CN109766554A - A kind of land sky call based on interactive mode is rehearsed error classification method - Google Patents

A kind of land sky call based on interactive mode is rehearsed error classification method Download PDF

Info

Publication number
CN109766554A
CN109766554A CN201910027246.0A CN201910027246A CN109766554A CN 109766554 A CN109766554 A CN 109766554A CN 201910027246 A CN201910027246 A CN 201910027246A CN 109766554 A CN109766554 A CN 109766554A
Authority
CN
China
Prior art keywords
matching
rehearsing
vector
rehearsed
matrix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201910027246.0A
Other languages
Chinese (zh)
Inventor
贾桂敏
程方圆
杨金锋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Civil Aviation University of China
Original Assignee
Civil Aviation University of China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Civil Aviation University of China filed Critical Civil Aviation University of China
Priority to CN201910027246.0A priority Critical patent/CN109766554A/en
Publication of CN109766554A publication Critical patent/CN109766554A/en
Pending legal-status Critical Current

Links

Landscapes

  • Machine Translation (AREA)

Abstract

A kind of land sky call based on interactive mode is rehearsed error classification method.It includes production corpus: according to civil aviaton's transmission standard and the dedicated vocabulary of language material storage preparation and obtaining word one-hot term vector, obtains enhancing one-hot term vector to one-hot vector simple modifications；The sentence matrix of each two sentences of centering of rehearsing is generated using enhancing one-hot term vector；Each two sentence matrixes of centering of rehearsing are interacted to obtain rehearse pair between matching matrix；Matching vector between the matching position and matching semantic feature for extracting matching matrix obtain rehearsing pair；Classified using full softmax layer of connection to matching vector, thus completes land sky and converse error classification of rehearsing.The invention has the advantages that the advantage that CNN learns matching characteristic automatically is utilized, the expense of artificial design features can be removed from, can realize that sky call in land is rehearsed error classification automatically, thus the present invention do not need it is for statistical analysis to mass data.

