A deep disaster emergency plan intelligent semantic matching method based on FECF

Abstract

The present invention relates to a deep disaster plan intelligent semantic matching method based on FECF, which is a method for semantic matching of disaster plans, belongs to the field of natural language processing and deep learning, and is characterized by adopting the following steps: (1) modeling the disaster plan problem; (2) determining the feature vector of a word; (3) determining the lexical context encoding, the implicit state output of the t position; (4) determining the attention information of the text; (5) determining the matching score of two sentences; (6) determining the model training of the sentence. The present invention effectively solves the problem of long disaster plan matching time, greatly saves disaster plan matching time, improves the efficiency of disaster plans, effectively improves the problem of inaccurate semantic matching of disaster plan texts, greatly improves the accuracy of disaster plan matching, and provides a scientific and effective basis for post-disaster processing.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and deep learning, and is a method for semantic matching of disaster emergency plans. Background Art

[0002] Disasters are highly uncertain and unpredictable, so the focus of disaster treatment is on scientific aftermath after the disaster. As long as a complete disaster emergency response plan is established in advance, the serious consequences of sudden disasters can be effectively reduced. The reasonable implementation of the post-disaster plan depends largely on the accuracy of the disaster plan selection. By implementing a reasonable disaster plan, the serious consequences of disasters can be greatly reduced. In addition, the possible occurrence of disasters has temporal uncertainty and randomness of location. The consequences of disasters will produce different results with the changes of different time and space. Therefore, how to choose a suitable disaster plan has become a very important issue. Semantic matching of disaster plans can assist in the accurate selection of disaster plans. Text semantic matching is usually achieved by segmenting keywords and then analyzing and judging the degree of semantic correlation between two texts. Traditional text matching based on machine learning requires manual definition and extraction of text features, with fewer parameters, poor generalization ability, and low efficiency, which cannot meet the needs of high efficiency and accuracy. There is a high probability of mismatching.

[0003] In the problem of intelligent semantic matching of disaster plans, the focus is on the efficiency and accuracy of semantic matching. Due to the diversity and uncertainty of disaster situations, higher requirements are placed on the semantic matching of disaster plans. Therefore, in order to achieve accurate semantic matching of disaster plans and ensure the efficiency and stability of post-disaster plan implementation, an intelligent semantic matching method based on FECF is proposed, which reduces the time of semantic matching and improves the matching accuracy of disaster plans, greatly saving the time from the occurrence of disasters to the reasonable selection of plans. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a deep disaster emergency plan intelligent semantic matching method based on FECF, and the specific process is as follows: Figure 1 shown.

[0005] The technical solution implementation steps are as follows:

[0006] (1) Modeling the disaster response plan problem:

[0007] score(X k ,Y k )=F(φ(X k ),φ(Y k ))

[0008] In the formula, and The text X k and Y k where φ(·) represents the function of word vector representation of text, F(·) is the matching model for calculating the semantic matching score of two texts, m is the length of sentence X, and n is the length of sentence Y.

[0009] (2) Determine word x i The eigenvector of

[0010]

[0011] In the formula, i is the counting unit, E tr is the trainable word vector matrix, E fix is a non-trainable word vector matrix, Conv is a one-dimensional convolution operation, and the above steps are used to extract the vocabulary feature vector for each word in the sentence. is a trainable word vector, is a fixed word vector, is the word preceding encoding vector, is the word context encoding vector.

[0012] (3) Determine the lexical context encoding, the hidden state output h at position t t :

[0013]

[0014] In the formula, is the forward output of the Bi-LSTM model at position t, is the reverse output of the Li-LSTM model at position t, x t is the input at position t, (·) T is the transpose operation.

[0015] (4) Determine the attention information of the text

[0016]

[0017] In the formula, e i,j =cos(h xi ,h yj ), n is the length of statement Y, k is the counting unit, e i,k For x i Position and y k Cosine similarity of position implicit output, e i,j For x iPosition and y j Cosine similarity of position implicit output, and x i Position and y j Implicit output of position, α i,j For x i Relative y j The attention score, For x i Attention vector representation relative to Y. Concatenating attention-based context vectors and after triggering As the output of the next layer Similarly, determine the next level output of sentence Y Modeling the interactive information between different locations.

