A method for detecting the consistency of air traffic control instruction restatement based on intelligent completion

Abstract

This invention discloses a method for detecting consistency alarms in regulatory instruction paraphrasing based on intelligent completion. The method includes: acquiring the current dialogue; classifying and filtering the regulatory text in the current dialogue using an SVM model to obtain regulatory-paraphrased text; identifying regulatory text elements in the regulatory-paraphrased text to obtain the identified regulatory elements; completing the simplified and omitted information in the regulatory-paraphrased text; and using a matching-comparison mechanism to check the consistency of each element in the completed regulatory-paraphrased text. If the paraphrased text is inconsistent with the regulatory text, the probability of the inconsistent text information is calculated to facilitate confirmation of paraphrasing inconsistency alarms. This invention reduces misjudgments of paraphrasing consistency caused by simplification and omission, improving the effectiveness of the technology. While providing the paraphrasing consistency judgment result, it also provides the judgment criteria and probability assessment, adapting to the paraphrasing consistency detection needs of corresponding scenarios and improving reliability.

Description

Technical Field

[0001] This invention relates to the field of air traffic control automation technology, specifically to a method for detecting consistency alarms in control command rephrase based on intelligent completion. Background Technology

[0002] Air traffic control is a highly demanding mental activity primarily based on subjective decision-making by air traffic controllers. Lacking specific, precise decision-making models and procedures, it places extremely high demands on the reliability of controllers. Compared to machines, controllers can only handle lower workloads, shorter durations, slower reaction times, and have limited memory capacity. Their reliability is easily affected by the environment, their efficiency is unstable, and errors are inevitable. With the continuous development of civil aviation and general aviation, the number and scale of flights are constantly increasing, placing an ever-growing burden on manual control and creating an increasingly urgent need for intelligent and automated air traffic control information systems.

[0003] In air traffic control, according to civil aviation control procedures, pilots must correctly repeat control instructions before they can begin executing them. With controllers facing increasing workloads and multiple flights being dispatched simultaneously, the accuracy of pilots' repeating instructions is prone to errors.

[0004] With the continuous development of technologies such as artificial intelligence and speech recognition, the demand for speech recognition and processing capabilities in the field of intelligent air traffic control is increasing. Current speech recognition technology is relatively mature and has achieved good results in various applications. Building upon speech recognition, and considering the unique characteristics of civil aviation control communications and the crucial importance of control repeatability for civil aviation flight safety, researching and implementing intelligent repeatability consistency alarm detection can help identify inconsistencies in repeatability in a timely manner, preventing pilots from executing incorrect control instructions and thus avoiding flight safety risks. This will also contribute to improving civil aviation air traffic control capabilities and efficiency, supporting the sustainable development of the civil aviation industry.

[0005] Existing technologies include those employing multimodal models for detection, which address the inherent inaccuracies of speech recognition but fail to consider the simplification or omission issues present in actual air traffic control. Simplified or omitted rephrases can differ significantly in modal features from the original instructions, leading to misidentification. Other technologies generate repeatability instruction texts based on standard civil aviation rules, but this is only for training purposes and cannot solve the problem of repeatability consistency verification in actual operations. Still other technologies directly generate word vectors from instruction texts and compare the word vectors of the instruction and its repetition, without considering the simplification or omission issues present in actual air traffic control. Simplified or omitted rephrases can differ significantly in word vector encoding from the original instructions, leading to misidentification. Finally, some technologies extract element information from instruction texts based on air-to-ground communication rules and directly match keywords and key sentence structures to construct the required structured information element data before performing repetition checks. However, these technologies lack clear methods for constructing extraction and matching rules and do not consider the simplification or omission issues present in actual air traffic control.

[0006] In summary, the shortcomings of the existing technology include: (1) it does not consider that the dialogue mode in the actual environment is not all instructions and rephrase, but also includes reports, inquiries, etc., and does not provide a method for classifying and judging the instruction scenario; (2) it is based on standard control call rules and does not consider the problem of simplified and omitted expressions in the actual scenario; (3) it relies on the text feature generalization ability of machine learning algorithms and does not consider that under the omission of expressions, different parameters may have the same value, resulting in similar generalization features. Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting consistency alarms in air traffic control instruction paraphrasing based on intelligent completion. This method classifies and identifies information about the instruction and paraphrasing patterns from the text of air traffic control instructions and pilot paraphrasing instructions. It intelligently completes simplified and omitted information. Instead of directly providing a judgment, it calculates the probability of paraphrasing inconsistencies to accommodate different potential text errors caused by different speech recognition models. This invention reduces false judgments of paraphrasing consistency caused by simplification and omission, improving the effectiveness of the technology. While providing the paraphrasing consistency judgment result, it also provides the judgment basis and probability assessment, which is beneficial for operational personnel in making decisions and processing the judgment results. It can adapt to the paraphrasing consistency detection needs of corresponding scenarios, improving reliability.

[0008] This invention is achieved through the following technical solution:

[0009] A method for detecting consistency alarms in control instruction paraphrasing based on intelligent completion, the method comprising:

[0010] The current dialogue is obtained, and the SVM model is used to classify and filter the control text in the current dialogue to obtain the control-repetition text.

[0011] The regulatory text is subjected to regulatory text element identification to obtain the identified regulatory elements;

[0012] Based on the control-restatement text and the control elements, the simplified and omitted information in the control-restatement text is supplemented with text information to obtain the supplemented control-restatement text;

[0013] A matching-comparison mechanism is used to check the consistency of each element in the completed control-review text. If the corresponding check results are inconsistent, that is, the review text is inconsistent with the control text, the probability of inconsistent text information is calculated to facilitate personnel to confirm the review inconsistency alarm. If the corresponding check results are consistent, that is, the review text is consistent with the control text, no alarm is issued. Each element includes control elements and review elements.

[0014] Furthermore, the training steps of the SVM model are as follows:

[0015] Training samples: Obtain raw control speech text as training samples; label the training samples as control-repetition type and other types; the raw control speech text is collected from airports / control centers based on actual application environments;

[0016] Text preprocessing: The training samples are rapidly segmented using a segmentation method based on prepositions, punctuation, and type to obtain the first segmentation result; the first segmentation result is further segmented and analyzed using the forward maximum matching method to obtain the second segmentation result.

[0017] Text representation: Based on the second segmentation result, the training samples are vectorized and used as the training input dataset;

[0018] Classification model training: Construct an SVM model, input the training dataset into the SVM model for model training, and obtain a trained SVM model;

[0019] The segmentation method based on prepositions, punctuation, and type is as follows: segmenting content according to sentence breaks and punctuation; identifying Chinese function words and further segmenting at the positions of Chinese function words; identifying the boundaries of numbers, characters, and text and segmenting at the boundaries; identifying significantly long segments and segmenting them according to 2 characters;

[0020] The positive maximum matching method is as follows: merge the parameterized elements and their parameter values ​​after the first segmentation result into a single word segmentation, and correct the first segmentation result so that the word segmentation of the first segmentation result is more in line with the controlled text features.

[0021] Furthermore, the steps of the forward maximum matching method are as follows:

[0022] Step 11, obtain the maximum word length L of the dictionary;

[0023] Step 12, let the cut length N = L, and the cut start at I = 0;

[0024] Step 13: Starting from position I of the first segmentation result, extract a string of length N. The string is a continuous string with a total initial segmentation length of N starting from I. If the total length of the i segments is less than N and the total length of the i+1 segments is greater than N, then only I segments are extracted.

[0025] Step 14: Determine if the string is in the dictionary. If it is, proceed to step 15; otherwise, proceed to step 16.

[0026] Step 15: Include the extracted string in the word segmentation results, then proceed to Step 18;

[0027] Step 16: Is the value of N already l0? N = l0 means that the current string is composed of only one initial segmentation. l0 is the initial segmentation length. If yes, proceed to step 15; otherwise, proceed to step 17.