Description

A kind of land sky call based on interactive mode is rehearsed error classification method
Technical field
The invention belongs to sky calls in land in AIRLINE & AIRPORT to rehearse error classification technical field, more particularly to one kind based on handing over The land sky call of mutual mode is rehearsed error classification method.
Background technique
It is well known that aviation flight safety is always positioned at first place in civil aviaton's cause.In AIRLINE & AIRPORT, accident is caused to be sent out Raw factor can be summarized as three classes: natural cause, mechanical breakdown and human factor.With the development of science and technology, by natural cause The trend that declines to a great extent is presented with accident rate caused by mechanical breakdown, but accident rate is not as caused by human factor It significantly decreases.
It is shown according to the survey report of US Federal Aviation Administration (NASA), land sky Talker Communication Error is the main original in human factor Cause.Wherein, land sky call rehearses wrong proportion close to half.It is rehearsed aviation caused by mistake to reduce as the call of land sky The incidence of unsafe incidents, we must grasp the tendentiousness of rehearse error type and error type of rehearsing, this is for improving Communication quality between controller and pilot improves civil aviaton's flight safety and has important practical significance.Previous research is equal It is by for statistical analysis with error analysis work of completing to rehearse to a large amount of land sky communicating data, this not only needs to expend greatly The man power and material of amount, and it is easy to appear when data volume is excessive the situation of error analysis.Therefore, an automatic land is established Sky call error classification model of rehearsing is particularly important.
Convolutional neural networks (CNN) in deep learning have good application in computer vision field, in recent years Also constantly it is used in natural language processing field, such as text matches, text classification etc..Since CNN has parameter sharing Advantage, thus it is few compared to the time of other methods consumption, meet and detects the characteristics of rehearsing error type in time.But at present simultaneously Do not find CNN for the report that sky call in land is rehearsed in terms of error classification task.
Summary of the invention
To solve the above-mentioned problems, the land sky call that the purpose of the present invention is to provide a kind of based on interactive mode is rehearsed difference Misclassification method.
In order to achieve the above object, the land sky call provided by the invention based on interactive mode is rehearsed error classification method packet Include the following steps carried out in order:
Step 1) makes corpus；
Transmission standard and above-mentioned language material storage preparation dedicated vocabulary and the one-hot word that obtains word of the step 2) according to civil aviaton Vector, and one-hot vector is simply improved and obtains enhancing one-hot term vector；
Step 3) generates the sentence matrix of each two sentences of centering of rehearsing using enhancing one-hot term vector；
Each two sentence matrixes of centering of rehearsing are interacted the matching matrix for obtaining rehearsing between by step 4)；
Step 5) extracts the matching position and matching semantic feature and the matching that obtains rehearsing between of above-mentioned matching matrix Vector；
Step 6) classifies to above-mentioned matching vector using full softmax layers of a connection, thus completes the call of land sky It rehearses error classification.
In step 1), the method for the production corpus includes the following steps:
Step 1.1) picks out the voices of multiple types of rehearsing to as rehearsing pair, often from true land sky calling record It is a rehearse to the instruction and pilot that are sent by controller rehearse instruction two sentences form, be then store as text txt format；
Step 1.2) is by above-mentioned multiple rehearsing pair to mistake of as positive sample, rehearsing of correctly rehearsing of rehearsing in of rehearsing As negative sample, in negative sample include common error type of rehearsing: course information rehearse mistake, runway information rehearse mistake, Aircraft call sign information rehearse mistake, elevation information rehearse mistake, partial information missing errors；
Step 1.3) is segmented two sentences of centering of rehearsing as unit of each word, by all after segmenting It rehearses to composition corpus.
In step 2), the transmission standard and the dedicated vocabulary of above-mentioned language material storage preparation according to civil aviaton simultaneously obtains word One-hot term vector, and one-hot vector is simply improved obtain enhancing one-hot term vector method Include the following steps:
Step 2.1) is according to one dedicated vocabulary of transmission standard and above-mentioned language material storage preparation of civil aviaton；
Step 2.2) obtains the one-hot term vector of word and enhancing one- in corpus using above-mentioned dedicated vocabulary Hot term vector；Wherein, only one position is " 1 " in one-hot term vector, and other positions are all " 0 ", the dimension of term vector For the length of dedicated vocabulary；Meanwhile flag bit " 1 " being added after one-hot term vector and obtains enhancing one-hot term vector.
In step 3), the sentence that each two sentences of centering of rehearsing are generated using enhancing one-hot term vector The method of matrix is: respectively carrying out the enhancing one-hot term vector of all words in each two sentences of centering of rehearsing by row Arrangement obtains two sentence matrix SPAnd SATC。
