An intelligent semantic understanding method for land-air calls based on a multi-task hierarchical network

Abstract

The present invention relates to the technical field of air traffic control, and in particular to a method for intelligent semantic understanding of land-to-air conversations based on a multi-task hierarchical network. The method comprises the following steps: encoding land-to-air conversation corpus data to obtain word-level representation, sentence-level representation, and word-level mask representation; performing knowledge distillation processing to obtain slot representation; performing role-attention processing and intent-attention processing on the sentence-level representation, respectively, and then performing role-intent fusion to obtain role representation and intent representation; performing semantic hierarchical bidirectional enhanced interactive processing on the slot representation, role representation, and intent representation to obtain word-level features and sentence-level features; and using a multi-task decoder to decode the word-level features and sentence-level features to obtain slots, predicted roles, and predicted intentions predicted for the land-to-air conversation corpus data. The present invention can achieve comprehensive and detailed intelligent semantic understanding of land-to-air conversations.

Description

Technical Field

[0001] The present invention relates to the technical field of air traffic control, and in particular to an intelligent semantic understanding method for air-land calls based on a multi-task hierarchical network. Background Art

[0002] Air traffic control (ATC) relies heavily on air-to-ground communications between controllers and pilots. Controllers relay precise situational awareness instructions to pilots via radio, who then read back these instructions and operate their aircraft accordingly. However, the rapid growth of air traffic has significantly increased the workload of both controllers and pilots, increasing the risk of accidents and incidents caused by miscommunication between controllers and pilots. With the widespread adoption of artificial intelligence (AI), developing intelligent systems that can understand controller-pilot conversations is crucial to assisting operators and improving flight safety and efficiency.

[0003] To address this urgent need, speech in air-to-ground conversations is first converted to text. The main purpose and key information in the converted text are then extracted and identified, achieving semantic understanding. Traditionally, semantic understanding relies on regular expressions and predefined rules to match structured keywords with text instructions. However, hand-crafted templates created by ATC experts are difficult to adapt to ever-changing application scenarios. Consequently, pre-trained language models (PLMs) have attracted attention from both academia and industry. Various semantic understanding tasks have been explored at different granularities, ranging from word-level named entity recognition and slot filling to sentence-level sentence meaning understanding and intent recognition. Furthermore, role detection, also at the sentence level, clarifies the distinct operational responsibilities of air traffic controllers and pilots, ensuring accurate communication and mission execution. While these studies have demonstrated performance improvements, the underlying relationships between different granularities, particularly regarding the comprehensive utilization of role information, have yet to be fully revealed. Accurately interpreting the meaning of specialized terms and phrases in utterances is crucial, while discerning the speaker's role and intent in different contexts is also a crucial component of higher-level semantic understanding. However, text representations at different semantic levels—words, syntax, sentences, and concepts—in land-air conversations offer insights from diverse dimensions, and effectively integrating them can enhance understanding. This includes word-level slot filling tasks, as well as sentence-level role detection and intent recognition tasks. These tasks are inherently heterogeneous. Slot filling involves mapping each word or phrase in a sentence to a specific semantic label, thus residing at the word level; role detection distinguishes different speakers by identifying the word order structure of a sentence, thus residing at the syntactic level; and intent recognition captures higher-level semantic intent by analyzing the entire sentence and its key actions to determine the purpose or meaning behind the utterance, thus residing at the sentence and concept levels. Therefore, how to mine and integrate the relationships between tasks at different granularities is a key challenge. Summary of the Invention

[0004] In view of the above problems, the present invention provides an intelligent semantic understanding method for land-to-air calls based on a multi-task hierarchical network, which solves the technical problem that the existing technology fails to mine and integrate tasks of different granularities in the semantic recognition task of land-to-air calls.

[0005] The present invention provides a method for intelligent semantic understanding of land-air conversations based on a multi-task hierarchical network. The multi-task hierarchical network processes land-air conversation corpus data including the following steps:

[0006] Step S1: Encode the land-air communication corpus data at the word level and sentence level to obtain word level representation and sentence level representation;

[0007] Step S2: performing word-level mask encoding on the land-air communication corpus data to obtain a word-level mask representation, and performing knowledge distillation based on the word-level representation and the word-level mask representation to obtain a slot representation;

[0008] Step S3: performing role attention processing and intention attention processing on the sentence-level representation, and then performing role-intention fusion to obtain role representation and intention representation;

[0009] Step S4: performing semantic-level bidirectional enhancement interactive processing on the slot representation, role representation, and intention representation to obtain word-level features and sentence-level features;

[0010] Step S5: Use a multi-task decoder to decode the word-level features and sentence-level features to obtain the predicted slot, predicted role, and predicted intention of the land-air communication corpus data.

