Attention-based power equipment standard question and answer method, system, and storage medium
By constructing retrieval and reader models and combining them with attention mechanisms, the problems caused by the large number of standard documents for power equipment, multiple versions, and cross-references of content were solved, resulting in an efficient and accurate question-answering method and system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ELECTRIC POWER RES INST OF STATE GRID ZHEJIANG ELECTRIC POWER COMAPNY
- Filing Date
- 2022-12-23
- Publication Date
- 2026-05-12
AI Technical Summary
The large number of power equipment standard documents, their multiple versions, and the complex cross-references between them result in low efficiency in question-and-answer processes, low accuracy in answer extraction, and a poor user experience.
We construct a retrieval model and a reader model, combine them with an attention mechanism, process the question description information, obtain the answer description context, and predict the answer span.
It improved the efficiency of question and answering, enhanced the accuracy of answer extraction, and improved the user experience.
Smart Images
Figure CN115905495B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an attention-based standard question-answering method, system, and storage medium for power equipment, belonging to the field of standard question-answering technology for power equipment. Background Technology
[0002] Against the backdrop of the national "dual carbon" goals and enterprise digital transformation, the State Grid Corporation of China is committed to promoting the deep integration of the traditional power grid industry with digital technologies, vigorously developing "new digital infrastructure," and building a modern equipment management system for the State Grid Corporation of China. Power equipment standard texts, as important data resources, are a key means and important acceptance condition for ensuring the standardized construction of "digital field operations."
[0003] Because power equipment standard documents are a type of unstructured question-and-answer information source, they are complex and challenging due to their large quantity, multiple versions, and cross-references. This makes real-time querying during production management difficult, affecting question-and-answer efficiency and hindering widespread adoption.
[0004] Furthermore, due to the diversity and complexity of Chinese expressions, the same standard question may have multiple descriptions, resulting in a low accuracy rate for answer extraction. Repeated manual question answering would consume a lot of manpower and resources, leading to a poor user experience. Summary of the Invention
[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for answering power equipment standards by constructing a retrieval model and a reader model to process question description information and obtain the required answer description context. This method effectively alleviates the problems caused by the large number of power equipment standard documents, multiple versions, and cross-references, improving question-and-answer efficiency and facilitating widespread use. Simultaneously, by integrating an attention mechanism to process the answer description context and question description information, the method obtains the prediction result of the answer span, completing an attention-based question-and-answer system for power equipment standards. This improves the accuracy of answer extraction. The solution is scientific, reasonable, feasible, and provides a good user experience.
[0006] To address the shortcomings of existing technologies, the second objective of this invention is to provide an attention-based question-and-answer system for power equipment standards. This system, by setting up a question acquisition module, a retrieval module, and a reader module, processes the question description information to obtain the required answer description context C. This effectively alleviates the problems caused by the large number of power equipment standard documents, multiple versions, and cross-references of content, improves question-and-answer efficiency, facilitates widespread use, and increases the accuracy of answer extraction. The solution is scientific, reasonable, feasible, and provides a good user experience.
[0007] To address the shortcomings of existing technologies, the third objective of this invention is to provide an attention-based question-and-answer method, system, and storage medium for power equipment standards that effectively alleviates the problems caused by the large number of power equipment standard documents, multiple versions, and cross-references of content, improves question-and-answer efficiency, facilitates widespread use, and improves the accuracy of answer extraction. The solution is scientific, reasonable, feasible, and provides a good user experience.
[0008] To achieve one of the above objectives, the first technical solution of the present invention is as follows:
[0009] An attention-based standard question-answering method for power devices.
[0010] Retrieve the given problem description information;
[0011] The question description information is processed by a pre-built retrieval model to obtain the required answer description context;
[0012] By utilizing a pre-built reader model and integrating an attention mechanism, the answer description context and question description information are processed to obtain the prediction result of the answer span;
[0013] Based on the predicted answer span, complete the attention-based standard question answering for electrical equipment.
[0014] Through continuous exploration and experimentation, this invention constructs a retrieval model and a reader model to process the question description information and obtain the required answer description context. This effectively alleviates the problems caused by the large number of power equipment standard documents, multiple versions, and cross-references, improves question-and-answer efficiency, and facilitates widespread use.
