WO2021109690A1 - 多类型问题智能问答方法、系统、设备及可读存储介质 - Google Patents

多类型问题智能问答方法、系统、设备及可读存储介质 Download PDF

Info

Publication number
WO2021109690A1
WO2021109690A1 PCT/CN2020/117816 CN2020117816W WO2021109690A1 WO 2021109690 A1 WO2021109690 A1 WO 2021109690A1 CN 2020117816 W CN2020117816 W CN 2020117816W WO 2021109690 A1 WO2021109690 A1 WO 2021109690A1
Authority
WO
WIPO (PCT)
Prior art keywords
question
target
target question
type
information
Prior art date
Application number
PCT/CN2020/117816
Other languages
English (en)
French (fr)
Inventor
张圣
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2021109690A1 publication Critical patent/WO2021109690A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3329—Natural language query formulation or dialogue systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G06F16/355—Class or cluster creation or modification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • This application relates to the field of artificial intelligence technology, and in particular to an intelligent question answering method, system, equipment and computer-readable storage medium for multiple types of questions.
  • the core of the existing intelligent question answering system is that it can accurately understand the semantics of the question and return high-quality answers.
  • the inventor realizes that most of the question-and-answer systems in the medical field such as vaccines currently fall into two categories. One is based on frequently asked questions (FQA, Frequently Asked Questions). Asked Questions) technical question answering system can only cover questions with deterministic answers; the other type is based on knowledge base question and answer (KB-QA, knowledge The base question answering) technical question answering system can only cover some simple questions that do not have a unique answer, while the existing intelligent question answering system is more complicated in other semantics, with more diverse answers, or needs to be obtained after obtaining multiple relevant information. When responding to a question that can be answered, it is often difficult to provide an accurate answer to the question proposer, which leads to technical problems with low accuracy of question response in the existing intelligent question answering system.
  • the main purpose of this application is to provide an intelligent question answering method, system, equipment and computer readable storage medium for multiple types of questions, aiming to solve the technical problem of the existing intelligent question answering system that the accuracy of question response is not high.
  • this application provides a method for intelligent question answering with multiple types of questions.
  • the method for intelligent question answering with multiple types of questions includes the following steps:
  • the target question is the first question type
  • entity recognition and multi-level semantic analysis are performed on the first target question to obtain the question template and multi-level semantics of the first target question, and according to the question template And multi-level semantics, generating response information to the first target question, the first target question being a target question of the first question type;
  • the related question set of the second target question is determined, and based on the multiple rounds of question and answer technology, the client's multiple related answers to the related question set are received, so as to be based on the multiple related questions.
  • this application also provides a multi-type question intelligent question answering system
  • the multi-type question intelligent question answering system includes:
  • the intention classification module is configured to, when a target question is received, perform intent classification on the target question based on a preset deep network model, so as to determine the question type of the target question;
  • the first question answering module is used to perform entity recognition and multi-level semantic analysis on the first target question when it is detected that the target question is the first question type, so as to obtain the question template and multi-level of the first target question Semantics, and based on the question template and multi-level semantics, generate response information to the first target question, where the first target question is a target question of the first question type
  • the second question answering module is used to determine the associated question set of the second target question when it is detected that the target question is of the second question type, and receive multiple associations of the client to the associated question set based on multiple rounds of question and answer technology Reply, to generate reply information for the second target question based on the multiple associated replies, and the second target question is a target question of a second question type.
  • the present application also provides a multi-type question intelligent question answering device
  • the multi-type question intelligent question answering device includes a processor, a memory, and a memory that is stored on the memory and can be executed by the processor.
  • a multi-type question intelligent question answering program wherein when the multi-type question intelligent question answering program is executed by the processor, the following steps are implemented:
  • the target question is the first question type
  • entity recognition and multi-level semantic analysis are performed on the first target question to obtain the question template and multi-level semantics of the first target question, and according to the question template And multi-level semantics, generating response information to the first target question, the first target question being a target question of the first question type;
  • the related question set of the second target question is determined, and based on the multiple rounds of question and answer technology, the client's multiple related answers to the related question set are received, so as to be based on the multiple related questions.
  • the present application also provides a computer-readable storage medium having a multi-type question intelligent question answering program stored on the computer-readable storage medium, wherein the multi-type question intelligent question answering program is executed by the processor.
  • the target question is the first question type
  • entity recognition and multi-level semantic analysis are performed on the first target question to obtain the question template and multi-level semantics of the first target question, and according to the question template And multi-level semantics, generating response information to the first target question, the first target question being a target question of the first question type;
  • the related question set of the second target question is determined, and based on the multiple rounds of question and answer technology, the client's multiple related answers to the related question set are received, so as to be based on the multiple related questions.
  • This application makes it possible to accurately reply to the second target question that requires multiple related information, so that the system can accurately answer questions that require more complex semantics and more diverse answers, or that require multiple related information to be answered. Reply, which solves the technical problem of low accuracy of the existing intelligent question answering system.