[0028] Step 17, let N = Nl i ,l i Given the length of the last initial segment that constitutes the current string, return to step 13;

[0029] Step 18: Determine if the word segmentation has ended. If not, set N = L, shift the truncation position N positions to the right, I = I + N, and return to step 13. If it has ended, proceed to step 19.

[0030] Step 19: Output the second segmentation result and end the word segmentation.

[0031] Furthermore, the SVM model is used to classify the controlled text of the current dialogue. The specific steps are as follows:

[0032] Load the current conversation, i.e., the regulatory text information to be categorized;

[0033] The controlled text information is preprocessed to obtain word segmentation information;

[0034] Based on the word segmentation information, the controlled text information is represented as a vector;

[0035] Load the SVM model and hyperplane parameters;

[0036] Calculate the distance between the representation vector and the hyperplane in the SVM model;

[0037] If the distance is positive, the output is classified as Controlled-Retelling Text; otherwise, the output is classified as Other Types of Text.

[0038] Furthermore, the regulatory text is subjected to regulatory text element identification to obtain the identified regulatory elements, including:

[0039] Step 21: Obtain a word segment from the controlled-review text, and retrieve the word segment from each dictionary in turn;

[0040] Step 22: If the word segmentation is detected, the word segmentation is labeled as the corresponding category, and proceed to step 26; otherwise, proceed to step 23.

[0041] Step 23: Determine whether the word segmentation consists entirely of numerical values ​​(which may be an expression that omits attribute names). If so, proceed to step 24; otherwise, proceed to step 25.

[0042] Step 24: Retain the word segmentation, mark it as pending, and proceed to step 26;

[0043] Step 25: Calculate the similarity probability between the word segment and all the defined rules of the regulatory element category. If there is a regulatory element category that exceeds the defined probability threshold, take the category with the highest probability as the recognition result; otherwise, determine that the word segment is meaningless, delete the word segment, and proceed to step 26.

[0044] Step 26: Determine whether the control element identification is complete. If yes, end the element identification process; otherwise, return to step 21.

[0045] Furthermore, the similarity probability calculation in step 35 involves defining a set of possible structural rules for each type of regulatory element and calculating the matching degree between the unidentified word segment and all rules for each regulatory element; if a certain type of regulatory element i has k possible structural rules, the matching degree between the word segment and each rule is p. ik Then, the average matching degree of the two structural rules with the highest comprehensive matching degree is taken as the matching probability p of word segmentation for this type of regulatory element rule. i If a certain type of regulatory element has a very high probability of matching a certain structural rule, then it is likely that the element is of that type; this avoids errors in probability calculation caused by excessive differences in structural rules between similar elements.

[0046] Among them, the matching probability p between unidentified word segmentation and control element i is... i The calculation formula is:

[0047] p max =max(p i1 P i2 ...p ik )

[0048] p dmax =max[(P i1 p i2 ...p ik )], p i1p i2 ...p ik Excluding p max

[0049] p i =(p max +p dmax ) / 2

[0050] In the formula, p i1 p i2 ...p ik Let p be the matching probability of the k possible structures of the position control element i. max and P dmax p represents the maximum and second-largest probability values ​​in the matching probability. i To identify the matching probability between word segmentation and control element i.

[0051] Furthermore, based on the control-restatement text and the control elements, text information completion is performed on simplified and omitted information in the control-restatement text, including:

[0052] Based on the control-restatement text and the control elements, the controller completes the control section by supplementing the simplified and omitted information in the control-restatement text. The specific steps are as follows:

[0053] Step 311: Take one element from the control instruction text elements and proceed to step 312;

[0054] Step 312: Based on the air-to-ground communication standard, determine whether a simplified description exists according to the element type to which the element belongs. If it exists, proceed to step 313; otherwise, proceed to step 314.

[0055] Step 313: Determine the simplification type and complete it. The completion is based on the air-to-ground communication standard and common simplifications in actual control to build a corresponding mapping relationship. For example, "down" is uniformly completed as "down to", and "altitude 21" is completed as "altitude 2100".

[0056] Step 314: Determine if the completion is complete. If complete, end the process; otherwise, return to step 311.

[0057] Furthermore, based on the control-restatement text and the control elements, text information completion for simplified and omitted information in the control-restatement text also includes:

[0058] Based on the control-restatement text and the control elements, the pilot's restatement portion is completed by supplementing the simplified and omitted information in the control-restatement text. The specific steps are as follows:

[0059] Step 321: Take one element from the elements of the restated instruction text and proceed to step 322;

[0060] Step 322: Determine if it is a pure numerical value. If yes, enter the pending group and return to step 321; otherwise, proceed to step 323.

[0061] Step 323: Based on the air-to-ground communication standard, determine whether a simplified description exists. If it exists, proceed to step 324; otherwise, proceed to step 325.

[0062] Step 324: Determine the simplification type and complete it; the completion is based on the air-to-ground communication standard and common simplifications in actual control to build a corresponding mapping relationship, such as completing "down" to "down to", and "altitude 21" to "altitude 2100", then proceed to step 325;

[0063] Step 325: Calculate the similarity γ between the element and all elements in the control section of the control-reply text using a text similarity calculation method. Select the most similar element; if a match is found, proceed to step 326; otherwise, proceed to step 327. To prevent mismatches caused by adding information during reply, set a similarity threshold K, where i represents the i-th element in the instruction text. The matching rule is: max(γ) i )>K,γ i The similarity between the element and the control section of the control-restatement text is the similarity of the i-th element.

[0064] Step 326: Keep the information in the features unchanged, and complete the information by matching the information in the features. After completion, proceed to step 327:

[0065] Let the element be K, with length k; let the matched control element be L, with length l; l > k; the first character of K is in the same position as i in L, and the completion rule is: K = L0..L i-1 KL l-i-k ..L l , where L i-1 This refers to the content located at position i-1 in the control element;

[0066] Step 327: Check if the completion is complete. If complete, end the process; otherwise, return to step 321.

[0067] Furthermore, a matching-comparison mechanism is used to perform consistency checks on each element of the completed control-restatement text, including:

[0068] Step 41: Take one element from the restatement instruction text elements and proceed to step 42;

[0069] Step 42: Match the control instruction text elements, with the rule that the element identifiers must match, and if they are attribute elements, the attribute names must match; if a match is found, proceed to step 43; if a match fails, proceed to step 44.

[0070] Step 43: Compare the element information, including whether the numerical information and non-numerical information are consistent. If they are consistent, proceed to step 45; otherwise, proceed to step 46.

[0071] Step 44: Move the selected features from the paraphrase instruction text features to the pending feature group, and proceed to step 47;

[0072] Step 45: Remove the element information from the paraphrased instruction text elements and the control instruction text elements respectively, and proceed to step 47;

[0073] Step 46: Remove the element information from the paraphrased instruction text elements and the control instruction text elements respectively, and save them as inconsistent element pairs, then proceed to step 47;

[0074] Step 47: Determine if the text elements of the restatement instruction are empty. If not empty, return to step 41; otherwise, proceed to step 48.

[0075] Step 48: Determine if the group of elements to be determined is empty. If it is not empty, proceed to step 49; otherwise, proceed to step 410.

[0076] Step 49: Determine if the instruction text element is empty. If it is empty, move the pending element to the inconsistent element and proceed to step 413; otherwise, proceed to step 411.

[0077] Step 410: Determine whether the control instruction text element is empty; otherwise, proceed to step 412; otherwise, proceed to step 413.

[0078] Step 411: Perform similarity detection between the undetermined element and the remaining instruction elements, construct inconsistency information pairs according to the most similar matching method and store them in the inconsistency element, then proceed to step 412.

[0079] Step 412: Place the last remaining instruction elements into the inconsistency elements individually, and proceed to step 413;

[0080] Step 413: Detect whether there are any inconsistent elements. If so, it is determined that the restatement is inconsistent and the inconsistent information is output; otherwise, it is determined that the restatement is consistent.

