Big data AI-based intelligent questioning and answering method and system for electric power audit knowledge

Abstract

The invention belongs to the technical field of natural language models, and provides an electric power audit knowledge intelligent question-answering method and system based on big data AI, aiming at different question-answering habits, text modes, question-answering purposes and the like of different users, the satisfaction degree of the users to replies is recognized through a feedback recognition model, and the user experience is improved. And a one-way three-level question and answer network is constructed through the pre-extraction model, the entity generation model and the relationship generation model, so that the occupation of a single question and answer session on computing resources can be flexibly adjusted, and the question and answer accuracy and efficiency are improved.

Description

Technical Field

[0001] The present application relates to the field of natural language model technology, and in particular to a method and system for intelligent question-answering of power audit knowledge based on big data AI. Background Art

[0002] The statements in this section merely provide background technical information related to the present application and do not necessarily constitute prior art.

[0003] The original audit report preparation, question answering, and rectification supervision rely on manual labor, and lack vertical training solutions for specific audit fields. Based on the current situation, our company has developed and built a power audit knowledge base and knowledge graph that combines "regulations library, question library, case library, and report library" using natural language models. It uses advanced semantic analysis technology and innovative cross-attribute knowledge fusion and expression technology to achieve the transformation of power audit knowledge from fragmentation to systematization, from implicit to explicit, and constructs a knowledge-intensive and structured power audit intelligent knowledge graph, and provides a question-and-answer system based on natural language models.

[0004] In the actual application of knowledge graphs, it is found that different users have different Q&A writing habits, Q&A purposes, and ability to conceive and summarize. They also have a lack of trust in the adaptability of large language models and the intelligent fusion responses of knowledge graphs, which results in excessive redundancy in the question language or deviation of the question semantics from the actual question being asked.

[0005] Typically, large sections of text are directly copied into the question-answering system, which then contains a large amount of invalid information, and then a concise question is asked at the end. The question-answering system uniformly parses and semantically understands all the question data, resulting in a waste of computing resources and easily leading to a deviation in the semantics of the question, reducing the accuracy and efficiency of the question-answering. Summary of the invention

[0006] In order to solve the problems in the background technology, this application proposes an intelligent question-answering method and system for power audit knowledge based on big data AI.

[0007] This application provides an intelligent question-answering method for power audit knowledge based on big data AI, including the following steps: Step 1: Establish a question-answering session and start receiving natural language question data; Step 2: Use the pre-extraction model to extract core questions from the latest question data, extract entities and relationships based on the core questions, generate pre-retrieval triple sequences, query the knowledge graph based on the pre-retrieval triple sequences, and output the answers; Step 3: Use the feedback recognition model to perform satisfaction semantic recognition on the latest input reply feedback. If the recognition semantics is positive satisfaction, jump to step 2; if the recognition semantics is negative satisfaction, jump to step 4; Step 4: Use the entity generation model to generate entity candidates from the latest question data, extract the question semantics based on the relationship generation model, generate relationship candidates, use the entity candidates and relationship candidates to query the knowledge graph, obtain the answer triples and output them; Step 5: Use the feedback recognition model to evaluate the satisfaction of the latest input response feedback. According to the recognition result, associate the question data in the previous text with the latest question data, and jump to step 4 in a loop until the session ends.

[0008] Preferably, in step 2, the specific method of extracting entities and relationships in the question data by the pre-extraction model is: Preprocess the question data and extract the smallest natural language paragraph L that is at the end and conforms to the question pattern; Extract entity naming words in L, convert entity naming words into word vectors, use word vectors to obtain semantic features, and obtain entities based on semantic features and associated knowledge graphs; Retrieve entity types to obtain a list of common relations, extract relation feature words in L and perform semantic matching with elements in the list of common relations, and take the element with the highest matching value as the target relation; Based on the acquired entity and target relationship, the knowledge graph is queried to acquire the target entity, and then the answer triple is acquired; The pre-extraction model is a composite model consisting of a variant neural network combining LSTM and Linear layers and a RoBERTa-large semantic recognition model.

[0009] Preferably, the feedback recognition model is a composite model consisting of a variant neural network combining LSTM and Linear layers and a RoBERTa-large semantic recognition model.

[0010] Preferably, in step 3, the specific method of the feedback recognition model to perform satisfaction semantic recognition on the reply feedback is: Perform word segmentation on question data and answer feedback to obtain a question word segmentation list, a feedback word segmentation list, and a feedback feature word list. Convert each word segmentation list into a word vector, use the word vector to obtain semantic features, match the preset satisfaction evaluation feature word library based on the semantic features of the feedback feature words, obtain the satisfaction evaluation with the highest matching value, evaluate the correlation between the semantic features of the question word segmentation and the semantic features of the feedback word segmentation, and judge the satisfaction based on the satisfaction evaluation and the correlation evaluation.