[0015] Furthermore, this invention integrates an attention mechanism to process the context of the answer description and the information of the question description, thereby obtaining the prediction result of the answer span and completing the attention-based standard question answering for power equipment. This can improve the accuracy of answer extraction, and the solution is scientific, reasonable, feasible, and provides a good user experience.
[0016] As a preferred technical measure:
[0017] The retrieval model processes the question description information in the following ways:
[0018] Based on the problem description information, identify the name of the search subject and determine the problem type;
[0019] By using a pre-established coarse- and fine-grained retrieval graph structure and based on the name of the retrieval subject, the core standard document is located, and several standard documents within the retrieval range are used as supplementary answer documents.
[0020] The core standard document and the supplementary answer document are merged to obtain a candidate document set;
[0021] Based on the problem type, classify and determine the category to which the problem description information belongs;
[0022] Locate the required answer description context based on the category of the question description information and the candidate document set.
[0023] As a preferred technical measure:
[0024] The method for identifying the name of the search subject is as follows:
[0025] For a given problem description information Q, a vectorized representation in character units is obtained by vector space encoding using a pre-trained language model, as shown in the following expression:
[0026] E Q =Encoder([w1,w2,…,w n ])
[0027] Among them, E Q For vectorized representation, w i For the i-th character;
[0028] For the obtained vectorized representation E Q After extracting contextual features through a bidirectional long short-term memory neural network layer, the sequence feature representation H is obtained. Q Its expression is as follows:
[0029] H Q =BILSTM([e1,e2,…,e n ])
[0030] Among them, e i The i-th context feature is vectorized into a representation;
[0031] The feature representation H is then used. Q The input conditional random field layer is used for sequence decoding to obtain the complete name of the retrieval device, N, which is included in the problem description information Q. Q Its expression is as follows:
[0032] P NQ =CRF(H Q )
[0033] The output of the CRF layer of the Conditional Random Field is labeled using the BIO annotation method. The annotation content is the retrieval main information, where B represents the start position of the retrieval main content, I represents the middle position of the retrieval main content, and O represents the remaining content in the question description information Q.
[0034] As a preferred technical measure:
[0035] The methods for determining the type of problem are as follows:
[0036] Vectorize the problem description information into E Q The input is fed into several independent fully connected (FC) layers for feature extraction and vector space mapping, and then classified and discriminated using a SoftMax activation function layer. The resulting problem description information belongs to the category T. Q Its expression is as follows:
[0037] T Q =SoftMax(FC(E Q ))
[0038] The calculation method for the SoftMax activation function is as follows:
[0039]
[0040] Where i is the output value of the i-th node, and K is the number of output nodes, i.e. the number of categories.
[0041] As a preferred technical measure:
[0042] The method to obtain the answer description context C is as follows:
[0043] To address the interoperability of power equipment standard documents, the coarse-grained and fine-grained retrieval graph structure is configured with coarse-grained retrieval graph units and fine-grained graph structures.
[0044] Coarse-grained retrieval graph units are used to form a graph structure from the candidate document set according to reference relationships.
[0045] The expression for the graph structure is as follows:
[0046] G L ={N;L};
[0047] Where N is a node, representing the name of a power equipment standard document; L is the relationship between nodes, representing the reference relationship between power equipment standards, and the type is based on;
[0048] For graph structure G L ={N;L}, the fine-grained graph structure is divided into different types of description nodes N′ according to its description content, thus forming a fine-grained graph structure G. F ={G L ;N′};
[0049] Based on the name of the search subject and combined with the graph structure G L = {N; L}, using the maximum matching method to quickly locate the core standard document, while all standard documents within one hop of the query are used as supplementary documents for the answer, finally obtaining the candidate document set D. C ;
[0050] Candidate document set D C It should include at least one core standard document and several candidate supplementary documents, and its expression is as follows:
[0051]
[0052] Based on the problem description information type T Q and candidate document set D C and fine-grained graph structure G F ={G L ;N′} quickly locates the required answer description context C.