[0081] Furthermore, the similarity calculation formula for similarity detection in step 411 is as follows:

[0082]

[0083]

[0084] Where x and y represent the two elements in the inconsistent element pair; i and j identify the corresponding character subscripts of x and y; lev x,y(i,j) refers to the edit distance between the first i characters of element x and the first j characters of element y; min(i,j) = 0, meaning that one of the corresponding character indices i and j of x and y is zero; max(i,j) means taking the larger value between the corresponding character indices i and j of x and y as the edit distance between the first i characters of element x and the first j characters of element y.

[0085] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0086] This invention discloses a control instruction repetition consistency alarm detection method based on intelligent completion. For control instructions from controllers and repetitions from pilots, it classifies and identifies information about the instruction and repetition patterns from the control communication instruction text; it intelligently completes simplified and omitted expressions; and it calculates the probability of repetition inconsistency instead of directly providing a judgment, thus adapting to different potential text errors brought about by different speech recognition models. This invention reduces misjudgments of repetition consistency caused by simplification and omission, improving the effectiveness of the technology; while providing the repetition consistency judgment result, it also provides the judgment basis and probability assessment, which is beneficial for operational personnel to make decisions and process the judgment results; it can adapt to the repetition consistency detection needs of corresponding scenarios, improving reliability. Attached Figure Description

[0087] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0088] Figure 1 This is a flowchart of a control instruction rephrase consistency alarm detection method based on intelligent completion according to the present invention;

[0089] Figure 2 This is a flowchart illustrating the text classification and filtering process of the present invention.

[0090] Figure 3 This invention provides a fast initial segmentation process and example diagram based on preposition, punctuation, and type segmentation methods.

[0091] Figure 4 This is a flowchart of the forward maximum matching method of the present invention;

[0092] Figure 5 This is a flowchart illustrating the model training and text representation process of this invention.

[0093] Figure 6 This is a flowchart illustrating the classification model training and text classification process of this invention.

[0094] Figure 7 This is a flowchart illustrating the continuous classification and judgment process for controlled calls in this invention.

[0095] Figure 8This is a flowchart of the text element recognition process of the present invention;

[0096] Figure 9 This is a flowchart to complete the control and paraphrase text elements of the present invention;

[0097] Figure 10 This is a flowchart of the consistency check process for the present invention. Detailed Implementation

[0098] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0099] The existing control instruction detection methods have the following defects: (1) They do not consider that the actual environment is not all dialogue mode consisting of instructions and rephrases, but also includes reports, inquiries, etc., and do not provide a method for classifying and judging the instruction scenario; (2) They are based on standard control call rules and do not consider the problem of simplified and omitted expressions in actual scenarios; (3) They rely on the text feature generalization ability of machine learning algorithms and do not consider that under the omission of expressions, different parameters may have the same value, resulting in similar generalization features.

[0100] Therefore, to address the above problems, this invention designs a control instruction repetition consistency alarm detection method based on intelligent completion. For control instructions from controllers and repetitions from pilots, it classifies and identifies information about the instruction and repetition patterns from the control communication instruction text; it intelligently completes simplified and omitted expressions; and it calculates the probability of repetition inconsistency instead of directly providing a judgment, thus adapting to different potential text errors brought about by different speech recognition models. This invention reduces misjudgments of repetition consistency caused by simplification and omission, improving the effectiveness of the technology; while providing the repetition consistency judgment result, it also provides the judgment basis and probability assessment, which is beneficial for operational personnel to make decisions and process the judgment results; it can adapt to the repetition consistency detection needs of corresponding scenarios, improving reliability.

[0101] This invention classifies and identifies information about instructions and repetition patterns from the text of control call instructions; it intelligently completes simplified and omitted expressions; and it calculates the probability of inconsistencies in repetitions instead of directly providing a judgment, in order to adapt to different potential text errors brought about by different speech recognition models.

[0102] Specifically, this invention comprises four parts: classification and filtering of regulatory texts, identification of regulatory text elements, intelligent text information completion, and paraphrase consistency detection. The overall data flow is as follows: Figure 1 As shown.

[0103] Control text classification and filtering: Used to determine whether the current dialogue is a control-repetition process, to avoid using report-confirmation, inquiry-answer, etc., as input for subsequent judgments, and to issue anomaly alarms; Control text element identification: Performs joint analysis on the instruction-repetition text, identifies and marks information elements, and removes prepositions, conjunctions, and other information that does not affect the meaning of the instruction; Intelligent text information completion: Identifies simplified and omitted information in instruction-repetition and completes it to prevent inconsistencies in identified features during comparison; Repetition consistency detection: First, checks the consistency of instruction-repetition. For inconsistencies, calculates the probability of inconsistency to facilitate personnel confirmation of repetition inconsistency alarms.

[0104] Example

[0105] like Figure 1 As shown, this invention provides a control instruction paraphrase consistency alarm detection method based on intelligent completion. The method includes:

[0106] Step 1: Obtain the current dialogue, and use the SVM model to classify and filter the control text in the current dialogue to obtain the control-repetition text;

[0107] Step 1 is used to determine whether the current dialogue is a control-repetition process, so as to avoid using report-confirmation, inquiry-response, or other similar content as input for subsequent judgments and to issue an abnormal alarm. The communication results between the controller and the pilot in a dialogue constitute a text information pair of "controller information - pilot information". Without considering the identification of other control communication types, this information pair can be simply classified into "control-repetition" type and "other types".

[0108] like Figure 2 As shown, the process of classifying and filtering controlled text includes:

[0109] (1) Training samples: Obtain the original control speech text as training samples; label the training samples as control-repetition type and other types; the original control speech text is collected from airports / control centers based on the actual application environment;

[0110] (2) Text preprocessing: The training samples are rapidly segmented using a segmentation method based on prepositions, punctuation, and type to obtain the first segmentation result; the first segmentation result is further segmented using the forward maximum matching method to obtain the second segmentation result;

[0111] (3) Text representation: Based on the second segmentation result, the training samples are vectorized and used as the training input dataset;

[0112] (4) Classification model training: Construct an SVM model, input the training input dataset into the SVM model for model training, and obtain a trained SVM model;

[0113] (5) Classification judgment.

[0114] It should be noted that when classifying and filtering regulatory texts, alternatives to the SVM classification framework chosen in this invention can be used, such as Naive Bayes classifiers, neural network algorithms, decision trees, and Ensemble algorithms. Since regulatory calls are subject to basic call protocol constraints, effective results can still be obtained.

[0115] Specifically as follows:

[0116] (1) Obtain training samples

[0117] The training samples use speech recognition results of actual air traffic control voices. Given the differences in the habits of air traffic controllers at different airports / control centers, original air traffic control voice texts can be collected according to the actual application environment. Then, all texts are arranged into dialogue samples according to the "controller-pilot" two-sentence format, and the samples are labeled as "control-repetition" type and "other types".

[0118] (2) Text preprocessing - controlled word segmentation

[0119] Text preprocessing primarily involves word segmentation. Due to the specific nature of regulated texts, attributes and their corresponding parameter values ​​should be grouped into a single word, such as flight number "CA4110" and landmark "P10". Understanding and statistics facilitate the separation of numerical values ​​from preceding non-numerical values. Based on the fundamental constraints of air traffic control communication (not entirely free-flowing, possessing relative determinism), a combination of forward maximum matching dictionary segmentation and regular expression matching is employed. The dictionary includes commonly used regulated terms such as heading, altitude, speed, and correction pressure, as well as flight numbers (extracted from civil aviation flight plans) and names of surrounding surface points (defined during planning or update, and can be directly entered). Based on the dictionary, keywords in the regulated text can be identified. Furthermore, regular expression matching of parameters with attributes or instructions merges cases where parameters are separated.