In step 4), described interact each two sentence matrixes of centering of rehearsing obtains rehearsing between The method of matching matrix is:
Above-mentioned two sentence matrix S is calculated using similarity functionPAnd SATCIn two-by-two the similarity between term vector and obtain To a two-dimensional matrix, referred to as matching matrix；
The similarity function is Indicator function and Dot Product function, the calculating of Indicator function Shown in formula such as formula (1):
In above formula, MijRepresent viAnd wjSimilarity value, viAnd wjRespectively represent sentence matrix SPAnd SATCMiddle position For the term vector of the i-th row and the word of jth row,Indicate the cosine similarity between word, | | | | it indicates Vector
Dot Product function directly calculates the inner product between two term vectors, the calculation formula of Dot Product function As shown in formula (2):
Mij=(vi)T(wj)(2)。
In step 5), the matching position and matching semantic feature of the described above-mentioned matching matrix of extraction and obtaining is rehearsed pair Between the method for matching vector be:
The matching position of matching matrix is extracted using the convolution sum pondization operation of CNN and is matched semantic feature and is rehearsed Matching vector between；
The CNN is made of input layer, convolutional layer and the pond Max layer, and the input of input layer is the matching in step 4) Matrix, convolutional layer using size be " k × k " convolution kernel to matching matrix carry out matching position and matching semantic feature extraction and Characteristic pattern is obtained, the maximum value of data in the region " p × p " in the keeping characteristics figure of the pond Max is then carried out by the pond Max layer；CNN Convolutional layer carry out matching position and match semantic feature extraction formula such as formula (3) shown in:
In above formula, zlIndicate l layers of output；σ indicates excitation function；N indicates the selection of input feature vector figure, wherein z0Table Show the matching matrix M of input layer output；Wl indicates l layers of convolution kernel；blIndicate l layers of biasing；
Shown in the formula such as formula (4) for carrying out the pond Max in the pond the Max layer of CNN:
Finally, by matching matrix by the operation of convolution sum pondization later obtained characteristic pattern connect into one it is one-dimensional to Amount, this vector is matching vector.
In step 6), described classifies to above-mentioned matching vector using full softmax layers of a connection, thus complete It is at the rehearse method of error classification of land sky call:
Full softmax layers of connection is error classification of being rehearsed using the full Connection Neural Network of single layer and the completion of softmax function, Shown in the formula such as formula (5) of the full Connection Neural Network of single layer:
X=(x1,x2,...,x6)T=W1·z+b1(5)
In above formula, the full Connection Neural Network output x of single layer indicates that input is rehearsed to the matching score for belonging to every one kind, W1Table Show the weight matrix of full softmax layers of connection；Z indicates the matching vector of CNN output；b1Indicate the inclined of full softmax layers of connection It sets；
In above formula, and p (y=j | x) indicate that input is rehearsed to the probability for belonging to matching classification j；The output of softmax function is One-dimensional vector, using the corresponding index value of most probable value in this vector as the label of prediction.
Land sky call error classification method of rehearsing provided by the invention based on interactive mode has the advantages that
The advantage that CNN learns matching characteristic automatically is utilized, the expense of artificial design features can be removed from, it can be automatically real Existing land sky call is rehearsed error classification, thus the present invention do not need it is for statistical analysis to a large amount of data.Meanwhile present invention benefit With the advantage of CNN parameter sharing, reduce the parameter of model and the complexity of model, to reduce runing time, favorably In the purpose for realizing real-time detection.
Detailed description of the invention
Fig. 1 is that the land sky call provided by the invention based on interactive mode is rehearsed error classification method flow diagram.
Fig. 2 is the matching matrix based on one-hot term vector.
Fig. 3 is the matching matrix based on enhancing one-hot term vector.
Fig. 4 is CNN basic block diagram.
Fig. 5 is that the land sky call based on interactive mode is rehearsed error classification model.
Fig. 6 is influence of the convolution kernel size for experimental result.
Fig. 7 is the test accuracy rate of 30 random experiments based on interactive mode.
Specific embodiment
The land sky call to provided by the invention based on interactive mode is rehearsed mistake in the following with reference to the drawings and specific embodiments Classification method is described in detail.
As shown in Figure 1, the land sky call provided by the invention based on interactive mode is rehearsed, error classification method includes by suitable The following steps that sequence carries out:
Step 1) makes corpus:
Step 1.1) picks out the voices of multiple types of rehearsing to as rehearsing pair, often from true land sky calling record It is a rehearse to the instruction and pilot that are sent by controller rehearse instruction two sentences form, be then store as text txt format；