  • FIG. 1 is a schematic diagram of the hardware structure of a multi-type question intelligent question answering device involved in the solution of the embodiment of the application;
  • FIG. 2 is a schematic flowchart of a first embodiment of an intelligent question answering method for multiple types of questions in this application;
  • FIG. 3 is a schematic diagram of the KBQA technical framework integrating question templates and multi-level semantic analysis in a specific embodiment of the multi-type question intelligent question answering method of this application.
  • the multi-type question intelligent question answering method involved in the embodiments of the present application is mainly applied to a multi-type question intelligent question answering device.
  • the multi-type question intelligent question answering device may be a device with display and processing functions such as a PC, a portable computer, and a mobile terminal.
  • FIG. 1 is a schematic diagram of the hardware structure of the intelligent question answering device for multiple types of questions involved in the solution of the embodiment of the application.
  • the intelligent question answering device for multiple types of questions may include a processor 1001 (for example, a CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005.
  • the communication bus 1002 is used to realize the connection and communication between these components;
  • the user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard);
  • the network interface 1004 may optionally include a standard wired interface, a wireless interface (Such as WI-FI interface);
  • the memory 1005 can be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a disk memory.
  • the memory 1005 can optionally also be a storage device independent of the aforementioned processor 1001 .
  • FIG. 1 does not constitute a limitation on the intelligent question answering device for multiple types of questions, and may include more or less components than shown, or a combination of certain components, or different Component arrangement.
  • the memory 1005 in FIG. 1 as a computer-readable storage medium may include an operating system, a network communication module, and a multi-type question intelligent answering program.
  • the network communication module is mainly used to connect to the server and perform data communication with the server; and the processor 1001 can call the intelligent question answering program for multiple types of questions stored in the memory 1005, and perform the following operations:
  • the target question is the first question type
  • entity recognition and multi-level semantic analysis are performed on the first target question to obtain the question template and multi-level semantics of the first target question, and according to the question template And multi-level semantics, generating response information to the first target question, the first target question being a target question of the first question type;
  • the related question set of the second target question is determined, and based on the multiple rounds of question and answer technology, the client's multiple related answers to the related question set are received, so as to be based on the multiple related questions.
  • the step of classifying the target question by intention based on a preset deep network model to determine the question type of the target question includes:
  • the step of generating response information to the first target question includes:
  • a preset entity recognition model is used to identify the entity information of the first target question, and the question template is obtained based on the entity information;
  • the first target question is converted into a structured query of the knowledge graph to obtain the answer information of the first target question, and the answer information of the first target question is stored in In the blockchain.
  • step of performing multi-level semantic analysis on the first target question to obtain the multi-level semantics of the first target question includes:
  • Conceptual mapping is performed on the entity information of the first target question based on the question template, and the question-level semantics of the first target question is obtained, and the entity-level semantics, phrase-level semantics, and question-level semantics are used as the multiple Hierarchical semantics.
  • the related question set of the second target question is determined, and based on multiple rounds of question and answer technology, the client's multiple related answers to the related question set are received, so as to
  • the step of generating response information for the second target question based on the multiple association responses includes:
  • the target question When it is detected that the target question is of the second question type, determine the related question set from a preset related question library, and conduct multiple rounds of question and answer with the client around the related question set;
  • the multiple association responses are converted into usable information, so as to use the usable information to generate answer information for the second target question, and the second target
  • the answer to the question is stored in the blockchain.
  • the step of receiving the multiple association replies sent by the client in multiple rounds of question and answer, and performing slot filling according to the multiple association replies includes:
  • the method further includes:
  • the target question of the third question type is taken as the third target question, and the third target question and the standard question set constructed based on the preset clustering algorithm are obtained. Similarity between problems;
  • the third target question is mapped to similar standard questions in the standard question set, and standard responses to the similar standard questions are obtained, so as to generate response information for the third target question based on the standard responses.
  • the similarity standard problem has the highest degree of similarity between the standard problem set and the target problem.
  • Fact-based questions are mainly knowledge-based questions with deterministic answers, such as "What diseases can be prevented by MMR vaccine?"; opinion-based questions are generally questions for which there is no single standard answer, such as "Is the vaccine free or self-financed?" Is it good?”; Task-type questions generally need to obtain the information required by the task from the continuous dialogue of the user. For example, "Scheduled vaccination” requires information such as the type of vaccination, the scheduled vaccination time, and the location of vaccination to complete the related questions.
  • the core of the existing intelligent question answering system is that it can accurately understand the semantics of the question and return high-quality answers.
  • most of the question and answer systems in the medical field such as vaccines are of two types.
