A method and device for detecting power grid dispatching events based on semantic feature capture

Abstract

The present invention discloses a method and device for detecting power grid dispatching events based on semantic features. The method includes obtaining text information with domain vocabulary and common vocabulary; training the domain vocabulary representation information and common vocabulary representation information to obtain basic data text information; processing the vocabulary in the basic data text information to obtain processed text information; algorithmically verifying the information involving Chinese in the processed text information and training it through a hybrid representation algorithm; inputting the preliminarily classified vocabulary information into a pre-constructed convolutional neural network to obtain the final detection result. The present invention can identify specific types of event instances in specific types in plain text. It can help power grid dispatching management to extract events and entities in work through automation and intelligent means at a deeper level, thereby improving the work efficiency of staff.

Description

Technical Field

[0001] The invention relates to a method and a device for detecting power grid dispatching events based on semantic features, belonging to the technical field of semantic recognition. Background Art

[0002] With the gradual deepening and research on the requirements for the construction of dual-carbon and new power systems, the construction of various information and intelligent systems for power grid dispatching has also made rapid progress. However, during its construction process, the volume and types of various trace documents of power grid dispatching have gradually increased, and more automated and intelligent means are needed to help power grid dispatching management generate events and entities in its work, thereby improving the work efficiency of staff.

[0003] In the current development history of artificial intelligence, event extraction is one of the important tasks of natural language processing, and event detection (ED) is one of the key steps in event extraction. It uses deep learning to identify event trigger words and further realizes event classification, aiming to identify specific types of event instances with specific types in plain text.

[0004] To date, many methods have been proposed [1-2] and have achieved good performance. Among them, document-based event monitoring methods in traditional media mainly detect events through text similarity and clustering. Yang [3-4] et al. proposed the basic steps of document-based event detection, including text preprocessing, data representation, data organization or clustering. These steps are still the basic components of many event detection methods. Salton [5] used TFIDF in his paper method to weight important words in documents according to the corpus. This method has also been widely used by subsequent event detection methods. In order to improve the shortcomings of term vectors and bag-of-words models, Kumaran [6] et al. proposed a VSM text vector model that integrates named entities to weight important word features and make up for the shortcomings of TFIDF.

[0005] However, existing event extraction methods find it challenging to capture sufficient semantic information from plain text since words may have different meanings in different sentences.

[0006] For example, in the two sentences "Sentence 1: The total installed capacity of thermal power stations in Province A is xxxkW" and "Sentence 2: Standardized requirements for the behavior of power supply employees when standing", the word "stand" has different meanings in different sentence scenarios; in addition, in sentence 1, "installed capacity" is a complete word in the power grid business scenario. In common word segmentation tools, "installed capacity" will be split into two words, "installed" and "capacity", and information cannot be extracted correctly and completely. Therefore, it is difficult to fully utilize these prompt words for effective information extraction through traditional word embedding. Summary of the invention

[0007] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and device for capturing power grid dispatch event detection based on semantic features, which can identify specific types of event instances in specific types in plain text. It can help power grid dispatch management to extract events and entities in work through automation and intelligent means at a deeper level, thereby improving the work efficiency of staff.

[0008] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0009] In a first aspect, the present invention provides a method for detecting power grid dispatching events based on semantic features, comprising:

[0010] Get text information with domain vocabulary and common vocabulary;

[0011] Performing training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information;

[0012] Training the domain vocabulary representation information and general vocabulary representation information to obtain basic data text information;

[0013] Processing the words in the basic data text information to obtain processed text information; wherein the processing content includes marking the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head;

[0014] Performing algorithm verification on the information related to Chinese in the processed text information, and training it through a hybrid representation algorithm;

[0015] Divide the information related to Chinese into words and perform preliminary classification of the words;

[0016] The initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result.

[0017] Furthermore, the text information is trained to obtain domain vocabulary representation information and general vocabulary representation information, including:

[0018] Acquire features from domain vocabulary and common vocabulary based on Token-Level neural network;

[0019] Let T = t 1 , t 2 , ..., t n where t i is the number of tokens in the sentence, and let x i Yes i The embedding of , introduces a convolutional layer with a window size of s to capture the composition semantics, the formula is as follows:

[0020]

[0021] Formula (1) shows the convolution process, where w i is the filter of the convolutional layer, x j：j+s-1 is the embedding layer from x j to x j +s-1 series, b i is bias;

[0022] Formula (2) provides important signals for different parts of a sentence by using dynamic multi-pooling, where Related c The aggregated results on the left are Related c The aggregated result on the right; by connecting and Get the representation of domain vocabulary f bword ; By using the same process on common word-level sequences, we obtain the representation f of common words nword .

