CN112651246B

CN112651246B - Service demand conflict detection method integrating deep learning and workflow modes

Info

Publication number: CN112651246B
Application number: CN202110065804.XA
Authority: CN
Inventors: 姜波; 陈骏武; 汪烨; 宋师哲
Original assignee: Zhejiang Gongshang University
Current assignee: Zhejiang Gongshang University
Priority date: 2021-01-19
Filing date: 2021-01-19
Publication date: 2023-08-22
Anticipated expiration: 2041-01-19
Also published as: CN112651246A

Abstract

The invention discloses a service demand conflict detection method integrating deep learning and workflow modes, which comprises the following steps: s1, analyzing a service demand document to extract outer layer information of a workflow-based process language specification; s2, extracting inner layer information of a workflow-based process language specification through the textCNN, wherein the method comprises the following steps of: s21, inputting natural language into an embedded layer to perform vector conversion; s22, inputting the vector into a convolution layer, and extracting the characteristics; s23, using a pooling function for each feature map to enable the dimensions of the feature maps to be the same; s24, obtaining probability values of each class by using an activation function according to the result of the full connection layer splicing pooling, and obtaining a workflow mode corresponding to the probability values according to the probability values; s3, converting service requirements into workflow-based process language specifications through the inner layer information and the outer layer information, so as to model the information; s4, automatically detecting the service demand conflict by adopting a detection rule.

Description

Service demand conflict detection method integrating deep learning and workflow modes

Technical Field

The invention relates to the technical field of conflict detection, in particular to a service demand conflict detection method integrating deep learning and workflow modes.

Background

The cloud computing is centered on the Internet, and quick and safe cloud computing services and data storage are provided on websites, so that everyone using the Internet can use huge computing resources and data centers on the network.

Service systems are currently deployed in cloud environments. The quality of the service system depends on the quality of the service requirement document to a great extent, but most service requirements are manually recorded in a natural language form at present, so that the problems of inaccuracy, easy error, lack of integral structure and the like easily exist. Deep learning is a new field developed in machine learning research, and can build a neural network for simulating human brain to perform analysis learning, namely, simulate a mechanism of human brain to interpret data such as images, sounds and texts, and has the advantage that non-supervised or semi-supervised feature learning and hierarchical feature extraction efficient algorithm is used for replacing manually acquired features. The Convolutional Neural Network (CNN) is one of representative algorithms for deep learning, has characteristic learning capability, can carry out translation invariant classification on input information according to a hierarchical structure, ensures that the convolutional neural network can carry out latticed characteristics with smaller calculation amount due to the sharing of convolutional kernel parameters in hidden layers and the sparsity of interlayer connection, has the advantages of stable effect, no additional characteristic engineering requirement on data and the like, and can lead the textCNN technology based on CNN to have more excellent performance on the problem of text classification.

Kamalrudin, hosking and Grundy propose a technology for improving the quality of a service requirement document by using an EUC interaction mode, develop a CASE tool to support the operation of the technology, firstly convert the service requirement in a natural language form into a half-formal requirement, then extract abstract interaction from the EUC to derive an EUC model, and finally compare and verify the derived EUC model with the EUC interaction model. In verifying service requirements, sarmiento, leite and Almentero developed a set of 3Cs patterns of aggregate correctness, consistency, and integrity to automatically perform scene verification, 3Cs patterns organize non-functional requirements related to correctness, consistency, and integrity using NFR methods and act as NFRs directories, but their methods can only evaluate the properties of a scene. Gacitua, sawyer and Gervasi propose a RAI technique called association driven abstract recognition for recognizing keywords in an abstract field from natural language documents.

