An intelligent question-answering method and system based on aviation equipment

Abstract

The present invention discloses an intelligent question-answering method and system based on aviation equipment, which includes extracting entity and intent candidate sets, and sorting the entity and intent candidate sets according to priority; performing first entity error correction and proofreading based on user intent; performing intent extraction, proofreading and reasoning based on a structured question-answering knowledge base and a syntax analysis tree; filling slots with entities, intents, and question-answer types; determining entity types, and performing multi-step queries and reasoning based on central entity information; extracting answers online based on user intent and entity information; recording questions, answers, entities, and intent-related information; and providing answers to users. The system includes an entity and intent acquisition module, an entity error correction and proofreading module, an intent extraction, proofreading and reasoning module, a slot filling module, a matching and reasoning module, a retrieval and information extraction module, an information recording module, and a providing module.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing and search technology, and in particular to an intelligent question-answering method and system based on aviation equipment. Background Art

[0002] The question-answering system is an advanced form of information retrieval system that can accurately and concisely answer questions raised by users in natural language. Traditional question-answering systems are often implemented in a template-based and neural network manner, and there is a strong division of question-answering fields. The template-based question-answering system can accurately match simple user questions, but requires a large number of manually constructed question templates, and the templates are not transferable. The neural network-based question-answering system can accurately identify user intent and accurately obtain user information by filling in entity slots, but requires a large amount of corpus for model training, and when the questions are more complex, the quality of the model response will deteriorate. Therefore, the present application proposes an intelligent question-answering method and device based on aviation equipment. The method can get rid of the dependence of model training on data, effectively extract user intent and entity information, and perform multi-step queries and reasoning based on central entity information; at the same time, the full text can be extracted online for questions without answers, which improves the accuracy and efficiency of question-answering. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide an intelligent question-answering method and system based on aviation equipment. The method and system can get rid of the strong dependence of model training on data, effectively extract user intentions and entity information, and perform multi-step queries and reasoning based on central entity information; at the same time, full-text online extraction can be performed for questions without answers, thereby improving the accuracy and efficiency of question answering.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] An intelligent question-answering method based on aviation equipment, comprising:

[0006] Step 10 extracts entity and intent candidate sets, and sorts the entity and intent candidate sets according to priority levels;

[0007] Step 20 performs first entity error correction and proofreading based on user intent;

[0008] Step 30 performs intention extraction, proofreading and reasoning based on the structured question-answering knowledge base and the syntax analysis tree;

[0009] Step 40: Fill the slots with entities, intents, and question and answer types;

[0010] Step 50 determines the entity type and performs multi-step query and reasoning based on the central entity information;

[0011] Step 60 extracts answers online from the entire text based on user intent and entity information;

[0012] Step 70 records questions, answers, entities, and intent related information;

[0013] Step 80 provides the answer to the user.

[0014] An intelligent question-answering system based on aviation equipment includes: an entity and intent acquisition module, an entity error correction and proofreading module, an intent extraction, proofreading and reasoning module, a slot filling module, a matching and reasoning module, a retrieval and information extraction module, an information recording module and a providing module;

[0015] Entity and intent acquisition module, the user obtains the user input sequence entity and intent candidate set;

[0016] Entity error correction and proofreading module, which is used to correct the user input sequence entities and proofread the entities;

[0017] Intent extraction, proofreading, and reasoning module, used to obtain the true intent of the input sequence;

[0018] Slot filling module, used to normalize input entity information, intent information, and question and answer types;

[0019] Matching and reasoning module, used to obtain answer information of questions;

[0020] Retrieval and information extraction module, used to extract relevant answers to user questions online;

[0021] Information recording module, used to record user question and answer information, collect training corpus and provide comparative data for intent extraction, proofreading and reasoning modules;

[0022] A module is provided for providing answer information to the user.