[0011] Preferably, in step 5, when the feedback recognition model semantically recognizes the latest input reply feedback satisfaction as negative satisfaction, the semantic features of the question data in the above text are associated with the semantic features of the latest question, and the word segmentation of the associated previous question is added to the word segmentation list of the latest question and deduplication is performed.

[0012] Preferably, in step 4, the entity generation model is a GRU-CRF model, and the entity extraction method is specifically as follows: Perform word segmentation on natural language questions to obtain word sequences, and perform part-of-speech tagging on each word to obtain a part-of-speech sequence; Convert the part-of-speech sequence into word vectors, and input the word vectors into the annotation model to obtain the optimal annotation sequence; Link the labeled optimal solution to the knowledge graph to obtain entity candidates.

[0013] Preferably, in step 4, the relationship generation model is an LSTM variant neural network with an attention mechanism, and the relationship extraction method is specifically as follows: Preprocess natural language questions to exclude entity names or replace them with placeholders; The preprocessed question is input into the LSTM decoder to obtain the relationship list and relationship score, and the highest score is selected as the candidate relationship.

[0014] Preferably, the original question and the preprocessed question are input into the LSTM decoder to obtain the relationship lists L1 and L2 respectively, sort L1 and L2 based on the relationship scores, and select the element with the highest ranking in L1 and matching the element in L2 as the candidate relationship.

[0015] This application also provides an intelligent question-answering system for power audit knowledge based on big data AI, including: conversation module, pre-extraction model, feedback recognition model, entity generation model, and relationship generation model. The functions of each module are as follows: Conversation module: establishes and ends conversations, receives natural language input from users during conversations, and outputs responses; Pre-extraction model: intercepts the last paragraph of the question in the question data, performs lightweight entity and relationship extraction, and quickly generates answers; Feedback recognition model: Identify the user's satisfaction with the reply, switch the extraction model, and form a unidirectional three-level recognition network; Entity generation model and relationship generation model: Process the latest question data as a whole to extract entities and relationships, and query the knowledge graph to obtain answers.

[0016] Compared with the prior art, the beneficial effects of this application are: This application targets different users' different question-and-answer habits, writing styles, and question-and-answer purposes, and uses a feedback recognition model to identify users' satisfaction with responses. It then constructs a unidirectional three-level question-and-answer network through a pre-extraction model, an entity generation model, and a relationship generation model, which is conducive to flexibly adjusting the computing resources occupied by a single question-and-answer session and improving the accuracy and efficiency of question-and-answering. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0018] Figure 1 This is a schematic diagram of the system composition of this application. Figure 2 Schematic diagram of the method flow of this application. DETAILED DESCRIPTION

[0019] The present application is further described below in conjunction with the accompanying drawings and embodiments.

[0020] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0021] In the present disclosure, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom" and the like indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are relational words determined only for the convenience of describing the structural relationships of the various parts or elements of the present disclosure, and do not specifically refer to any part or element in the present disclosure and should not be understood as limitations on the present disclosure.

[0022] Example 1 like Figure 1 to Figure 2 As shown, this application provides an intelligent question-answering method for power audit knowledge based on big data AI, including the following steps: Step 1: Establish a question-answering session and start receiving natural language question data; Step 2: Use the pre-extraction model to extract core questions from the latest question data, extract entities and relationships based on the core questions, generate pre-retrieval triple sequences, query the knowledge graph based on the pre-retrieval triple sequences, and output the answers; Step 3: Use the feedback recognition model to perform satisfaction semantic recognition on the latest input reply feedback. If the recognition semantics is positive satisfaction, jump to step 2; if the recognition semantics is negative satisfaction, jump to step 4; Step 4: Use the entity generation model to generate entity candidates from the latest question data, extract the question semantics based on the relationship generation model, generate relationship candidates, use the entity candidates and relationship candidates to query the knowledge graph, obtain the answer triples and output them; Step 5: Use the feedback recognition model to evaluate the satisfaction of the latest input response feedback. According to the recognition result, associate the question data in the previous text with the latest question data, and jump to step 4 in a loop until the session ends.

[0023] The session is a chat window, which is actively established by the user or automatically established after inputting initial question and answer data, and the session ends when the user actively operates to end it.