[0053] As a preferred technical measure:
[0054] The reader model processes the answer description context C and the question description information to obtain the answer span prediction result as follows:
[0055] By encoding the answer description context C and the question description information using a pre-trained language model, a context vector representation E in Chinese characters is obtained. C With the problem description information vector representation E Q ;
[0056] The obtained context vector is represented as E C With the problem description information vector representation E Q The fusion yields a fusion vector representation;
[0057] Then, the fused vector is passed through a BILSTM layer to obtain a hidden layer vector that integrates the contextual features of the answer and the question description information;
[0058] Then, the hidden layer vector is input into the attention layer, thereby removing the name of the electrical equipment in the question and obtaining information including the content of the question object. This information is represented by a vector that matches the content of the object in the context of the answer, and the attention weight score is calculated.
[0059] Furthermore, when passing through the attention layer, the question description information is masked using zero-one vectors to facilitate subsequent prediction of the answer span;
[0060] The vector representation in the answer context is updated based on the calculated attention weight score to obtain the updated vector representation HA. The vector representation HA does not include the subject and object content of the question.
[0061] The vector representation HA is processed by two independent activation functions, Sigmoid, to predict the start and end positions of the correct answer, respectively.
[0062] As a preferred technical measure:
[0063] The formula for calculating the hidden layer vector is as follows:
[0064]
[0065] Among them, E CQ This is a fusion vector representation.
[0066] The formula for calculating the weighted score is as follows:
[0067]
[0068]
[0069]
[0070] Among them, A S A is a set consisting of characters representing the object's content. S C For its complement, e i a is a vector representation of characters. ij For weighted scores, This is the vector representation after multiplying by the weighted score.
[0071] As a preferred technical measure:
[0072] The formula for calculating the starting position is as follows:
[0073] p1 = Sigmoid(K1H) A +b1)
[0074] The formula for calculating the ending position is as follows:
[0075] p2 = Sigmoid(K2H) A +b2).
[0076] Where p1 and p2 are the start and end positions, respectively, and k i With b i H is a hyperparameter. n This is the updated input vector after weight calculation.
[0077] The formula for calculating the sigmoid function is as follows:
[0078]
[0079] To achieve one of the above objectives, the second technical solution of the present invention is as follows:
[0080] An attention-based standard question-answering system for electrical equipment.
[0081] It includes a question acquisition module, a search engine module, and a reader module;
[0082] The question acquisition module is used to obtain the given question description information; the retrieval module is used to process the question description information to obtain the required answer description context C.
[0083] The reader module is used to integrate the attention mechanism to process the answer description context C and the question description information to obtain the prediction result of the answer span;
[0084] Based on the predicted answer span, complete the attention-based standard question answering for electrical equipment.
[0085] Through continuous exploration and experimentation, this invention, by setting up a question acquisition module, a retrieval module, and a reader module, processes the question description information to obtain the required answer description context C. This effectively alleviates the problems caused by the large number of power equipment standard documents, multiple versions, and cross-references, improves question-and-answer efficiency, and facilitates widespread use.
[0086] Furthermore, this invention integrates an attention mechanism to process the context of the answer description and the information of the question description, thereby obtaining the prediction result of the answer span and completing the attention-based standard question answering for power equipment. This can improve the accuracy of answer extraction, and the solution is scientific, reasonable, feasible, and provides a good user experience.
[0087] To achieve one of the above objectives, the third technical solution of the present invention is as follows:
[0088] A computer-readable storage medium,
[0089] It stores a computer program that, when executed by a processor, implements the aforementioned attention-based standard question-answering method for power devices.
[0090] Compared with the prior art, the present invention has the following beneficial effects:
[0091] Through continuous exploration and experimentation, this invention constructs a retrieval model and a reader model to process the question description information and obtain the required answer description context. This effectively alleviates the problems caused by the large number and multiple versions of power equipment standard documents, improves question-and-answer efficiency, and facilitates widespread use.
[0092] Furthermore, this invention integrates an attention mechanism to process the context of the answer description and the information of the question description, thereby obtaining the prediction result of the answer span and completing the attention-based standard question answering for power equipment. This can improve the accuracy of answer extraction, and the solution is scientific, reasonable, feasible, and provides a good user experience.