Step 1.2) is by above-mentioned multiple rehearsing pair to mistake of as positive sample, rehearsing of correctly rehearsing of rehearsing in of rehearsing As negative sample, in negative sample include common error type of rehearsing: course information rehearse mistake, runway information rehearse mistake, Aircraft call sign information rehearse mistake, elevation information rehearse mistake, partial information missing errors；
Step 1.3) all is rehearsed to segmenting to above-mentioned: in order to facilitate the semanteme of descriptive statement, being needed to rehearsing pair It is segmented, is segmented two sentences of centering of rehearsing as unit of each word, rehearsed pair by all after segmenting Constitute corpus；
Transmission standard and above-mentioned language material storage preparation dedicated vocabulary and the one-hot word that obtains word of the step 2) according to civil aviaton Vector, and one-hot vector is simply improved and obtains enhancing one-hot term vector:
Step 2.1) is according to one dedicated vocabulary of transmission standard and above-mentioned language material storage preparation of civil aviaton；
Step 2.2) obtains the one-hot term vector of word and enhancing one- in corpus using above-mentioned dedicated vocabulary Hot term vector；Wherein, only one position is " 1 " in one-hot term vector, and other positions are all " 0 ", the dimension of term vector For the length of dedicated vocabulary；Meanwhile flag bit " 1 " being added after one-hot term vector and obtains enhancing one-hot term vector, Enhancing one-hot term vector can increase the Words similarity between keyword and keyword abbreviation.Keyword abbreviation exists It is ever-present in the call of controller and pilot, such as: amendment/sea is pressed, pilot or controller may Contracting reads to repair/press.
Step 3) generates the sentence matrix of each two sentences of centering of rehearsing using enhancing one-hot term vector:
Respectively the enhancing one-hot term vector of all words in each two sentences of centering of rehearsing arrange by row To two sentence matrix SPAnd SATC, this sentence matrix is similar to one " picture "；
Step 4) is by each two sentence matrix S of centering that rehearsePAnd SATCInteract the matching for obtaining rehearsing between Matrix；
Above-mentioned two sentence matrix S is calculated using similarity functionPAnd SATCIn two-by-two the similarity between term vector and obtain To a two-dimensional matrix, referred to as matching matrix；
The similarity function is Indicator function and Dot Product function, the calculating of Indicator function Shown in formula such as formula (1):
In above formula, MijRepresent viAnd wjSimilarity value, viAnd wjRespectively represent sentence matrix SP, SATCMiddle position is The term vector of the word of i-th row and jth row,Indicate word between cosine similarity, | | | | indicate to Amount
Dot Product function directly calculates the inner product between two term vectors, the calculation formula of Dot Product function As shown in formula (2):
Mij=(vi)T(wj) (2)
Step 5) extracts the matching position and matching semantic feature and the matching that obtains rehearsing between of above-mentioned matching matrix The method of vector is:
The matching position of matching matrix is extracted using the convolution sum pondization operation of CNN and is matched semantic feature and is rehearsed Matching vector between；
The CNN is made of input layer, convolutional layer and the pond Max layer, and the input of input layer is the matching in step 4) Matrix, convolutional layer using size be " k × k " convolution kernel to matching matrix carry out matching position and matching semantic feature extraction and Characteristic pattern is obtained, the maximum value of data in the region " p × p " in the keeping characteristics figure of the pond Max is then carried out by the pond Max layer；CNN Convolutional layer carry out matching position and match semantic feature extraction formula such as formula (3) shown in:
In above formula, zlIndicate l layers of output；σ indicates excitation function；N indicates the selection of input feature vector figure, wherein z0Table Show the matching matrix M of input layer output；WlIndicate l layers of convolution kernel；blIndicate l layers of biasing；
Shown in the formula such as formula (4) for carrying out the pond Max in the pond the Max layer of CNN:
The pond Max is the maximum value of data in the region " p × p " in keeping characteristics figure, this can be extracted to matching relationship The best part is contributed, finally, the characteristic pattern that matching matrix obtains after the operation of convolution sum pondization is connected into one one The vector of dimension, this vector are matching vector；
Step 6) classifies to above-mentioned matching vector using full softmax layers of a connection, thus completes the call of land sky It rehearses error classification；
Full softmax layers of connection is error classification of being rehearsed using the full Connection Neural Network of single layer and the completion of softmax function, Shown in the formula such as formula (5) of the full Connection Neural Network of single layer:
X=(x1,x2,...,x6)T=W1·z+b1 (5)
In above formula, the full Connection Neural Network output x of single layer indicates that input is rehearsed to the matching score for belonging to every one kind, W1Table Show the weight matrix of full softmax layers of connection；Z indicates the matching vector of CNN output；b1Indicate the inclined of full softmax layers of connection It sets.
In above formula, and p (y=j | x) indicate that input is rehearsed to the probability for belonging to matching classification j；The output of softmax function is One-dimensional vector, using the corresponding index value of most probable value in this vector as the label of prediction.
