wherein the output vector of each text unit contains both its semantic role in the forward time sequence and its logical position in the reverse dependence.

[0054] The technical effect of S400 is that: (1) The gating mechanism of Bi-LSTM can effectively solve the long-distance dependence forgetting problem of traditional recurrent neural networks, accurately capture the implicit association of logical relationships such as construction time at the beginning of the notice and restricted height at the end, according to the characteristics of the time coordinate accuracy and event causality of the navigation notice; (2) The 256-dimensional hidden layer is adapted to the dimension of the enhanced text representation, ensuring that the key semantic and glyph information is retained while the features are compressed, and providing high-quality input for the subsequent attention module to focus on time, coordinates, and restriction conditions; (3) The bidirectional modeling capability can identify logical contradictions in the text (such as RWYCLSD without specifying the closing time), provide abnormal features at the logical level for subsequent error detection, and improve the comprehensiveness of the notice text verification.

[0055] In S500, the target weight of each word and each sentence in the time sequence feature is obtained, and the words and sentences in the time sequence feature are weighted in combination with the target weight to obtain a weighted word representation and a weighted sentence representation.

[0056] In the embodiment of the application, the attention mechanism is used to obtain the target weight of each word and each sentence in the time sequence feature, and the words and sentences in the time sequence feature are weighted in combination with the target weight to obtain a weighted word representation and a weighted sentence representation.

[0057] In the embodiment of the present application, the attention mechanism is a rule-guided hierarchical attention mechanism, which is different from the traditional general attention mechanism (such as the Transformer attention) which only relies on data-driven weight allocation. The mechanism is designed in two stages of data learning initial weight + domain rule correction weight, which accurately focuses on the time, coordinates, height and other core safety elements in the navigation notice text, and solves the problem of attention dispersion of traditional mechanisms on key information in the field. The hierarchical attention mechanism is divided into two layers of word-level attention and sentence-level attention, which correspond to the word unit and sentence unit of the time sequence feature respectively.

[0058] In the embodiment of the present application, the attention mechanism is used to obtain the weight of each word and each sentence in the time sequence feature, which specifically includes: S501, based on the semantic association of the time sequence feature, the word initial weight and the sentence initial weight are respectively generated through the preset nonlinear transformation component and the activation function component, so as to ensure that the weight reflects the semantic importance of the text itself.

[0059] S501 specifically can include: For the word unit in the time sequence feature, the hidden state vector corresponding to each word unit is obtained, and after the hidden state vector is input into the word-level nonlinear transformation component for feature mapping, the first activation function component is accessed, and the word initial weight of each word unit is output.

[0060] In the embodiment of the present application, the word-level nonlinear transformation component includes a trainable feature conversion matrix and a bias term, specifically, the word initial weight of each word unit satisfies the following condition: alpha i initial =softmax((W q ·h i +b q ) / (d k 1 / 2 )).

[0061] Wherein, alpha i initial is the word initial weight of the i-th word unit in the time sequence feature, i is 1 to n, n is the number of word units in the time sequence feature. h i is the hidden state vector of the i-th word unit in the time sequence feature, with a dimension of 256, representing the bidirectional context information of the i-th word; W q and b q are trainable word-level feature conversion parameters, specifically, W q is a word-level feature conversion matrix with a dimension of 256x256, b q is a word-level bias term with a dimension of 256. d k 1 / 2 is a scaling factor, d k= 256, to avoid the value being too large after feature conversion, which leads to the softmax output tending to be extreme (some weights tend to 1 and the rest tend to 0), and to ensure that the initial weights can reasonably distinguish the semantic importance of different words (for example, the initial weights of “CLSD” and “120600Z” are higher than those of “THE” and “A”).

[0062] For the sentence units in the time sequence feature, a pooling operation is performed on the word hidden state vectors contained in each sentence unit to obtain a sentence vector, the sentence vector is input into a sentence-level nonlinear transformation component for feature mapping, then a second activation function component is accessed, and sentence initial weights of each sentence unit are output.

[0063] In the embodiment of the application, the sentence-level nonlinear transformation component includes a trainable feature conversion matrix and a bias term, and specifically, each sentence initial weight satisfies the following condition: beta j initial = sigmoid (W s + b j ). s ).