[0011] Preferably, the step S1 specifically includes:

[0012] Step S1-1: Encode the land-air communication corpus data through a shared encoder to obtain a shared representation; the shared encoder is a BERT model;

[0013] Step S1-2: Encode the shared representation through a feature-specific encoder to obtain a word-level representation and a sentence-level representation; the feature-specific encoder is a bidirectional LSTM network.

[0014] Preferably, the step S1-1 specifically includes: forming an input sentence word sequence x from the text of the land-air conversation content, where x={x1, x2, ..., x n}, where x n is the nth word of the input sentence; the BERT model is used to generate a shared representation, denoted as E = {e [CLS] , e1, e2, ..., en , e [SEP]}, where e n is the shared representation corresponding to the nth word in x, e [CLS] and e [SEP] are respectively sentence boundaries; the step S1-2 specifically comprises: inputting the shared representation E into the first bidirectional LSTM network and the second bidirectional LSTM network respectively, and obtaining the word level representation H w and sentence-level representation H u .

[0015] Preferably, the step S2 specifically includes:

[0016] Step S2-1: Establish an external knowledge base based on standard terminology for air traffic control scenarios; establish an internal knowledge base based on historical land-air communication corpus data; and combine the internal knowledge base and the external knowledge base to form a candidate mask word set;

[0017] Step S2-2: After processing the land-air communication corpus data using the masked language model and the shared encoder, the data is input into a masked feature-specific encoder for encoding to obtain a word-level masked representation;

[0018] Step S2-3: perform knowledge distillation on the word-level representation and the word-level mask representation based on the teacher model and the student model to obtain a slot representation.

[0019] Preferably, in step S2-1, the specific steps of establishing the internal knowledge base based on the historical land and air communication corpus data include: using TF-IDF word frequency processing technology to divide the text in the historical land and air communication corpus data into words, and sorting the divided words according to the frequency of word occurrence, and selecting a portion of the words with the highest ranking from the sorted words as the internal knowledge base;

[0020] Step S2-2 specifically includes: using a masked language model to process the land-air communication corpus data to obtain a masked word sequence, x′={x1,[MASK],...,x n}, [MASK] represents the mask mark of the replacement, and the shared encoder is used to process x′ to obtain E′, E′={e′ [CLS] ,e′1,e′ [MASK］ ,...,e′ n , e′ [SEP]}, where e′ n is the shared representation corresponding to the nth word in x′, e′ [MASK] is the shared representation corresponding to the mask mark, e′ [CLS] and e′ [SEP]are mask sentence boundaries respectively; E′ is calculated using the third bidirectional LSTM network to obtain the word-level mask representation H m .

[0021] Step S2-3 specifically includes: the teacher model converts H w With H m Splice and process to generate teacher representation F t The student model is H w Processing is performed to generate slot representation F s , the expression is:

[0022]

[0023] F s =Student(H w )

[0024] in, represents the concatenation operation, Teacher(·) represents the teacher model, and Student(·) represents the student model.

[0025] Preferably, the step S3 specifically includes:

[0026] Step S3-1: Calculate based on the sentence-level representation to obtain a role attention weight and an intention attention weight; obtain a role context and an intention context based on the role attention weight and the intention attention weight, respectively;

[0027] Step S3-2: Use the role-intent fusion layer to project the role context into the intent latent space to obtain a fused projection, concatenate the intent context and the fused projection to obtain a role representation; map the role representation to the intent label space to obtain an intent representation.

[0028] Preferably, the step S3-1 specifically includes: calculating the character attention weight A by the following formula: r and the intention attention weight A i :

[0029] A r =softmax(W r H u +b r )

[0030] A i =softmax(W i H u +b i )

[0031] Among them, W r and b rRepresent the weight and bias of the character's attention respectively, and the softmax function normalizes the calculation result to the character's attention weight A r , W i and b i Represent the weight and bias of the intention attention respectively;

[0032] By adding H u With transposed attention weights and Multiply them together to get the role context C r and intent context C i :

[0033]

[0034]

[0035] in, and Denote the generated weighted role and intention representations, V r and V i are the dimensional conversion matrices of roles and intentions respectively;

[0036] The step S3-2 specifically includes: using the role-intention fusion layer to calculate and obtain the fusion projection C f : The expression is:

[0037] C f =fusion(C r )

[0038] Where fusion represents the role-intention fusion layer;

[0039] Character representation F r and intention expression F i The expression is:

[0040]

[0041] F i =W u F r +b u

[0042] Among them, W u and b u Represent the weight and bias of the sentence respectively.