[0093] Through continuous exploration and experimentation, this invention, by setting up a question acquisition module, a retrieval module, and a reader module, processes the question description information to obtain the required answer description context C. This effectively alleviates the problems caused by the large number and multiple versions of power equipment standard documents, improves question-and-answer efficiency, facilitates widespread use, and enhances the accuracy of answer extraction. The solution is scientific, reasonable, feasible, and provides a good user experience. Attached Figure Description
[0094] Figure 1 This is a flowchart of a standard question-and-answer method for power equipment according to the present invention;
[0095] Figure 2 This is another flowchart of the standard question-and-answer method for power equipment of the present invention;
[0096] Figure 3 This is a flowchart of one processing step of the retrieval model of the present invention;
[0097] Figure 4 This is a flowchart of a reader model of the present invention;
[0098] Figure 5 This is a flowchart illustrating one aspect of the masking operation in this invention. Detailed Implementation
[0099] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0100] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.
[0101] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention.
[0102] like Figure 1 As shown, this is the first specific embodiment of the attention-based question-answering method for standard electrical equipment of the present invention:
[0103] An attention-based standard question-answering method for power devices.
[0104] Retrieve the given problem description information;
[0105] The question description information is processed by a pre-built retrieval model to obtain the required answer description context;
[0106] By utilizing a pre-built reader model and integrating an attention mechanism, the answer description context and question description information are processed to obtain the prediction result of the answer span;
[0107] Based on the predicted answer span, complete the attention-based standard question answering for electrical equipment.
[0108] like Figure 2 As shown, this is the second specific embodiment of the attention-based question-answering method for standard electrical equipment of the present invention:
[0109] An attention-based standard question answering method for power equipment has evolved from a three-stage architecture of question analysis, document retrieval, and answer extraction to a "retrieval-reader" architecture with the development of deep learning. It can be formally described as follows: For a given question answering task T, the question Q and the text set D are taken as input, and the output is the question answer A with the highest probability of being correct in the text set D, which is either "retrieved" or "generated".
[0110] This invention relates to the Electrical Equipment Standards Question Answering (EES-QA) model, an automatic retrieval and extraction model that primarily comprises a retrieval model and a reader model.
[0111] The retrieval model retrieves relevant text fragments based on the characteristics of standard texts for power equipment and the questions raised by users; the reader model predicts the answer range from the combined document content.
[0112] like Figure 3 As shown, a specific embodiment of the retrieval model of the present invention is as follows:
[0113] New power equipment standards are based on national or previous standards, adding additional descriptive content or further subdividing the equipment to regulate its use. The existence of cross-references between different power equipment standards increases the difficulty of text retrieval. However, as standardized documents, power equipment standards maintain consistency between their content descriptions and hierarchical headings, offering the advantage of a strong logical structure.
[0114] The document retrieval model aims to find content descriptions related to the problem from numerous power equipment standard documents.
[0115] Extracting the retrieval subject can be formally described as follows: For a given question description information Q, a vectorized representation in character units is obtained by using a pre-trained language model for vector space encoding, as shown in formula (2).
[0116] E Q =Encoder([w1,w2,…,w n (2)
[0117] For the obtained vectorized representation E Q Contextual features are extracted using a bidirectional long short-term memory (BILSTM) neural network layer, further yielding the sequence feature representation H. Q As shown in formula (3). Then the feature representation H... Q The input conditional random field (CRF) layer is used for sequence decoding to obtain the complete retrieval device entity name N included in question Q. Q As shown in formula (4).
[0118] H Q =BILSTM([e1,e2,…,e n (3)
[0119] P NQ =CRF(H Q (4)
[0120] The output of the Conditional Random Field (CRF) layer is labeled using the BIO annotation method. The annotation content is the retrieval subject information, where "B" indicates the start position of the retrieval subject content, "I" indicates the middle position of the retrieval subject content, and "O" indicates other content in question Q.
[0121] Because power equipment standards are characterized by clear document structure and distinct content types, this paper summarizes the problem description information types by consulting experts and summarizing the document content methods of power equipment standards, as shown in Table 1. Based on the problem type, problem classification is formally described as follows: For the vector representation E of the problem description information obtained by formula (2)... Q The input is fed into several independent fully connected layers (FC) for further feature extraction and vector space mapping. Finally, a SoftMax layer is used for classification and discrimination to determine the category T to which the problem description information belongs. Q As shown in formula (5).