[0064] Wherein, beta j initia is the sentence initial weight of the jth sentence unit in the time sequence feature, s j is the sentence vector of the jth sentence unit in the time sequence feature, which is obtained by averaging and pooling all word hidden state vectors contained in the jth sentence, has a dimension of 256, and represents the global semantics of the jth sentence; j has a value of 1 to m, m is the number of sentences in the time sequence feature, W s and b s are trainable sentence-level feature conversion parameters, specifically, W s is a sentence-level feature conversion matrix with a dimension of 256x128, and b s is a sentence-level bias term with a dimension of 128. In the embodiment of the application, the trainable parameter initialization adopts a Xavier normal distribution to ensure the stability of the gradient in the training process.

[0065] In the embodiment of the application, sigmoid is used instead of softmax because there is no constraint that the sum of weights between sentences is 1 (for example, both the “event description sentence” and the “restriction explanation sentence” in the announcement need high weights), which can simultaneously improve the attention proportion of key sentences.

[0066] S502, based on the preset domain rule, the word initial weights and the sentence initial weights are adjusted through a dynamically configurable rule storage component to obtain target weights of each word and each sentence in the time sequence feature.

[0067] In the embodiment of the present application, the preset field rule is a core field weight adjustment rule formulated according to field normative documents, and the core field at least includes a time type field, a coordinate type field, a height type field and a facility state type field representing key attributes of the field, and different types of core fields correspond to preset weight addition ratios. Specifically, according to the influence degree of elements on aviation safety, the weight addition ratios are divided: First addition (+30%): time field (such as "120600Z" and "121200Z"), coordinate field (such as "N30°24'E120°18'"), and height field (such as "FL350" and "1500FT"), such element errors directly cause safety risks and need the highest priority attention; Second addition (+15%): facility state field (such as "RWYCLSD" and "NAVAIDOUT"), and operation instruction field (such as "ATCCLEARANCE" and "MAINTAIN"), such elements are related to operation decisions and need secondary priority attention; Third addition (+5%): airport / airspace identification field (such as "PEK" and "ZBAA"), such elements ensure the positioning accuracy of information and need basic priority attention.

[0068] In the embodiment of the present application, 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. The structured data format includes but is not limited to JSON configuration file, XML configuration file, YAML configuration file, etc. In one specific embodiment, the rule storage component is a JSON configuration file.

[0069] In the weight adjustment process, the core field contained in the time sequence feature is first identified by the field matching component to determine the word unit and sentence unit whose weight needs to be adjusted, and then the preset addition ratio in the rule storage component is used to correct the corresponding word initial weight and sentence initial weight. After correction, the weight value is subjected to boundary constraint processing to ensure that the target weight is within the preset effective range. The specific adjustment process can include: Field matching: the above core fields are identified in the time sequence feature by using regular expressions to determine the word unit and sentence unit whose weight needs to be corrected (such as identifying "120600Z", then locating the corresponding word unit and the sentence unit in which it is located); Weight superposition: multiplying the initial weight of the matched word unit by (1+addition ratio), such as the weight after first addition is α i initial ×1.3; multiplying the initial weight of the sentence unit containing the core field by (1+the highest level addition ratio contained in the sentence), such as the sentence contains time and height fields, and is corrected to βj initial ×1.3; Boundary constraint: if the corrected weight exceeds 1, it is truncated to 1 (to avoid weight overflow leading to subsequent weighted calculation distortion), and if it is lower than 0.1, it is raised to 0.1 (to ensure that non-key elements still retain basic attention).

[0070] Further, the word hidden state vector h i is multiplied element by element with the corresponding corrected weight h i to obtain the weighted word representation, highlighting the semantic features of the key words. The word weighting formula is h i w = a i · h i , h i w is the weighted word representation of the i-th word.

[0071] Further, the sentence vector s j is multiplied element by element with the corresponding corrected weight β j to obtain the weighted sentence representation, strengthening the global semantics of the key sentences. The sentence weighting formula is s j w = β j · s j , s j w is the weighted sentence representation of the j-th sentence.

[0072] The final output of the weighted word representation and the weighted sentence representation not only retains the semantic association of the text itself, but also strengthens the features of the safety-critical elements through domain rules, providing key highlighted feature input for subsequent error detection and correction.

[0073] S600, performing correction processing on the weighted word representation and the weighted sentence representation to generate a corrected content carrier.