[0120] Ordinary word segmentation breaks a sentence into individual characters. For example, "continue approaching, runway 36R, pressure repair 1008, east 123" is segmented into "continue approaching runway 36R pressure repair 1008, east 123". Then, forward maximum matching or other similar algorithms are used to determine whether adjacent characters constitute a word. In the scheme of this invention, based on the special characteristics of controlled speech, a segmentation method based on prepositions, punctuation and type is used for fast initial segmentation. The initial segmentation method is optimized to improve the efficiency of forward maximum matching.

[0121] Perform a quick initial segmentation based on prepositions, punctuation, and type segmentation methods, which generally include: segmenting the content according to sentence-ending punctuation; identifying Chinese function words and further segmenting at the positions of Chinese function words; identifying the boundaries of numbers, characters, and texts and segmenting at the boundaries; identifying significantly long segments and segmenting them into 2 characters; the process is as Figure 3 shown. The specific steps of the segmentation method based on prepositions, punctuation, and types are as follows:

[0122] 1) Segment the content according to sentence-ending punctuation;

[0123] 2) Identify Chinese function words, including conjunctions, auxiliary words, modal particles, etc., such as "de", "he", etc., and further segment at the positions of function words;

[0124] 3) Identify the boundaries of numbers, letters, and texts in the text and segment at the boundaries;

[0125] 4) Identify significantly long segments and segment them into "2 characters" (to fit the characteristic that words in regulatory terms are generally composed of 2 characters); when identifying significantly long segments, the existing number of segments is n, and the length of the i-th segment is l i , and the average segment length is l n , then the significantly long segment satisfies:

[0126] l i > 2 and l i > 2 * l n

[0127] For example, the initial segmentation result of "Continue approach, runway 36R, altitude 1008, east 123" is "Continue|approach|runway|36|R|altitude|1008|east|123". Compared with the initial segmentation method of ordinary word segmentation, the number of segments is less, significantly reducing the subsequent matching calculation times.

[0128] The result of the initial segmentation is further analyzed for word segmentation using the forward maximum matching method. Correspondingly, in the forward maximum matching, the determination and change of the string length are related to the initial segmentation of the string that makes up the string, rather than the usual continuous change with a step size of 1. The forward maximum matching method is: Merge the parameterized elements and their parameter values after the first segmentation result into one word segmentation, and correct the first segmentation result to make the word segmentation of the first segmentation result more in line with the text characteristics of regulation. The process is as Figure 4 shown. The specific steps of the forward maximum matching method are:

[0129] Step 11, obtain the maximum word length L of the dictionary;

[0130] Step 12, let the intercepted length N = L, and the interception start at I = 0;

[0131] Step 13: Starting from position I of the first segmentation result, extract a string of length N. The string is a continuous string with a total initial segmentation length of N starting from I. If the total length of the i segments is less than N and the total length of the i+1 segments is greater than N, then only I segments are extracted.

[0132] Step 14: Determine if the string is in the dictionary. If it is, proceed to step 15; otherwise, proceed to step 16.

[0133] Step 15: Include the extracted string in the word segmentation results, then proceed to Step 18;

[0134] Step 16: Is the value of N already l0? N = l0 means that the current string is composed of only one initial segmentation. l0 is the initial segmentation length. If yes, proceed to step 15; otherwise, proceed to step 17.

[0135] Step 17, let N = Nl i ,l i Given the length of the last initial segment that constitutes the current string, return to step 13;

[0136] Step 18: Determine if the word segmentation has ended. If not, set N = L, shift the truncation position N positions to the right, I = I + N, and return to step 13. If it has ended, proceed to step 19.

[0137] Step 19: Output the second segmentation result and end the word segmentation.

[0138] The purpose of the above regular expression matching is to merge the parameterized elements and their parameter values ​​after word segmentation into a single word segment, correcting the segmentation result and making the segmentation more consistent with the controlled text features. Regular expression matching establishes a corresponding regular expression description for each parameterized element:

[0139] The heading, altitude, speed, and corrected sea pressure are all followed by consecutive numbers, and the corresponding rules are as follows:

[0140] Heading [d] + Speed ​​[d] + Height [d] + Pressure repair [d] +

[0141] [d] + It represents one or more consecutive numbers.

[0142] For example, the communication during an air traffic control approach, "Continue approach, runway 36R, pressure adjustment 1008, East 123", can be segmented into "Continue approach|runway|36|R|pressure adjustment|10008|East 123" after positive maximum matching on the initial segmentation, where "East 123" is the flight number as a whole in the dictionary. After regular expression matching correction, it becomes: "Continue approach|runway 36R|pressure adjustment 10008|East 123". Given that the content of the air traffic control dialogue is relatively fixed, the combination of matching and correction can improve the accuracy of word segmentation, reduce computational complexity, and simultaneously lower the dimensionality of the segmentation results.

[0143] (3) Text representation - Single sentence representation vector

[0144] Text representation is the conversion of text into a computer-processable representation, namely vector representation. Given that air traffic control has standardized communication protocols and professional terminology, but simplified or omitted expressions may exist in actual control communications, the Skip-gram model (a neural network-based word vector generation method) is used to extract feature information from control text. The SkipGram model can consider synonyms, near-synonyms, and abstract and concrete concepts when learning word vectors, perfectly meeting the needs of control communication text representation.

[0145] First, model training is performed, and the general steps are as follows:

[0146] 1) Construct a Skip-gram model;

[0147] 2) Use dictionaries as the thesaurus, including commonly used terms in the control, such as heading, altitude, speed, correction pressure, etc., as well as flight numbers (extracted from civil aviation flight plans) and names of surface points around the control (defined during planning or updating of plans, which can be directly entered).

[0148] 3) Obtain regulated voice data in the target environment as the raw training data (the larger the amount of data, the better the training results);

[0149] 4) Use the controlled word segmentation method in text preprocessing to segment the training data (including three main steps: initial segmentation, forward maximum matching, and regular expression correction);

[0150] 5) Train the Skip-gram model with the training data to obtain the representation model.

[0151] The trained representation model ranks all feature words in a text sample based on their relevance to relevant categories. Taking Tengchong, Yunnan as an example, the average daily number of inbound and outbound flights is about 30, there are 52 airlines nationwide, and there are 10 landmarks in the surrounding area used for navigation. Assuming the vocabulary is regulated, the number of feature words in the model is less than 200. Taking a feature word count of T=200 as an example.

[0152] Secondly, the regulated text is vectorized based on a representation model. In the trained representation model, the feature words are the dimensions of the representation vector. After the text to be represented is segmented, it is compared with the trained model. If a feature word is matched, its value in the vector is set to 1, and the others are set to 0. Each regulated text forms a vector of dimension t. The general process of vectorization is as follows:

[0153] 1) Obtain the text information to be represented;

[0154] 2) Perform controlled word segmentation on text information;

[0155] 3) Construct a t-dimensional zero vector and compare it with the trained Skip-gram feature word model;

[0156] 4) Obtain a feature word;

[0157] 5) Detection: If a feature word exists, the corresponding position in the vector is 1; otherwise, it is zero.

[0158] 6) If the detection is complete, output the vector; otherwise, return to step 4).

[0159] For each training sample, each text is represented by a vector and then used to form the training input dataset in the manner of "vector-classification". The classification type "control-repetition" is represented by "TRUE" and "other types" is represented by "FALSE".

[0160] Figure 5 (a) is a flowchart of the model training process. Figure 5 (b) is a text representation of the flowchart.

[0161] (4) Classification model training

[0162] SVM is a binary classifier, well-suited for classifying regulatory instruction texts into "regulation-repetition" and "other" types. Its margin-maximizing learning strategy allows the model to classify new data more accurately, making the classifier more stable. The goal of training a classification model is to obtain a hyperplane that separates data types after building the SVM model. To improve the model's accuracy, multiple rounds of training and optimization can be performed using the training dataset. The general training process is as follows:

[0163] 1) Construct the SVM model;

[0164] 2) Set the model penalty parameter C = 0.90. Based on the fact that there are basic normative constraints on controlled calls, but there are also simplifications and omissions, a certain degree of fault tolerance is allowed (C takes a value between 0 and 1. The smaller the value, the stronger the generalization (fault tolerance) ability, but the lower the classification accuracy).

