CN113515930A - Heterogeneous equipment body matching method fusing semantic information - Google Patents

Abstract

A heterogeneous equipment body matching method fusing semantic information comprises the following steps: (1) set sk of instruction fragments₁And a common semantic ontology set sk₂Inputting all elements in the instruction understanding model one by one to generate an instruction intention characterization vector; (2) screening small part of instruction fragment set sk from training data set₁'for making an exact match dataset F'; (3) calculating a similarity matrix S; (4) calculating a mapping matrix F; (5) calculating an objective function f 1; (6) updating the mapping matrix F, and calculating an objective function F2; (7) and (5) repeating the steps (2) to (6) for multiple times until the mapping matrix F achieves accurate matching sk'₁The resulting mapping matrix F will be used to generate the instruction intent dictionary.

Description

Heterogeneous equipment body matching method fusing semantic information

Technical Field

The invention relates to a heterogeneous equipment body matching method fusing semantic information, belongs to the technical field of Internet of things, and particularly belongs to the technical field of heterogeneous equipment body matching.

Background

In the field of internet of things, different devices use different instruction languages to express the same instruction intention, because suppliers intentionally design instruction grammars that are quite different from competitors to increase the switching cost of customers. In addition, the instruction syntax of the device is also strictly protected by the patent. Thus, there is no clear one-to-one correspondence between instruction statements from different vendors, and even different representations for the same term. This results in that it is extremely difficult to manage the network when heterogeneous devices are included in the network. Therefore, how to match the instruction fragments of different devices with a general instruction intention set (namely a semantic ontology) reduces the difficulty in managing heterogeneous device networks, and becomes a technical problem which needs to be solved urgently in the technical field of the internet of things at present.

Disclosure of Invention

In view of this, the invention aims to provide a method for matching an instruction fragment of internet of things equipment with a general semantic ontology based on a deep learning model. In order to achieve the purpose, the invention provides a heterogeneous equipment body matching method fusing semantic information, which is used for matching an instruction fragment of Internet of things equipment with a general semantic body; the semantic ontology refers to a general instruction intention set, and the instruction fragment is a specific instruction when the Internet of things equipment executes an intention; is provided with N₁Instruction fragment set of individual elements

And has N₂Individual element universal semantic ontology set

The method finds sk₁And sk₂Pairs of elements with the same intent in between, each element can only be matched once;

the method specifically comprises the following operation steps:

step S100, set sk of instruction fragment₁And a common semantic ontology set sk₂All elements in the system are input into an instruction understanding model one by one to generate respective instruction intention characterization vectors; the input instruction understanding model comprises an intention description coder and an instruction fragment coder;

step S200, screening a small part of instruction fragment sets sk from the whole training data set₁'for making an exact match dataset F';

step S300, calculating a similarity matrix based on the instruction intention characterization vector

Step S400, calculating a mapping matrix based on the similarity matrix S

Step S500, calculating an objective function F1 based on the similarity matrix S and the mapping matrix F, wherein the calculation result is used for back propagation to update the mapping matrix F;

step S600, updating a corresponding part of a mapping matrix F by using the accurate matching data set F', calculating an objective function F2, and using a calculation result for back propagation to update an instruction understanding model;

step S700, the steps S200 to S600 are repeated for multiple times until the mapping matrix F realizes accurate matching sk'₁(ii) a The resulting mapping matrix F will be used to generate an instruction intent dictionary.

The step S100 includes the following operation sub-steps:

step S110, for the fingerLet the fragment set sk₁And a common semantic ontology set sk₂Each element in the element, and the description text content and the instruction content in the element are respectively input into an intention description coder and an instruction fragment coder;

and step S120, connecting the coding results of the two coders into a vector, wherein the vector is used as an instruction intention characterization vector of the element.

