CN113468875A

CN113468875A - MNet method for semantic analysis of natural language interaction interface of SCADA system

Info

Publication number: CN113468875A
Application number: CN202110840951.XA
Authority: CN
Inventors: 王书琴
Original assignee: Shen Yanyi
Current assignee: Shen Chunshan
Priority date: 2021-07-26
Filing date: 2021-07-26
Publication date: 2021-10-01

Abstract

The invention discloses an ideogram network MNet method for semantic analysis of a natural language interactive interface of an SCADA system. The method comprises the following steps: step (1), constructing an ideogram network MNet element relation; step (2), establishing a relation in the intention network MNet; step (3) constructing an external relation of the ideogram network MNet; and (4) mapping the instruction from the semantic network MNet to the SCADA system call interface. The invention converts the semantic analysis process of the natural language interface into the construction process of MNet, gives consideration to the internal semantics of the structure of the language unit and the foreign semantics of the language background, and can effectively analyze the natural language control interface instruction of the SCADA system.

Description

MNet method for semantic analysis of natural language interaction interface of SCADA system

Technical Field

The invention relates to a natural language interface semantic understanding method, in particular to a semantic analysis MNet method for a natural language interaction interface of an SCADA system, and belongs to the technical field of natural language processing.

Background

Semantic analysis is a key process of natural language processing and is a broad concept. When facing different language units, the tasks are different, such as word meaning disambiguation at word level, character labeling at sentence level, and reference disambiguation layer at discourse level.

In recent years, semantic analysis mainly focuses on rule-based methods based on a series of linguistic rules to generate linguistics and statistical methods based on analysis of large-scale corpora to adopt probability and data-driven methods. The rule method usually starts from establishing "relationship between predicate arguments", such as first-order predicate calculus, semantic network, concept dependency graph, and frame-based representation method. With the development towards the deep semantic analysis direction, concepts and applications such as semantic dependency analysis trees, dependency analysis graphs, abstract semantic representation (AMR), combined domain logic, knowledge maps and the like are gradually evolved. Such methods typically perform semantic analysis on a combinatorial basis, which tends to be lexical, considering vocabulary as the center of descriptive languages, such as sememe analysis, semantic fields, HowNet, etc. Statistical methods consider that human use of language relies on a large amount of past practical experience of language. This results in a probabilistic and data-driven based approach, currently represented by deep networks based mainly on distributed learning (also known as word embedding, word vector representation). Vector representation of language units including word embedding and sentence embedding obtains good effects in some shallow semantic analysis, such as named entity recognition, word relation extraction, text classification, automatic term extraction and the like. The method based on rational rules and empirical statistics is also continuously fused in practical application, and good effect is achieved.

From the spatiotemporal process of semantic analysis, the semantic analysis processing and the syntactic analysis can be carried out sequentially or synchronously, thereby leading to different semantic representation methods. The parsing can be performed first and then semantic analysis, also known as syntax-driven semantic analysis, and corresponding semantic representation methods such as parse trees with semantic attachments; the syntactic analysis and the semantic analysis can also be performed synchronously, which is also called as a syntactic-semantic integration method, such as a combination category grammar merged into lambda calculus. From the perspective of efficiency, the syntax and semantic synchronization analysis has certain advantages.

From the practical application effect, when different natural language understanding tasks are oriented, the key point of semantic analysis is to convert a natural language into formal description logic of a certain target system. The design of natural language Control interface of distributed SCADA (Supervisory Control And Data Acquisition) requires converting the speech or text expressed by the user into an executable computer system calling interface. The core is to solve the following two problems: (1) formalized semantic function description of a natural control language, analyzing control intentions and input and output call relations among complex control sequences; (2) mapping between manipulation sequences and computer system call interfaces. Just like the software modeling approach, where step (1) is a fundamental task, a complex natural language concept structure is created that consists of abstract semantic entities.

If the natural language is said to model the objective world scene descriptively, the semantic analysis in the natural language processing field is the modeling of the natural language, but the underlying semantic analysis is the modeling of the objective world scene. The nature of an objective world scenario is to describe a combination of static and dynamic features, i.e. composition and their interrelationships. The static characteristics describe the composition relationship among the various component parts; the dynamic features describe how the components interact cooperatively. The semantic analysis model generally incorporates static and dynamic modeling methods: composition of semantic entities and their interrelationships. The evaluation of the model can be described in terms of visualization, concurrency, executability, complexity, consistency, and variability, among other aspects.