Disclosure of Invention

In order to solve the defects of the prior art and realize the purpose of improving the quality of service demand documents and finally improving the quality of service systems in cloud environments by mining demand defects to reduce service demand conflicts, the invention adopts the following technical scheme:

the service demand conflict detection method integrating the deep learning and workflow modes comprises the following steps:

s1, analyzing a service demand document to extract outer layer information of a workflow-based process language specification;

s2, extracting inner layer information based on process language specifications of a workflow through the textCNN, processing natural language through the textCNN, judging a corresponding workflow mode of the obtained probability value, and comprising the following steps:

s21, inputting natural language into an embedding layer to perform vector conversion, wherein the purpose is to convert non-computable and unstructured words into computable and structured vectors, namely converting word problems in the current world into vector problems in mathematics, so as to facilitate convolution processing;

s22, inputting vectors into a convolution layer, extracting features by using one-dimensional convolution, wherein the one-dimensional convolution only carries out convolution in one direction of a sequence, the width of a convolution kernel is fixed to be the dimension of a word vector, the height is a super parameter, window sliding step length is set, and convolution operation is carried out on each window formed by sentence words to obtain a feature map;

s23, using pooling functions for each feature map to enable the dimensions of the feature maps to be the same, and using convolution kernels with different heights in the convolution layer process, wherein the sizes of features (feature maps) obtained by the convolution kernels with different sizes are different, so that the dimensions of the feature_maps are the same by using pooling functions for each feature_map [1..n ];

s24, the result is spliced and pooled through the full connection layer, so that main characteristics of the result are reserved, the number of parameters is reduced, the dimension of the result is reduced, probability values belonging to each class are obtained through an activation function, a workflow mode corresponding to the probability values is obtained through the probability values, and therefore inner layer information required by a process language specification based on the workflow mode is extracted;

s3, converting service requirements into workflow-based process language specifications through the inner layer information and the outer layer information, so as to model the information;

and S4, automatically detecting the service demand conflict by adopting a detection rule, so as to prevent the problems of inconsistency, inaccuracy, incompleteness and the like in the service demand document.

Further, S2 is performed first, the document is analyzed to obtain the outer layer information of each service requirement through the inner layer information obtained in S2, and then S3 is executed.

Further, the step S22 specifically includes the following steps:

s221, a convolution kernel, a matrix w with a width d and a height h, wherein the matrix w has h x d parameters to be updated, and for a sentence, the matrix A, A [ i ] is obtained after passing through an embedding layer: j represents the ith to jth rows of a, and the convolution operation is represented by the following formula:

o _i ＝w·A[i：i+h-1]，i＝1，2，...，s-h+1

s represents: and (3) performing word segmentation on the input sentence to obtain s words, namely the number of rows of the matrix A.

S222, setting the sliding step length of a window to be 1, sliding the window one line downwards in the next convolution calculation, and finally obtaining a feature map [1..n ];

s223, superposing the bias b on the feature map, and activating by using an activation function f to obtain the feature to be extracted, wherein the formula is as follows:

c _i ＝f(o _i +b)

one convolution kernel gets the feature c, s-h+1 features in total;

s224, adopting a plurality of convolution kernels with different heights to obtain richer feature expression.

Further, in the step S23, a 1-Max-pooling function is adopted, and a maximum value is selected from feature_map [1..n ] as an output, so as to inhibit the phenomenon of mean shift of estimation caused by network parameter errors, so that feature information (the most important feature represented by the maximum value is artificial) can be better extracted, and each convolution kernel obtains a feature which is a value.

Further, the activating function of S24 adopts sigmoid as the activating function, and the function expression is:

the sigmoid function is continuous, smooth, and strictly monotonic, and can be a good threshold function.

Further, in the step S3, the service requirement is modeled by a workflow-based process language (WPPL), so as to obtain a document conforming to the expected service requirement, and an operation language is used to represent the operation of the service requirement and the object of the operation, where the description of the service requirement is divided into an inner layer structure and an outer layer structure, the inner layer describes the process flow in the service, and the outer layer describes the key attribute of the service.

Further, the inner layer, each service requirement is one or more business processes composed of a series of operations connected by a control flow, the control flow being represented by a workflow pattern.

Further, the relationship attributes of the outer layer description include: input and output data are represented by entities by states, and the pre and post conditions are represented by states for subsequent requirement verification based on the IOPE concept.