[0024] This application uses a feedback recognition model to recognize the user's reply feedback, and divides the semantic parsing operation of the question into a unidirectional three-level parsing network based on the pre-extraction model, entity generation model, and relationship generation model. First, the pre-extraction model is used to perform lightweight entity and relationship extraction on the core questions in the question data. If it is recognized that the user is satisfied with the reply, the pre-extraction model is used to extract entities and relationships. When it is recognized that the user is dissatisfied with the reply, the entity generation model and the relationship generation model are used to extract entities and relationships for the latest question as a whole. The operation is maintained until the feedback recognition model recognizes that the user is dissatisfied with a certain round of replies. The context is then associated with the latest question data and the entity generation model and the relationship generation model are used to extract entities and relationships until the end of the session. The increase in the "complexity" of the question data is unidirectional and irreversible in each session, which is conducive to flexibly adjusting the computing resource occupancy of a single session and improving the accuracy and efficiency of question and answering.

[0025] Specifically, in step 2, the pre-extraction model is a composite model consisting of a variant neural network combining LSTM and Linear layers and a RoBERTa-large semantic recognition model, and the specific method for extracting entities and relationships in question data is: First, the question data is preprocessed to extract the smallest natural language paragraph L that is at the end and conforms to the question pattern. The neural network is trained to recognize and extract typical question formats such as "why", "what" and "how". The smallest natural language paragraph that is at the end and conforms to the question pattern in the latest question data is intercepted and considered as the core question.

[0026] Extract the entity naming words in the question L to generate the word segmentation sequence SL=(s1,s2,...,s n ), s nFor words, embed SL and convert it into a vector to obtain the embedding vector EL=(e1,e2,...,e m ), m is the maximum value of the element number, the vector EL is input into the language model RoBERTa-large to extract semantic features, and the semantic features are input into the LSTM and Linear layers for BIO sequence labeling to obtain the entity and its position in the question.

[0027] The entity type is obtained according to the preset entity type table, the entity type is retrieved to obtain a list of common relations, the relation feature words in L are extracted for semantic matching with the elements in the common relation list, the element with the highest matching value is taken as the target relation, and the knowledge graph is queried based on the above-obtained entities and target relations to obtain the target entity, and then the reply triple is obtained.

[0028] As an optional technical solution, due to the complexity of Chinese, in order to avoid special cases where the relationship feature words are vague or missing or the common relationship list is not entered, the above-mentioned method of extracting entities is used to extract relationships as an alternative or parallel technical solution, as follows: The question is converted into a vector, and the semantic features are extracted and input into the neural network model for BIO sequence labeling to obtain the relationship and its position in the question.

[0029] The LSTM model is a long short-term memory network, a special type of recurrent neural network (RNN) designed specifically to solve the gradient vanishing or gradient exploding problems encountered by traditional RNNs when processing long sequence data. LSTM can effectively learn long-term dependency information through its unique "gate" mechanism and performs well in a variety of tasks involving sequence data, such as natural language processing, speech recognition, and time series prediction. The Linear layer is a fully connected layer in the neural network. RoBERTa-large is a large pre-trained language model based on the Transformer architecture. It is a large-scale model in the RoBERTa (RobustlyOptimized BERT approach) model series and is widely used in semantic feature extraction and entity extraction.

[0030] Specifically, the feedback recognition model is a composite model consisting of a variant neural network combining LSTM and Linear layers and a RoBERTa-large semantic recognition model. In step 3, the specific method for semantically recognizing satisfaction with the reply feedback is as follows: Perform word segmentation on question data and answer feedback to obtain a question word segmentation list, a feedback word segmentation list, and a feedback feature word list. Convert each word segmentation list into a word vector, use the word vector to obtain semantic features, match the semantic features of the feedback feature words with the preset satisfaction evaluation feature word library, and obtain the satisfaction evaluation with the highest matching value. At the same time, perform correlation evaluation on the semantic features of the question word segmentation and the semantic features of the feedback word segmentation, and make a satisfaction judgment based on the satisfaction evaluation and the correlation evaluation.

[0031] The method for obtaining semantic features is consistent with the above-mentioned entity extraction method and will not be repeated here. The focus of the satisfaction recognition method is on two aspects. One is to train the neural network to recognize the accuracy of feedback feature words. The feedback feature words typically include negative words, such as "wrong", "error", "not", etc. However, in the actual intelligent question-answering system, after not getting the desired answer, the user may not express feedback directly, but modify the wording of the question or add more conditions to obtain an accurate answer. Therefore, identifying the repetitiveness between two questions asked by the user (the correlation mentioned above) can also obtain the satisfaction with the previous question. When training the neural network model, the matching of semantic features can be quantified through preset conditions to obtain quantitative satisfaction evaluation and correlation evaluation.