[0023] Further, the words in the basic data text information are processed to obtain processed text information, including:

[0024] Use POS to tag the vocabulary, where POS is the part of speech of the word;

[0025] Analyze the dependency relationship between words through DR;

[0026] The distance between each word and Head is calculated using the DIS algorithm.

[0027] Furthermore, the information related to Chinese in the processed text information is subjected to algorithm verification and trained by a hybrid representation algorithm, the formula is as follows:

[0028] αTi=s(WTif bword +UTif nword +VTif′F+bTi) (3)

[0029] αTc=s(WTcf bword +UTcf nword +VTcf′F+bTc) (4)

[0030] Where s is the sigmoid function, W∈Rd'×d', U∈Rd'×d' and V∈Rd'×d' are weight matrices, and b is the bias;

[0031] Construct an 82-dimensional vector f'F, obtaining the concatenation of features and words as a new representation: f′h =[f′ F f bword ]; Based on the grid introduced into the trigger kernel generator and the event type classifier, the final vector as input is obtained as follows:

[0032] fTi=αTif bword +(1-αTi)f′h (5)

[0033] fTc=αTcf nword +(1-αTc)f′h (6)

[0034] Where fTi is the mixed feature of trigger identification, fTc is the mixed function of event type classifier, αTi and (1-αti) represent the f in trigger identification respectively. nword and f' h In terms of importance, αTc plays a similar role in the event type classifier.

[0035] Furthermore, the initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result, including:

[0036] Get the concatenation of the embedding and the relative position xc, concatenate xc to a pre-defined feature vector, and then use the vocabulary-level features as input to the convolutional layer to capture the compositional semantics and obtain the feature map;

[0037] Depending on the number of convolution kernels in the sentence, the feature map output c j Divided into i parts, through dynamic multi-pooling, the final output of a filter is composed of p ji =max(c ji ) is given to obtain the p of each feature map ji , and all p ji Connect to the final result formed.

[0038] Furthermore, it also includes: connecting the feature vector and the vocabulary feature into a vector Fword, obtaining a character-level vector Fchar, and generating two mixed representations for the trigger recognition and trigger type classification layers.

[0039] In a second aspect, the present invention provides a device for detecting power grid dispatching events based on semantic features, comprising:

[0040] An acquisition unit, used for acquiring text information with domain vocabulary and common vocabulary;

[0041] A first training unit is used to perform training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information;

[0042] A second training unit is used to train the domain vocabulary representation information and the general vocabulary representation information to obtain basic data text information;

[0043] A processing unit, used to process the words in the basic data text information to obtain processed text information; wherein the processing content includes annotating the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head;

[0044] A hybrid representation training unit; used for algorithmically verifying the information related to Chinese in the processed text information and training it through a hybrid representation algorithm;

[0045] A classification unit, used to divide the information related to Chinese into words and perform preliminary classification of words;

[0046] The detection result acquisition unit is used to input the preliminarily classified vocabulary information into a pre-built convolutional neural network to obtain a final detection result.

[0047] In a third aspect, the present invention provides a device for detecting power grid dispatching events based on semantic feature capture, comprising a processor and a storage medium;

[0048] The storage medium is used to store instructions;

[0049] The processor is used to operate according to the instructions to execute the steps of any of the methods described above.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention adopts a hybrid representation method for learning information from words, characters and dependencies. By using a Token-Level neural network to learn two separate character-level and word-level representations; and obtaining dependency information from a dependency parser, generating its representation by read-hot encoding, through technical method research, and related experimental determination, it is possible to identify specific types of event instances in specific types in plain text, and can help power grid dispatching management generate events and entities in work through automation and intelligent means at a deeper level, thereby improving the work efficiency of staff. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a dependency graph provided by an embodiment of the present invention;

[0054] Figure 2Schematic diagram of a scheduling text semantic event detection framework provided by an embodiment of the present invention;

[0055] Figure 3 It is a schematic diagram of a trigger detection framework provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0057] Example 1

[0058] This embodiment introduces a method and device for detecting power grid dispatching events based on semantic features, including:

[0059] Get text information with domain vocabulary and common vocabulary;

[0060] Performing training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information;

[0061] Training the domain vocabulary representation information and general vocabulary representation information to obtain basic data text information;

[0062] Processing the words in the basic data text information to obtain processed text information; wherein the processing content includes marking the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head;

[0063] Performing algorithm verification on the information related to Chinese in the processed text information, and training it through a hybrid representation algorithm;

[0064] Divide the information related to Chinese into words and perform preliminary classification of the words;

[0065] The initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result.