[0122] T Q =SoftMax(FC(EQ (5)
[0123] SoftMax is used as the activation function, and its calculation method is shown in formula (6).
[0124]
[0125] For power equipment documents, using traditional information retrieval algorithms to match questions with documents results in slower response times as the number of documents increases. Furthermore, to address the inter-application characteristics of power equipment standard documents, this paper designs a coarse-grained and fine-grained retrieval graph structure. Coarse-grained refers to forming a graph structure G based on the reference relationships of power equipment standard documents. L ={N;L}, where node N is the name of the power equipment standard document, and the relationship L between nodes represents the reference relationship between power equipment standards, with the type "based on". For example... Figure 3 As shown in the lower right part, the national standard GB / T 6451-2015 "Technical Parameters and Requirements for Oil-Immersed Power Transformers" is based on the national standard GB 1094.1-2013 "Power Transformers Part 1: General Rules" and the industry standard JB / T10088-2004 "Sound Level of 6kV~500kV Power Transformers". Therefore, a coarse-grained retrieval map can be generated according to the citation relationship. The retrieval subject name obtained according to formula (4) can be used to quickly locate the core standard by using maximum matching. At the same time, all standards within one hop of the query are used as supplementary documents for the answer, and finally the candidate document set D is obtained. C This includes a core standard document and several candidate supplementary documents, as shown in formula (7).
[0126]
[0127] Based on the coarse-grained retrieval graph, each standard file is divided into different types of description nodes N′ according to its description content, ultimately forming a fine-grained graph structure G. F ={G L ;N′}. The problem description information type T obtained according to formula (5) Q and candidate document set D C Quickly locate the context of the answer description you need (C).
[0128] like Figure 4 As shown, a specific embodiment of the reader model of the present invention is as follows:
[0129] Based on the answer description context C obtained through the retrieval system, this paper designs a reader model that incorporates an attention mechanism to predict the span of the answer to question Q. The model structure of the reader part is as follows: Figure 4As shown, it is mainly divided into three parts: feature encoding of context and question description information, attention vector fused with location, and answer span prediction.
[0130] The reader component can be formally described as follows: For a given question description information Q, the retrieval system obtains candidate context content descriptions C, and the question description information Q is represented by a character vector E. Q The problem description information includes the complete name of the power equipment, N. Q Candidate contexts are encoded into vectorized representations E, using Chinese character units, through a pre-trained language model. C As shown in formula (8).
[0131] E C =Encoder([w1,w2,…,w s ])(8)
[0132] The resulting vector is represented as E. C With the problem description information vector representation E Q The fusion yields the fusion vector representation E. CQ =concatenate(E C ||E Q ), where || represents the vector concatenation operation. Before vector fusion, a certain degree of masking is performed using the obtained full names of the power equipment, that is, the vector representation of the power equipment name in the question description information is multiplied by a masking threshold τ (0 << τ << 1). Since after document retrieval, it is necessary to focus on extracting the content most relevant to the question description information from the answer context, the importance of the power equipment name as the sole main content in the answer context decreases after locating it by name. For example... Figure 5 As shown, the question is "What is the contact capacity of a gas relay?". In this case, the key to extracting the answer is the content following "contact capacity". In short, the masking operation improves the accuracy of answer extraction by reducing the weight of the name of the power equipment.
[0133] Then, the fusion vector is passed through the BILSTM layer to obtain the hidden layer vector that integrates the contextual features of the answer and the question description information, as shown in Equation (9).
[0134]
[0135] Then, the hidden layer vector is input into the attention layer, so that after removing the name of the electrical equipment in the question, information including the content of the question object can be obtained. The attention weight score is calculated using the vector representation in the answer context that matches the content of the object. The weight score is calculated as shown in formulas (10), (11), and (12), where A S A is a set consisting of characters representing the object's content.S C For its complement, e i The vector representation is in character units. The updated vector representation H is obtained by updating the vector representation in the answer context representation based on the calculated attention weight score. A The update vector does not include the subject and object content of the problem.