Semantic analysis methods require further research in terms of complexity, consistency, variability, visualization, etc. Such as AMR, relaxes the constraints on the nodes on the basis of the dependency graph, and the nodes can be supplemented. The nodes can be words in sentences or words which are not contained in special original sentence units such as PropBankFramest and the like, but the semantics of some sentences which do not accord with the combination principle can not be well expressed, a series of special rules need to be added, the consistency is not good enough, and the application is difficult. The semantic network also integrates a plurality of concept nodes which are not available in the natural language, and the consistency of the model structure is poor, so that the model is not convenient to understand. In the aspect of statistics, based on a large-scale corpus, the semantic parser is obtained by training natural language sentences and corresponding logic expressions, so that the visualization and variability are poor, and the embedded expression of paragraphs and chapters is difficult.

Therefore, there is a need for an efficient semantic analysis model to solve semantic parsing problems such as SCADA system natural language interaction interfaces.

Disclosure of Invention

The invention provides an ideogram network MNet method for analyzing the natural language interactive interface semantic of an SCADA system, which is slightly integrated and extended on the basis of the existing concepts of absorbing a semantic network, a deep network, dependency analysis and the like. MNet is composed of the meaning elements, the internal relation, the external relation and the characteristic attributes, is defined in a hierarchical recursive mode, is oriented to the overall semantic space description from phrases, sentences to sections, and can effectively solve the semantic analysis problem of the natural language interaction interface of the SCADA system.

Intention network MN for semantic analysis of natural language interaction interface of SCADA system_ethe method comprises the following steps:

step (1): constructing an MNet element relation of the ideogram network;

step (2): establishing a relation in the intention network MNet;

and (3): constructing an external relation of the MNet of the ideogram network;

and (4): mapping the MNet to the command of the SCADA system call interface;

the composition of MNet is:

define the ideogram network MNet: is an ordered group, MNet ═ (n)₁，n₂，...，n_m(ii) a R, P), wherein:

n_im is also an ideogram network, which is a sub-ideogram network of N (this is a recursive definition, which is especially important, and also is an important feature for distinguishing other semantic networks), m is more than or equal to 1, and N is under the framework of natural language description_iAnd may be in any form from words, phrases, to sentences and chapters.

r is the set of inner relations of N, r ═ r_ijI, j ≠ m, and i ≠ j };

r is an outer set of relationships of N, R ═ R_ik|i＝1..m，k＝1..∞}；

Set of P attributes, P ═ P_i|i＝1..∞}；

Defining the internal relations of the ideogram network MNet: r is_ijIs a quadruple (n)_i，n_jRelation, P), where relation is n_iPoint of n_jThe name of a relationship of (1), wherein n_i∈MNet，n_j∈MNet。

Define the outer relations of the ideogram network MNet: r_ikIs a quadruple (n)_i，n_kP) is n_iPoint of n_kThe name of a relationship of (1), wherein n_i∈MNet，

Defining a property set P: p is a radical of_iIs a bigram (Attributame, AtttriValue), i.e., consists of an attribute name and an attribute value.

Defining the meta-relation: if n is_i，n_jIs an independent word, then r_ijOr R_ikAre meta-relationships.

MNet construction comprises meta-relation construction, inner-relation construction and outer-relation construction.

The result of the inner relation construction is a semantic dependency analysis tree, which is called MNet tree for short, and the MNet tree construction steps are as follows:

inputting: a sentence S containing n words, the ith word from left to right is marked as word_i，

And (3) outputting: semantic dependency D (is a tree) of sentence S, D being initially null

(1) And initializing:

the word set word is a word segmentation result of the natural language instruction S; all words in word are marked as "unresolved";

the variable n is assigned as the number of words in S;

the aggregate workend record does not determine the sequence number of the dependent word: word _ undep ═ 1.. n };

the dependency tag set head is array [ n ], and the initial value is-1;

a dependency type set deprel ═ array [ n ]; initial value of-1, -1 indicates dependence on ROOT;

(2) calculating all the word which is not dependent on the item in the S according to the ideogram network N_iProximity dependency probability P_ij(or P)_ji)， i∈word_undep，word_jSatisfies the following conditions:

(1) j belongs to { k | k ═ i-1, or k ═ i +1, or word_kIs word_jAll upper level dependencies (parent nodes) }, and simultaneously:

word_jand word_iThere is no determined dependency relationship between them.

word_jThe label is ` Unresolved `.