[0023] Compared with the prior art, one or more embodiments of the present invention may have the following advantages:

[0024] The use of natural language processing technology and search technology allows users to build a question-and-answer system without relevant corpus foundation and when the questions and answers are relatively complex, reducing the difficulty of template building and corpus-based training models, and improving the accuracy of questions and answers, which is of great significance to the construction of the question-and-answer system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the intelligent question-answering method based on aviation equipment;

[0026] Figure 2 A schematic diagram of extracting intent based on structured data in one embodiment of an intelligent question-answering method based on aviation equipment;

[0027] Figure 3 It is a schematic diagram of intention extraction based on grammatical analysis in one embodiment of an intelligent question-answering method based on aviation equipment;

[0028] Figure 4 It is a diagram of graph query and reasoning of an embodiment of an intelligent question-answering method based on aviation equipment;

[0029] Figure 5 This is the structural diagram of the intelligent question-answering system based on aviation equipment. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with embodiments and drawings.

[0031] like Figure 1 The figure shows the process of intelligent question-answering method based on aviation equipment, including:

[0032] Step 10 extracts entity and intent candidate sets based on fixed sentence patterns, professional dictionary segmentation technology, AC automaton technology, neural network entity recognition technology, neural network intent extraction technology, and regular methods, and sorts the entity and intent candidate sets according to their priority levels.

[0033] Specifically, the user inputs a text sequence, and each entity acquisition module and intent acquisition module performs entity recognition and intent extraction respectively, and sorts the entity candidate set according to a predefined priority level.

[0034] The levels of entity extraction are divided into: AC automaton technology > = regular expression > = fixed sentence pattern > = professional dictionary word segmentation technology > = neural network entity recognition technology. Among them, AC automaton is aimed at extracting complex and long text entity data, such as "determination method of nitriding and nitrocarburizing layer depth of aviation steel parts" and "solid material 60-370K specific heat capacity test method". Regular expression is aimed at extracting national standard data, such as "GJB 367A-2001" and "GJB 150.11A-2009". Fixed sentence pattern is a commonly used question syntax template, such as "Who proposed the 10KeV X-ray irradiation reinforcement test method for semiconductor devices?" The entity is "10KeV X-ray irradiation reinforcement test method for semiconductor devices", and the intention is "person" + "proposed". Professional dictionary word segmentation technology loads local dictionaries through Jieba to extract entities, such as "classification standard of the current status of military civil air defense engineering technology" and "revision of high extreme pressure lubricant specifications for helicopters". Neural network entity recognition identifies various types of term entities.

[0035] The intent extraction is divided into the following levels: fixed sentence pattern >= professional dictionary word segmentation technology >= neural network intent extraction technology. The fixed sentence pattern "Who proposed the 10KeV X-ray irradiation hardening test method for semiconductor devices?" has the intent of "person" + "proposed" = "proposer".

[0036] Step 20 performs first entity error correction and proofreading based on user intent;

[0037] Specifically, the first candidate entity with the highest probability is determined based on the user's intention. For example, in "Is the modification of the specification of high extreme pressure lubricant oil for helicopters feasible?", the candidate set entities are "modification of the specification of high extreme pressure lubricant oil for helicopters" and "modification of the specification of high extreme pressure lubricant oil", and the intention is the validity of "modification of the specification of high extreme pressure lubricant oil for helicopters", not whether the "modification of the specification of high extreme pressure lubricant oil for helicopters" is feasible.

[0038] The first entity error correction and proofreading are implemented by ES search technology and Faiss semantic search technology respectively. ES search technology and Faiss obtain entity candidate sets from the string level and semantic level respectively. If the similarity between the ES candidate set and the first entity is greater than 0.95, the first candidate entity is replaced with the first entity information. There is no similar entity in ES, and the Faiss search engine is used to calculate the entity semantic similarity based on word embedding. The entity text training in the Faiss search engine adopts the IndexIVFFlat method, and the word vector calculation adopts the sentence transformer neural network. When the similarity value is greater than 0.85, the entity error correction and replacement are performed. The first entity error correction and proofreading adopts the ES search engine and the Faiss search engine to proofread the entity from the character and semantic levels respectively, which is not limited to a specific method. For example, the word embedding training can adopt "sentence transformer neural network", "Bert", "Genism", etc.; the "IndexIVFFlat" training search engine can adopt "IndexIVFFlat", "IndexFlatL2", etc.