[0136]
[0137]
[0138]
[0139] Furthermore, when passing through the attention layer, the question description information is masked using a zero-one vector to facilitate subsequent prediction of the answer span. Finally, this paper uses two independent sigmoid activation functions to predict the start and end positions of the correct answer, respectively. The formula for calculating the start position is shown in formula (13), and the formula for calculating the end position is shown in formula (14).
[0140] p1 = Sigmoid(K1H) A +b1) (13)
[0141] p2 = Sigmoid(K2H) A +b2) (14).
[0142] This patent proposes an EES-QA retrieval-reader model for power equipment standards, which effectively retrieves relevant power equipment standard context and extracts answers for a given question Q. Compared to other question-answering models, the EES-QA model proposes a retrieval and extraction method tailored to the domain characteristics of power equipment standards, providing a new approach for building a question-answering system in the power field.
[0143] Furthermore, this invention is a retrieval and location method for power equipment standards, which can effectively alleviate the problems caused by the large number and multiple versions of power equipment standards.
[0144] Furthermore, this invention proposes an attention mechanism for both the subject and object that combines occlusion operations, which can effectively improve the accuracy of answer extraction. As a result, this invention can outperform other models in question-answering accuracy on both public datasets and independently constructed standard datasets of power equipment.
[0145] A system embodiment applying the method of the present invention:
[0146] An attention-based question-answering system for power equipment standards includes a question acquisition module, a retrieval module, and a reader module. The question acquisition module is used to obtain given question description information. The retrieval module is used to process the question description information to obtain the required answer description context C. The reader module is used to integrate an attention mechanism to process the answer description context C and the question description information to obtain a prediction result of the answer span. Based on the prediction result of the answer span, the attention-based question-answering system for power equipment standards is completed.
[0147] An embodiment of a device applying the method of the present invention:
[0148] A computer device comprising:
[0149] One or more processors;
[0150] Storage device for storing one or more programs;
[0151] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described attention-based standard question-answering method for power devices.
[0152] An embodiment of a computer medium applying the method of the present invention:
[0153] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described attention-based standard question-answering method for power devices.
[0154] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0155] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.
[0156] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0157] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. An attention-based standard question-answering method for electrical equipment, characterized in that, Retrieve the given problem description information; The question description information is processed by a pre-built retrieval model to obtain the required answer description context; The retrieval model processes the question description information in the following ways: Based on the problem description information, identify the name of the search subject and determine the problem type; By using a pre-established coarse- and fine-grained retrieval graph structure and based on the name of the retrieval subject, the core standard document is located, and several standard documents within the retrieval range are used as supplementary answer documents. The core standard document and the supplementary answer document are merged to obtain a candidate document set; Based on the problem type, classify and determine the category to which the problem description information belongs; Locate the required answer description context based on the category of the question description information and the candidate document set; By utilizing a pre-built reader model and integrating an attention mechanism, the answer description context and question description information are processed to obtain the prediction result of the answer span; The reader model processes the answer description context C and the question description information to obtain the answer span prediction result as follows: By encoding the answer description context C and the question description information using a pre-trained language model, a context vector representation E in Chinese characters is obtained. C With the problem description information vector representation E Q ; The obtained context vector is represented as E C With the problem description information vector representation E Q The fusion yields a fusion vector representation; Then, the fused vector is passed through a BILSTM layer to obtain a hidden layer vector that integrates the contextual features of the answer and the question description information; Then, the hidden layer vector is input into the attention layer, thereby removing the name of the electrical equipment in the question and obtaining information including the content of the question object. This information is represented by a vector that matches the content of the object in the context of the answer, and the attention weight score is calculated. Furthermore, when passing through the attention layer, the question description information is masked using zero-one vectors to facilitate subsequent prediction of the answer span; The vector representation in the answer context is updated based on the calculated attention weight score to obtain the updated vector representation HA. The vector representation HA does not include the subject and object content of the question. The vector representation HA is processed by two independent sigmoid activation functions to predict the start and end positions of the correct answer, respectively. Based on the predicted answer span, complete the attention-based standard question answering for electrical equipment.