Further, the workflow mode includes five sub-services of sequential mode, parallel split mode, synchronous mode, exclusive selection mode and simple merge mode, and the business operation, business service or a combination service mapped to atoms are cooperated through a set of predefined workflows;

the sequential mode is that in one flow example, each activity is sequentially extruded and executed;

the parallel splitting mode is that in a flow example, two or more execution paths are executed in parallel, no correlation exists between the parallel paths, and the execution of the parallel paths has no definite sequencing relation;

the synchronous mode is that in a flow example, the execution of an activity depends on the execution results of the previous paths;

the exclusive selection mode is that in a flow example, different execution paths exist according to different conditions;

the simple merge mode is that in one flow instance, two or more execution paths merge on one active node.

Further, the embedding layer in step S21 performs word embedding by using a word2vec method to form a two-dimensional matrix, which is used as an input of the convolution layer. The word2vec method is selected because the method has the advantages of small dimension, high mapping speed, strong universality and the like.

The invention has the advantages that:

the invention uses a workflow mode, can provide a general circulation flow structure set, so as to make up the problem that the EUC interaction mode is not highly applicable; the present invention focuses on verifying service requirements with multiple flows; the present invention focuses on the use of workflow patterns to obtain flow information from natural language text rather than the field of abstract concepts.

Drawings

Fig. 1 is a UML2.0 activity diagram based on workflow patterns.

Fig. 2 is an overview of the method of the present invention.

FIG. 3 is an overview of the architecture of the service demand modeling process of the present invention.

Fig. 4 is a TextCNN model framework diagram in the present invention.

FIG. 5 is a graph of Word enhancement results in the present invention.

FIG. 6 is a schematic diagram of convolution calculations in the present invention

Fig. 7 is a sigmoid function diagram in the present invention.

FIG. 8 is a graph of the results of using the 1-Max-pooling function in the present invention.

Fig. 9 is a schematic representation of softmax in the present invention.

Detailed Description

The following describes specific embodiments of the present invention in detail with reference to the drawings. It should be understood that the detailed description and specific examples, while indicating and illustrating the invention, are not intended to limit the invention.

The invention merges the workflow mode and the deep learning algorithm in the cloud computing environment. The core concept of the workflow mode is to abstract and describe an actual business processing flow, which has 21 common mode types, and the algorithm uses the most basic five modes as basic references to extract process messages between services and evaluate the quality of a required document, which are respectively 1) a Sequence order mode, namely, in a flow example, each activity is sequentially executed by sequential extrusion; 2) A Parallel split mode, i.e. in one flow instance, there are two or more execution paths executing in Parallel, but there is no association between these Parallel paths, and execution of Parallel paths has no definite sequencing relationship; 3) The Synchronization mode, that is, in a process instance, the execution of an activity depends on the execution results of the previous paths; 4) Exclusive choice Exclusive selection mode, i.e. in a flow instance, there are different execution paths according to different conditions; 5) Simple merge mode, i.e., in one flow instance, two or more execution paths merge on one active node. As shown in fig. 1.

1. Method frame

To get a high quality service requirement document, we model the service requirement using workflow pattern based procedural language (WPPL), as shown in figure 2. To improve the efficiency of the conversion, we use workflow patterns and TextCNN text classification techniques modified from convolutional neural networks to semi-automatically extract the inner layer process information from the service requirement document.

The conversion process is divided into an inner layer information and an outer layer information which need to be extracted and analyzed:

1. firstly, the outer layer information is analyzed and obtained from a service demand document by a demand analyzer;

2. the inner layer information uses the textCNN technology, and the corresponding workflow mode is judged through the finally obtained probability value. After the inner layer information and the outer layer information are obtained through the two steps, the service requirement can be converted into the WPPL standard so as to model the information;

3. finally, in order to prevent the problems of inconsistency, inaccuracy, incompleteness and the like in the service demand document, the automatic detection of the service demand conflict is carried out by adopting a self-defined detection rule.

The conversion process is demonstrated with a trade order example, as shown in FIG. 3.