[0066] The method and device for detecting power grid dispatching events based on semantic features provided in this embodiment specifically involve the following steps in their application process:

[0067] The proposed model is divided into two stages and processes the mixed representation through a dynamic multi-pooling convolutional neural network; the first stage is trigger recognition, which uses the character composition structure of the trigger to capture the potential trigger semantics containing the relevant role by using a Token-Level neural network. The second stage is used to determine the specific type of event, which is called trigger type classification. We use the dependency information extraction from the dependency parser to generate feature representations through a Token-Level neural network in these two stages; finally, it is combined with the word feature representation to obtain a mixed representation.

[0068] The architecture of the power grid dispatch semantic event detection method consists of four parts, including the representation of input sequence, feature representation based on dependency parser, hybrid representation, and dynamic multi-polling convolutional neural network. Figure 2 As shown:

[0069] First, trigger recognition-input sequence representation, this paper uses two-level embedding, namely domain vocabulary embedding and common vocabulary level embedding. To further improve performance, this paper uses pre-trained weights to initialize embedding. And two Token-Level neural networks are used to obtain features from domain vocabulary and common vocabulary. The network architecture is similar to NPNS. Let T = t 1 , t 2 , ..., t n where t i is the number of tokens in the sentence, and let x i Yes i The embedding (domain vocabulary or common words) of is introduced into a convolutional layer with a window size of s to capture the composition semantics. Its hidden layer is identified as follows:

[0070]

[0071] Formula (1) shows the convolution process, where w i is the filter of the convolutional layer, x j：j+s-1 is the embedding layer from x j to x j +s-1 series, b i is bias.

[0072] Equation (2) provides important signals for different parts of a sentence by using dynamic multi-pooling, where Related c The aggregated results on the left are Related c The aggregated result on the right; by connecting and Get the representation of domain vocabulary f bword By using the same process on the common word-level sequences, we can also obtain the common word representation fnword The methods include:

[0073] 1) Input text information with domain vocabulary and common vocabulary for training;

[0074] 2) Perform training processing through formula (1) and formula (2) to obtain domain vocabulary representation information and general vocabulary representation information.

[0075] 3) Train mature and available domain vocabulary representation information and general vocabulary representation information, and store them in the power grid information corpus as the basic data information of the dependency analyzer algorithm;

[0076] Second, trigger identification - feature representation based on dependency analysis. The dependency parser is an important part of dependency parsing based on dependency relations. Syntactic dependencies can be used to obtain deep semantic information. Syntactic dependencies are used in neural network models by directly incorporating them into embeddings. In this work, we use three different feature abstraction layers to represent three features:

[0077] POS: Part-of-speech tagging.

[0078] DR: Dependency Analysis.

[0079] DIS: Distance from Head

[0080] POS is the part of speech of a word, which plays an important role in natural language, such as the recognition of named entities, grammatical analysis, and event extraction. A noun or a pronoun can be the subject of a sentence, but they cannot be transformed into each other. This is because the grammatical component has restrictions on the phonetic component. Therefore, POS is suitable for abstract features to express different characteristics of text semantic information. This paper uses POS as a feature to strengthen word-based features. There are about 52 common parts of speech in Chinese. This paper will use a 52-dimensional unique hot vector to represent the part of the sentence POS. This means that the POS of each word in the sentence can be represented as a 52-dimensional feature vector. Each dimension of the 52-dimensional vector represents a part of the POS. Indicates the position of a word if one of the values ​​is 1 and the remaining 52 dimensions are 0.

[0081] Dependencies express the semantic relationship between sentence components. For event detection tasks, trigger words are usually predicates (i.e., verbs). Generally, in the power grid dispatch corpus, the role of trigger words as verb objects accounts for 19%. Therefore, this paper believes that dependencies can be used to improve trigger detection. At the feature layer of dependencies, the vector dimension is 23 (22 relationship types and one 'other' type). We found that 22 types of dependencies often use syntactic dependencies, and in order to reduce the complexity of feature representation, we classify other dependencies as "other" types.