The step S200 specifically includes the following operation substeps:

step S210, calculating the instruction fragment set according to the following formula

Information entropy E (p) of each element in (1)_i)：

Wherein p is_i，jIs an instruction fragment p_iThe frequency of occurrence of the word j in (b) in the whole instruction fragment set;

step S220, calculating the instruction fragment set sk according to the following formula₁All elements in and the general semantic ontology set sk₂Redundancy of all elements in (d):

R_i，j＝max(0，cos(p_i，q_j))

in the above formula, cos (p)_i，q_j) Represents the instruction fragment set sk₁Middle element p_iWith the general semantic ontology set sk₂Middle element q_jThe instructions of the two are intended to represent the cosine distance between the vectors;

step S230, selecting a value quantity index RE using the ratio of the information entropy to the redundancy as a quantization sample, wherein a calculation formula is as follows:

wherein B is a collection of data;

step S240, selecting partial data sk according to the quantized sample selection value index RE₁' an exact match dataset F ' is made, the dataset F ' including the sk₁' exact match result for each instruction fragment in.

The step S300 includes the following operation sub-steps:

step S310, calculating the instruction fragment set sk₁The instruction intention characterization vector and the general semantic ontology set sk of each element₂The instructions of each element in (a) are intended to characterize the euclidean distance between the vectors;

step S320, all calculation results form a similarity matrix

Wherein the element s_i，je.S represents the instruction fragment set sk₁Element p in (1)_iWith the general semantic ontology set sk₂Middle element q_jThe euclidean distance between them.

The specific content of step S400 is as follows: solving the optimal mapping matrix F using a classical algorithm Sinkhorn in optimal transmission theory, the Sinkhom algorithm defining a mapping matrix F ═ diag (u) K · diag (v), where K: e ═ e^-λSIn the formula, λ is a hyper-parameter used for controlling the difference between the elements of the mapping matrix F, and the larger λ is, the smaller the difference is; s is the similarity matrix, diag (u) and diag (v) refer to the remaining 0 squares with vector u and vector v as diagonals, respectively. The vector u and the vector v are calculated based on an iterative mode, and are initialized to be vectors with all 1, and the calculation method comprises the following steps:

wherein

Representing the division of vector elements item by item, a and b respectively representing the normalized vectors of the sum of the similarity matrix S calculated according to rows and columns; the Sinkhorn solution process will iteratively solve for F until convergence or maximum number of steps is reached; mapping matrix

For the element f_i，jE.g. F, if the instruction fragment set sk₁Element p in (1)_iWith the general semantic ontology set sk₂Middle element q_jShould be aligned then f_i，j1, otherwise f_i，j＝0。

The calculation formula of the objective function f1 in step S500 is:

wherein the first item

Is to minimize the sum of euclidean distances between matched segments; second term lambda (-f)_i，jlogf_i，j) Is an additional entropy regularization target, such that even f_i，jNot an integer, namely: f. of_i，j∈[0，1]It can still be made as close to 0 or 1 as possible.

The step S600 includes the following operation sub-steps:

step S610, updating the mapping matrix F by the accurate matching result in the accurate matching data F', and updating the corresponding F in the mapping matrix F_i，jSet to 1, set the other tags to 0;

in step S620, a triple is constructed for each instruction fragment, i.e. for instruction fragment p_iConstruction of a triplet (p)_i，q_pos，q_neg) Wherein p is_iAnd q is_posAre alignable positive samples obtained based on a mapping matrix F, and q_negIs from a general semantic ontology set sk₂Negative samples sampled from, i.e. from, the general semantic ontologyCollection sk₂In order to randomly extract a non-q_posA fragment;

step S630, on the basis of the triplet, by minimizing the target f₂Training the intention description encoder and the instruction fragment encoder simultaneously, the objective function f₂The calculation formula of (2) is as follows:

where dis (·) is a similarity measure between vector representations, which can be computed using euclidean distances.

And step S640, updating the instruction understanding model according to the calculation result.

The instruction intention dictionary in step S700 includes a one-to-one correspondence between the general instructions in the semantic ontology and the specific instruction fragments of a single internet of things device.