Experiment and analysis
Experimental data of the invention includes 5000 and rehearses pair, will wherein 2500 rehearse and correctly rehearse to as positive sample This, remaining 2500 mistakes of rehearsing are rehearsed to including that following 5 class is common in the negative sample for mistake of as negative sample, rehearsing Rehearse error type: course information rehearse mistake, aircraft call sign information of mistake, runway information of rehearsing is rehearsed mistake, elevation information Rehearse mistake, partial information missing errors；Wherein, the sample number of every class type of error is 500.It is described according to step 2) One-hot term vector and enhancing one-hot term vector, the present inventor depict two kinds of differences in table 1 rehearse type based on One-hot term vector and based on enhancing one-hot term vector matching matrix, as shown in Figures 2 and 3.It is described to be based on interaction Mode land sky call rehearse error classification method model as shown in figure 5, the present inventor is referred to as Match-CNN model.This Inventor selects 4000 to rehearse to for training above two model (wherein to rehearse and correctly rehearse to as positive sample for 2000 This, 2000 wrong rehearse to as negative sample of rehearsing), 1000 are rehearsed to for testing above-mentioned model, (500 are rehearsed just True to rehearse to as positive sample, 500 mistakes of rehearsing are rehearsed to as negative sample).
In order to more accurately measure the performance of above two model, present inventor has performed 30 random training and survey Examination, training and test randomly extract 4000 and rehearse to being trained every time, and 1000 are rehearsed to for testing.Pass through 30 After secondary experiment, the present inventor assesses effectiveness of the invention using the average test accuracy rate of 30 random experiments.Test Shown in the formula such as formula (7) of accuracy rate (Acc.):
In order to verify the stability of model, the present inventor calculates the mean square error of the test accuracy rate of 30 random experiments Difference.Shown in mean square error (MSE) formula such as formula (8):
Meanwhile in order to analyze the classifying quality of different error types, the present inventor calculates Different matching class using formula (9) Other F value.
In above formula, P (Precision) is measuring accuracy, and R (Recall) is recall rate.
The present inventor tests the method for the present invention, explores convolution kernel size during the experiment and experiment is tied The influence of fruit, experimental result are as shown in Figure 6.After determining convolution kernel size, by adjusting other network parameters, it is determined that more excellent Network model.Land sky call based on interactive mode is rehearsed the experimental result such as table 2, table 3 and Fig. 7 institute of error classification method Show.In table 3, semantic congruence that Correct of rehearsing is rehearsed pair；Type of error 1 respectively indicates following 5 class mistake to type of error 5 Accidentally: course information rehearse mistake, runway information rehearse mistake, aircraft call sign information rehearse mistake, elevation information rehearse mistake, portion Divide loss of learning mistake.
Table 1. is rehearsed to citing
The average test accuracy rate of 2. 30 random experiments of table
The F value of the classification of respectively rehearsing of land sky call error classification method of the table 3. based on interactive mode
When it can be seen from Fig. 3 and Fig. 4 by the semanteme of one-hot term vector expression word, the matching of different types of rehearsing Matrix may be consistent, and when the semanteme of use enhancing one-hot term vector expression word, the matching matrix of different types of rehearsing Otherness is bigger.
As seen from Figure 5: (1) convolution kernel size has a certain impact for test accuracy rate of the invention, Er Qiejuan Product core size increases, and the test accuracy rate of model increases.This is because the size of convolution kernel increases, receptive field increases, CNN The feature of extraction is more abundant.(2) it is all unfavorable for model that convolution kernel is oversized or too small.This is because working as convolution When core is smaller, trained model can be too sensitive to local feature and introduces noise；And when convolution kernel is excessive, trained model The detail section of matching matrix can be ignored.Therefore, the convolution kernel that the present invention designs is having a size of " 12 × 12 ".(3) using enhancing The experimental result that the semanteme of one-hot term vector expression word obtains is better than the semanteme using one-hot term vector expression word Obtained experimental result, this is because enhancing one-hot term vector not only can preferably express the semanteme of word, but also can be with Increase the otherness between different classes of.
The land sky call provided by the invention based on interactive mode is rehearsed error classification it can be seen from table 2, table 3 and Fig. 7 Method can effectively complete error classification task of rehearsing, and relatively stable by the model of the method training.Meanwhile it can be with The test accuracy rate for finding out that enhancing one-hot term vector and Dot Product function combine is higher, this is because Dot Product function can extract richer semantic feature.Demonstrate through a large number of experiments the method for the present invention practicability and Feasibility.