[0039] Step 30 performs intention extraction, proofreading and reasoning based on the structured question-answering knowledge base and the syntax analysis tree;

[0040] Specifically, Figure 2 As shown in the figure, the structured question-answering knowledge base contains information related to previous user questions, answers, entities, and intentions. The questions that mask entity information are similar to the questions that mask entity information in the structured question-answering knowledge base to extract user intentions. At the same time, the syntactic analysis is performed on the words and parts of speech of the input question sequence to extract user intentions. Figure 3As shown, the input sequence obtains the words and parts of speech of the sequence through word segmentation technology, and builds a grammar tree according to the parts of speech. Among them, "is" is the predicate and the root node, "rated speed" is the subject, and "what" is the pronoun. There is no object, so the intention is "rated speed". The user's intention can be obtained through the syntax tree. The intention level based on the structured question and answer knowledge base is greater than the intention level obtained based on the syntax analysis tree, and a new intention is obtained. The new intention and the original intention are proofread with the Faiss search engine. When the proofread intent is not the same intention, the proofread new intention replaces the original intention, and the entity information is proofread again. When the proofread intent is the same intention, the proofread new intention replaces the original intention, and the question and answer type is determined and inferred.

[0041] The question-answering type is divided into general question sentences and judgment sentences. Among them, general question sentences are missing components in the question, and the question-answering system will reply. For example, the question "Who proposed the revision of the specification of high-extreme-pressure lubricant oil for helicopters?" lacks an object, so the question-answering system will reply. Judgment sentences are complete components in the question, and the question-answering system will proofread the answer. For example, "Is the person who proposed the revision of the specification of high-extreme-pressure lubricant oil for helicopters Zhang Wei?" The question-answering system will compare and reply based on the answers obtained.

[0042] Step 40: Fill the slots with entities, intents, and question and answer types;

[0043] Step 50 determines the entity type and performs multi-step query and reasoning based on the central entity information;

[0044] Specifically, the entity index ID and entity center node information are obtained through the ES search engine, and all the relationships corresponding to the entity node in the graph database are obtained according to the index ID. When the entity information is a non-center node entity and the relationship corresponding to the entity does not contain relevant intent, the entity is transferred to the center entity and queried again, such as Figure 4 When an entity node corresponds to multiple central nodes, a multi-hop query is performed based on the central node correlation description. At the same time, when the question-answer type is a judgment sentence, entity information determination is performed.

[0045] Step 60 extracts answers online from the entire text based on user intent and entity information;

[0046] Specifically, when there is no relevant answer in the database, the question-answering system performs online reasoning and extraction of all aviation-related standards based on the ES search engine. For example, the question "What are the life-end conditions of explosives?" corresponds to the entity "explosives" and the user intent is "life-end conditions". The question-answering system performs an ES similarity search based on the entity "explosives" and the user intent "life-end conditions", obtains the entity information set: "Safety Regulations for Storage of Explosives", "General Specifications for Standard Materials of Explosives", "Regulations for Finalization of Explosives", and locates the intent and extracts the answer through a string matching algorithm.

[0047] Step 70 records the question, answer, entity, and intent related information;

[0048] Specifically, the question-answering system records questions, answers, entities, and intent information in a database, and processes the data in the database asynchronously, such as Figure 2 Structured knowledge base module. At the same time, the entity information in the question is stored in a mask and duplicates are removed to provide data support for intent extraction.

[0049] Step 80 the answer is provided to the user.

[0050] This embodiment also provides an intelligent question-answering system based on aviation equipment, such as Figure 5 As shown, the system includes: entity and intention acquisition module 101; entity error correction and proofreading module 102; intention extraction, proofreading and reasoning module 103; slot filling module 104; matching and reasoning module 105; retrieval and information extraction module 106; information recording module 107; providing module 108;

[0051] The entity and intent acquisition module 101 is used to acquire the user input sequence entity and intent candidate set. Specifically, the user inputs a text sequence, and each entity acquisition module and intent acquisition module respectively performs entity recognition and intent extraction, and sorts the entity candidate set according to a predefined priority level.