The intention description encoder is constructed by a transform layer; each Transformer layer comprises a self-attention layer and a feedforward neural network layer, and the attention mechanism of the self-attention layer is utilized to automatically learn the characteristics with the resolving power so as to meet the training target; the input is a sequence of descriptors with a placeholder "[ cls ]" added at its forefront; for each input word, the output is its vectorized representation obtained by the graph description encoder; the vector representation using "[ cls ]" is characterized as a description;

the instruction fragment encoder uses the same settings and operations as the intent description encoder, the only difference being that the instruction fragment content cannot use a pre-trained language model; the sentence contains many special characters and abbreviations and therefore cannot be parsed directly using a pre-trained language model.

The invention has the beneficial effects that: according to the invention, the instruction fragments of various heterogeneous devices are matched with the universal semantic ontology on the basis of the intention description, so that the difficulty in managing the heterogeneous device network is reduced; according to the method, the model is understood through the training instruction, the similarity of vector representations among elements which can be matched is increased, and the similarity of vector representations of elements which cannot be matched is reduced; the method is used for manufacturing the matching data set by screening a small amount of data, so that the matching performance of the model is ensured while the manufacturing difficulty is reduced.

Drawings

Fig. 1 is a flowchart of a heterogeneous device ontology matching method fusing semantic information according to the present invention;

FIG. 2 is a comparison graph of experimental results of the impact of different sample selection methods on matching accuracy in an embodiment of the present invention;

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the accompanying drawings.

Referring to fig. 1, the invention provides a heterogeneous device ontology matching method fusing semantic information, which matches an instruction fragment of an internet of things device with a general semantic ontology; the semantic ontology refers to a general instruction intention set, and the instruction fragment is a specific instruction when the Internet of things equipment executes an intention; is provided with N₁Instruction fragment set of individual elements

And has N₂Individual element universal semantic ontology set

The method finds sk₁And sk₂Pairs of elements with the same intent in between, each element can only be matched once; table 1 is an example of an instruction fragment set and a common semantic ontology set.

TABLE 1

The method specifically comprises the following operation steps:

step S100, set sk of instruction fragment₁And a common semantic ontology set sk₂All elements in the system are input into an instruction understanding model one by one to generate respective instruction intention characterization vectors; the input instruction understanding model comprises an intention description coder and an instruction fragment coder;

step S200, screening a small part of instruction fragment sets sk from the whole training data set₁'for making an exact match dataset F'; the implementation details are as follows: b batches of sample data are extracted from all instruction fragments, wherein B is set to be 10, namely only 10 instruction fragments are extracted for manufacturing an accurate matching data set; and finding out a universal semantic ontology matched with each instruction fragment according to candidate configuration matching in a Common Configuration Tree (CCT).

Step S300, calculating a similarity matrix based on the instruction intention characterization vector

Step S400, calculating a mapping matrix based on the similarity matrix S

Step S500, calculating an objective function F1 based on the similarity matrix S and the mapping matrix F, wherein the calculation result is used for back propagation to update the mapping matrix F;

step S600, updating a corresponding part of a mapping matrix F by using the accurate matching data set F', calculating an objective function F2, and using a calculation result for back propagation to update an instruction understanding model;

step S700, the steps S200 to S600 are repeated for multiple times until the mapping matrix F realizes accurate matching sk'₁(ii) a The resulting mapping matrix F will be used to generate an instruction intent dictionary.

The step S100 includes the following operation sub-steps:

step S110, for the instruction fragmentCollection sk₁And a common semantic ontology set sk₂Each element in the element, and the description text content and the instruction content in the element are respectively input into an intention description coder and an instruction fragment coder;

and step S120, connecting the coding results of the two coders into a vector, wherein the vector is used as an instruction intention characterization vector of the element.

The step S200 specifically includes the following operation substeps:

step S210, calculating the instruction fragment set according to the following formula

Information entropy E (p) of each element in (1)_i)：

Wherein p is_i，jIs an instruction fragment p_iThe frequency of occurrence of the word j in (b) in the whole instruction fragment set;

step S220, calculating the instruction fragment set sk according to the following formula₁All elements in and the general semantic ontology set sk₂Redundancy of all elements in (d):

R_i，j＝max(0，cos(p_i，q_j))

in the above formula, cos (p)_i，q_j) Represents the instruction fragment set sk₁Middle element p_iWith the general semantic ontology set sk₂Middle element q_jThe instructions of the two are intended to represent the cosine distance between the vectors;

step S230, selecting a value quantity index RE using the ratio of the information entropy to the redundancy as a quantization sample, wherein a calculation formula is as follows:

wherein B is a collection of data;

step S240, selecting partial data sk according to the quantized sample selection value index RE₁' an exact match dataset F ' is made, the dataset F ' including the sk₁' exact match result for each instruction fragment in.