2. The attention-based question-answering method for electrical equipment as described in claim 1, characterized in that, The method for identifying the name of the search subject is as follows: For a given problem description information Q, a vectorized representation in character units is obtained by vector space encoding using a pre-trained language model, as shown in the following expression: E Q =Encoder([w1,w2,…,w n ]) Among them, E Q For vectorized representation, w i For the i-th character; For the obtained vectorized representation E Q After extracting contextual features through a bidirectional long short-term memory neural network layer, the sequence feature representation H is obtained. Q Its expression is as follows: H Q =BILSTM([e1,e2,…,e n ]) Among them, e i The i-th context feature is vectorized into a representation; Then the feature representation H Q The input conditional random field layer is used for sequence decoding to obtain the complete name of the retrieval device, N, which is included in the problem description information Q. Q Its expression is as follows: P NQ =CRF(H Q ) The output of the CRF layer of the Conditional Random Field is labeled using the BIO annotation method. The annotation content is the retrieval main information, where B represents the start position of the retrieval main content, I represents the middle position of the retrieval main content, and O represents the remaining content in the question description information Q.
3. The attention-based question-answering method for electrical equipment as described in claim 1, characterized in that, The methods for determining the type of problem are as follows: Vectorize the problem description information into E Q The input is fed into several independent fully connected (FC) layers for feature extraction and vector space mapping, and then classified and discriminated using a SoftMax activation function layer. The resulting problem description information belongs to the category T. Q Its expression is as follows: T Q =SoftMax(FC(E Q )) The calculation method for the SoftMax activation function is as follows: Where i is the output value of the i-th node, and K is the number of output nodes, i.e. the number of categories.
4. The attention-based question-answering method for electrical equipment as described in claim 1, characterized in that, The method to obtain the answer description context C is as follows: To address the interoperability of power equipment standard documents, the coarse-grained and fine-grained retrieval graph structure is configured with coarse-grained retrieval graph units and fine-grained graph structures. Coarse-grained retrieval graph units are used to form a graph structure from the candidate document set according to reference relationships. The expression for the graph structure is as follows: G L ={N;L}; Where N is a node, representing the name of a power equipment standard document; L is the relationship between nodes, representing the reference relationship between power equipment standards, and the type is based on; For graph structure G L ={N;L}, the fine-grained graph structure is divided into different types of description nodes N′ according to its description content, thus forming a fine-grained graph structure G. F ={G L ;N′}; Based on the name of the search subject and combined with the graph structure G L = {N; L}, using the maximum matching method to quickly locate the core standard document, while all standard documents within one hop of the query are used as supplementary documents for the answer, finally obtaining the candidate document set D. C ; Candidate document set D C It should include at least one core standard document and several candidate supplementary documents, and its expression is as follows: Based on the problem description information type T Q and candidate document set D C and fine-grained graph structure G F ={G L ;N′} quickly locates the required answer description context C.
5. The attention-based question-answering method for electrical equipment as described in claim 1, characterized in that, The formula for calculating the hidden layer vector is as follows: Among them, E CQ This is a fusion vector representation; The formula for calculating the weighted score is as follows: Among them, A S A is a set consisting of characters representing the object's content. S C For its complement, e i a is a vector representation of characters. ij For weighted scores, This is the vector representation after multiplying by the weighted score.
6. The attention-based question-answering method for electrical equipment as described in claim 5, characterized in that, The formula for calculating the starting position is as follows: p1=Sigmoid(K1H A +b1) The formula for calculating the ending position is as follows: p2=Sigmoid(K2H A +b2); Where p1 and p2 are the start and end positions, respectively, and k i With b i H is a hyperparameter. A This is the updated input vector after weight calculation.
7. An attention-based standard question-answering system for electrical equipment, characterized in that, The application includes an attention-based standard question-answering method for power equipment as described in any one of claims 1-6, comprising a question acquisition module, a retrieval module, and a reader module; The problem retrieval module is used to obtain the given problem description information; The retrieval module is used to process the question description information to obtain the required answer description context C; The reader module is used to integrate the attention mechanism to process the answer description context C and the question description information to obtain the prediction result of the answer span; Based on the predicted answer span, complete the attention-based standard question answering for electrical equipment.
8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements an attention-based standard question-answering method for power devices as described in any one of claims 1-6.