(2) If j is empty, let m, n be the most distant descendant node from j left and right sides respectively, take:

j belongs to { k | k ═ m-1, or k ═ n +1, or word_kIs word_jAll upper level dependencies (parent nodes) };

(3) selecting the largest P_ij(or P)_ji) To the corresponding word_iAnd word_jTo determine the dependency relationship, is recorded as

D is added.

Representing word_iRelying on word_jAnd vice versa represent word_jRelying on word_i。

(4) And updating D:

word_undep＝word_undep-{j}；

head[i]＝j；

deprel[i]＝word_irelying on word_jA relationship type name of (1);

word is to be_iAnd word_jWord in between (excluding word)_iAnd word_j) Marked as ` digested'

(5) And processing in the step 2 until the length len of the word _ undep (word _ undep) is 1.

(6) And marking the remained unique undetermined dependency items as Root nodes, wherein the mark dependency is Root.

(7) And D is output as a dependency analysis result after finishing.

The external relation is an MNet network, and the construction method of the MNet network comprises the following steps: on the basis of MNet tree construction, according to the characteristics of SCADA system instructions, instruction triples { objects, parameters and actions } are extracted and added to the MNet tree as an external relation, and an MNet network structure is formed.

The instruction mapping process is to translate the instruction triples extracted by the MNet network into the call interfaces corresponding to the SCADA system by using the instruction triples filled with parameters to be put back to the temple.

Drawings

Fig. 1 is an overall flow diagram of the MNet method;

fig. 2MNet method example;

FIG. 3 Bi-GRU in combination with character and sentence level attention mechanism

FIG. 4 MNetSParser analysis process

FIG. 5 dependency computation selection Process

FIG. 6 MNet-SCADA-NLI added MNet network

Detailed Description

The invention is further illustrated by the following figures and examples.

1. Composition of MNet:

the ideogram network MNet is an ordered group, MNet ═ n₁，n₂，...，n_m(ii) a R, P), wherein:

n_im is also an ideogram network, which is a sub-ideogram network of N (this is a recursive definition, which is particularly important, and also is an important feature to distinguish other semantic networks), m ≧ 1, for example, N is defined under the framework of natural language description_iAnd may be in any form from words, phrases, to sentences and chapters.

r is the set of inner relations of N, r ═ r_ijI, j ≠ m, and i ≠ j };

r is an outer set of relationships of N, R ═ R_ik|i＝1..m，k＝1..∞}；

Set of P attributes, P ═ P_i|i＝1..∞}；

Internal relationships of the ideogram network MNet: r is_ijIs a quadruple (n)_i，n_jRelation, P), where relation is n_iPoint of n_jThe name of a relationship of (1), wherein n_i∈MNet，n_j∈MNet。

External relations of the ideogram network MNet: r_ikIs a quadruple (n)_i，n_kP) is n_iPoint of n_kThe name of a relationship of (1), wherein n_i∈MNet，

Attribute set P: p is a radical of_iIs a bigram (Attributame, AtttriValue), i.e., consists of an attribute name and an attribute value.

Meta-relation: if n is_i，n_jIs an independent word, then r_ijOr R_ikAre meta-relationships.

Description of the drawings: (1) MNet considers the minimum semantic unit as a word, and the relation between the words is called as element relation; (2) MNet only defines binary relations, there are multivariate relations between actual semantic units, for example, the instruction in SCADA natural language interface parsing of section 4.3 is a ternary relation, and the general ternary relation can be converted into a binary relation, and is not expanded in detail for the sake of simplicity. (3) The foreign relationship can be understood as the context of semantic units, or some common sense relationships, such as the sentence "turn on the table lamp in the bedroom", "table lamp" and "appliance" are generic relationships, but the "appliance" is not a component of the sentence.

Fig. 2 shows the MNet solving method process of the instruction task of extracting "close bedroom window" from the user natural language "please turn on the bedroom desk lamp first and then close the window". Establishing corresponding mapping relations between the bedroom and the Location, between the bedroom and the Action, and between the window and the Object, wherein the mapping relations can be understood as external relations; "first" depends on "open, the dependency name attribute is" time ", these are internal relationships.

Fig. 1 shows that the instruction mapping of the SCADA system is realized by using the result of the MNet construction, so that the SCADA system with physical manipulation of the natural language interactive interface is realized, and the implementation process is described in detail later.

2. MNet meta-relationship construction

The MNet element relation can directly adopt WordNet and HowNet knowledge base tools, but when the problems of the characteristic field are met, the MNet element relation needs to be constructed in advance through sample training. The relationships between words are trained from the samples using bi-directional GRUs and word vectors, in combination with sub-character level and sentence level attention mechanisms.