The step S300 includes the following operation sub-steps:

step S310, calculating the instruction fragment set sk₁The instruction intention characterization vector and the general semantic ontology set sk of each element₂The instructions of each element in (a) are intended to characterize the euclidean distance between the vectors;

step S320, all calculation results form a similarity matrix

Wherein the element s_i，je.S represents the instruction fragment set sk₁Element p in (1)_iWith the general semantic ontology set sk₂Middle element q_jThe euclidean distance between them.

TABLE 2

Based on the instruction fragments and the general semantic ontology in table 1, the similarity matrix obtained by calculation according to the above steps is shown in table 2, where the number of rows of the matrix is 10 and is equal to the number of instruction fragments, and the number of columns is 10 and is equal to the number of semantic ontologies. Element value s in ith row and jth column of Table 2_i，jRepresenting the Euclidean distance between the ith instruction fragment and the jth semantic ontology. The minimum distance value in each row is shown in bold.

The specific content of step S400 is as follows: solving the optimal mapping matrix F using a classical algorithm Sinkhorn in optimal transmission theory, the Sinkhom algorithm defining a mapping matrix F ═ diag (u) K · diag (v), where K: e ═ e^-λSIn the formula, λ is a hyper-parameter used for controlling the difference between the elements of the mapping matrix F, and the larger λ is, the smaller the difference is; the hyper-parameter λ is set to 0.1 in the present embodiment.

S is the similarity matrix, diag (u) and diag (v) refer to the remaining 0 squares with vector u and vector v as diagonals, respectively. The vector u and the vector v are calculated based on an iterative mode, and are initialized to be vectors with all 1, and the calculation method comprises the following steps:

wherein

Representing the division of vector elements item by item, a and b respectively representing the normalized vectors of the sum of the similarity matrix S calculated according to rows and columns; the Sinkhorn solution process will iteratively solve for F until convergence or maximum number of steps is reached; mapping matrix

For the element f_i，jE.g. F, if the instruction fragment set sk₁Element p in (1)_iWith the general semantic ontology set sk₂Middle element q_jShould be aligned then f_i，j1, otherwise f_i，j＝0。

The calculation formula of the objective function f1 in step S500 is:

wherein the first item

Is to aim at the Euclidean distance between the segments to be matchedMinimizing the sum of the distances; second term lambda (-f)_i，jlogf_i，j) Is an additional entropy regularization target, such that even f_i，jNot an integer, namely: f. of_i，j∈[0，1]It can still be made as close to 0 or 1 as possible.

The step S600 includes the following operation sub-steps:

step S610, updating the mapping matrix F by the accurate matching result in the accurate matching data F', and updating the corresponding F in the mapping matrix F_i，jSet to 1, set the other tags to 0;

in step S620, a triple is constructed for each instruction fragment, i.e. for instruction fragment p_iConstruction of a triplet (p)_i，q_pos，q_neg) Wherein p is_iAnd q is_posAre alignable positive samples obtained based on a mapping matrix F, and q_negIs from a general semantic ontology set sk₂Negative samples sampled from, i.e. from, the common semantic ontology set sk₂In order to randomly extract a non-q_posA fragment;

step S630, on the basis of the triplet, by minimizing the target f₂Training the intention description encoder and the instruction fragment encoder simultaneously, the objective function f₂The calculation formula of (2) is as follows:

dis (·) is a similarity measure between vector representations, which can be calculated using euclidean distance;

and step S640, updating the instruction understanding model according to the calculation result.

The instruction intention dictionary in step S700 includes a one-to-one correspondence between the general instructions in the semantic ontology and the specific instruction fragments of a single internet of things device.