[0043] Preferably, the step S4 specifically includes:

[0044] The slot representation, role representation and intention representation are sequentially determined as the semantic level sequence of the semantic level bidirectional enhanced interaction processing, that is, S1=Fs , S2=F r , S3=F i , S l is the input representation of the lth layer;

[0045] The interaction score is calculated within the hierarchy using a bidirectional bilinear transformation, expressed as:

[0046]

[0047]

[0048] in, is the reverse interaction score of layer l, is the forward interaction score of the lth layer, Score(·,·) represents the interaction score between the calculation layers, is the transpose of the input representation of layer l, is the reverse score matrix, is the forward score matrix;

[0049] Calculate the backward and forward pass vectors for layer l and The expression is:

[0050]

[0051]

[0052] in, and For the backward and forward transfer weights, and Bias is transferred for reverse and forward directions;

[0053] By combining and and and Recursively aggregate the calculations to obtain word-level features and sentence-level features, expressed as:

[0054]

[0055]

[0056] in, represents the aggregated word-level features from layer L to layer l, where L represents the total number of layers. represents the aggregated sentence-level features from level 1 to level l;

[0057] Ultimately, and They are respectively used as the word-level features and sentence-level features.

[0058] Preferably, the step S5 specifically includes:

[0059] Step S5-1: Input to the slot classifier, and then pass through a CRF layer to obtain the predicted slot;

[0060] Step S5-2: With F r Concatenate and input into the role classifier to obtain the predicted role;

[0061] Step S5-3, F i Input into the intent quantity classifier to get the predicted number of intents. Input to the intent classifier to obtain the intent probability, and from the predicted number of intents, select the top c with the highest probability * The intention is used as the predicted intention.

[0062] Preferably, by training the multi-task hierarchical network, with the goal of minimizing the joint target training loss, back propagation is performed to update the model parameters, and after obtaining the optimal parameters of the model, the trained multi-task hierarchical network is used for intelligent semantic understanding of land-air calls; the joint target training loss of the multi-task hierarchical network is minimized. The expression is:

[0063]

[0064] in, is the slot filling loss, Detect loss for the character, To identify the loss of intention, Slot distillation loss, λ, μ and η are loss weights;

[0065]

[0066] in, is the true label of the k-th slot of the i-th word, is the predicted slot probability of the kth slot of the i-th word, n is the total number of words, and K is the total number of slot labels;

[0067]

[0068] in, and are the true binary labels for the controller and pilot roles, respectively; and are the predicted probabilities for the controller and pilot roles, respectively;

[0069]

[0070] in, is the Kronecker delta function, Indicates the probability of predicting the number of intentions to be c. When the number of real intentions c is matched is 1, otherwise is 0, represents the binary true label of the mth intent, is the predicted probability of the mth intention, M is the total number of intentions;

[0071]

[0072]

[0073] in, is the predicted probability distribution of the i-th word, m i Indicates whether the i-th word is masked, z i is the unique hot representation of the i-th word;

[0074]

[0075] in, and are the probability distributions of teacher representation and student representation, respectively. express and The KL divergence of is a finite set space, χ is The elements in the collection, They represent the target probability and approximate probability corresponding to x respectively.

[0076] Compared with the prior art, the present invention has at least the following beneficial effects:

[0077] (1) The present invention generates features of the vocabulary in the sentence by inputting the ground-air conversation text into a corpus encoder composed of a shared encoder and a feature-specific Bi-LSTM encoder. It can simultaneously perform slot filling, role detection and intent recognition, achieve a comprehensive understanding of the conversation between the controller and the pilot, and improve the efficiency and accuracy of the overall semantic understanding.

[0078] (2) This paper introduces word-level knowledge mask slot distillation, utilizes knowledge-guided MLM, guides slot representation through teacher-student distillation, builds a dedicated ATC knowledge base, focuses on keywords and enhances the generalization ability of the model, thereby improving the model's ability to recognize key terms and enhancing its adaptability to different scenarios.

[0079] (3) The present invention adopts a sentence-level role-perceived intention attention mechanism to extract label space features and integrate roles into the intention space. Subsequently, through bidirectional enhanced interaction at the semantic level, information is propagated and aggregated between multiple semantic understanding tasks, which can effectively distinguish and integrate semantic levels of different granularities, and improve the recognition accuracy of slot filling, role detection, and intent recognition. In addition, due to the use of a recursive formula structure, the framework has good scalability. The current model supports four levels of semantic hierarchy, corresponding to three tasks. On this basis, more levels and tasks can be added without modifying the existing module framework. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The drawings are only for purposes of illustrating particular embodiments and are not to be considered limiting of the invention.

[0081] Figure 1 This is the overall framework diagram of the intelligent semantic understanding method for land-to-air calls based on a multi-task hierarchical network provided by the present invention.