[0082] The distance from the head is dependent on the length of the path. Specifically, if a word is directly related to the head, the distance to the head is defined as 1 in this paper. If the path includes an intermediate dependency, the distance to the head is 2. For example, in Figure 1 the sentence, the dependency path for the misoperation of the secondary circuit triggering the heavy gas alarm is as follows:

[0083] Misoperation of the secondary circuit → Trigger → Heavy gas alarm

[0084] where "misoperation of the secondary circuit" is the head in this sentence, and "trigger" creates an intermediate dependency between the head and "heavy gas alarm".

[0085] The method includes:

[0086] 1) As Figure 1 , input the text information;

[0087] 2) Based on the "input sequence representation" step, obtain the vocabulary in the text information, including domain vocabulary and common vocabulary;

[0088] 3) Perform part-of-speech tagging on the vocabulary through POS;

[0089] 4) Analyze the dependency relationships between the vocabulary through DR;

[0090] 5) Calculate the distance of various types of vocabulary from the Head through the DIS algorithm;

[0091] 6) The information such as the vocabulary, part-of-speech analysis, dependency relationships, and distances in the first two steps is used as the information input for the "hybrid representation learning" step.

[0092] Third, hybrid representation learning. For Chinese event detection, using only domain vocabulary representation or common vocabulary-level representation cannot obtain sufficient information. For example, if understanding "trigger" in the character-level representation, it is a trigger composed of '引' and '发'. In the word-level representation, the word-level sequence can provide more explicit information to distinguish the semantics of '引'.

[0093] After the embedding layer, the word-level feature representation f bword , a character-level representation f nword and a feature representation f′ F can be obtained. In this paper, two gates are learned to simulate the information flow of the trigger recognition and event type classifier, and the formula is as follows:

[0094] αTi = s(WTif bword + UTif nword + VTif′F + bTi) (3)

[0095] αTc = s(WTcf bword + UTcfnword +VTcf′F+bTc) (4)

[0096] Where s is the sigmoid function, W∈Rd'×d', U∈Rd'×d' and V∈Rd'×d' are weight matrices, and b is the bias. Figure 2 With three feature layers in , we can get three feature representations. We construct an 82-dimensional vector f'F by connecting these three feature representations. Finally, we can get the concatenation of features and words as a new representation: f' h =[f′ F f bword ]; Based on the grid introduced into the trigger buckle kernel generator and event type classifier, we can obtain the final vector as input.

[0097] fTi=αTif bword +(1-αTi)f′h (5)

[0098] fTc=αTcf nword +(1-αTc)f′h (6)

[0099] Where fTi is the mixed feature of trigger identification, and fTc is the mixed function of event type classifier. αTi and (1-αti) represent the f in trigger identification respectively. nword The importance of f'h and αTc plays a similar role in the event type classifier.

[0100] Methods include:

[0101] 1) Input the processing results of the first two steps involving Chinese information into this step;

[0102] 2) Perform algorithm verification on the vocabulary information divided in the first step to avoid decomposing vocabulary-level information into character-level information;

[0103] 3) Implementing hybrid representation algorithm training through formula (3), formula (4), formula (5), and formula (6);

[0104] 4) Divide Chinese-related information into more reasonable vocabulary;

[0105] 5) The event type classifier is used to preliminarily classify the vocabulary as the data information input for the next step of "dynamic multi-pooling convolutional neural network".

[0106] Fourth, dynamic multi-pooling convolutional neural network. Traditional convolutional neural networks only use one pooling layer to implement the maximum operation. This means that traditional convolutional neural networks only capture the most important information in the representation of a sentence. In event detection, a sentence may contain two or more events, and parameters can use different triggers to obtain different roles. However, traditional convolutional neural networks only capture the most useful information of the entire sentence and lose other information in the same sentence. In order to solve the above problems, a team used a dynamic multi-pooling convolutional neural network (DMCNN) to obtain more valuable information without losing the maximum pool.

[0107] In this study, we will use a similar neural network, such as Figure 3 The architecture of its trigger detection is described.