Until accurate matching sk 'is realized by mapping matrix F'₁。

TABLE 3

0	0	0	1	0	0	0	0	0	0
										1	0	0	0	0	0	0	0	0	0
0	1	0	0	0	0	0	0	0	0
										0	0	0	0	0	0	1	0	0	0
0	0	1	0	0	0	0	0	0	0
										0	0	0	0	0	0	0	0	1	0
0	0	0	0	0	0	0	1	0	0
										0	0	0	0	0	0	0	0	0	1
0	0	0	0	0	1	0	0	0	0
										0	0	0	0	1	0	0	0	0	0

Mapping matrix of the invention

Showing the matching relation between the instruction fragment and the general semantic ontology, wherein the element f of the ith row and the jth column_i，j1 means that the ith instruction fragment matches the jth common semantic ontology, f_i，jAnd 0 indicates that the ith instruction fragment does not match the jth general semantic ontology. Since each element can only be matched once, each row can only have one value of 1. For the instruction fragments and the general semantic ontology in table 1, the mapping matrix F obtained by final calculation is shown in table 3, where each row of the matrix corresponds to one instruction fragment, and each column corresponds to one semantic ontology, and there are 10 rows and 10 columns in total.

In step S700, the mapping matrix F finally obtained is used to generate an instruction intent dictionary. In the present embodiment, the instruction intention dictionary in step S700 contains a one-to-one correspondence relationship between the general instructions in the general semantic ontology and the specific instruction fragments of a single device.

TABLE 4

The instruction intention dictionary obtained from table 3 is shown in table 4, the instruction fragment of each row in table 4 is matched with the universal semantic ontology on the right side thereof, and the accuracy of the matching result is 100%.

The intention description encoder is constructed by a transform layer; each Transformer layer comprises a self-attention layer and a feedforward neural network layer, and the attention mechanism of the self-attention layer is utilized to automatically learn the characteristics with the resolving power so as to meet the training target; the input is a sequence of descriptors with a placeholder "[ cls ]" added at its forefront; for each input word, the output is its vectorized representation obtained by the graph description encoder; the vector representation using "[ cls ]" is characterized as a description; it is intended to describe that the encoder uses a Pre-trained BERT model (BERT: Pre-training of Deep Bidirectional transducers for length estimation) BERT-small, with 4 layers of transducers, each layer having 512 heads.

The instruction fragment encoder uses the same setup and operation as the intent description encoder. The only difference is that the instruction fragment content cannot use the pre-trained language model. The statement contains many special characters and abbreviations and therefore cannot be parsed directly using a pre-trained language model, the model variables being initialized randomly.

During training, an Adam optimizer with a learning rate of 10-5 is used.

The inventors used instruction fragments from 689 configuration files from four vendors (Cisco, Huashi, Xinhua san, Sharp) as the instruction fragment set for training and testing. There were 304 profiles from cisco, 186 from hua ye, 151 from xinhua san, and 67 from sharia. All the vendor profiles come from data centers that support the same service and the devices perform the same network architecture role.

The experimental results show that the alignment accuracy of the present invention can be 100% for different suppliers, which demonstrates the robustness of the method of the present invention to a variety of different environments. Because the invention considers the correlation between the intention description and the instruction fragment, even though the instruction fragments of different suppliers may be greatly different, the invention can still realize the matching of the instruction fragments of various heterogeneous devices with the common semantic ontology by taking the intention description as the reference.

In step S200, the invention adopts a mechanism for automatically screening sample data, so the inventors also studied the influence of different sample selection methods on the matching precision. In addition to the sample selection method of the present invention, we also evaluated the performance of two other sample selection methods (i.e., stochastic and entropy-based only) on four data sets. The random method is to randomly select samples in the learning process, and the method based on the information amount is to calculate the information entropy of all the data in the batch according to the formula in step S210, and then select the sample with higher information entropy. The results of the experiment are shown in FIG. 2.