The input expression uses character embedding and position embedding information vectors of a pair of words of a relation to be predicted, a character level attention mechanism is added, and a bidirectional GRU and the character attention mechanism are combined to form the embedding vector expression of the sentence.

For example, there are related relationship pairs (word)₁，word₂) N sentences s_iAnd (i-1.. n), the information on whether the relation r is included in the embedded expression vector information of each sentence. To utilize the information of all sentences, a word-pair relationship pair (word)₁，word₂) When the relation r is predicted, a sentence-level attention mechanism is added, the n sentence sets are expressed by using the characteristic vectors containing all sentence embedded expression information, and then the whole training is carried out. This has the advantage that the noise contribution of erroneous standard data can be reduced. The model architecture is shown in fig. 3.

There is no uniform definition of the relationship between words, including WordNet, HowNet, ConceptNet, etc., but none of them can be complete. For example, an original word relation sample can be constructed from a Chinese semantic dependency analysis evaluation data packet "evsam05. zip" disclosed in natural language processing and Chinese computation conference (NLP & CC 2013), and trained by using the model of FIG. 3.

The format of evsam05.zip adopts a Chinese dependency corpus in CoNLL format, from which experimental data comes from the Qinghua library. To facilitate the training of the model, the data samples are uniformly converted into the format shown in table 1.

Table 1 partial sample data display

Table 2 Some sample data display

The dependency relationship types between words in the Qinghua semantic dependency tree library are listed in 69 categories, as shown in Table 2.

For example, in the sentence "eighth odds of the world appear", the relationship of "world" depending on "odds" is "limit". And meanwhile, negative samples are added, and the relationship type is NULL, which indicates that no dependency relationship occurs between the words in the group. The data is divided into a training set and a test set. Two cases, no negative examples and negative examples, were tested separately, with 80% of the training set containing negative examples being negative examples. The addition of the negative sample can improve the efficiency and accuracy of training.

TABLE 2 partial word relationship types

Table 3 Types of partial word relations

3. MNet inner relation construction

Semantic dependency tree analysis methods mainly include transfer-based and graph-based methods. In order to better accord with the thinking habit of human language, the concurrency performability is considered, a method based on transfer and a graph is integrated, and a semantic dependency tree is constructed by adopting a bottom-up adjacent word competition and combination dependency mode on the basis of obtaining the relation probability between words in a sentence. The difference from the traditional transfer-based approach is that: (1) the method is not limited by the input sequence of sentences any more, and is not related to the priority of left combination dependence or the priority of right combination dependence; (2) for the words of which the dependency objects are not determined, not only the dependency relationship is checked from the left adjacent words and the right adjacent words, but also the words on which the adjacent words depend are included; (3) and optimizing the multi-subtree phenomenon generated in the construction process. The specific algorithm is as follows:

algorithm 1 MNetSParser

Inputting: a sentence S containing n words, the ith word from left to right is marked as word_iIf S is equal to "please open the desk lamp in the bedroom first and then close the window"

And (3) outputting: semantic dependency D (which is a tree) of sentence S, D is initially empty, and the data form is as follows:

and (3) treatment:

BEGIN

STEP0, initialization:

word is the segmentation result of S; all words in word are marked as "unresolved";

n is assigned as the number of words in S (len (word);

word _ undep ═ 1.. n }; // recording the number of words on which no dependency is determined

head ═ array [ n ]; initial value of-1

deprel ═ array [ n ]; initial value of-1, -1 indicates dependence on ROOT

STEP 1, calculating all the words without dependent items in S according to the ideogram network N_iProximity dependency probability P_ij(or P)_ji)，i∈word_undep， word_jSatisfies the following conditions:

word_jand word_iThere is no determined dependency relationship between them.

word_jThe label is ` Unresolved `.

STEP 2, select the largest P_ij(or P)_ji) To the corresponding word_iAnd word_jTo determine the dependency relationship, is recorded as

D is added.

Representing word_iRelying on word_jOtherwise, wor is indicatedd_jRelying on word_i。

STEP3, update D: // determining word_iRelying on word_j

word_undep＝word_undep{j}；

head[i]＝j；

deprel[i]＝word_iRelying on word_jA relationship type name of (1);

STEP 4, go to STEP 1 to process until len (word _ undep) is 1.

STEP 5, marking the remained unique undetermined dependence items as Root nodes, and marking the dependence as Root.