[0165] 3) Choosing the Gaussian kernel function maps the input vector to a higher-dimensional space. Based on its good performance with both large and small samples, it can adapt to the model training needs of scenarios with varying business volumes, such as approximately 30 daily flights in Tengchong, Yunnan, and approximately 530 daily flights in Shuangliu, Chengdu. The Gaussian kernel function is expressed as: xi is the center of the kernel function, ‖x-xi‖ 2 Let σ be the Euclidean distance between vectors x and xi. σ controls the range of influence of the Gaussian kernel function. The larger the value of σ, the greater the local influence range of the Gaussian kernel function.

[0166] 4) Based on the number of feature words, set the Gaussian kernel coefficient χ = ​​1 / T;

[0167] 5) Load the training dataset;

[0168] 6) Initialize the parameters of SMO (Sequence Minimum Optimization Algorithm);

[0169] 7) Obtain the vector representation and identifier type of a data sample;

[0170] 8) Use a kernel function to map the vector to a high-dimensional space to obtain the input X;

[0171] 9) Use X to optimize the parameters of the SMO so that the hyperplane correctly classifies X into the specified type;

[0172] 10) Check if the training is over; otherwise, return to step 7); otherwise, proceed to step 11.

[0173] 11) Output and save the hyperplane parameters (i.e., the trained SVM model).

[0174] The training dataset in step 5) above was obtained through annotation of the control call text classification training data. This involves generating "sentence 1-sentence 2" relationship pairs between adjacent sentences in the original training dataset according to the time sequence, and labeling them as either "control-repetition" type. The original training dataset consists of N sentences of call text, with each sentence forming a relationship pair with the preceding and following sentences, resulting in a total of n-1 relationship pairs. After annotation, N-1 valid training samples are formed.

[0175] In step 7) above, the sample vector representation is generated by using a trained representation model to generate representation vectors for sentence 1 and sentence 2 in the relation pair, and then concatenating them; if the dimension of a single sentence representation vector is t, then the dimension of the sample representation vector is 2t; this design can retain the correlation information between sentence 1 and sentence 2 in the effective samples and improve the accuracy of classification.

[0176] Figure 6 (a) is a flowchart of the classification model training process. Figure 6 (b) is a flowchart of text classification.

[0177] (5) Classification judgment

[0178] The classification decision is based on a trained SVM model. It calculates the geometric margin between the input data's representation vector and the trained separating hyperplane, defined as: w T x is the transpose of the parameter vector w. (i) and y (i) Let w be the target value corresponding to the i-th sample and the i-th sample value, and let ||w|| represent the 2-norm of the parameter vector w.

[0179] If the sample is classified as "Control-Restatement", the value is positive; if the sample is classified as "Other Types", the value is negative.

[0180] Specifically, the SVM model is used to classify the controlled text of the current dialogue. The specific steps are as follows:

[0181] Load the current conversation, i.e., the regulatory text information to be categorized;

[0182] The controlled text information is preprocessed to obtain word segmentation information;

[0183] Based on the word segmentation information, the controlled text information is represented as a vector;

[0184] Load the SVM model and hyperplane parameters;

[0185] Calculate the distance between the representation vector and the hyperplane in the SVM model;

[0186] If the distance is positive, the output is classified as control-repetition text; otherwise, the output is classified as other types of text.

[0187] The categorization is complete.

[0188] In the context of air traffic control classification, given that air traffic control involves one-to-many communication between controllers and pilots on the same frequency, the text identified by a single sentence will not simultaneously belong to two "control-reply" pairs; nor are control and reply communications necessarily directly adjacent, nor are they too far apart. The present invention constructs a candidate level for air traffic control classification based on these scenario characteristics, and designs the air traffic control classification process as follows: Figure 7 As shown, the description is as follows:

[0189] 1) Obtain a control call text, check if there is a cached previous unclassified call text, if not, proceed to step 2), if so, proceed to step 3);

[0190] 2) If only one statement fails to build a category candidate, store the text in the uncategorized cache and return to step 1);

[0191] 3) The controlled call text and the cached previous unclassified call texts are respectively formed into controlled statement pairs to form a candidate set. If the number of cached call texts is n, the number of classified candidate statement pairs constructed is also n, and then proceed to step 4);

[0192] 4) Take a statement pair from the candidate set, perform statement word segmentation and vector representation, and submit for classification. If it is "control - repetition", proceed to step 6); otherwise, proceed to step 5);

[0193] 5) Determine whether all statement pairs in the candidate set have completed classification judgment. If not, repeat step 4). If so, it means that all candidate statement pairs are not of the "control - repetition" type. Store the current text in the unclassified cache and return to step 1);

[0194] 6) Remove the cached statement that forms "control - repetition" with the current statement from the cache. At the same time, delete the cached statements before the corresponding cached statement time as expired data. Store the two statements in the "control - repetition" cache in a paired manner, and repeat step 1).

[0195] For example, if a total of 5 controlled call text statements A, B, C, D, and E are input in sequence, and B - D and C - E form "control - repetition", the cache, candidate set, and recognition results for each input statement are shown in the following table:

[0196] Serial Number enter cache Candidate set Result set 1 A - - - 2 B A AB - 3 C AB AC,BC - 4 D ABC AD, BD, CD BD 5 E C CE CE

[0197] Step 2: Identify the control text elements of the said control - repetition text to obtain the identified control elements;

[0198] In step 2, the control text element identification is to perform a joint analysis on the instruction - repetition text, identify and mark the information elements, and at the same time remove the information such as prepositions and conjunctions that do not affect the instruction meaning. In the word - segmentation stage of the control text classification, the control word - segmentation of the text has been completed. The element identification is carried out based on the word - segmentation. Each word - segment is type - labeled based on the dictionary library, and the words that cannot be labeled are removed as invalid information. When constructing the dictionary, the simplified representations should also be included in the dictionary. For example, both "corrected sea level pressure" and its simplified expression "corrected pressure" should be included in the dictionary.

[0199] Specifically, as Figure 8 shown, identify the control text elements of the said control - repetition text to obtain the identified control elements, including:

[0200] Step 21: Obtain a word - segment from the said control - repetition text and sequentially retrieve the word - segment in each dictionary;

[0201] Step 22: If the word - segment is detected, label the word - segment as the corresponding classification and proceed to step 26; otherwise, proceed to step 23;

[0202] Step 23: Determine whether the word segmentation consists entirely of numerical values ​​(which may be an expression that omits attribute names). If so, proceed to step 24; otherwise, proceed to step 25.

[0203] Step 24: Retain the word segmentation, mark it as pending, and proceed to step 26;

[0204] Step 25: Calculate the similarity probability between the word segment and all the defined rules of the regulatory element category. If there is a regulatory element category that exceeds the defined probability threshold, take the category with the highest probability as the recognition result; otherwise, determine that the word segment is meaningless, delete the word segment, and proceed to step 26.

[0205] Step 26: Determine whether the control element identification is complete. If yes, end the element identification process; otherwise, return to step 21.

[0206] The similarity probability calculation in step 35 above involves defining a set of possible structural rules for each type of regulatory element and calculating the matching degree between the unidentified word segment and all rules for each regulatory element. If a certain type of regulatory element i has k possible structural rules, the matching degree between the word segment and each rule is p. ik Then, the average matching degree of the two structural rules with the highest comprehensive matching degree is taken as the matching probability p of word segmentation for this type of regulatory element rule. i If a certain type of regulatory element has a very high probability of matching a certain structural rule, then it is likely that the element is of that type; this avoids errors in probability calculation caused by excessive differences in structural rules between similar elements.