[0074] Further, S600 specifically includes: S601, performing error detection operation on the weighted word representation and the weighted sentence representation, and outputting error position label based on BIO tagging system, the error position label indicating at least one of the following error types: semantic error, syntax error, format error.

[0075] S601 specifically includes: S6011, establishing the association relationship of the word representation and the sentence representation in the logical order of the text: concatenating 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 being used to cover both local word-level abnormalities and global sentence-level logical contradictions in error detection.

[0076] S6012, input the sentence-word level feature matrix into the error detection classifier, and output the error position label.

[0077] The error detection classifier includes a 3-layer fully connected network, and each layer is configured as follows: The first layer has a hidden dimension of 512 and uses a ReLU activation function, is used for nonlinear transformation of the sentence-word level feature matrix, and extracts general error features; The second layer has a hidden dimension of 256 and embeds a civil aviation field error feature extraction submodule, which is used to capture field-specific error patterns such as facility state conflict and time format anomaly; The third layer has an output dimension of 4, corresponding to a BIO label space and an error type code. The BIO label space includes B-error start, I-error continuation, O-no error, and E-error end. The error type code includes semantic error = 1, syntax error = 2, and format error = 3. The label probability distribution of each word unit is output by a softmax function.

[0078] The error detection classifier identifies the error type as follows: For 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 “facility state conflict” and “time logic contradiction” and other issues, and label them as “B-semantic error”, “I-semantic error”, and “E-semantic error”; For syntax errors: according to the specification that navigation notice element description should contain a complete subject-predicate-object structure, identify incomplete components and improper collocations (such as “CLSD” missing the subject “RWYXX”), and label them as B-format error, I-format error, and E-format error; For format errors: compare the civil aviation time (DDHHMMZ format), coordinate format (NXX°XX′E XX°XX′ format), and height format (XXXFT / FLXXX format) standards, identify missing symbols and character transposition issues (such as “120600” missing “Z”), and label them as B-format error, I-format error, and E-format error; Word units with no errors are uniformly labeled as “O”.

[0079] S6013, integrate all BIO labels and error type codes of the word units to generate a label sequence containing text position-error type, which provides accurate positioning basis for error correction in S602.

[0080] S602, use a sequence-to-sequence model to perform correction operations on the error content indicated by the error position label, and generate the corresponding corrected text.

[0081] In the embodiment of the present application, the sequence-to-sequence model adopts the architecture of a bidirectional GRU encoder-unidirectional GRU decoder + domain attention mechanism, which is suitable for the characteristics of short error fragments and strong correction regularity of the navigation notice text, wherein: Encoder: a bidirectional GRU network is adopted, the hidden dimension is 256, the input is the splicing features of the error position label sequence and the weighted word representation, and the semantic association of the context before and after the error position (such as the keywords “FROM” and “TO” before and after the error time “120600”) is encoded; Decoder: a unidirectional GRU network is adopted, the hidden dimension is 256, the context vector output by the encoder is received, the “domain attention mechanism” (preferentially paying attention to the error position and the associated civil aviation core field) is combined, and the correction candidate content is generated word by word; Output constraint layer: access to a civil aviation professional term dictionary (including time, coordinate, facility code, and other standard expressions) and a softmax function, so as to ensure that the generated correction content meets the requirements of the domain format and semantics, and avoid new standard conflicts.

[0082] S602 specifically includes: S6021, the error position label sequence, the weighted word representation, and the weighted sentence representation output by S601 are spliced according to a preset dimension to form input features of the sequence-to-sequence model; the input features include error position information, word-level semantic features, and sentence-level logical features, which are used to provide error context association information for the model.

[0083] S6022, based on the error type indicated by the error position label, a preset domain rule resource is called to perform targeted correction: For semantic errors: the domain knowledge graph in the verification module of the integrated domain rule is called, the mapping relationship between civil aviation entities and standard values stored in the knowledge graph is used to retain the state expression conforming to the latest timestamp for facility state conflict type errors; for logical missing type errors, the associated mandatory elements are supplemented (such as supplementing the closing time interval for “RWYCLSD”); For syntax errors: the standard sentence pattern template library of the navigation notice is called, the subject, predicate, or object is completed for component missing type errors (such as supplementing the subject “RWY09” for “CLSD”); the word order is adjusted according to the template of “facility code + state description + restriction condition” for syntax disorder type errors; For format errors: the standard format library of the regular expression engine is called, the time zone identifier “Z” is completed and the “DDHHMMZ” format is verified for time field errors; the angle symbols “°” and “′” are supplemented and the “NXX°XX′EXXX°XX′” format is verified for coordinate field errors; the unit “FT” or “FL” is supplemented and the numerical range is verified for height field errors.