Claims (7)

  1. A kind of error classification method 1. land sky call based on interactive mode is rehearsed, it is characterised in that: the method includes pressing The following steps that sequence carries out:
    Step 1) makes corpus；
    Step 2) according to the transmission standard and the dedicated vocabulary of above-mentioned language material storage preparation of civil aviaton and obtain the one-hot word of word to Amount, and one-hot vector is simply improved and obtains enhancing one-hot term vector；
    Step 3) generates the sentence matrix of each two sentences of centering of rehearsing using enhancing one-hot term vector；
    Each two sentence matrixes of centering of rehearsing are interacted the matching matrix for obtaining rehearsing between by step 4)；
    Step 5) extracts the matching position and matching semantic feature and the matching vector that obtains rehearsing between of above-mentioned matching matrix；
    Step 6) classifies to above-mentioned matching vector using full softmax layers of a connection, thus completes the call of land sky and rehearses Error classification.
  2. The error classification method 2. the land sky call according to claim 1 based on interactive mode is rehearsed, it is characterised in that: In step 1), the method for the production corpus includes the following steps:
    Step 1.1) picks out the voices of multiple types of rehearsing to as rehearsing pair from true land sky calling record, Mei Gefu Reading aloud to rehearse to the instruction and pilot that are sent by controller instructs two sentences to form, and is then store as text txt format；
    Step 1.2) is by above-mentioned multiple rehearsing to conduct to mistake of as positive sample, rehearsing of correctly rehearsing of rehearsing in of rehearsing Negative sample includes common error type of rehearsing in negative sample: rehearse mistake, runway information of course information is rehearsed mistake, aircraft Call sign information rehearse mistake, elevation information rehearse mistake, partial information missing errors；
    Step 1.3) is segmented two sentences of centering of rehearsing as unit of each word, is rehearsed by all after segmenting To composition corpus.
  3. The error classification method 3. the land sky call according to claim 1 based on interactive mode is rehearsed, it is characterised in that: In step 2), the transmission standard and the dedicated vocabulary of above-mentioned language material storage preparation according to civil aviaton and the one-hot for obtaining word Term vector, and simply being improved one-hot vector the method for obtaining enhancing one-hot term vector includes following step It is rapid:
    Step 2.1) is according to one dedicated vocabulary of transmission standard and above-mentioned language material storage preparation of civil aviaton；
    Step 2.2) obtains the one-hot term vector of word and enhancing one-hot word in corpus using above-mentioned dedicated vocabulary Vector；Wherein, only one position is " 1 " in one-hot term vector, and other positions are all " 0 ", and the dimension of term vector is dedicated The length of vocabulary；Meanwhile flag bit " 1 " being added after one-hot term vector and obtains enhancing one-hot term vector.
  4. The error classification method 4. the land sky call according to claim 1 based on interactive mode is rehearsed, it is characterised in that: In step 3), the method for the sentence matrix that each two sentences of centering of rehearsing are generated using enhancing one-hot term vector It is: is arranged to obtain two by row by the enhancing one-hot term vector of all words in each two sentences of centering of rehearsing respectively A sentence matrix SPAnd SATC。
  5. The error classification method 5. the land sky call according to claim 1 based on interactive mode is rehearsed, it is characterised in that: It is described that each two sentence matrixes of centering of rehearsing are interacted into the matching matrix for obtaining rehearsing between in step 4) Method is:
    Above-mentioned two sentence matrix S is calculated using similarity functionPAnd SATCIn two-by-two the similarity between term vector and obtain one A two-dimensional matrix, referred to as matching matrix；