[0207] Among them, the matching probability p between unidentified word segmentation and control element i is... i The calculation formula is:

[0208] p max =max(p i1 ,p i2 ...p ik )

[0209] p dmax =max[(p i1 ,p i2 ...p ik )],p i1 ,p i2 ...p ik Excluding p max

[0210] p i =(p max +p dmax ) / 2

[0211] In the formula, p i1 ,p i2...p ik Let p be the matching probability of the k possible structures of the position control element i. max and p dmax p represents the maximum and second-largest probability values ​​in the matching probability. i To identify the matching probability between word segmentation and control element i.

[0212] For example, the word segmentation result of the control text “Continue Approach|Runway 36R|Pressure Repair 10008|Eastern 123” after element recognition is “

Instruction

Geography

Attribute

Flight Number

[0213] Step 3: Based on the control-restatement text and the control elements, complete the simplified and omitted information in the control-restatement text to obtain the completed control-restatement text;

[0214] Step 3 identifies and completes simplified and omitted information in the instruction paraphrasing to prevent inconsistencies in features identified during comparison. Analysis of actual air traffic control datasets reveals that omitted or simplified information primarily occurs when pilots omit attribute names or partial instruction information during paraphrasing. For example, "Southern 3580 direct flight to POMOK resumes autonomous navigation" is paraphrased as "Fly to POMOK autonomous navigation Southern 3580," and "Air China 4239 continuously descends to 3000, correcting sea pressure to 1009" is paraphrased as "30001009 Air China 4239." Direct feature extraction for detection reveals inconsistencies. Alternatively, information may be simplified in altitude attribute parameter values, such as "Spring Airlines 6617 flying to Nanxun at 6000" being paraphrased as "Fly to Nanxun at 6000, Spring Airlines 6617."

[0215] Text information completion is based on the fundamental rule that paraphrasing corresponds to air traffic control instructions. It requires constructing common simplified expression mapping relationships based on air-to-ground communication standards and actual communication data, such as "pressure correction - corrected sea pressure," "up - up to," etc. Altitude descriptions are usually three digits or more, but civil aviation altitudes are divided into whole numbers of 100. Sometimes, the tens and units digits are omitted during control and paraphrasing, representing a two-digit number. This is generally seen in altitude parameters or altitude adjustment instructions, such as "up to 48," "continuously descend to 23," "altitude 65," etc. The consecutive numbers immediately following the altitude or altitude adjustment instruction are represented by "S." The simplified altitude detection model is as follows:

[0216] {S|S is the height value && length(S)<3}

[0217] The `length()` function calculates the number of characters. Based on the constructed mapping library, the elements in the "Control" section are first completed, and then the corresponding "Restatement" section is completed.

[0218] like Figure 9 As shown in (a), based on the control-restatement text and the control elements, the controller completes the control portion of the simplified and omitted information in the control-restatement text. The specific steps are as follows:

[0219] Step 311: Take one element from the control instruction text elements and proceed to step 312;

[0220] Step 312: Based on the air-to-ground communication standard, determine whether a simplified description exists according to the element type to which the element belongs. If it exists, proceed to step 313; otherwise, proceed to step 314.

[0221] Step 313: Determine the simplification type and complete it. The completion is based on the air-to-ground communication standard and common simplifications in actual control to build a corresponding mapping relationship. For example, "down" is uniformly completed as "down to", and "altitude 21" is completed as "altitude 2100".

[0222] Step 314: Determine if the completion is complete. If complete, end the process; otherwise, return to step 311.

[0223] like Figure 9 As shown in (b), the completion of the paraphrase first restores the elements that conform to common simplified mappings, and then compares them with the elements in "control" to perform analogical completion while keeping its own information unchanged. Specifically, based on the control-paraphrase text and the control elements, the pilot's paraphrase is completed by supplementing the simplified and omitted information in the control-paraphrase text. The specific steps are as follows:

[0224] Step 321: Take one element from the elements of the restated instruction text and proceed to step 322;

[0225] Step 322: Determine if it is a pure numerical value. If yes, enter the pending group and return to step 321; otherwise, proceed to step 323.

[0226] Step 323: Based on the air-to-ground communication standard, determine whether a simplified description exists. If it exists, proceed to step 324; otherwise, proceed to step 325.

[0227] Step 324: Determine the simplification type and complete it; the completion is based on the air-to-ground communication standard and common simplifications in actual control to build a corresponding mapping relationship, such as completing "down" to "down to", and "altitude 21" to "altitude 2100", then proceed to step 325;

[0228] Step 325: Calculate the similarity γ between the element and all elements in the control section of the control-reply text using a text similarity calculation method. Select the most similar element; if a match is found, proceed to step 326; otherwise, proceed to step 327. To prevent mismatches caused by adding information during reply, set a similarity threshold K, where i represents the i-th element in the instruction text. The matching rule is: max(γ) i )>K,γ i The similarity between the element and the control section of the control-restatement text is the similarity of the i-th element.

[0229] Step 326: Keep the information in the features unchanged, and complete the information by matching the information in the features. After completion, proceed to step 327:

[0230] Let the element be K, with length k; let the matched control element be L, with length l; l > k; the first character of K is in the same position as i in L, and the completion rule is: K = L0..L i-1 KL l-i-k ..L l , where L i-1 This refers to the content located at position i-1 in the control element;

[0231] Step 327: Check if the completion is complete. If complete, end the process; otherwise, return to step 321.

[0232] Feature completion is a rule-based reverse matching method. At the feature identification point, a set of possible structural rules is defined for each type of feature. Among the k structural rules for feature type i, exactly one is a complete rule, and the others are rules corresponding to various simplified and omitted expressions.

[0233] Simplified judgment: The simplified judgment in the control part completion step 2) and the restatement part completion step 3) is to match the text with all the rules to which the corresponding element type belongs. If the complete rule is matched, it is not a simplified expression; otherwise, it is a simplified expression.

[0234] Element completion: The element completion steps 3) of the control part completion and 4) of the restatement part completion are to compare the simplified expression rules of text matching with the complete expression rules and complete them according to the complete expression rules.

[0235] Complete the structural consistency of the paraphrased elements: Steps 5) and 6) of the paraphrased parts are mainly used to improve the consistency of the expression structure of the same element between the paraphrased text and the control text in the "control-paraphrasing" pair, and reduce the error rate / false alarm rate of subsequent consistency judgment.

[0236] For example, the "control-restatement" structure built using Air China as an example is as follows:

[0237] "Air China Flight 4239 continuously descended to 3000, sea pressure adjusted to 1009-3000, 1009 Air China Flight 4239"

[0238] After word segmentation element identification, the result is:

[0239] "

Flight Number

Instruction

Attribute

Flight Number

[0240] After completion:

[0241] "

Flight Number

Instruction

Attribute

Flight Number

[0242] For example, “Southern 3580 direct flight to POMOK resumes autonomous navigation” can be restated as “Southern 3580 autonomously navigates to POMOK”. After word segmentation element recognition, the result is:

[0243] "[Flight Number] China Southern 3580 | [Instruction] Direct flight to POMOK | [Instruction] Resume autonomous navigation - [Instruction] Fly to POMOK | [Instruction] Autonomous navigation | [Flight Number] China Southern 3580"

[0244] The preliminary completion is as follows:

[0245] "[Flight Number] China Southern 3580 | [Instruction] Direct flight to POMOK | [Instruction] Resume autonomous navigation - [Instruction] Direct flight to POMOK | [Instruction] Autonomous navigation | [Flight Number] China Southern 3580"

[0246] The final version is:

[0247] "[Flight Number] China Southern 3580 | [Instruction] Direct flight to POMOK | [Instruction] Resume autonomous navigation - [Instruction] Direct flight to POMOK | [Instruction] Resume autonomous navigation | [Flight Number] China Southern 3580"

[0248] Step 4: Use a matching-comparison mechanism to check the consistency of each element in the completed control-review text. If the corresponding check results are inconsistent, that is, the review text is inconsistent with the control text, the probability of inconsistent text information is calculated to facilitate personnel to confirm the review inconsistency alarm; if the corresponding check results are consistent, that is, the review text is consistent with the control text, no alarm is issued; where each element includes control elements and review elements.