2. The method comprises the following specific steps

1. Workflow-based procedural language (WPPL) overview

We model the service requirements using WPPL, i.e. a workflow-mode-based process language, to arrive at a document that meets the expected service requirements. The service requirements are described in terms of business processes consisting of workflow patterns, each having a unique name, each name being defined in an operational form similar to an operational language, such as action (identity 1, identity 2, …), which is a language designed to implement modeling operations and effects. The semantics of the operation language are based on the assumption that "things remain unchanged until a change occurs". This assumption is the same as the service requirement assumption we represent in WPPL. Thus, we employ an operation language to represent the operation of a service requirement and the object of the operation. The description of each service requirement is divided into an inner layer structure and an outer layer structure, wherein the inner layer describes the process flow in the service, and the outer layer describes key attributes such as input, output, pre-condition, external condition and the like of the service.

In the outer layer, input and output data are represented by entries, and pre-and post-conditions are represented by states (entries), which means states of entities. We use states to represent pre-and post-conditions for subsequent verification of demand based on the IOPE concept. At the inner layer, each service requirement may be considered as one or more business processes consisting of a number of operations connected by a series of control flows, which are represented by a workflow pattern. The following is the back-Naur paradigm syntax of WPPL:

outer layer

<name>::＝name Action

<input>::＝input Entity+

<output>::＝output Entity+

<precondition>::＝precondition State+

<postcondition>::＝postcondition State+

Inner layer

<process>::＝process<pattern>|<pattern>；<process>

<pattern>::＝<pattern_name>(<action_set>)

<action_set>::＝Action,Action|Action,<action_set>

<pattern_name>::＝seq|par_split|sync|excl_choice|merge

Trade orders are a representative service in a financial transaction service system, so we use trade orders as an example to illustrate the method, and the service requirements are: "The trader enters a new order to the system.the operator would then match the new order with existing opposite-side orders in the order book if the new order fails to be matched, the new order should be added to the order book; other wise, the printing agent will print the trade to the public, after the trade is successfully printed, the system will notify the trader with the new trading information, including the traded price and trading sizes, and so on. "i.e." trader inputs a new trade order into the system, then the operator matches the new trade order with the existing counter-face order in the order book, if the new order fails to match, then the new order is added to the order book, otherwise the information disclosure agent will disclose the trade order to the public. After the order has been successfully printed, the system will communicate new trade information to the trader, including the price of the order, the size of the order, etc. The name of this service requirement is defined as track (order), the precondition of track (order) is the initial order that has been prepared, it is defined as pre (initial_order), the post condition is that the order has been traded or that the order has been added to the order book, it is defined as track (order) and updated (order_book), respectively, and it is logically or logically connected, denoted V.

In the inner layer, the flow contains five sub-services, which can be mapped to atomic business operations, business services, or a composite service. These sub-services cooperate with a predefined set of workflows, which can be expressed as:

seq(enter(order),match(order))；excl_choice(match(order),add(order),print(trade))；seq(print(trade),notify(trader))

this means that in order to achieve the track (order) goal, this order needs to enter the system (i.e., enter) and then match with the order in the order book (i.e., match). If this order is not found in the flow, then a new order is added to the order book (i.e., add), otherwise the trade order (i.e., print) is output and the trader is notified (i.e., notify (trader)). To avoid the use of the same key as a pattern name for certain service requirements, abbreviations are used in WPPL to represent each pattern, e.g., seq represents sequence and par_split represents parallel split.

2. Extracting inner and outer layer information based on textCNN

In order to accurately and comprehensively extract inner-layer information from service requirements, a textCNN algorithm technology in deep learning is fused on the basis of a workflow mode to process natural language so as to extract information. The textCNN is improved on the basis of a convolutional neural network algorithm, namely, the input embedded layer is deformed, text data is used as input, and because the text data is one-dimensional data, only one-dimensional convolution is needed in the convolutional layer, and the text classification performance is excellent under the condition of simple network structure. As shown in fig. 4.

1) Embedding layer (input embedded layer)

Word spotting is performed at this layer to convert non-computable, unstructured words into computable, structured vectors, i.e., to convert word problems in the current world into vector problems in mathematics. The example sentence 'The trader entera new order to The system' is used for demonstration, the word is divided into 'The/track/enter/a/new/order/to/The/system', a mapping is needed to be constructed for judging each character of The word, but a common mathematical model only can accept numerical input, but The invention processes natural language and is abstract summary of human beings, so The model is converted into numerical form, namely, word2vec method is selected by word embedding in The method, each word in The word division result is mapped into a 5-dimensional word vector, and The natural language is quantized to facilitate convolution processing, and The result is shown in figure 5.