[0018] (5) Determine the matching score r of the two statements:

[0019] The obtained interactive information is passed through the pooling layer and the fully connected layer, and the matching score value is output through a nonlinear function:

[0020] r=f(W r q+b r ),s=W s q+b s

[0021] Where f(·) is the nonlinear activation function, W r , W s are weight functions, b r , b s is the bias vector, q is the new vector composed of the pooled vector arranged in descending order, and s is the score matching value output after linear transformation.

[0022] (6) Determine the model training of the sentence:

[0023] Build a convolutional neural network, input the above preprocessed data into the network, extract features through convolutional layer, pooling layer, and fully connected layer operations, and determine the scores of the two sentences score(X,Y):

[0024] score(X,Y)=s+λs'

[0025] Where s is the matching score of the deep multi-view semantic module, s' is the score of the semantic matching module based on the upper and lower features of the entity, and λ is the balancing factor that adjusts the score ratio of the two modules.

[0026] The present invention has the following advantages over the prior art:

[0027] (1) The present invention effectively solves the problem of long disaster plan matching time, greatly saves disaster plan matching time, improves the efficiency of disaster plans, and provides a solution for post-disaster processing.

[0028] (2) The present invention effectively improves the problem of inaccurate semantic matching of disaster plan texts, greatly improves the accuracy of disaster plan matching, and provides a scientific and effective basis for post-disaster processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to better understand the present invention, further description is given below in conjunction with the accompanying drawings.

[0030] Figure 1 It is a flowchart of the steps to establish the intelligent semantic matching method of deep disaster emergency plan based on FECF;

[0031] Figure 2 It is a flow chart of the intelligent semantic matching method for deep disaster emergency plans based on FECF;

[0032] Figure 3 It is a schematic diagram of the structure of a convolutional neural network;

[0033] Figure 4 It is the result of semantic matching of four groups of disaster emergency plans using the present invention; Specific implementation plan

[0034] The data used in this implementation case comes from a typical demonstration area of ​​underground sheltered space, with a total of 1,000 sets of samples, including 5 scenes: underground tunnels, railway tunnels, underground shopping malls, underground parking lots, and subway stations. Each scene has 200 sets of data. A random sampling method is used to extract 140 sets of samples from each of the 5 scenes as training sets, and the rest as test sets. In the end, the total number of samples used as training sets is 700, and the total number of samples used as test sets is 300.

[0035] The overall flow chart of the deep disaster emergency plan intelligent semantic matching method provided by the present invention is as follows: Figure 1 As shown, the specific steps are as follows:

[0036] (1) Modeling the disaster response plan problem:

[0037] score(X k ,Y k )=F(φ(X k ),φ(Y k ))

[0038] In the formula, and The text X k and Y kwhere φ(·) represents the function of word vector representation of text, F(·) is the matching model for calculating the semantic matching score of two texts, m is the length of sentence X, and n is the length of sentence Y, which are 35 and 55 respectively.

[0039] (2) Determine word x i The eigenvector of

[0040]

[0041] In the formula, i is the counting unit, E tr is the trainable word vector matrix, E fix is a non-trainable word vector matrix, Conv is a one-dimensional convolution operation, and the above steps are used to extract the vocabulary feature vector for each word in the sentence. is a trainable word vector, is a fixed word vector, is the word preceding encoding vector, is the word context encoding vector.

[0042] (3) Determine the lexical context encoding, the hidden state output h at position t t :

[0043]

[0044] In the formula, is the forward output of the Bi-LSTM model at position t, is the reverse output of the Li-LSTM model at position t, x t is the input at position t, (·) T is the transpose operation.

[0045] (4) Determine the attention information of the text

[0046]

[0047] In the formula, n is the length of statement Y, the size is 35, k is the counting unit, e i,k For x i Position and y k Cosine similarity of position implicit output, e i,j For x i Position and y j Cosine similarity of position implicit output, and x i Position and y j Implicit output of position, αi,j For x i Relative y i The attention score, For x i Attention vector representation relative to Y. Concatenating attention-based context vectors and after triggering As the output of the next layer Similarly, determine the next level output of sentence Y Modeling the interactive information between different locations.

[0048] (5) Determine the matching score r of the two statements:

[0049] The obtained interactive information is passed through the pooling layer and the fully connected layer, and the matching score value is output through a nonlinear function:

[0050] r=f(W r q+b r ),s=W s q+b s

[0051] Where f(·) is the nonlinear activation function, W r , W s are weight functions, b r , b s is the bias vector, q is the new vector composed of the pooled vector arranged in descending order, and s is the score matching value output after linear transformation.