[0082] Figure 2 Schematic diagram of the knowledge shielding slot distillation construction process provided by the present invention.

[0083] Figure 3 Schematic diagram of the character-aware intention attention mechanism provided by the present invention.

[0084] Figure 4 This is a schematic diagram of the semantic-level bidirectional enhanced interaction provided by the present invention.

[0085] Figure 5 This is a flow chart of the intelligent semantic understanding method for land-to-air calls based on a multi-task hierarchical network provided by the present invention. DETAILED DESCRIPTION

[0086] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0087] This paper proposes an intelligent semantic understanding method for air-to-land calls based on a multi-task hierarchical network, introducing word-level knowledge mask slot distillation to improve word-level sensitivity and generalization. This goal is achieved by using a masked language model and knowledge distillation guided by ATC knowledge. It also involves a sentence-level role-aware intent attention mechanism, using two attention mechanisms to extract and explicitly fuse role and intent features. The attention mechanism minimizes semantic ambiguity by mapping the same corpus to the role and intent label space, and the fusion layer uses role information as context to enrich the intent representation. Finally, the semantic level is bidirectionally enhanced to capture the relationship between the hierarchical semantic representations. Specifically, the bidirectional interaction between the layers is learned to recursively generate the final multi-task representation.

[0088] In order to illustrate the effectiveness of the method proposed by the present invention, the above technical solution of the present invention is described in detail below through a specific embodiment. Figure 1 、 Figure 5 As shown, a method for intelligent semantic understanding of land-air calls based on a multi-task hierarchical network is disclosed. The specific implementation steps are as follows:

[0089] Step S1: Encode the land-air communication corpus data at the word level and sentence level to obtain word level representation and sentence level representation;

[0090] The ground-to-air conversation data in this embodiment of the present invention is derived from ground-to-air conversations between air traffic controllers and pilots during air traffic control. The ground-to-air conversations include both the controller's and the pilot's speeches. The original audio of the controller's and pilot's speeches is transcribed to generate multiple segments of text representing the ground-to-air conversations, which serve as the ground-to-air conversation data.

[0091] Regarding the encoding process of the land-air communication corpus data, the present invention encodes the corpus data successively through a shared encoder and a feature-specific encoder to obtain word-level representation and sentence-level representation. The specific process is as follows:

[0092] (1) Shared encoder

[0093] The shared encoder of the present invention adopts the BERT model, which is a widely accepted pre-trained language model (PLM). The BERT model is obtained by pre-training on a large-scale corpus. The sentences in the text of the land-air call content form an input sentence word sequence x, where x = {x1, x2, ..., x n}, where x n is the nth word of the input sentence. Using the knowledge of the BERT model pre-trained on a large-scale corpus, the generated shared representation is denoted as E = {e [CLS] , e1, e2, ..., e n , e[SEP]}, where e n is the shared representation corresponding to the nth word in x, e [CLS] and e [SEP] Statement boundaries added at the beginning and end of a statement respectively.

[0094] (2) Feature-specific encoder

[0095] The feature-specific encoder of the present invention is a bidirectional LSTM (Bi-LSTM) network, which inputs the shared representation E into two bidirectional LSTM (Bi-LSTM) networks of specific features to obtain word-level representations H w and sentence-level representation H u .

[0096] H w =Bi-LSTM w (E) (1)

[0097] H u =Bi-LSTM u (E) (2)

[0098] Among them, Bi-LSTM w (·) is a word-level bidirectional LSTM network, Bi-LSTM u (·) is a sentence-level bidirectional LSTM network.

[0099] Step S2: performing word-level mask encoding on the land-air communication corpus data to obtain a word-level mask representation, and performing knowledge distillation based on the word-level representation and the word-level mask representation to obtain a slot representation;

[0100] In air traffic control scenarios, accurate keyword recognition is crucial to ensuring the effectiveness and clarity of air-to-ground communication, which is called the slot filling task. Specifically, when processing an air-to-ground communication instruction from a pilot or controller, the slot filling task requires assigning a slot label to each word in the sentence. Therefore, the present invention provides a word-level knowledge mask slot distillation processing method, such as Figure 2 shown.

[0101] First, for air-land communication corpus data, the present invention provides a knowledge-guided masked language model (MLM) to mask specific keywords and predict their content, thereby enhancing the model's sensitivity to these keywords while reducing interference from irrelevant words. Specifically, the present invention uses TF-IDF word frequency processing technology to divide the text in the historical air-land communication corpus data into words, sorts the divided words according to their frequency of occurrence, and selects the highest-ranked words from the sorted words as the internal knowledge base. In addition, based on standard terminology for air traffic control scenarios, an external knowledge base is established, which includes letters, numbers, and air traffic control term abbreviations.