The use of different emmbeddings can have a great influence on the result, and because the word2vec method has the advantages of small dimension, high mapping speed, strong universality and the like, the method uses the word2vec to form a 9*5 two-dimensional matrix, and the two-dimensional matrix is used as an input for convolution calculation in a convolution layer.

2) Convolition layer (convolutional layer)

The convolutional neural network applies a two-dimensional convolution when processing an image, but since text data is one-dimensional, textCNN techniques apply a one-dimensional convolution when convolving.

The matrix obtained from the upper layer of the matrix is regarded as an image, and the convolutional neural network is used for extracting the characteristics. Since the relevance of adjacent words in a sentence is always high, one-dimensional convolution can be used, i.e. text convolution differs from image convolution in that the convolution is done only in one direction (perpendicular) of the text sequence, the width of the convolution kernel being fixed to the dimension d of the word vector, i.e. 5. The height is an superparameter, and the method sets the window sliding step length to 1. And carrying out convolution operation on each possible window of the sentence word to obtain a feature_map [1..n ].

Now, assuming that a convolution kernel is a matrix w with a width d and a height h, then w has h×d parameters to be updated. For a sentence, the matrix a can be obtained after passing through the embedding layer. A [ i: j represents the ith to jth rows of a, then the convolution operation may be expressed by the following formula:

o _i ＝w·A[i：i+h-1]，i＝1，2，...，s-h+1

s represents the number of rows of the matrix A, which is obtained by segmenting the input sentence into s words; for example, the word "The/divider/entry/a" gets a matrix of 4*5, multiplies it by The convolution kernel by elements, and adds them together, i.e., the convolution operation, as shown in fig. 6.

Setting the sliding step length of the window as 1, sliding the window one line downwards in the next convolution calculation, and finally obtaining a feature map [1..n ], superposing the bias b, and activating by using an activation function f to obtain the required feature. The formula is as follows:

c _i ＝f(o _i +b)

for one convolution kernel, feature c can be obtained for a total of s-h+1 features. The method selects a plurality of convolution kernels of different heights (i.e., heights 4, 3) to obtain a richer feature expression. Since the sigmoid function is continuous, smooth and strictly monotonic, and is a good threshold function, the method uses the sigmoid as an activation function, and the function expression is as follows:

the functional image is shown in fig. 7.

3) Pooling (Pooling)

Since we use convolution kernels of different heights in the convolution layer process, the features (feature maps) obtained by the convolution kernels of different sizes are different in size, we use a pooling function for each feature_map [1..n ] to make their dimensions the same. The method selects a 1-Max-pooling function, namely, selects a maximum value from feature_map [1..n ] as output so as to inhibit the phenomenon of mean shift of estimation caused by network parameter errors, and can better extract feature information (the most important feature is represented by the maximum value). Thus, each convolution kernel obtains a characteristic which is a value, all convolution kernels are cascaded by using 1-max pooling, and final characteristic vectors can be obtained, so that main characteristics are reserved, the number of parameters is reduced, and the dimension of the final characteristic vectors is reduced. As shown in fig. 8.

4) Fully connected layer (Whole connecting layer)

The results of the upper layer pooling with the 1-Max-pooling function are stitched together, as shown in FIG. 9, and then the softmax activation function is used to get the probability of belonging to each class. And obtaining the corresponding workflow mode through the obtained probability value, namely, the sequence workflow mode is corresponding to the probability value of 1, and the synchronous workflow mode is corresponding to the probability value of-1, so that the inner layer information required by the WPPL standard is extracted.

After obtaining the inner information of the WPPL specification, we obtain the outer information of each service requirement by analyzing the document, i.e. including its own input, output, preconditions, post-conditions. Finally, by combining the obtained inner and outer layers of information, the service requirements can be converted into WPPL specifications so as to model the information, and thus, a high-quality service requirement document can be obtained.