[0032] Specifically, in step 4, the entity generation model is a GRU-CRF model, and the entity extraction method is as follows: Perform word segmentation on natural language questions to obtain a word sequence, perform part-of-speech tagging on each word segment to obtain a part-of-speech sequence, convert the part-of-speech sequence into a word vector, input the word vector into the tagging model to obtain the optimal tagging sequence, and link the optimal tagging solution to the knowledge graph to obtain entity candidates.

[0033] The GRU-CRF model is a recurrent neural network that integrates conditional random fields. It combines the advantages of the gated recurrent unit (GRU) and the conditional random field (CRF). Compared with LTSM, its model complexity is lower, the training speed is faster, and the ability to connect context is stronger. More importantly, the use of different models to extract entities and relationships is conducive to improving recognition accuracy and preventing model confusion.

[0034] After obtaining a question, the GRU layer performs word segmentation on the question. Preferably, part-of-speech tagging (noun, verb, adjective) can be performed to generate word vectors, which are then transferred to the CRF layer transfer probability matrix to generate a possible set of sequence labels and their corresponding probabilities. The optimal label sequence is selected as the entity labeling result. In the process, the BIO (head, middle, outer) mode is used for sequence labeling.

[0035] For example, the score of each annotation result is as follows: Among them, X indicates the question, X=(x1,x2,...,x n ), x is the word in the question, y=(y1,y2,...,y m ) represents the labeling result, P is the state feature matrix of the conditional random field, A is the state transfer matrix, Ai,j indicates the score of transferring from the i-th label to the j-th label, for all possible labeling result sets Y X , use the softmax regression function to get the probability of each annotation: The training goal of the model is to maximize the log-conditional probability of the correct label sequence: In practice, the Viterbi algorithm is used to solve the optimal labeling sequence.

[0036] Specifically, in step 4, the relationship generation model is an LSTM variant neural network with an attention mechanism, and the relationship extraction method is specifically as follows: Preprocess natural language questions to exclude entity names or replace them with placeholders; The preprocessed question is input into the LSTM decoder to obtain the relationship list and relationship score, and the highest score is selected as the candidate relationship.

[0037] The relationship generation model consists of a two-layer LSTM encoder and an LSTM decoder with an attention mechanism. Given a question, after encoding by the encoder (including word segmentation, vector processing, etc.), the decoder can directly obtain the relationship set corresponding to the question and the relationship corresponding probability (score). In this process, the semantic contribution of entity words is not great, and they are replaced by placeholders. The attention mechanism performs weight-based screening on the context, allowing the neural network to more accurately find information related to the current output, thereby improving the accuracy of the output.

[0038] As an optional technical solution, considering the redundant negative contribution of entity words and the possibility of semantic deviation after removing entity words in reverse, the original question and the preprocessed question are input into the LSTM decoder to obtain the relationship lists L1 and L2 respectively, and L1 and L2 are sorted based on the relationship scores. The element with the highest ranking in L1 and matching the element in L2 is selected as the candidate relationship. Matching the element in L2 means that the corresponding relationship exists in the L2 list.

[0039] Preferably, in step 5, when the feedback recognition model semantically recognizes the latest input reply feedback satisfaction as negative satisfaction, the semantic features of the question data in the above text are associated with the semantic features of the latest question, and the word segmentation of the associated previous question is added to the word segmentation list of the latest question and deduplication is performed.

[0040] The training feedback recognition model recognizes the correlation of semantic features. When the answer is inaccurate, the word segmentation data in the previous text is added to the latest question. In the continuous extraction of entities and relations, the recognized semantic features are retained to facilitate their correlation and matching with the semantic features of the latest question.

[0041] Example 2 On the basis of Example 1, the present application also provides an intelligent question-answering system for power audit knowledge based on big data AI, including: a conversation module, a pre-extraction model, a feedback recognition model, an entity generation model, and a relationship generation model. The functions of each module are as follows: Conversation module: establishes and ends conversations, receives natural language input from users during conversations, and outputs responses; Pre-extraction model: intercepts the last paragraph of the question in the question data, performs lightweight entity and relationship extraction, and quickly generates answers; Feedback recognition model: Identify the user's satisfaction with the reply, switch the extraction model, and form a unidirectional three-level recognition network; Entity generation model and relationship generation model: Process the latest question data as a whole to extract entities and relationships, and query the knowledge graph to obtain answers.

[0042] The above are only preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0043] Although the above describes the specific implementation methods of the present application in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present application. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present application, various modifications or deformations that can be made by technical personnel in this field without creative work are still within the scope of protection of the present application.