[0102] The internal and external knowledge together constitute a candidate mask word set. The mask language model is used to process the original land-air communication corpus data to generate an input sentence mask word sequence x′ containing mask words and original words, x′={x1, [MASK], ..., x n}, where some original words are replaced with [MASK] tags, [MASK] represents the replaced mask tags, and the shared representation corresponding to x′ is recorded as E′, E′={e′ [CLS] ,e′1,e′ [MASK] ,...,e′ n , e′ [SEP]}, where e′ n is the shared representation corresponding to the nth word in x′, e′ [MASK] is the shared representation corresponding to the mask mark, e′ [CLS] and e′ [SEP] These are the masked sentence boundaries added at the beginning and end of the masked sentence, respectively.

[0103] Use another bidirectional LSTM network to calculate E′ and obtain the word-level mask representation H m , as shown below:

[0104] H m =Bi-LSTM m (E′) (3)

[0105] Among them, Bi-LSTM m (·) is a bidirectional LSTM network with word-level masking.

[0106] The loss function of the knowledge-guided masked language model The expression is:

[0107]

[0108] in, is the predicted probability distribution of the i-th word, mi Indicates whether the i-th word is masked, z i is the unique hot representation of the i-th word. This is done by applying weights w and bias b to the encoder features e′ i And the predicted probability distribution of the i-th word is obtained after normalization by the softmax function. If the word is masked, then m i =1, otherwise 0. Loss function of masked language model is the true distribution z at the mask position i and the predicted distribution By minimizing this loss function, the model is trained to accurately predict the masked words based on the context.

[0109] After obtaining the word-level mask representation, the present invention provides a knowledge distillation method for subsequent processing. Specifically, the knowledge distillation method is based on a teacher model and a student model. The teacher model converts H w With H m Splice and process to generate teacher representation F t The student model is H w Processing is performed to generate slot representation F s The expression is:

[0110]

[0111] F s =Student(H w ) (6)

[0112] in, In order to align the representation and promote the effective distillation from the teacher model to the student model, F t and F s Unify the dimensions. In addition, use softmax to F t and F s Perform normalization to obtain its probability distribution and

[0113] The goal of the knowledge distillation method based on the teacher model and the student model is to minimize the target distribution and approximate distribution The Kullback-Leibler (KL) divergence between the two, the objective function of the knowledge distillation method The expression is:

[0114]

[0115] in, express and KL divergence, χ is a finite set space, χ is an element in the χ set, They represent the target probability and approximate probability corresponding to χ respectively.

[0116] Finally, the overall tank distillation loss in this step The expression is:

[0117]

[0118] Step S3: Perform role attention processing and intention attention processing on the sentence-level representation, and then perform role-intention fusion to obtain role representation and intention representation.

[0119] In air-to-ground conversations, the speeches of air traffic controllers and pilots usually have different semantic functions. Air traffic controllers mainly issue operational instructions, while pilots confirm or respond to these instructions. Therefore, the role detection task is an important task in the intelligent semantic understanding task of air-to-ground conversations, which is used to identify and distinguish the identities of different speakers (air traffic controllers, pilots). The intention recognition task analyzes the dialogue input of pilots and air traffic controllers to determine the operations they expect to perform or the important information they want to convey. In addition, different roles show different tendencies towards certain intentions. For example, air traffic controllers often issue instruction intentions such as "altitude intentions", while pilots are more inclined to "request intentions" and "status report intentions".

[0120] The present invention provides a sentence-level role-aware intention attention mechanism that integrates the contextual relevance of the role with its corresponding intention, and can simultaneously complete the role detection task and the intention recognition task, such as Figure 3 By leveraging role information, we aim to achieve more accurate intent prediction.

[0121] Specifically, the present invention utilizes two independent attention mechanisms to represent H from the same sentence level. u The role representation and intention representation are extracted separately in , thus ensuring a clear distinction between them. For the role attention processing, the role attention weight A r The calculation expression is:

[0122] A r =softmax(W r H u +b r ) (9)

[0123] Where W r and b rRepresent the weight and bias of the character's attention respectively, and the softmax function normalizes the calculation result to the character's attention weight A r .

[0124] Similarly, the intention attention weight A i The calculation expression is:

[0125] A i =softmax(W i H u +b i ) (10)

[0126] Where W i and b i denote the weight and bias of the intention attention respectively.

[0127] Then, by adding H u With transposed attention weights and Multiply them together to get the role context C r and intent context C i :

[0128]

[0129]

[0130] in, and Denote the generated weighted role and intention representations, V r and V i The dimension conversion matrices are role and intent respectively.