    The similarity function is Indicator function and Dot Product function, the calculation formula of Indicator function As shown in formula (1):
    In above formula, MijRepresent viAnd wjSimilarity value, viAnd wjRespectively represent sentence matrix SPAnd SATCMiddle position is i-th The term vector of capable and jth row word,Indicate the cosine similarity between word, | | | | indicate vector
    Dot Product function directly calculates the inner product between two term vectors, the calculation formula such as formula of Dot Product function (2) shown in:
    Mij=(vi)T(wj) (2)。
  6. The error classification method 6. the land sky call according to claim 1 based on interactive mode is rehearsed, it is characterised in that: In step 5), the matching position and matching semantic feature of the described above-mentioned matching matrix of extraction and the matching that obtains rehearsing between The method of vector is:
    The matching position of matching matrix is extracted using the convolution sum pondization operation of CNN and is matched semantic feature and is obtained rehearsing to it Between matching vector；
    The CNN is made of input layer, convolutional layer and the pond Max layer, and the input of input layer is the matching matrix in step 4), Convolutional layer carries out matching position and matching semantic feature extraction to matching matrix using the convolution kernel that size is " k × k " and obtains Then characteristic pattern is carried out the maximum value of data in the region " p × p " in the keeping characteristics figure of the pond Max by the pond Max layer；The volume of CNN Lamination carries out matching position and matches shown in the formula such as formula (3) of semantic feature extraction:
    In above formula, zlIndicate l layers of output；σ indicates excitation function；N indicates the selection of input feature vector figure, wherein z0Indicate defeated Enter the matching matrix M of layer output；WlIndicate l layers of convolution kernel；blIndicate l layers of biasing；
    Shown in the formula such as formula (4) for carrying out the pond Max in the pond the Max layer of CNN:
    Finally, the characteristic pattern that matching matrix obtains after the operation of convolution sum pondization is connected into an one-dimensional vector, this A vector is matching vector.
  7. The error classification method 7. the land sky call according to claim 1 based on interactive mode is rehearsed, it is characterised in that: In step 6), described classifies to above-mentioned matching vector using full softmax layers of a connection, thus completes the call of land sky The method for error classification of rehearsing is:
    Full softmax layers of connection is error classification of being rehearsed using the full Connection Neural Network of single layer and the completion of softmax function, single layer Shown in the formula such as formula (5) of full Connection Neural Network:
    X=(x1,x2,...,x6)T=W1·z+b1 (5)
    In above formula, the full Connection Neural Network output x of single layer indicates that input is rehearsed to the matching score for belonging to every one kind, W1Indicate complete The weight matrix of softmax layers of connection；Z indicates the matching vector of CNN output；b1Indicate the biasing of full softmax layers of connection；
    In above formula, and p (y=j | x) indicate that input is rehearsed to the probability for belonging to matching classification j；The output of softmax function is one-dimensional Vector, using the corresponding index value of most probable value in this vector as the label of prediction.
CN201910027246.0A 2019-01-11 2019-01-11 A kind of land sky call based on interactive mode is rehearsed error classification method Pending CN109766554A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910027246.0A CN109766554A (en) 2019-01-11 2019-01-11 A kind of land sky call based on interactive mode is rehearsed error classification method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910027246.0A CN109766554A (en) 2019-01-11 2019-01-11 A kind of land sky call based on interactive mode is rehearsed error classification method