STEP 6, end, output D is the dependency analysis result.

END

Taking a specific sentence "please turn on the table lamp in the bedroom first and then close the window", the calculation process of the algorithm 1 "mnetspearser" is shown in fig. 4.

For example, the step "viii" calculates the dependency probability of the related words including "table lamp", "open", "please" for the word "close", and selects "close" to depend on "please" in the manner of 'result event' according to the probability maximization principle from the unlabeled related broken line in fig. 5.

The evaluation can adopt three indexes to evaluate the tested system, which are respectively:

dependency marking accuracy (LAS)

Dependence accuracy (Unlabeled Attachment Score, UAS)

Labeling Accuracy (Labeled Accuracy, LA)

Assuming that the total number of words contained in the whole test corpus is N, the triple for dependence of any word<word_i，word_j，deprel_ij>And (4) showing. Wherein word_iBeing words themselves, words_iIn the relation deprel_ijDependent on word_j. All word_jCorrect word_iIs N_uas(ii) a All deprel_ijCorrect word data is N_la(ii) a All word_jAnd deprel_ijNumber of words being all correct N_las. Then, the calculation method of the test index is as shown in equations (1) to (3):

3. MNet outer relation construction

The content of the MNet network construction is different due to the difference of semantic understanding tasks. The essence is that on the basis of tree construction, words and the mutual relations thereof are labeled again according to different tasks of downstream natural language processing. The secondary labeling here can be performed after the semantic dependency tree parsing is completed, or during the semantic dependency tree parsing. Therefore, the MNet can continuously optimize the analysis of the natural language, namely, the semantics can be continuously and deeply understood, and the method has certain advantages compared with a neural network mode.

The process of constructing a specific MNet network is described by taking a natural language interactive interface application of a SCADA (System Control and Data Acquisition) System as an example. The SCADA system control mode mainly comprises an inquiry command and a control command, wherein the inquiry command acquires state data of a field process or equipment, and the control command changes parameters of the field equipment or the process. SCADA system steering instructions generally include three parts: actions, objects, and parameters. The most typical parameter is position, indicating a specific location of the manipulation object. For example, after the natural language instruction "turn on the desk lamp in the bedroom", and the natural language processing program is converted into the intermediate language of the data structure "{ Object ═ desk lamp, Location ═ bedroom, and Action ═ turn on }", the call instruction of the SCADA system can be produced through the formalized rule. The analysis principle of the natural language control instruction of the SCADA system is shown in FIG. 1.

The algorithm 1 'MNetSParser' is used for carrying out semantic dependency analysis on the sentence 'turning on the table lamp in the bedroom', and the result is as follows:

on the basis of the dependency analysis result, in combination with the grammar and the usage habit of the specific natural language in the specific environment, the rule set for the target task can be extracted, so that the structure "{ Object, Location, Action }" can be extracted relatively easily. For example, the following rules may be defined:

rule 1: if word_iRelying on word in ' relationship ' subject ' context_jIf the word is not true, then the Action can be extracted as word_j，Object＝word_i；

Rule 2: if word_iDependency on word by relationship' constraint_jAnd word_jHas been extracted as Object, then Location can be extracted as word_i。

Rule 3: if there is a connection dependency relationship with the resolved Object and it is noun, then it is also identified as Object and the Action and Location of both are consistent.

Thus, a simple SCADA system natural language manipulation interface can be described with the algorithm 2 "MNet-SCADA-NLI". Of course, the actual natural language manipulation system may be varied, such as "i want you to turn on the bedroom desk lamp and air conditioner". Aiming at more complex natural language control sequences, the method can solve the problem by adding rules, and the principle is the same.

Algorithm 2MNet-SCADA-NLI

Inputting: MNet semantic dependency analysis trees D, D are defined as algorithm 1

And (3) outputting: command Object, Location, Action

And (3) treatment:

in order to complete the target task more efficiently, the process of constructing a specific MNet network can be integrated into the process of constructing the MNet tree. In this way, STEP3 of algorithm 1 "mnetspearser" may be modified to meet application requirements. After processing by the algorithm 2MNet-SCADA-NLI, the MNet network can be schematically shown in FIG. 6.

Aiming at natural language interaction scenes in the field of smart homes, a small-range questionnaire survey mode is adopted, enough common language control instructions are constructed, and MNet-SCADA-NLI of the invention is respectively adopted to carry out intermediate language identification of the SCADA system of the instructions.

Table 3 shows a part of the common natural language instructions and the corresponding result of analyzing the intermediate manipulation instruction of the SCADA system. Table 4 is the corresponding actual instruction resolution result.