[0131] In order to effectively fuse the role and intent information, a role-intent fusion layer is used to transform the role context C r Projected into the intention latent space, the fusion projection C is obtained f , the expression of the role-intention fusion layer calculation process is:

[0132] C f =fusion(C r ) (13)

[0133] Where fusion represents the role-intention fusion layer. The role-intention fusion layer converts C r Projection to C f , and activate each role through the sigmoid function.

[0134] Finally, to ensure that the role and intent representations are tightly coupled and better adapt to the sentence-based feature distribution, we define It represents the role.i It is obtained by the following expression:

[0135] F i =W u F r +b u (14)

[0136] Among them, W u and b u Represent the weight and bias of the sentence respectively. Through such calculation, F r is mapped to the intent label space to obtain F i .

[0137] Step S4: Perform semantic-level bidirectional enhancement interactive processing on the slot representation, role representation, and intention representation to obtain word-level and sentence-level features.

[0138] Figure 4 The process of semantic level bidirectional enhancement interaction provided by the present invention is shown. The input includes representations S1, S2, S3 at multiple different semantic levels, and the final output is the aggregated word-level features. and sentence-level features

[0139] The present invention provides a semantic hierarchical bidirectional enhanced interaction design, which explores the semantic relationship between multiple tasks by using bidirectional interaction scores. Through the bidirectional propagation layer, the present invention combines the bidirectional interaction in the hierarchical semantics, thereby ultimately forming aggregated word-level and sentence-level features. Semantic hierarchy sequence S = (S1, ..., S L ) as input, where L represents the total number of levels. In the practical application of the present invention, the semantic level sequence is defined as: S = (F s , F r , F i ), corresponding to slot, role, and intent representation respectively. This semantic hierarchy sequence provides a detailed and comprehensive semantic representation.

[0140] The hierarchical semantics of S have bidirectional interdependencies. The information flow from fine-grained semantic details to the overall conceptual meaning is different from the information flow from the overall conceptual meaning to fine-grained semantic details. Therefore, the present invention calculates the interaction score within the hierarchy by using a bidirectional bilinear transformation, which is expressed as:

[0141]

[0142] in, is the reverse interaction score of layer l, is the forward interaction score of the lth layer, Score(·,·) represents the interaction score between the calculation layers, S l is the input representation of the lth layer, is the transpose of the input representation of layer l, is the reverse score matrix, is the forward score matrix.

[0143] The present invention then applies backward and forward semantic transfers at each level to compute the backward and forward transfer vectors for layer l and The expression is:

[0144]

[0145] in, and For the backward and forward transfer weights, and Bias is transferred for both reverse and forward directions.

[0146] Next, by combining the transfer vector and Interaction score and Recursively calculate the aggregated word-level and sentence-level features. The aggregated word-level features are obtained as follows:

[0147]

[0148] in, represents the aggregated word-level features from layer L to layer l, where L represents the total number of layers.

[0149] Similarly, the aggregated sentence-level features are obtained as follows:

[0150]

[0151] in, represents the aggregated sentence-level features from level 1 to level l.

[0152] final, and As the final aggregated word-level and sentence-level features respectively. In the practical application of the present invention, L=3, that is, using and Specifically, we gradually shift from intent features to role and slot features, and ultimately form word-level features that integrate sentence features. Starting from the slot features, role and intent features are gradually integrated, and finally sentence-level features are formed that integrate word features.

[0153] Step S5: Use a multi-task decoder to decode the word-level and sentence-level features to obtain the predicted slot, predicted role, and predicted intent.

[0154] The multi-task decoder includes decoders for the following three tasks:

[0155] (1) Slot decoder:

[0156] is fed into a slot classifier and then passed through a Conditional Random Field (CRF) layer to obtain the predicted slot probability The CRF layer determines the most likely sequence of slot labels by considering the probability of a single slot and the transition between consecutive slots. The loss function for the slot filling loss is the cross entropy loss

[0157]

[0158] in, is the true label of the k-th slot of the i-th word, is the predicted slot probability of the kth slot of the uth word, n is the total number of words, and K is the total number of slot labels.

[0159] (2) Role decoder:

[0160] The present invention will With F r Concatenate and input into the role classifier to obtain the predicted role probability The loss function of character detection loss is defined as follows:

[0161]

[0162] in, and are the true binary labels of the controller and pilot roles respectively. The predicted probabilities of the controller and pilot roles are expressed as and And their relationship is

[0163] (3) Intent decoder:

[0164] The present invention will be intended to represent F i Input to the intent number classifier, which normalizes the number of each possible intent and outputs the predicted number of intents in the sentence where C represents the maximum number of intentions considered.