Publications (1)

Publication Number Publication Date
CN109766554A true CN109766554A (en) 2019-05-17

Family

ID=66452695

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910027246.0A Pending CN109766554A (en) 2019-01-11 2019-01-11 A kind of land sky call based on interactive mode is rehearsed error classification method

Country Status (1)

Country Link
CN (1) CN109766554A (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106156848A (en) * 2016-06-22 2016-11-23 中国民航大学 A kind of land based on LSTM RNN sky call semantic consistency method of calibration
CN106227718A (en) * 2016-07-18 2016-12-14 中国民航大学 Land based on CNN sky call semantic consistency method of calibration
CN108520298A (en) * 2018-04-09 2018-09-11 中国民航大学 A kind of land sky call semantic consistency method of calibration based on improvement LSTM-RNN
CN108920464A (en) * 2018-07-03 2018-11-30 中国民航大学 A kind of land sky call based on CNN is rehearsed error classification method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106156848A (en) * 2016-06-22 2016-11-23 中国民航大学 A kind of land based on LSTM RNN sky call semantic consistency method of calibration
CN106227718A (en) * 2016-07-18 2016-12-14 中国民航大学 Land based on CNN sky call semantic consistency method of calibration
CN108520298A (en) * 2018-04-09 2018-09-11 中国民航大学 A kind of land sky call semantic consistency method of calibration based on improvement LSTM-RNN
CN108920464A (en) * 2018-07-03 2018-11-30 中国民航大学 A kind of land sky call based on CNN is rehearsed error classification method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
LIANG PANG 等: "Text Matching as Image Recognition", 《ARXIV》 *
李丹 等: "陆空通话复诵语义自动化校验BiLSTM模型", 《信号处理》 *
杨金锋 等: "基于深度 CNN 的陆空通话语义一致性校验", 《中国民航大学学报》 *

Similar Documents

Publication Publication Date Title
CN109918491B (en) Intelligent customer service question matching method based on knowledge base self-learning
CN108717406B (en) Text emotion analysis method and device and storage medium
CN109213999A (en) A kind of subjective item methods of marking
US10922991B2 (en) Cluster analysis of participant responses for test generation or teaching
CN106156848B (en) A kind of land sky call semantic consistency method of calibration based on LSTM-RNN
CN110135459B (en) Zero sample classification method based on double-triple depth measurement learning network
CN110019843A (en) The processing method and processing device of knowledge mapping
CN110826337A (en) Short text semantic training model obtaining method and similarity matching algorithm
CN106407333A (en) Artificial intelligence-based spoken language query identification method and apparatus
CN107704456B (en) Identification control method and identification control device
CN109902285B (en) Corpus classification method, corpus classification device, computer equipment and storage medium
CN110222560B (en) Text person searching method embedded with similarity loss function
CN106227718A (en) Land based on CNN sky call semantic consistency method of calibration
CN110096711A (en) The natural language semantic matching method of the concern of the sequence overall situation and local dynamic station concern
CN109711465A (en) Image method for generating captions based on MLL and ASCA-FR
CN105701514A (en) Multi-modal canonical correlation analysis method for zero sample classification
CN110717341B (en) Method and device for constructing old-Chinese bilingual corpus with Thai as pivot
CN110765254A (en) Multi-document question-answering system model integrating multi-view answer reordering
CN110276069A (en) A kind of Chinese braille mistake automatic testing method, system and storage medium
CN104899335A (en) Method for performing sentiment classification on network public sentiment of information
CN107145573A (en) The problem of artificial intelligence customer service robot, answers method and system
CN109800309A (en) Classroom Discourse genre classification methods and device
CN106227836B (en) Unsupervised joint visual concept learning system and unsupervised joint visual concept learning method based on images and characters
CN110096572A (en) A kind of sample generating method, device and computer-readable medium
CN106776880A (en) A kind of paper based on picture and text identification reviews system and method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20190517

WD01 Invention patent application deemed withdrawn after publication