TABLE 3 Natural language instruction in the field of Smart homes and intermediate manipulation instruction analysis example of SCADA System

TABLE 4 results of the MNetParser analysis

Evaluating the efficiency of the algorithm by using the accuracy (P), the recall rate (R) and the F value, and defining the following formulas (4) to (6):

the precision ratio is as follows: p reflects the ratio of correct parsing parameters in the intermediate language of the prediction.

The recall ratio is as follows: p reflects the ratio of correctly resolved parameters in the natural manipulation language sample.

F value:

in the formula:

n: number of parameter indicators analyzed correctly;

Total_P: the number of parameter indexes in the algorithm prediction result;

Total_R: the number of parameter indexes in the expected result of the original natural manipulation language sample.

Claims

1. An ideogram network MNet method for semantic analysis of natural language interactive interface of SCADA system, comprising:

step (1): constructing an MNet element relation of the ideogram network;

step (2): establishing a relation in the intention network MNet;

and (3): constructing an external relation of the MNet of the ideogram network;

and (4): instruction mapping of the sement network MNet to the SCADA system call interface.

2. The ideographic network MNet method according to claim 1, characterized in that said ideographic network MNet is an ordered group, MNet ═ (n)₁，n₂，...，n_m(ii) a R, P), wherein:

n_im is also an ideogram network, which is a sub-ideogram network of N (defined recursively), m ≧ 1, N_iCan be in any form from words, phrases to sentences and chapters;

r is a set of inner relations for N,r＝{r_iji, j ≠ m, and i ≠ j }, with an inner relation r_ijIs a quadruple (n)_i，n_jRelation, P), where relation is n_iPoint of n_jThe name of a relationship of (1), wherein n_i∈MNet，n_j∈MNet；

R is an outer set of relationships of N, R ═ R_ikI 1.. m, k 1.. infinity, and an extrinsic relationship R_ikIs a quadruple (n)_i，n_kP) is n_iPoint of n_kThe name of a relationship of (1), wherein n_i∈MNet，

Set of P attributes, P ═ P_i1 |, infinity, attribute p_iIs a binary (Attributame, AtttriValue), i.e., consists of an attribute name and an attribute value;

if n is_i，n_j，n_kIs an independent word, then r_ijOr R_ikAre meta-relationships.

3. The method according to claim 1, wherein the inner relationship building result in step (2) is a semantic dependency analysis tree, abbreviated as MNet tree, and the construction steps of the MNet tree are as follows:

And (3) outputting: semantic dependency D (being a tree) of the sentence S, D being initially null;

(1) and initializing:

the variable n is assigned as the number of words in S;

the set word undep records the sequence number of the word on which no dependency is determined: word _ undep ═ 1.. n };

the dependency tag set head is array [ n ], and the initial value is-1;

(2) calculating all the word which is not dependent on the item in the S according to the ideogram network N_iProximity dependency probability P_ij(or P)_ji)，i∈word_undep，word_jSatisfies the following conditions:

a.j belong to { k | k ═ i-1, or k ═ i +1, or word_kIs word_jAll upper level dependencies (parent nodes) }, and simultaneously:

word_jand word_iNo determined dependency relationship is established between the two;

word_jlabeled as 'not digested';

b. if j is empty, let m, n be the most distant descendant node from j left and right sides respectively, take:

(or

) And D is added.

Representing word_iRelying on word_jAnd vice versa represent word_jRelying on word_i；

(4) And updating D:

word_undep＝word_undep-{j}；

head[i]＝j；

deprel[i]＝word_irelying on word_jA relationship type name of (1);

word is to be_iAnd word_jWord in between (excluding word)_iAnd word_j) Marked as 'digested';

(5) processing in the step 2 until the length len of the word _ undep (word _ undep) is 1;

(6) marking the remaining unique undetermined dependency items as Root nodes, and marking the dependency as Root;

(7) and D is output as a dependency analysis result after finishing.

4. The method according to claim 1, wherein the foreign relation in step (3) is an MNet network, and the MNet network is constructed by: on the basis of MNet tree construction, triples { objects, parameters and actions } corresponding to the SCADA system instructions are extracted from the natural language instructions and are added to the MNet tree as external relations to form an MNet network structure.

5. The entary network MNet method according to claim 1, wherein the instruction mapping procedure of step (4) is to translate the instruction triplets extracted by the MNet network into the call interfaces of the corresponding SCADA system.