[0165] The present invention will Input to the intent classifier to obtain the intent probability. Select the top c with the highest probability * multi-label intent set

[0166] Intention recognition loss This includes binary cross entropy loss for multi-label intent recognition and cross entropy loss for determining the number of intents:

[0167]

[0168] in, is the Kronecker delta function, when When the number of real intentions c is matched is 1, otherwise is 0, represents the probability that the number of intentions is predicted to be c, represents the binary true label of the mth intent, is the predicted probability of the mth intent, and M is the total number of intents.

[0169] Joint objective training loss of the model of the present invention Filling loss for slots Character Detection Loss Intention recognition loss and tank distillation losses The weighted sum of , where λ, μ, and η are loss weights:

[0170]

[0171] By learning the loss function during training and then backpropagating the model parameters, the model's optimal parameters are obtained, effectively representing the aforementioned multiple tasks. After training, the model serves as a language model for air traffic control, enabling comprehensive and in-depth intelligent semantic understanding of ground-to-air conversations.

[0172] The present invention discloses a method for intelligent semantic understanding of air-to-ground conversations based on a multi-task hierarchical network. This method addresses the problem of frequent air-to-ground conversations in air traffic control, which places heavy workloads on controllers and pilots, leading to communication misunderstandings and potentially causing accident symptoms or even accidents. The method proposes a novel end-to-end multi-task hierarchical network. This network intelligently achieves comprehensive and accurate semantic understanding through a multi-task framework, integrating slot filling, role detection, and intent recognition tasks. This method integrates comprehensive semantic understanding of words, syntax, sentences, and concepts at different levels of air-to-ground conversations, ensuring clear and consistent communication between controllers and pilots, reducing misunderstandings and improving safety and operational efficiency. Compared to existing technologies, the present invention effectively handles multiple tasks to achieve comprehensive and detailed intelligent semantic understanding of air-to-ground conversations, ensuring flight safety and improving flight efficiency.

[0173] Although the specific embodiments of the present invention depict various actions or steps in a specific order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in a sequential order, or requiring that all illustrated actions or steps should be performed to obtain the desired results. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although a number of specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of this disclosure. Certain features described in the context of separate embodiments can also be implemented in combination in a single implementation. Conversely, the various features described in the context of a single implementation can also be implemented in multiple implementations individually or in any suitable sub-combination. The above is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered within the scope of protection of the present invention.

[0174] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent semantic understanding of land-air calls based on a multi-task hierarchical network, characterized by: The processing of the land-air communication corpus data by the multi-task hierarchical network comprises the following steps: Step S1: Encode the land-air communication corpus data at the word level and sentence level to obtain word level representation and sentence level representation; Step S2: performing word-level mask encoding on the land-air communication corpus data to obtain a word-level mask representation, and performing knowledge distillation based on the word-level representation and the word-level mask representation to obtain a slot representation; Step S3: performing role attention processing and intention attention processing on the sentence-level representation, and then performing role-intention fusion to obtain role representation and intention representation; Step S4: performing semantic-level bidirectional enhancement interactive processing on the slot representation, role representation, and intention representation to obtain word-level features and sentence-level features; Step S5: using a multi-task decoder to decode the word-level features and sentence-level features to obtain a predicted slot, a predicted role, and a predicted intent for the land-air communication corpus data; The step S2 specifically includes: Step S2-1: Establish an external knowledge base based on standard terminology for air traffic control scenarios; establish an internal knowledge base based on historical land-air communication corpus data; and combine the internal knowledge base and the external knowledge base to form a candidate mask word set; Step S2-2: After processing the land-air communication corpus data using a masked language model and a shared encoder, the data is input into a masked feature-specific encoder for encoding to obtain a word-level masked representation; Step S2-3: performing knowledge distillation on the word-level representation and the word-level mask representation based on the teacher model and the student model to obtain a slot representation; In step S2-1, the specific steps of establishing the internal knowledge base based on the historical land and air communication corpus data include: using TF-IDF word frequency processing technology to divide the text in the historical land and air communication corpus data into words, and sorting the divided words according to the frequency of word occurrence, and selecting the highest-ranked words from the sorted words as the internal knowledge base; Step S2-2 specifically includes: using a masked language model to process the land-air communication corpus data to obtain a masked word sequence, , Represents the replacement mask tag, using the shared encoder pair Process and obtain , ,in, for Middle n The shared representation corresponding to each word, is the shared representation corresponding to the mask mark, and are mask sentence boundaries respectively; use the third bidirectional LSTM network to Perform calculations to obtain word-level mask representation ; Step S2-3 specifically includes: the teacher model will and Splice and process to generate teacher representation , the student model Process and generate slot representation , the expression is: in, Represents a splicing operation, represents the teacher model, represents the student model; The step S3 specifically includes: Step S3-1: Calculate based on the sentence-level representation to obtain a role attention weight and an intention attention weight; obtain a role context and an intention context based on the role attention weight and the intention attention weight, respectively; Step S3-2: Project the role context into the intent latent space using the role-intent fusion layer to obtain a fused projection, concatenate the intent context and the fused projection to obtain a role representation; and map the role representation into the intent label space to obtain an intent representation. The step S3-1 specifically includes: calculating the character attention weight by the following formula: and intention attention weight : in, and Represent the weight and bias of the character's attention respectively, The function normalizes the calculation result to the character attention weight , and Represent the weight and bias of the intention attention respectively; By With transposed attention weights and Multiply them together to get the role context and intent context : in, and denote the generated weighted role and intention representations, respectively. and are the dimensional conversion matrices of roles and intentions respectively; The step S3-2 specifically includes: using the role-intention fusion layer to calculate and obtain the fusion projection : The expression is: in Represents the role-intention fusion layer; Character Representation and expression of intent The expression is: in, and Represent the weight and bias of the sentence respectively; The step S4 specifically includes: The slot representation, role representation and intention representation are sequentially determined as the semantic level sequence of the semantic level bidirectional enhanced interaction processing, that is, , For the The input representation of the layer; The interaction score is calculated within the hierarchy using a bidirectional bilinear transformation, expressed as: in, For the The reverse interaction score of the layer, For the The forward interaction score of the layer, represents the interaction score between calculation levels, For the the transpose of the layer’s input representation, is the reverse score matrix, is the forward score matrix; Calculate the Backward and forward pass vectors of the layer and , the expression is: in, and For the backward and forward transfer weights, and Bias is transferred for reverse and forward directions; By combining and and and , recursively aggregate and calculate to obtain word-level features and sentence-level features, the expression is: in, Indicates that from Layer to Aggregated word-level features of the layer, Indicates the total number of levels, From level 1 to level Aggregate sentence-level features at the level; Ultimately, and They are respectively used as the word-level features and sentence-level features.