[0052] The entity error correction and proofreading module 102 is used for error correction of input sequence entities and proofreading of entities. Specifically, the module performs entity proofreading based on user intent, ES search engine, and Faiss search engine. According to user intent, the first candidate entity with the highest probability in each candidate set is determined. Entity proofreading is performed through ES search engine and Faiss search engine to ensure that the corresponding entity information exists in the graph database. Among them, ES search engine and Faiss search engine are character and semantic-based search engines, and this module is not limited to ES search engine and Faiss search engine.

[0053] The intention extraction, proofreading and reasoning module 103 is used to obtain the true intention of the input sequence. Specifically, this module uses the historical question and answer data set to extract user intention by entity masking; at the same time, it uses syntax tree analysis to obtain user intention. Finally, it is proofread through the Faiss semantic search engine. Furthermore, when the original intention is different from the current intention, the entity information enters the entity error correction and proofreading module for re-verification and proofreading. At the same time, the question and answer type is determined based on the entity information and user intention.

[0054] The slot filling module 104 is used to normalize the input entity information, intent information, and question and answer type. Specifically, the module determines the final entity information sequence, intent information, and whether the question and answer need to be reasoned through slot filling.

[0055] The matching and reasoning module 105 is used to obtain the answer information of the question. Specifically, the slot information is converted into a graph database query statement, and the relationship corresponding to the main entity is obtained. The main entity relationship is matched with the user's intention. When the target node cannot be obtained, the main entity will jump to the central node for further query. At the same time, when the question and answer type is a judgment sentence, the target node information is compared with the entity information to be compared for similarity. If the threshold is met, a positive reply is given, otherwise a negative reply is given.

[0056] The retrieval and information extraction module 106 is used to extract answers related to user questions online. Specifically, it searches for relevant answer information based on relevant entity information and intent information and ES search engine, and extracts the corresponding answers using information extraction algorithms.

[0057] The information recording module 107 is used to record user question and answer information, collect training corpus and provide comparative data for the intention extraction, proofreading and reasoning modules. Specifically, user questions, answers, related entities and intentions are stored, and questions that cover up entities are clustered and stored.

[0058] The providing module 108 is used to provide answer information to the user.

[0059] Although the embodiments disclosed in the present invention are as above, the above contents are only embodiments adopted for facilitating the understanding of the present invention and are not intended to limit the present invention. Any technician in the technical field to which the present invention belongs can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed in the present invention, but the patent protection scope of the present invention shall still be subject to the scope defined in the attached claims.

Claims

1. An intelligent question-answering method based on aviation equipment, characterized in that: The method comprises the following steps: Step 10 extracts entity and intent candidate sets, and sorts the entity and intent candidate sets according to priority levels; Step 20 performs first entity error correction and proofreading based on user intent; Step 30 performs intention extraction, proofreading and reasoning based on the structured question-answering knowledge base and the syntax analysis tree; Step 40: Fill the slots with entities, intents, and question and answer types; Step 50 determines the entity type and performs multi-step query and reasoning based on the central entity information; Step 60 extracts answers online from the entire text based on user intent and entity information; Step 70 records questions, answers, entities, and intent related information; Step 80 provides the answer to the user; The step 20 determines the first candidate entity with the highest probability based on the user's intention, and the first entity error correction and proofreading are implemented by using ES search technology and Faiss semantic search technology respectively; ES search technology and Faiss obtain entity candidate sets from the string level and semantic level respectively; if the similarity between the ES candidate set and the first entity is greater than 0.95, the first candidate entity is replaced with the first entity information; if there is no similar entity in ES, the Faiss search engine is used to calculate the entity semantic similarity based on word embedding, wherein the entity text training in the Faiss search engine adopts the IndexIVFFlat method, and the word vector calculation adopts the sentencetransformer neural network; when the similarity value is greater than 0.85, the entity error correction and replacement are performed; the first entity error correction and proofreading adopts ES search engine and Faiss search engine to perform entity proofreading from the character and semantic levels respectively; In step 30, the structured question-answering knowledge base is information related to previous user questions, answers, entities, and intentions; the similarity between the questions that mask the entity information and the questions that mask the entity information in the structured question-answering knowledge base is calculated to extract the user intention; at the same time, the syntactic analysis is performed on the words and parts of speech of the input question sequence to extract the user intention; the input sequence obtains the words and parts of speech of the sequence through word segmentation technology, and a grammar tree is built according to the parts of speech; the user's intention is obtained through the syntactic tree; the intention level based on the structured question-answering knowledge base is greater than the intention level obtained based on the syntactic analysis tree, and a new intention is obtained; the new intention and the original intention are proofread with the Faiss search engine; when the proofread intention is not the same, the proofread new intention replaces the original intention, and the entity information is proofread again; when the proofread intention is the same, the proofread new intention replaces the original intention, and the question-answering type is determined and inferred; The step 50 obtains the entity index ID and entity central node information through the ES search engine, and obtains all relationships corresponding to the entity node in the graph database according to the index ID; when the entity information is a non-central node entity and the relationship corresponding to the entity does not contain relevant intent, the entity is transferred to the central entity and queried again. When the entity node corresponds to multiple central nodes, a multi-hop query is performed based on the central node correlation description; at the same time, when the question and answer type is a judgment sentence, the entity information is determined.