[0084] S6023, the decoder of the sequence-to-sequence model adopts a beam search strategy (beam size = 3) to filter the optimal candidate words from the civil aviation professional term dictionary based on the correction logic of S6022, and generates a corrected text sequence word by word; the civil aviation professional term dictionary contains standard terms such as facility codes, time formats, coordinate expressions and the like conforming to the ICAO standard, and is used to constrain the domain compliance of the correction content.

[0085] S6024, double-checking the generated corrected text: Error residual checking: re-input the corrected text into the error detection logic of S601 to check whether there is still an error position tag; Rule compliance checking: call the verification module integrated with the 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; If the double-checking passes, output the final corrected text; if there is an item that does not pass, feed back the uncorrected position and the corresponding error type to S6022, and re-execute the correction operation until a corrected text conforming to the civil aviation domain specification is generated.

[0086] S700, checking the correction content carrier to obtain a target content carrier.

[0087] In the embodiment of the application, the verification module integrated with the domain rules is used to check the corrected text to obtain a target content carrier.

[0088] In the embodiment of the application, the verification module integrated with the domain rules includes a domain knowledge graph, a regular expression engine, a knowledge graph incremental update interface, a regular rule hot loading unit and a conflict resolution unit, each of which is a software logic component, specifically: The domain knowledge graph is a software component that stores the mapping relationship between entities in the target domain and corresponding standard values, and internally pre-constructs a standardized entity-attribute-standard value triple structure, which is used to constrain the logicality and compliance of entity association in the correction content carrier. The triple covers civil aviation core business scenarios, for example: entity = RWYCLSD (runway closure) → attribute = associated elements → standard value = closure time interval (format DDHHMMZ-DDHHMMZ), affected airspace range (including airport code and coordinate boundary), entity = NAVAIDOUT (navigation facility failure) → attribute = restriction condition → standard value = affected flight height ≤ FL200, alternate airport identifier, entity = UAVRESTRICT (UAV restricted area) → attribute = flight restriction parameter → standard value = flight restriction time, flight restriction radius ≤ 5KM, to ensure that the mapping relationship between entities and standard values fully matches the requirements of official documents such as the "Civil Aviation Navigation Notice Management Method" and the "Airspace Use Management Regulations".

[0089] The regular expression engine is a software component for performing format compliance checks on specific format fields in the modified content carrier, including at least time type fields, coordinate type fields, and can be extended to height type fields, facility code type fields. The regular expression engine internally preloads a civil aviation exclusive format regular rule library, and each rule in the rule library corresponds to a civil aviation field standard format, for example: Time type field rule: ^\d{6}Z(-\d{6}Z)?$ (matches a single timestamp such as 120600Z, or a time interval such as 120600Z-121200Z); Coordinate type field rule: ^N\d{2}°\d{2}'E\d{3}°\d{2}'$ (matches north latitude, east longitude coordinates such as N30°24'E120°18'); Height type field rule: ^(FL\d{3}|\d{1,4}FT)$ (matches a flight altitude layer such as FL350, or a foot height such as 1500FT); Facility code type field rule: ^RWY\s+\d{1,2}([ / ]\d{1,2})?$ (matches a runway code such as RWY09, or RWY09 / 27); Quickly identify format exception fields through regular matching.

[0090] The knowledge graph incremental update interface is a software interface that supports dynamic iteration of domain knowledge, and is used to add, modify or delete entity-attribute-specification value mapping relationships in the domain knowledge graph through a pre-set network interface protocol (such as RESTful API, SOAP API). When civil aviation field rules are updated (such as adding new navigation facilities, adjusting airspace restriction parameters), users can send standardized requests (such as POST requests to add "entity = SBAS INOP (Satellite Based Augmentation System failure) → attribute = impact range → specification value = coverage area code") through the interface, and the interface automatically synchronizes to the knowledge graph database after receiving the request, and the update process takes effect in real time, without the need to rebuild the overall structure of the knowledge graph.