[0052] (6) Determine the model training of the sentence:

[0053] Build a convolutional neural network, input the above preprocessed data into the network, extract features through convolutional layer, pooling layer, and fully connected layer operations, and determine the scores of the two sentences score(X,Y):

[0054] score(X,Y)=s+0.35s'

[0055] In the formula, s is the matching score of the deep multi-view semantic module, s' is the score of the semantic matching module based on the upper and lower features of the entity, and λ is the balance factor, which is set to 0.35 to adjust the score ratio of the two modules. In the process of building the model for training, the number of convolutional layers and pooling layers is 300, the number of fully connected layers is 100, the probability of the Dropout layer is 0.2, the activation function uses the tanh function, the number of iterations is 1000, and the model is trained with the above parameters.

[0056] For the trained deep neural network, the model is tested using test set samples to achieve intelligent semantic matching of disaster plans, obtain matching results, and complete the FECF-based deep disaster plan intelligent semantic matching method.

[0057] In order to verify the accuracy of semantic matching of disaster emergency plans in the present invention, four groups of semantic matching experiments were conducted on the present invention. The experimental results are as follows: Figure 4 As shown. Figure 4 It can be seen that the semantic matching accuracy of the disaster plan established by the deep disaster plan intelligent semantic matching method of the present invention is maintained at more than 96%, which can achieve a high accuracy rate on the basis of ensuring stability, and the matching effect is good. This shows that the semantic matching method of the disaster plan established by the present invention is effective, and provides a better method for establishing an accurate disaster plan matching model, which has certain practicality.

Claims

1. A deep disaster plan intelligent semantic matching method based on FECF, characterized by: (1) modeling the disaster plan problem; (2) determining the feature vector of a word; (3) determining the lexical context encoding and the implicit state output at position t; (4) determining the attention information of the text; (5) determining the matching score of two sentences; (6) determining the model training of the sentence, which specifically includes the following six steps: Step 1: Model the disaster preparedness problem: score(X k ,Y k )=F(φ(X k ),φ(Y k )) In the formula, and The text X k and Y k The i-th and j-th words of , φ(·) represents the function of word vector representation of text, F(·) is the matching model for calculating the semantic matching score of two texts, m is the length of sentence X, and n is the length of sentence Y; Step 2: Determine word x i The eigenvector of In the formula, i is the counting unit, E tr is the trainable word vector matrix, E fix is a non-trainable word vector matrix, Conv is a one-dimensional convolution operation, and the above steps are used to extract the vocabulary feature vector for each word in the sentence. is a trainable word vector, is a fixed word vector, is the word preceding encoding vector, Encode the context vector for the word; Step 3: Determine the lexical context encoding, the hidden state output h at position t t : In the formula, is the forward output of the Bi-LSTM model at position t, is the reverse output of the Li-LSTM model at position t, x t is the input at position t, (·) T is the transpose operation; Step 4: Determine the attention information of the text In the formula, n is the length of statement Y, k is the counting unit, e i,k For x i Position and y k Cosine similarity of position implicit output, e i,j For x i Position and y j Cosine similarity of position implicit output, and x i Position and y j Implicit output of position, α i,j For x i Relative y j The attention score, For x i Attention vector representation relative to Y; concatenation of attention-based context vectors and after triggering As the output of the next layer Similarly, determine the next level output of sentence Y Modeling interactive information between different locations; Step 5: Determine the matching score r of the two statements: The obtained interactive information is passed through the pooling layer and the fully connected layer, and the matching score value is output through a nonlinear function: r=f(W r q+b r ),s=W s q+b s Where f(·) is the nonlinear activation function, W r , W s are weight functions, b r 、b s is the bias vector, q is the new vector composed of the pooled vectors arranged in descending order, and s is the score matching value output after linear transformation; Step 6: Determine the model training of the sentence: Build a convolutional neural network, input the above preprocessed data into the network, perform feature extraction through convolutional layer, pooling layer, and fully connected layer operations, and determine the scores of the two sentences score(X,Y): score(X,Y)=s+λs' Where s is the matching score of the deep multi-view semantic module, s' is the score of the semantic matching module based on the upper and lower features of the entity, and λ is the balancing factor that adjusts the score ratio of the two modules.

Images