3. Detecting collisions

To detect conflicts in the service requirement document, the method proposes 13 checking rules to verify the service requirement. As shown in table 1:

TABLE 1 conflict check rules for service demand

By utilizing defined checking rules for checking, the problems of inconsistency, inaccuracy, incompleteness and the like in service demand documents can be effectively avoided through examples of trade orders, applicability of the method is improved through wide application of deep learning and workflow modes in industry, and finally the method comprises the following steps:

1) The service requirements document is analyzed to extract the outer information in the WPPL specification.

2) TextCNN is used to extract the inner layer flow information in the WPPL specification, including sub-services and control flows between them.

3) The WPPL specification is generated by combining the results obtained in the first two steps.

4) The WPPL specification is verified by checking against workflow pattern matches and checking rules.

5) And outputting the verification result.

The above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced with equivalents; such modifications and substitutions do not depart from the spirit of the technical solutions according to the embodiments of the present invention.

Claims

1. The service demand conflict detection method integrating the deep learning and workflow modes is characterized by comprising the following steps of:

s21, inputting natural language into an embedded layer to perform vector conversion;

s23, using a pooling function for each feature map to enable the dimensions of the feature maps to be the same;

s24, obtaining probability values of each class by using an activation function according to the result of the full connection layer splicing pooling, and obtaining a workflow mode corresponding to the probability values according to the probability values;

s4, automatically detecting the service demand conflict by adopting a detection rule.

2. The method for detecting conflict of service demands by fusing deep learning and workflow modes as claimed in claim 1, wherein the step S2 is performed first, the document is analyzed to obtain the outer layer information of each service demand through the inner layer information obtained in the step S2, and then the step S3 is performed.

3. The service demand conflict detection method integrating deep learning and workflow patterns according to claim 1 or 2, wherein the step S22 specifically comprises the steps of:

o _i ＝w·A[i：i+h-1]，i＝1，2，...，s-h+1

s represents: the method comprises the steps of performing word segmentation on an input sentence to obtain s words, namely the number of rows of a matrix A;

c _i ＝f(o _i +b)

one convolution kernel gets the feature c, s-h+1 features in total;

s224, a plurality of convolution kernels of different heights are employed.

4. The method for detecting service demand conflict in combination with deep learning and workflow patterns according to claim 1 or 2, wherein in step S23, a 1-Max-pooling function is adopted, a maximum value is selected from the feature map as an output, and each convolution kernel obtains a feature as a value.

5. The service demand conflict detection method integrating deep learning and workflow modes according to claim 1 or 2, wherein the activation function of S24 adopts sigmoid as the activation function, and the function expression is:

6. the method for detecting conflict between service demands by combining deep learning and workflow modes according to claim 1 or 2, wherein in the step S3, service demands are modeled by a workflow-based process language, operation languages are used to represent operations of the service demands and objects of the operations, descriptions of the service demands are divided into an inner layer structure and an outer layer structure, an inner layer describes process flows in the service, and an outer layer describes key attributes of the service.

7. The method for detecting service demand conflict in combination with deep learning and workflow patterns as recited in claim 6, wherein said inner layer, each service demand is one or more business processes comprised of a plurality of operations connected by a series of control flows, said control flows being represented by workflow patterns.

8. The method for detecting service demand conflict in combination with deep learning and workflow patterns according to claim 6, wherein the relationship attributes of the outer layer description include: input and output data are represented by entities by states, and pre and post conditions are represented by states (entities).

9. The method for detecting service demand conflict in fusion of deep learning and workflow patterns according to claim 1, wherein the workflow patterns include five sub-services of sequential pattern, parallel split pattern, synchronous pattern, exclusive selection pattern and simple merge pattern, and the business operations, business services or a combination service mapped to atoms are cooperated by a set of predefined workflows;

10. The method for detecting service demand conflict in combination with deep learning and workflow patterns according to claim 1 or 2, wherein the embedding layer in step S21 performs word embedding by using word2vec method to form a two-dimensional matrix as input of the convolution layer.