[0108] First, the concatenation of the embedding and the relative position xc is obtained, and xc is concatenated to a pre-defined feature vector. The word-level features are then used as the input of the convolution layer to capture the compositional semantics and obtain the feature map. Specifically, the convolution operation generates new features by using the convolution kernel to scan a window of H words. Let x i :i+j refers to word x i , x i +1, ..., x i +j. The convolution kernel is applied to a window x of H words in the sentence. 1 :h,x 2 :h+1,...,x nh+1 :n to generate a feature map c i , where the index ranges from 1 to NH+1. A convolution kernel produces a feature i at a position i :c ij =σ(wj · xi ： i+h-1+b) where σ is nonlinear (Tanh is generally used), j ranges from 1 to m, and m is the number of convolution kernels.

[0109] Afterwards, the feature map output c is calculated based on the number of convolution kernels in the sentence. j For example, if a sentence has one convolution kernel, the sentence will be divided into two parts. When the sentence has two convolution kernels, the two triggers will divide the sentence into three parts. Through dynamic multi-pooling, the final output of a filter is composed of p ji =max(c ji ) is given to obtain the p of each feature map ji , and all p ji Connect to the final result formed.

[0110] Finally, the above feature vectors and vocabulary features are concatenated into a vector Fword. We adopt a similar approach to obtain the character-level vector Fchar, producing two hybrid representations for the trigger recognition and trigger type classification layers.

[0111] Finally, the training and classification of the algorithm. During the training process, this paper treats event detection as a multi-class classification problem. The mixed representation fN learned from the architecture described above is the input of the convolutional neural network. The input is the sentence after word segmentation. Dropout is added to prevent overfitting.

[0112] Deep learning can be further studied and utilized through experiments to identify event trigger words and further implement event classification, which aims to identify specific types of event instances in plain text.

[0113] Finally, with the goal of event recognition and classification, the architecture of the power grid dispatching semantic event detection method is constructed through four parts: input sequence representation, feature representation based on dependency analysis, hybrid representation learning, and dynamic multi-pooling convolutional neural network. At a deeper level, through automation and intelligent means, it helps power grid dispatching management to extract events and entities in work, thereby improving the work efficiency of staff.

[0114] Example 2

[0115] This embodiment provides a device for detecting power grid dispatching events based on semantic features, including:

[0116] An acquisition unit, used for acquiring text information with domain vocabulary and common vocabulary;

[0117] A first training unit is used to perform training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information;

[0118] A second training unit is used to train the domain vocabulary representation information and the general vocabulary representation information to obtain basic data text information;

[0119] A processing unit, used to process the words in the basic data text information to obtain processed text information; wherein the processing content includes annotating the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head;

[0120] A hybrid representation training unit; used for algorithmically verifying the information related to Chinese in the processed text information and training it through a hybrid representation algorithm;

[0121] A classification unit, used to divide the information related to Chinese into words and perform preliminary classification of words;

[0122] The detection result acquisition unit is used to input the preliminarily classified vocabulary information into a pre-built convolutional neural network to obtain a final detection result.

[0123] Example 3

[0124] This embodiment provides a device for detecting power grid dispatching events based on semantic features, including a processor and a storage medium;

[0125] The storage medium is used to store instructions;

[0126] The processor is configured to operate according to the instructions to perform the steps of any of the following methods:

[0127] Get text information with domain vocabulary and common vocabulary;

[0128] Performing training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information;

[0129] Training the domain vocabulary representation information and general vocabulary representation information to obtain basic data text information;

[0130] Processing the words in the basic data text information to obtain processed text information; wherein the processing content includes marking the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head;

[0131] Performing algorithm verification on the information related to Chinese in the processed text information, and training it through a hybrid representation algorithm;

[0132] Divide the information related to Chinese into words and perform preliminary classification of the words;

[0133] The initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result.

[0134] Example 4

[0135] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of any of the following methods are implemented:

[0136] Get text information with domain vocabulary and common vocabulary;

[0137] Performing training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information;

[0138] Training the domain vocabulary representation information and general vocabulary representation information to obtain basic data text information;

[0139] Processing the words in the basic data text information to obtain processed text information; wherein the processing includes marking the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head;

[0140] Performing algorithm verification on the information related to Chinese in the processed text information, and training it through a hybrid representation algorithm;

[0141] Divide the information related to Chinese into words and perform preliminary classification of the words;

[0142] The initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result.

[0143] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting power grid dispatching events based on semantic features, characterized in that: include: Get text information with domain vocabulary and common vocabulary; Performing training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information; Training the domain vocabulary representation information and general vocabulary representation information to obtain basic data text information; Processing the words in the basic data text information to obtain processed text information; wherein the processing content includes marking the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head; Performing algorithm verification on the information related to Chinese in the processed text information, and training it through a hybrid representation algorithm; Divide the information related to Chinese into words and perform preliminary classification of the words; The initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result.