The accuracy in fig. 2 refers to the ratio of the number of fragments matching correctly to the total number of fragments, and the sample rate refers to the ratio of the number of samples used to make an accurate match data set to the total number of samples. From the experimental results, it can be seen that compared with other methods, the accuracy of the method of the present invention increases fastest with the increase of the number of samples, and only 10% of the labels are required on average to achieve 100% accuracy. This indicates that the swatches of the present invention can participate in achieving the best accuracy with the fewest swatch labels.

Furthermore, it can be seen from the experimental results that the increasing trend of accuracy is consistent for different suppliers, further illustrating the robustness of the method of the present invention to various environments.

The terms "comprises," "comprising," or other similar terms are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process or method.

So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the invention, a person skilled in the art can make equivalent changes or substitutions on the related technical features, and the technical solutions after the changes or substitutions will fall within the protection scope of the invention.

Claims (9)

1. A heterogeneous equipment body matching method fusing semantic information is characterized by comprising the following steps: matching the instruction fragment of the Internet of things equipment with the universal semantic ontology; the general semantic ontology refers to a general instruction and meaning set, and the instruction fragment isThe specific instruction when the Internet of things equipment executes a certain intention; is provided with N₁Instruction fragment set of individual elements

And has N₂Individual element universal semantic ontology set

The method finds sk₁And sk₂Pairs of elements with the same intent in between, each element can only be matched once;

the method specifically comprises the following operation steps:

step S100, set sk of instruction fragment₁And a common semantic ontology set sk₂All elements in the system are input into an instruction understanding model one by one to generate respective instruction intention characterization vectors; the input instruction understanding model comprises an intention description coder and an instruction fragment coder;

step S200, screening a small part of instruction fragment sets sk from the whole training data set₁'for making an exact match dataset F';

step S300, calculating a similarity matrix based on the instruction intention characterization vector

Step S400, calculating a mapping matrix based on the similarity matrix S

Step S500, calculating an objective function F1 based on the similarity matrix S and the mapping matrix F, wherein the calculation result is used for back propagation to update the mapping matrix F;

step S600, updating a corresponding part of a mapping matrix F by using the accurate matching data set F', calculating an objective function F2, and using a calculation result for back propagation to update an instruction understanding model;

step S700, the steps S200 to S600 are circulated for a plurality of times untilTo mapping matrix F to achieve accurate matching sk'₁(ii) a The resulting mapping matrix F will be used to generate an instruction intent dictionary.

2. The method for matching the heterogeneous device ontology with fused semantic information according to claim 1, wherein the step S100 comprises the following operation substeps:

step S110, for the instruction fragment set sk₁And a common semantic ontology set sk₂Each element in the element, and the description text content and the instruction content in the element are respectively input into an intention description coder and an instruction fragment coder;

and step S120, connecting the coding results of the two coders into a vector, wherein the vector is used as an instruction intention characterization vector of the element.

3. The method for matching the heterogeneous device ontology with fused semantic information according to claim 1, wherein the step S200 specifically includes the following operation substeps:

step S210, calculating the instruction fragment set according to the following formula

Information entropy E (p) of each element in (1)_i)：

Wherein p is_i,jIs an instruction fragment p_iThe frequency of occurrence of the word j in (b) in the whole instruction fragment set;

step S220, calculating the instruction fragment set sk according to the following formula₁All elements in and the general semantic ontology set sk₂Redundancy of all elements in (d):

R_i，j＝max(0，cos(p_i，q_j))

in the above formula, cos (p)_i,q_j) Represents the instruction fragment set sk₁Middle element p_iWith the general semantic ontology set sk₂Middle element q_jThe instructions of the two are intended to represent the cosine distance between the vectors;

step S230, selecting a value quantity index RE using the ratio of the information entropy to the redundancy as a quantization sample, wherein a calculation formula is as follows:

wherein B is a collection of data;

step S240, selecting partial data sk according to the quantized sample selection value index RE₁' an exact match dataset F ' is made, the dataset F ' including the sk₁' exact match result for each instruction fragment in.