[0249] Step 4 involves converting the text between the two parties in the "Control Instruction - Paraphrase Instruction" dialogue into a structured format after the element completion is completed. Based on this, a paraphrase consistency check can be performed. To ensure the reliability of the judgment, the probability of inconsistent information is calculated to measure the reliability of the result.

[0250] (1) Consistency checks can be performed by comparing each element one by one. Since the order of elements in the restated text may not be consistent with the order of elements in the instruction, a matching-comparison mechanism is used.

[0251] Specifically, such as Figure 10 As shown, a matching-comparison mechanism is used to check the consistency of each element in the completed control-restatement text, including:

[0252] Step 41: Take one element from the restatement instruction text elements and proceed to step 42;

[0253] Step 42: Match the control instruction text elements, with the rule that the element identifiers must match, and if they are attribute elements, the attribute names must match; if a match is found, proceed to step 43; if a match fails, proceed to step 44.

[0254] Step 43: Compare the element information, including whether the numerical information and non-numerical information are consistent. If they are consistent, proceed to step 45; otherwise, proceed to step 46.

[0255] Step 44: Move the selected features from the paraphrase instruction text features to the pending feature group, and proceed to step 47;

[0256] Step 45: Remove the element information from the paraphrased instruction text elements and the control instruction text elements respectively, and proceed to step 47;

[0257] Step 46: Remove the element information from the paraphrased instruction text elements and the control instruction text elements respectively, and save them as inconsistent element pairs, then proceed to step 47;

[0258] Step 47: Determine if the text elements of the restatement instruction are empty. If not empty, return to step 41; otherwise, proceed to step 48.

[0259] Step 48: Determine if the group of elements to be determined is empty. If it is not empty, proceed to step 49; otherwise, proceed to step 410.

[0260] Step 49: Determine if the instruction text element is empty. If it is empty, move the pending element to the inconsistent element and proceed to step 413; otherwise, proceed to step 411.

[0261] Step 410: Determine whether the control instruction text element is empty; otherwise, proceed to step 412; otherwise, proceed to step 413.

[0262] Step 411: Perform similarity detection between the undetermined element and the remaining instruction elements, construct inconsistency information pairs according to the most similar matching method and store them in the inconsistency element, then proceed to step 412.

[0263] Step 412: Place the last remaining instruction elements into the inconsistency elements individually, and proceed to step 413;

[0264] Step 413: Detect whether there are any inconsistent elements. If so, it is determined that the restatement is inconsistent and the inconsistent information is output; otherwise, it is determined that the restatement is consistent.

[0265] To support probability calculation, the consistency judgment completes the consistency judgment process for all elements, rather than judging it as a restatement inconsistency and ending the consistency detection as soon as a set of inconsistent elements is identified.

[0266] (2) Probability Calculation

[0267] After completing the consistency assessment, to ensure the reliability of the assessment and support controllers in making decisions based on the results, the probability of information deemed inconsistent is calculated. This probability primarily considers the possibility of misrecognition of speech as text; therefore, the general steps for probability calculation are as follows:

[0268] 1) Count the number A of all inconsistent information;

[0269] 2) Count the number of isolated messages (B) and the number of paired messages (C) respectively, where A = B + C;

[0270] 3) Calculate the inconsistency probability for each isolated term.

[0271] 4) For each pair of information, obtain the phonetic information of each character from the pronunciation database of the speech recognition model, construct a phonetic representation pair, and calculate the similarity γ(c) between the phonetic representation pairs. The more similar the phonetic representations, the greater the probability of speech recognition error, and the smaller the probability of paraphrasing inconsistency.

[0272] 5) Finally, calculate the total probability of inconsistency:

[0273]

[0274] The final output information includes the judgment result, the probability of inconsistency in the inconsistency information pair, and the restatement inconsistency.

[0275] Specifically, the similarity calculation formula in step 411 is as follows:

[0276]

[0277]

[0278] Where x and y represent the two elements in the inconsistent element pair; i and j identify the corresponding character subscripts of x and y; lev x,y (i,j) refers to the edit distance between the first i characters of element x and the first j characters of element y; min(i,j) = 0, meaning that one of the corresponding character indices i and j of x and y is zero; max(i,j) means taking the larger value between the corresponding character indices i and j of x and y as the edit distance between the first i characters of element x and the first j characters of element y.

[0279] The beneficial effects of the present invention are as follows: (1) The present invention can adapt to the paraphrase consistency detection requirements of the corresponding scenario by updating the corresponding dictionary and training with the corresponding data samples in different application scenarios; (2) The present invention can handle the situation where the rules in the call specifications are simplified or omitted in the actual control scenario, reduce the paraphrase consistency misjudgment caused by simplification and omission, and improve the effectiveness of the technology; (3) The present invention provides the judgment basis and probability assessment while giving the paraphrase consistency judgment result, which is conducive to business personnel to make decisions and process the judgment result; (4) The present invention uses the recognition results of different speech recognition models as input, and only needs to include the corresponding pronunciation model, which can reduce the impact of speech recognition error on the effectiveness of control instruction text and improve the reliability of paraphrase consistency judgment result.

[0280] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0281] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0282] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0283] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0284] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting the consistency of regulatory instruction paraphrasing based on intelligent completion, characterized in that, The method comprises: obtaining a current conversation, classifying and screening the current conversation by using an SVM model to obtain a regulation-repeating text; performing regulation text element identification on the regulation-repeating text to obtain identified regulation elements; performing text information completion on simplified and omitted information in the regulation-repeating text according to the regulation-repeating text and the regulation elements to obtain completed regulation-repeating text; performing element consistency checking on the completed regulation-repeating text by using a matching-comparing mechanism, if the corresponding checking result is inconsistent, i.e., the repeating text is inconsistent with the regulation text, calculating the probability of inconsistent text information to facilitate personnel to confirm the repeating inconsistency alarm; if the corresponding checking result is consistent, i.e., the repeating text is consistent with the regulation text, no alarm is performed; wherein the elements include regulation elements and repeating elements; performing element consistency checking on the completed regulation-repeating text by using a matching-comparing mechanism, comprising: step 41, taking an element from the repeating instruction text element, entering step 42; step 42, matching in the regulation instruction text element, the rule being that the element identifier is consistent, if it is an attribute element, the attribute name is consistent; if matching is successful, entering step 43; if matching fails, entering step 44; step 43, comparing element information, including whether the numerical information and non-numerical information are consistent, if consistent, entering step 45, if inconsistent, entering step 46; step 44, moving the selected element from the repeating instruction text element to the pending element group, entering step 47; step 45, removing the element information from the repeating instruction text element and the regulation instruction text element respectively, entering step 47; step 46, removing the element information from the repeating instruction text element and the regulation instruction text element respectively, and storing as an inconsistent element pair, entering step 47; step 47, judging whether the repeating instruction text element is empty, if not, returning to step 41, otherwise, entering step 48; step 48, judging whether the pending element group is empty, if not, entering step 49, otherwise, entering step 410; step 49, judging whether the instruction text element is empty, if empty, moving the pending element to the inconsistent element, entering step 413; otherwise, entering step 411; step 410, judging whether the regulation instruction text element is empty, if not, entering step 412, if yes, entering step 413; step 411, detecting the similarity of the pending element and the remaining instruction element, constructing an inconsistent information pair in the inconsistent element according to the most similar matching mode, and entering step 412; step 412, putting the last remaining instruction element into the inconsistent element, entering step 413; step 413, detecting whether there is an inconsistent element, if yes, judging that the repeating is inconsistent, and outputting the inconsistent information; otherwise, judging that the repeating is consistent.