2. The method for detecting power grid dispatching events based on semantic features according to claim 1 is characterized in that: The text information is trained to obtain domain vocabulary representation information and general vocabulary representation information, including: Acquire features from domain vocabulary and common vocabulary based on Token-Level neural network; Let T = t1, t2, ..., t n where t i is the number of tokens in the sentence, and let x i Yes i The embedding of , introduces a convolutional layer with a window size of s to capture the composition semantics, the formula is as follows: h ij =tanh(w i x j：j+s-1 +b i ) (1) Formula (1) shows the convolution process, where w i is the filter of the convolutional layer, x j：j+s-1 is the embedding layer from x j to x j+s-1 The series connection, b i is bias; Formula (2) provides important signals for different parts of a sentence by using dynamic multi-pooling, where Related c The aggregated results on the left are Related c The aggregated result on the right; by connecting and Get the representation fbword of the domain vocabulary; and get the representation fnword of the common vocabulary by using the same process on the common vocabulary level sequence.

3. The method for detecting power grid dispatching events based on semantic features according to claim 2 is characterized in that: Processing the words in the basic data text information to obtain processed text information includes: Use POS to tag the vocabulary, where POS is the part of speech of the word; Analyze the dependency relationship between words through DR; The distance between each word and Head is calculated using the DIS algorithm.

4. The method for detecting power grid dispatching events based on semantic features according to claim 3 is characterized in that: The information related to Chinese in the processed text information is verified by the algorithm and trained by the hybrid representation algorithm. The formula is as follows: αTi=s(WTifbword+UTifnword+VTif′F+bTi) (3) αTc=s(WTcfbword+UTcfnword+VTcf′F+bTc) (4) Where s is the sigmoid function, W∈Rd'×d', U∈Rd'×d' and V∈Rd'×d' are weight matrices, and b is the bias; Construct an 82-dimensional vector f'F, obtaining the concatenation of features and words as a new representation: f′ h =[f′ F f bword ]; Based on the grid introduced into the trigger kernel generator and the event type classifier, the final vector as input is obtained as follows: fTi=αTifbword+(1-αTi)f′h (5) fTc=αTcfnword+(1-αTc)f′h (6) Where fTi is the hybrid feature for trigger identification, fTc is the hybrid function for event type classifier, αTi and (1-αTi) represent the importance of fbword and f'h in trigger identification, respectively, and αTc plays a similar role in event type classifier.

5. The method for detecting power grid dispatching events based on semantic features according to claim 4 is characterized in that: The initially classified vocabulary information is input into a pre-built convolutional neural network to obtain the final detection result, including: Get the concatenation of the embedding and the relative position xc, concatenate xc to a pre-defined feature vector, and then use the vocabulary-level features as input to the convolutional layer to capture the compositional semantics and obtain the feature map; Depending on the number of convolution kernels in the sentence, the feature map output c j Divided into i parts, through dynamic multi-pooling, the final output of a filter is composed of p ji =max(c ji ) is given to obtain the p of each feature map ji , and all p ji Connect to the final result formed.

6. The method for detecting power grid dispatching events based on semantic features according to claim 5 is characterized in that: Also includes: The feature vectors and vocabulary features are concatenated into a vector Fword to obtain a character-level vector Fchar, generating two hybrid representations for the trigger recognition and trigger type classification layers.

7. A device for detecting power grid dispatching events based on semantic features, characterized in that: include: An acquisition unit, used for acquiring text information with domain vocabulary and common vocabulary; A first training unit is used to perform training processing on the text information to obtain domain vocabulary representation information and general vocabulary representation information; A second training unit is used to train the domain vocabulary representation information and the general vocabulary representation information to obtain basic data text information; A processing unit, used to process the words in the basic data text information to obtain processed text information; wherein the processing content includes annotating the words, analyzing the dependency relationship between the words, and calculating the distance between each word and the head; A hybrid representation training unit; used for algorithmically verifying the information related to Chinese in the processed text information and training it through a hybrid representation algorithm; A classification unit, used to divide the information related to Chinese into words and perform preliminary classification of words; The detection result acquisition unit is used to input the preliminarily classified vocabulary information into a pre-built convolutional neural network to obtain a final detection result.

8. A device for detecting power grid dispatching events based on semantic features, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images