4. The method for matching the heterogeneous device ontology with fused semantic information according to claim 1, wherein the step S300 comprises the following operation sub-steps:

step S310, calculating the instruction fragment set sk₁The instruction intention characterization vector and the general semantic ontology set sk of each element₂The instructions of each element in (a) are intended to characterize the euclidean distance between the vectors;

step S320, all calculation results form a similarity matrix

Wherein the element s_i,je.S represents the instruction fragment set sk₁Element p in (1)_iWith the general semantic ontology set sk₂Middle element q_jThe euclidean distance between them.

5. The method for matching the heterogeneous device ontology with fused semantic information according to claim 1, wherein the specific content of the step S400 is as follows: solving the optimal mapping matrix F, S by using the classical algorithm Sinkhorn in the optimal transmission theoryThe inkhorn algorithm defines a mapping matrix F ═ diag (u) · K · diag (v), where K: e ═ e^-λSIn the formula, λ is a hyper-parameter used for controlling the difference between the elements of the mapping matrix F, and the larger λ is, the smaller the difference is; s is the similarity matrix, diag (u) and diag (v) refer to the remaining 0 squares with vector u and vector v as diagonals, respectively. The vector u and the vector v are calculated based on an iterative mode, and are initialized to be vectors with all 1, and the calculation method comprises the following steps:

wherein

Representing the division of vector elements item by item, a and b respectively representing the normalized vectors of the sum of the similarity matrix S calculated according to rows and columns; the Sinkhorn solution process will iteratively solve for F until convergence or maximum number of steps is reached; mapping matrix

For the element f_i，jE.g. F, if the instruction fragment set sk₁Element p in (1)_iWith the general semantic ontology set sk₂Middle element q_jShould be aligned then f_i，j1, otherwise f_i，j＝0。

6. The method for matching the heterogeneous device ontology with fused semantic information according to claim 1, wherein the objective function f1 in step S500 is calculated by the following formula:

wherein the first item

Is to minimize the sum of euclidean distances between matched segments; second term lambda (-f)_i，jlogf_i，j) Is an additional entropy regularization target, such that even f_i，jNot an integer, namely: f. of_i，j∈[0，1]It can still be made as close to 0 or 1 as possible.

7. The method for matching the heterogeneous device ontology with fused semantic information according to claim 1, wherein the step S600 comprises the following operation substeps:

step S610, updating the mapping matrix F by the accurate matching result in the accurate matching data F', and updating the corresponding F in the mapping matrix F_i，jSet to 1, set the other tags to 0;

in step S620, a triple is constructed for each instruction fragment, i.e. for instruction fragment p_iConstruction of a triplet (p)_i，q_pos，q_neg) Wherein p is_iAnd q is_posAre alignable positive samples obtained based on a mapping matrix F, and q_negIs from a general semantic ontology set sk₂Negative samples sampled from, i.e. from, the common semantic ontology set sk₂In order to randomly extract a non-q_posA fragment;

step S630, on the basis of the triplet, by minimizing the target f₂Training the intention description encoder and the instruction fragment encoder simultaneously, the objective function f₂The calculation formula of (2) is as follows:

dis (·) is a similarity measure between vector representations, which can be calculated using euclidean distance;

and step S640, updating the instruction understanding model according to the calculation result.

8. The method as claimed in claim 1, wherein the instruction intent dictionary in step S700 includes a one-to-one correspondence relationship between general instructions in the semantic ontology and specific instruction fragments of a single internet of things device.

9. The heterogeneous device ontology matching method fusing semantic information according to claim 1, wherein:

the intention description encoder is constructed by a transform layer; each Transformer layer comprises a self-attention layer and a feedforward neural network layer, and the attention mechanism of the self-attention layer is utilized to automatically learn the characteristics with the resolving power so as to meet the training target; the input is a sequence of descriptors with a placeholder "[ cls ]" added at its forefront; for each input word, the output is its vectorized representation obtained by the graph description encoder; the vector representation using "[ cls ]" is characterized as a description;

the instruction fragment encoder uses the same settings and operations as the intent description encoder, the only difference being that the instruction fragment content cannot use a pre-trained language model; the sentence contains many special characters and abbreviations and therefore cannot be parsed directly using a pre-trained language model.

Images