2. The method of claim 1, wherein the method is based on intelligent completion of regulatory command restatement consistency alerting. The training step of the SVM model is: obtaining an original regulation voice text as a training sample; performing type labeling on the training sample, labeling as regulation-repeating type and other type; the original regulation voice text is collected according to the airport / regulation center in the actual application environment; The training sample is initially segmented based on a preposition, punctuation and type segmentation method to obtain a first segmentation result; the first segmentation result is further analyzed by a forward maximum matching method to obtain a second segmentation result; According to the second segmentation result, the training sample is vectorized and represented as a training input data set; An SVM model is constructed, and the training input data set is input into the SVM model for model training to obtain a trained SVM model; The preposition, punctuation and type segmentation method is to segment the content according to the punctuation; Chinese virtual words are identified, and further segmentation is performed at the positions of the Chinese virtual words; the boundaries of numbers, characters and texts are identified, and segmentation is performed at the boundaries; and significant long segmentation is identified, and segmentation is performed according to 2 characters. The forward maximum matching method is to combine the parametric elements and their parameter values after the first segmentation result into one word segmentation, and correct the first segmentation result, so that the word segmentation of the first segmentation result is more in line with the text features of the regulation.

3. The method for detecting consistency alarms in control instruction rephrase based on intelligent completion according to claim 2, characterized in that, The steps of the forward maximum matching method are as follows: Step 11, obtain the maximum word length L of the dictionary; Step 12, set the intercept length N = L, and the start of the intercept is I = 0; Step 13, start from the first segmentation result I position, intercept a string of length N; the string is a string of continuous initial segmentation with a total length of N starting from I, if the total length of i segmentations is less than N and the total length of i+1 segmentations is greater than N, only i segmentations are taken; Step 14, determine whether the string is in the dictionary, if yes, go to step 15, otherwise go to step 16; Step 15, add the intercepted string to the word segmentation result, and go to step 18; Step 16, whether the value of N has reached l0, N = l0 indicates that the current string is composed of only one initial segmentation, l0 is the initial segmentation length, if yes, go to step 15, otherwise go to step 17; Step 17, let N = N - 1 i , i for the length of the last initial cut that makes up the current string, go back to step 13; Step 18, determine whether the word segmentation is finished, if not, set N = L, and move the intercept position to the right by N positions I = I + N, and return to step 13; if yes, go to step 19; Step 19, output the second segmentation result, and end the word segmentation.

4. The method of claim 1, wherein the method is based on intelligent completion of regulatory command restatement consistency alerting. The current conversation is classified as a regulation text by using the SVM model, and the specific steps are as follows: Load the current conversation, i.e., the regulation text information to be classified; Preprocess the regulation text information to obtain word segmentation information; According to the word segmentation information, represent the regulation text information as a vector; Load the SVM model and the hyperplane parameters; Calculate the distance between the representation vector and the hyperplane under the SVM model; If the distance is positive, output the classification as regulation-repeating text, otherwise output the classification as other type text.

5. The method for detecting consistency alarms in control instruction rephrase based on intelligent completion according to claim 1, characterized in that, Identify the regulation elements from the regulation-repeating text to obtain the identified regulation elements, including: Step 21, obtain a word segmentation from the regulation-repeating text, and sequentially search the word segmentation in each dictionary; Step 22, if the word segmentation is detected, mark the word segmentation as the corresponding classification, and go to step 26; otherwise, go to step 23; Step 23, determine whether the word segmentation is all numerical values, if yes, go to step 24; otherwise, go to step 25; Step 24, mark the word segmentation as a number, and go to step 26; Step 24, the reserved word is marked as pending, and step 26 is entered; Step 25, calculate the similarity probability of the word with all regulatory element category definition rules, if there is a regulatory element category that exceeds the definition probability threshold, take the highest probability category as the recognition result; Otherwise, the word is judged to be meaningless, and the word is deleted, and step 26 is entered; Step 26, judge whether the regulatory element recognition is completed, if yes, end the element recognition; Otherwise, return to step 21.

6. The method of claim 5, wherein the method is based on intelligent completion of regulatory command restatement consistency alerting. The calculation of the similar probability in step 35 is to define a set of possible structure rules for each type of control element, and to calculate the matching degree of the unrecognized word segmentation and all rules of each control element; if a type of control element i has k possible structure rules, and the matching degree of the word segmentation and each rule is p ik , then the average matching degree of the two structure rules with the highest comprehensive matching degree is taken as the matching probability p i of the word segmentation to the rules of this type of control element; if a type of control element has a great probability of matching with a certain structure rule, then its great probability is the type of control element; Wherein, the matching probability p of the unrecognized word segmentation and the regulatory element i i The calculation formula is: p max = max(p i1 , p i2 ... p ik ) p dmax = max[(p i1 ,p i2 ...p ik )],p i1 ,p i2 ...p ik not including p max p i = (p max + p dmax ) / 2 where p i1 , p i2 ... p ik is the matching probability of the k possible structures of the regulatory element i, p max and p dmax are the maximum and second maximum values of the matching probabilities, p i is the matching probability of the identified token with the regulatory element i.

7. The method for detecting consistency alarms in control instruction rephrase based on intelligent completion according to claim 1, characterized in that, According to the regulation-repeating text and the regulatory element, the simplified and omitted information in the regulation-repeating text is text information completed, including: According to the regulation-repeating text and the regulatory element, the simplified and omitted information in the regulation-repeating text is completed by the controller's regulation part, and the specific steps are: Step 311, take an element from the regulation instruction text element, enter step 312; Step 312, according to the air-ground communication standard, judge whether there is a simplified description according to the element type to which the element belongs, if yes, enter step 313, otherwise enter step 314; Step 313, determine the simplified type and complete, the completion is based on the air-ground communication standard and the common simplification in actual regulation to build the corresponding mapping relationship; Step 314, judge whether the completion is completed, if yes, end, otherwise return to step 311.

8. The method for detecting consistency alarms in control instruction rephrase based on intelligent completion according to claim 7, characterized in that, According to the regulation-repeating text and the regulatory element, the simplified and omitted information in the regulation-repeating text is text information completed, including: According to the regulation-repeating text and the regulatory element, the simplified and omitted information in the regulation-repeating text is completed by the controller's regulation part, and the specific steps are: Step 321, take an element from the regulation instruction text element, enter step 322; Step 322, judge whether it is a pure numerical value, if yes, enter the pending group and return to step 321; Otherwise, enter step 323; Step 323, according to the air-ground communication standard, judge whether there is a simplified description, if yes, enter step 324, otherwise enter step 325; Step 324, determine the simplified type and complete; The completion is based on the air-ground communication standard and the common simplification in actual regulation to build the corresponding mapping relationship, enter step 325; Step 325, the similarity of the element and all elements in the regulation part of the regulation-repeating text is calculated by using a text similarity calculation method, the most similar element is selected, if there is, it goes to step 326, otherwise it goes to step 327; a similarity threshold K is set, i represents the i th element in the instruction text, and the matching rule is selected as: max(γ i )>K, γ i is the similarity of the element and the i th element in the repeated instruction. Step 326, save the information in the element unchanged, complete the information in the matching element in front and back, complete the information, and enter step 327 after completion: Let the element be K, length k; the matched regulatory element be L, length l; l>k; the first character of K is the same as the i-th character of L, the completion rule is: K=L0..L i-1 KL l-i-k ..L l Wherein, L i-1 is the content located at i-1 in the regulatory element; Step 327, judge whether the completion is completed, if yes, end, otherwise return to step 321.

9. The method of claim 1, wherein the method is a method of detecting compliance of a regulatory command restatement, based on intelligent completion. The similarity calculation formula of step 411 similarity detection is: wherein x, y represent two element information in the inconsistent element pair respectively; i, j represent corresponding character subscripts of x, y; lev x,y (i,j) means the edit distance of the first i characters of element x and the first j characters of element y, min(i,j)=0 means that one of the corresponding character subscripts i, j of x, y is zero, and max(i,j) means taking the larger value of the corresponding character subscripts i, j of x, y as the edit distance value of the first i characters of element x and the first j characters of element y.

Images