[0091] The regular rule hot loading unit is a software logic unit for realizing dynamic updating of regular check rules, and is used for updating a rule base of the regular expression engine through an editable configuration file (supporting structured formats such as JSON, XML, and YAML), and the updating process does not need to interrupt the running of the check system. For example, when a new standard of “DDHHMMSSZ” (containing a second-level precision) is added to the civil aviation time format, a user can edit a JSON configuration file, add a rule ^\d{8}Z(-\d{8}Z)?$ under a “time_rules” field, save the file, and the regular rule hot loading unit automatically reads and updates the content through a file listening mechanism, injects the new rule into the rule base of the regular expression engine, and meanwhile, retains historical rules, so as to ensure the compatibility of new and old format checks, and the whole process does not affect the check task of the current correction content carrier.

[0092] The conflict resolution unit is a software logic unit for processing inconsistency of multi-source check results. When the correction processing result of S602, the error detection result of S601, and the check result of the integrated domain rule check module (domain knowledge graph and regular expression engine) exist conflicts, the conflict resolution unit determines a final check result according to a preset priority strategy. In the preset priority strategy, the check result of the domain knowledge graph has the highest priority (because the entity association logic directly affects flight safety), the format check result of the regular expression engine has the second priority (format specification ensures information transmission accuracy), and the correction result of S602 has the last priority; for example, S602 corrects “RWY09CLSD” to “RWY09CLSD121200Z-120600Z” (time interval logic contradiction), the domain knowledge graph checks and marks the time interval as incorrect, and the regular expression engine checks and marks the format as compliant. At this time, the conflict resolution unit preferentially adopts the knowledge graph result, and determines that the time interval needs to be re-corrected, rather than accepting the result of format compliance.

[0093] Further, S700 specifically includes: 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: traversing civil aviation professional terms (such as “RWYCLSD” and “NAVAIDOUT”) in the content carrier, associating corresponding attribute information (such as closing time and impact range), forming an entity class field set, and inputting the domain knowledge graph; identifying and extracting format class fields: positioning time class, coordinate class, height class, and facility code class fields through keyword matching (such as “Z”, “°”, “FT”, and “FL”), removing redundant spaces and special characters (such as meaningless “_” and “-”), forming a format class field set, and inputting the regular expression engine; 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.

[0094] 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: (1) Entity association verification based on domain knowledge graph: 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"); 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. 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). 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".

[0095] (2) Format validation based on regular expression engine: 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; 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 '°''′'"); C. Output format verification result: generate a format verification report consistent with the entity verification report format to ensure that the subsequent conflict resolution can be uniformly processed.

[0096] S703, call the conflict resolution unit to process the conflict between the multi-source verification result output by S702 and the previous processing result, specifically including: (1) Conflict detection: compare the correction result of S602, the error detection result of S601, and the entity verification result and format verification result of S702 to identify inconsistent items (such as S602 determining that "RWY09CLSD121200Z-120600Z" is corrected, while the domain knowledge graph determines that its "standard value logic is contradictory"); (2) Priority determination: according to the priority of "domain knowledge graph result > regular expression engine result > S602 correction result > S601 error detection result", determine the conflict items; (3) Conflict processing: If the domain knowledge graph result conflicts with other results: take the domain knowledge graph result as the standard, generate a "field to be corrected - correction direction" instruction (such as "the closing time interval of RWY09CLSD needs to be adjusted to start time < end time, such as 120600Z-121200Z"); If the regular expression engine result conflicts with the S602 correction result: take the regular result as the standard, and modify the format (such as "time field 120600 / needs to supplement the time zone identifier 'Z', modify to 120600Z"); Output the resolved result: integrate the conflict-free verification passed fields and the modified instructions after conflict processing to form the "final verification report".

[0097] S704, if the final verification report contains modification instructions, feed the modification instructions back to S602, and based on the modification instructions, S602 re-executes error correction (such as modifying the closing time interval to "120600Z-121200Z") to generate a secondary correction content carrier; input the secondary correction content carrier back to S701-S703 to perform full-process verification, if all fields pass entity association verification and format verification (no error mark), determine that the content carrier is the target content carrier; if there are still errors, repeat step four until a completely compliant target content carrier is generated.