2. The intelligent semantic understanding method for land-air calls based on a multi-task hierarchical network according to claim 1 is characterized in that: The step S1 specifically includes: Step S1-1: Encode the land-air communication corpus data through a shared encoder to obtain a shared representation; the shared encoder is a BERT model; Step S1-2: Encode the shared representation through a feature-specific encoder to obtain a word-level representation and a sentence-level representation; the feature-specific encoder is a bidirectional LSTM network.

3. The intelligent semantic understanding method for land-air calls based on a multi-task hierarchical network according to claim 2 is characterized by: The step S1-1 specifically includes: forming an input sentence word sequence from the text of the land-air call content , ,in, The first n words; use the BERT model to generate shared representations, denoted as ,in, for Middle n The shared representation corresponding to each word, and They are statement boundaries; The step S1-2 specifically includes: Input them into the first bidirectional LSTM network and the second bidirectional LSTM network respectively to obtain word level representation and sentence-level representation .

4. The intelligent semantic understanding method for land-air calls based on a multi-task hierarchical network according to claim 3 is characterized in that: The step S5 specifically includes: Step S5-1: Input to the slot classifier, and then pass through a CRF layer to obtain the predicted slot; Step S5-2: and Concatenate and input into the role classifier to obtain the predicted role; Step S5-3: Input into the intent quantity classifier to get the predicted number of intents. Input to the intent classifier to obtain the intent probability, and select the top prediction with the highest probability from the number of predictions of the intent. The intention is used as the predicted intention.

5. The intelligent semantic understanding method for land-air calls based on a multi-task hierarchical network according to claim 4 is characterized by: By training the multi-task hierarchical network, with the goal of minimizing the joint target training loss, back propagation is performed to update the model parameters, and after obtaining the optimal parameters of the model, the trained multi-task hierarchical network is used for intelligent semantic understanding of land-air calls; the joint target training loss of the multi-task hierarchical network is minimized. The expression is: in, is the slot filling loss, Detect loss for the character, To identify the loss of intention, Tank distillation loss, 、 and is the loss weight; in, It is The first word The true label of each slot, It is The first word The predicted slot probability for each slot, is the total number of words, is the total number of slot labels; in, and are the true binary labels for the controller and pilot roles, respectively; and are the predicted probabilities for the controller and pilot roles, respectively; in, yes function, The number of intentions predicted is The probability of and the number of true intentions When matching is 1, otherwise is 0, Indicates the The binary ground-truth labels of the intents, It is The predicted probability of an intention, is the total number of intentions; in, For the The predicted probability distribution of words, Indicates the Whether the word is masked, It is The unique expression of a word; in, and are the probability distributions of teacher representation and student representation, respectively. express and The KL divergence of For a finite set of spaces, for The elements in the collection, Respectively The corresponding target probability and approximate probability.

Images