Claims

1. The intelligent question-answering method of power audit knowledge based on big data AI is characterized by: The steps include: Step 1: Establish a question-answering session and start receiving natural language question data; Step 2: Use the pre-extraction model to extract core questions from the latest question data, extract entities and relationships based on the core questions, generate pre-retrieval triple sequences, query the knowledge graph based on the pre-retrieval triple sequences, and output the answers; Step 3: Use the feedback recognition model to perform satisfaction semantic recognition on the latest input reply feedback. If the recognition semantics is positive satisfaction, jump to step 2; if the recognition semantics is negative satisfaction, jump to step 4; Step 4: Use the entity generation model to generate entity candidates from the latest question data, extract the question semantics based on the relationship generation model, generate relationship candidates, use the entity candidates and relationship candidates to query the knowledge graph, obtain the answer triples and output them; Step 5: Use the feedback recognition model to evaluate the satisfaction of the latest input response feedback. According to the recognition result, associate the question data in the previous text with the latest question data, and jump to step 4 in a loop until the session ends.

2. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 1 is characterized by: In step 2, the specific method of extracting entities and relationships in the question data by the pre-extraction model is: Preprocess the question data and extract the smallest natural language paragraph L that is at the end and conforms to the question pattern; Extract entity naming words in L, convert entity naming words into word vectors, use word vectors to obtain semantic features, and obtain entities based on semantic features and associated knowledge graphs; Retrieve entity types to obtain a list of common relations, extract relation feature words in L and perform semantic matching with elements in the list of common relations, and take the element with the highest matching value as the target relation; Based on the acquired entity and target relationship, the knowledge graph is queried to acquire the target entity, and then the answer triple is acquired; The pre-extraction model is a composite model consisting of a variant neural network combining LSTM and Linear layers and a RoBERTa-large semantic recognition model.

3. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 1 is characterized by: The feedback recognition model is a composite model consisting of a variant neural network combining LSTM and Linear layers and a RoBERTa-large semantic recognition model.

4. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 3 is characterized by: In step 3, the specific method of the feedback recognition model to perform satisfaction semantic recognition on the reply feedback is: Perform word segmentation on question data and answer feedback to obtain a question word segmentation list, a feedback word segmentation list, and a feedback feature word list. Convert each word segmentation list into a word vector, use the word vector to obtain semantic features, match the preset satisfaction evaluation feature word library based on the semantic features of the feedback feature words, obtain the satisfaction evaluation with the highest matching value, evaluate the correlation between the semantic features of the question word segmentation and the semantic features of the feedback word segmentation, and judge the satisfaction based on the satisfaction evaluation and the correlation evaluation.

5. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 4 is characterized by: In step 5, when the feedback recognition model semantically recognizes the latest input reply feedback satisfaction as negative satisfaction, the semantic features of the question data in the previous text are associated with the semantic features of the latest question, and the word segmentation of the associated previous question is added to the word segmentation list of the latest question and deduplication is performed.

6. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 1 is characterized by: In step 4, the entity generation model is a GRU-CRF model, and the entity extraction method is specifically as follows: Perform word segmentation on natural language questions to obtain word sequences, and perform part-of-speech tagging on each word to obtain a part-of-speech sequence; Convert the part-of-speech sequence into word vectors, and input the word vectors into the annotation model to obtain the optimal annotation sequence; Link the labeled optimal solution to the knowledge graph to obtain entity candidates.

7. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 6 is characterized by: In step 4, the relationship generation model is an LSTM variant neural network with an attention mechanism, and the relationship extraction method is specifically as follows: Preprocess natural language questions to exclude entity names or replace them with placeholders; The preprocessed question is input into the LSTM decoder to obtain the relationship list and relationship score, and the highest score is selected as the candidate relationship.

8. The intelligent question-answering method for power audit knowledge based on big data AI according to claim 7 is characterized by: The original question and the preprocessed question are input into the LSTM decoder to obtain the relationship lists L1 and L2 respectively. L1 and L2 are sorted based on the relationship scores, and the element with the highest ranking in L1 and matching the element in L2 is selected as the candidate relationship.

9. Intelligent question and answer of power audit knowledge based on big data AI, characterized by: include: Conversation module, pre-extraction model, feedback recognition model, entity generation model, relationship generation model, the functions of each module are as follows: Conversation module: establishes and ends conversations, receives natural language input from users during conversations, and outputs responses; Pre-extraction model: intercepts the last paragraph of the question in the question data, performs lightweight entity and relationship extraction, and quickly generates answers; Feedback recognition model: Identify the user's satisfaction with the reply, switch the extraction model, and form a unidirectional three-level recognition network; Entity generation model and relationship generation model: Process the latest question data as a whole to extract entities and relationships, and query the knowledge graph to obtain answers.

Images