2. The intelligent question-answering method based on aviation equipment according to claim 1, characterized in that: In the step 10, entity and intent candidate sets are extracted based on fixed sentence patterns, professional dictionary segmentation technology, AC automaton technology, neural network entity recognition technology, neural network intent extraction technology, and regular methods.

3. The intelligent question-answering method based on aviation equipment according to claim 1 or 2, characterized in that: The levels of entity extraction are divided into: AC automaton technology >= regular expression >= fixed sentence pattern >= professional dictionary word segmentation technology >= neural network entity recognition technology; The levels of intent extraction are: fixed sentence pattern >= professional dictionary word segmentation technology >= neural network intent extraction technology.

4. The intelligent question-answering method based on aviation equipment according to claim 1, characterized in that: In step 60, when there is no relevant answer in the database, the question-answering system performs online reasoning and extraction of all aviation-related standards based on the ES search engine.

5. An intelligent question-answering system based on aviation equipment, characterized in that: The system uses the method of claim 1, including an entity and intent acquisition module, an entity error correction and proofreading module, an intent extraction, proofreading and reasoning module, a slot filling module, a matching and reasoning module, a retrieval and information extraction module, an information recording module and a providing module; The entity and intent acquisition module is used to obtain the user input sequence entity and intent candidate set; Entity error correction and proofreading module, used for error correction of input sequence entities and proofreading of entities; Intent extraction, proofreading, and reasoning module, used to obtain the true intent of the input sequence; Slot filling module, used to normalize input entity information, intent information, and question and answer types; Matching and reasoning module, used to obtain answer information of questions; Retrieval and information extraction module, used to extract relevant answers to user questions online; Information recording module, used to record user question and answer information, collect training corpus and provide comparative data for intent extraction, proofreading and reasoning modules; Providing a module for providing answer information to a user; The entity error correction and proofreading module performs entity proofreading based on user intent, ES search engine, and Faiss search engine to ensure that the corresponding entity information exists in the graph database; The intention extraction, proofreading and reasoning module uses the historical question-answering dataset to extract user intentions through entity masking, and uses syntax tree analysis to obtain user intentions, and proofreads through the Faiss semantic search engine; The matching and reasoning module specifically includes converting slot information into a graph database query statement, and when obtaining entity correspondences, matching entity relationships with user intentions.

6. The intelligent question-answering system based on aviation equipment according to claim 5, characterized in that: Acquiring user input sequence entities and intent candidate sets specifically includes: the user inputs a text sequence, each entity acquisition module and intent acquisition module performs entity recognition and intent extraction respectively, and sorts the entity candidate sets according to a predefined priority level.

7. The intelligent question-answering system based on aviation equipment according to claim 5, characterized in that: The retrieval and information extraction module specifically includes searching for relevant answer information through relevant entity information, intent information and based on the ES search engine, and extracting the corresponding answer using an information extraction algorithm; The information recording module specifically stores user questions, answers, related entities, and intentions, and clusters and stores questions that mask entities.

Images