[0098] S705, arrange the target content carrier that passes the final verification according to the standard format of the navigation notice (such as including "notice number - release time - entity information - restriction condition" paragraph), output to the civil aviation navigation notice publishing system, and store the "error record - modification instruction - verification result" log in the verification process for subsequent tracing and auditing.

[0099] The semantic enhancement-based navigation notice text processing method provided by the embodiment of the present application has a semantic error recall rate of 92.7% and a format false alarm rate of 3.1% on a test set containing 50,000 navigation notices, and the processing delay of a single notice is less than 350 ms.

[0100] Based on the same inventive concept, the embodiment of the present application provides a semantic enhancement-based navigation notice text processing device, which comprises a text acquisition module, a feature extraction module, a feature fusion module, a time sequence modeling module, an attention weighting module, an error detection and correction module, and a domain rule verification module. Wherein: The text acquisition module is configured to acquire a text to be processed, wherein the text to be processed is a navigation notice text. The feature extraction module comprises a pre-trained language model unit adapted to the civil aviation field and a glyph feature extraction model unit, wherein the pre-trained language model unit is configured to perform semantic coding on the text to be processed to output a semantic vector, and the glyph feature extraction model unit is configured to perform glyph feature extraction on the text to be processed to output a glyph feature vector. The feature fusion module is configured to receive the semantic vector and the glyph feature vector, perform a splicing operation on the semantic vector and the glyph feature vector, and form an enhanced text representation that fuses semantic and glyph information. The time sequence modeling module comprises a time sequence modeling network, which is configured to receive the enhanced text representation, perform context modeling on the enhanced text representation through the time sequence modeling network, and output a context modeling text containing forward and backward logical relationships of the text. The attention weighting module comprises an attention mechanism, which is configured to receive the context modeling text, obtain a target word weight of each word unit and a target sentence weight of each sentence unit in the context modeling text through the attention mechanism, and combine the target word weight to weight the word units and the sentence units in the context modeling text to output a weighted word representation and a weighted sentence representation. The error detection and correction module is configured 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 a corrected text that conforms to the civil aviation field specifications. The domain rule verification module is configured to receive the corrected text, perform compliance verification on the corrected text, and output a final target text.

[0101] Further, the time sequence modeling network in the time sequence modeling module is a bidirectional long short-term memory network (Bi-LSTM) with a hidden dimension of 256, and the Bi-LSTM includes a forward LSTM unit for forward encoding of the enhanced text representation in a text order and a backward LSTM unit for reverse encoding of the enhanced text representation in a reverse text order, and an output of the Bi-LSTM is a context modeling text obtained by fusing a forward encoding result of the forward LSTM unit and a reverse encoding result of the backward LSTM unit.

[0102] Further, 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 configured to generate word initial weights and sentence initial weights through a nonlinear transformation component and an activation function component based on semantic relevance of the context modeling text, the activation function component includes a softmax function subcomponent for generating the word initial weights and a sigmoid function subcomponent for generating the sentence initial weights; and the weight correction unit is internally provided with a rule storage subunit based on a core field weight addition rule predefined by a civil aviation system specification document and an editable structured data format rule storage subunit, and is configured to dynamically adjust the word initial weights and the sentence initial weights through the rule storage subunits to obtain the target word weights and the target sentence weights.

[0103] Further, the error detection and correction module includes an error detection unit and an error correction unit; the error detection unit is internally provided with a BIO tagging system and an error detection classifier, the error detection classifier is a 3-layer fully connected network, and is configured to perform an error detection operation on the weighted word representation and the weighted sentence representation to output an error position label based on the BIO tagging system, the error type indicated by the error position label at least including semantic errors, syntax errors and format errors; and the error correction unit is internally provided with a sequence-to-sequence (Seq2Seq) model, the Seq2Seq model adopts an architecture of “bidirectional GRU encoder-single-direction GRU decoder+domain attention mechanism”, and is configured to perform a correction operation on error content based on the error position label to output the corrected text.

[0104] Further, 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 rule hot loading submodule and a conflict resolution submodule, each of the submodules is a software logic component.

[0105] The domain knowledge graph submodule is configured to store the mapping relationship between entities and specification values in the form of "entity-attribute-specification value" triples in the civil aviation domain, and perform entity association compliance checking on the entity type fields in the corrected text. The regular expression engine submodule is configured to have a civil aviation exclusive format regular rule library, which at least includes time type field rules, coordinate type field rules, height type field rules, and facility code type field rules, and is configured to perform format compliance checking on the format type fields in the corrected text. The knowledge graph incremental update interface submodule is a software interface supporting a preset network interface protocol, and is configured to add or modify the entity mapping relationship in the domain knowledge graph submodule. The regular rule hot loading submodule is configured to update the rule library of the regular expression engine submodule through an editable configuration file, and the updating process does not need to interrupt the system operation. The conflict resolution submodule is configured to process the conflicts between the processing results of the error detection and correction module and the checking results of the domain rule checking module, determine the final checking result according to a preset priority strategy, and the checking result of the domain knowledge graph submodule has a higher priority than other results in the preset priority strategy.

[0106] The embodiment of the present application also provides an electronic device, comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method of the embodiment of the present application.

[0107] The embodiment of the present application also provides a computer readable storage medium storing computer executable instructions, and the computer executable instructions are used to execute the method of the embodiment of the present application.

[0108] It should be understood that the steps shown above can be reordered, added or deleted. For example, the steps described in the present application can be executed in parallel, in sequence or in different orders, as long as the desired results of the technical solutions disclosed in the present application can be achieved, which are not limited herein.

[0109] The above specific embodiments do not constitute a limitation on the protection scope of the present application. 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 replacements and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

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 form feature vector of the to-be-processed content carrier; S300, generating an enhanced content carrier representation based on the semantic vector and the character form feature vector; S400, performing a time sequence feature extraction operation on the enhanced content carrier representation to acquire corresponding time sequence features; S500, acquiring target weights of each word and each sentence in the time sequence features, weighting the words and sentences in the time sequence features in combination with the target weights to obtain weighted word representations and weighted sentence representations; S600, performing a correction process on the weighted word representations and the weighted sentence representations to generate a corrected content carrier; S700, verifying the corrected content carrier to obtain a target content carrier.

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

3. The method of claim 1, wherein, In S500, the target weights of each word and each sentence in the time sequence features are acquired by using an attention mechanism, and specifically comprising: S501, 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; S502, adjusting the word initial weights and the sentence initial weights by a rule storage component which can be dynamically configured based on a preset field rule to obtain the target weights of each word and each sentence in the time sequence features.

4. The method of claim 3, wherein, S501 specifically comprises: For word units in the time sequence features, corresponding hidden state vectors of the word units are acquired, the hidden state vectors are input into a word-level nonlinear transformation component for feature mapping, then input into a first activation function component to output word initial weights of the word units; For sentence units in the time sequence features, word hidden state vectors contained in each sentence unit are subjected to a pooling operation to obtain a sentence vector, the sentence vector is input into a sentence-level nonlinear transformation component for feature mapping, then input into a second activation function component to output sentence initial weights of the sentence units.

5. The method of claim 4, wherein, The preset field rule is a core field weight adjustment rule formulated according to field normative documents, the core fields at least include time type fields, coordinate type fields, height type fields and facility state type fields which represent key attributes of the field, and different types of core fields correspond to preset weight addition ratios.

6. The method of claim 5, wherein, The rule storage component adopts an editable structured data format, supports adding, deleting and modifying the type division of the core fields and the corresponding weight addition ratios, and can complete rule iteration without adjusting the main architecture of the attention mechanism.

7. The method of claim 5, wherein, In the weight adjustment process, the core fields contained in the time sequence features are identified by a field matching component to determine the word units and the sentence units whose weights need to be adjusted, then the corresponding word initial weights and the corresponding sentence initial weights are corrected based on the preset addition ratios in the rule storage component, and the weight values after correction are subjected to boundary constraint processing to obtain corresponding target weights.

8. The method of claim 1, wherein, S600 specifically comprises: S601, performing an error detection operation on the weighted word representation and the weighted sentence representation, and outputting an error position label based on a BIO tagging system; S602, performing a correction operation on error content indicated by the error position label by using a sequence-to-sequence model, and generating a corresponding correction content carrier.

9. An electronic device, comprising: comprising a processor and a memory; the processor is configured to execute the steps of the method according to any one of claims 1 to 8 by invoking programs or instructions stored in the memory.

10. A computer-readable storage medium, characterized in that, the computer readable storage medium is configured to store programs or instructions, which cause the computer to execute the steps of the method according to any one of claims 1 to 8.

Images