  • One type is based on Frequently Asked Questions (FQA). Questions
  • FQA Frequently Asked Questions
  • KB-QA knowledge base question and answer
  • the base question answering) technical question answering system can only cover some simple questions that do not have a unique answer, while the existing intelligent question answering system is more complicated in other semantics, with more diverse answers, or needs to be obtained after obtaining multiple relevant information.
  • this application provides an intelligent question answering method for multiple types of questions, that is, the question type of the target question is determined through a deep network model, which improves the accuracy of problem positioning; through entity recognition and multi-level semantics for the first target question Analysis effectively reduces the complexity of the model and achieves a deep understanding of the semantics of the first goal question, so that it can respond to the first goal question with complex semantics efficiently and accurately; by launching multiple rounds with the client for the second goal question Questions and answers, until the client's related answers to the related question set are obtained, so that the second target question that needs to know multiple related information can be accurately answered, so that the system can be more complicated to other semantics, more diverse answers, or need to be in The question that can be answered only after obtaining multiple relevant information is answered accurately, thereby solving the technical problem of incomplete question types supported by the existing intelligent question answering system.
  • This application is applied to an intelligent question answering system for multiple types of questions.
  • FIG. 2 is a schematic flowchart of the first embodiment of the intelligent question answering method for multiple types of questions in this application.
  • the first embodiment of the present application provides an intelligent question answering method for multiple types of questions.
  • the intelligent question answering method for multiple types of questions includes the following steps:
  • Step S10 when a target question is received, perform intent classification on the target question based on a preset deep network model to determine the question type of the target question;
  • the target question is usually a question that needs to be answered by the user input to the multi-type question intelligent question answering system through natural language input or voice input through the client.
  • the preset deep network model incorporates the Bidirectional Encoder Representations from Transformer (BERT, Bidirectional Encoder Representations from Transformers) and Bi-directional Long Short-Term Memory Network (Bi-LSTM, Bi-directional Long Short-Term Memory).
  • Intent classification aims to classify the target questions raised by users through a deep network model.
  • the specific types of questions can include common questions, factual questions, task questions, and other irrelevant questions.
  • the multi-type question intelligent question answering system receives the target question proposed by the user that currently needs to be answered, it inputs the target question as a model to the BERT layer, and the BERT layer obtains the sequence representation vector of the word sequence corresponding to the user question. After the system obtains the sequence representation vector of the word sequence, it then inputs the sequence representation vector to the Bi-LSTM layer to obtain the hidden vector of the word sequence of the target problem. The system combines the sequence representation vector and the hidden vector to learn to obtain the target representation vector, and after the fully connected layer and the Softmax (the excitation function of the output layer) operation, the classification of the target problem can be obtained, which is the above classification result.
  • the system receives the target question proposed by the user that currently needs to be answered, it inputs the target question as a model to the BERT layer, and the BERT layer obtains the sequence representation vector of the word sequence corresponding to the user question. After the system obtains the sequence representation vector of the word sequence, it then inputs the sequence representation vector to the Bi-LSTM layer to
  • Step S20 When it is detected that the target question is the first question type, entity recognition and multi-level semantic analysis are performed on the first target question to obtain the question template and multi-level semantics of the first target question, and according to the The question template and multi-level semantics are used to generate response information to the first target question, where the first target question is a target question of the first question type;
  • the first question type may be factual, opinion, task or other types, and is preferably opinion.
  • KBQA technology adopts knowledge base question (KB-QA, knowledge base question answering) technology, through a given natural language question, the semantic understanding and analysis of the question, and then use the knowledge base to query and reason to get the answer.
  • knowledge base question and answer can be divided into open domain knowledge question and answer, such as encyclopedia knowledge question and answer, and specific domain knowledge question and answer, such as financial field, medical field, religious field, etc.
  • the system first recognizes the entity of the first target question, to obtain a question template related to the first target question based on the entity information of the first target question, and performs multi-level semantic recognition of the first target question to understand the first target deeply The semantics of the question. Finally, combining the obtained question template with the parsed multi-level semantics, the structured query of the knowledge graph is performed to generate the answer information for the first target question.
  • Step S30 When it is detected that the target question is of the second question type, determine the related question set of the second target question, and receive the client's multiple related answers to the related question set based on the multiple rounds of question answering technology, so as to be based on all the related questions.
  • the multiple-associated answer generates answer information for the second target question, where the second target question is a target question of a second question type.
  • the second question type can be factual, opinion, task or other types, and is preferably task type.
  • the multi-round question answering technology is to first obtain the initial information input by the user terminal, that is, the second target question, and identify the initial information, and determine the predetermined questions corresponding to different answers under different combinations of conditions, that is, the set of related questions. And began the first round of question and answer to the determined predetermined questions. In each round of question and answer, a question about an unknown condition for the determined predetermined question is output for the user to answer. When the condition that the user answers is obtained, that is, the multiple-associated answer, according to the predetermined condition combination and the corresponding relationship between the answer, it is judged whether there is a corresponding answer in the obtained condition combination.
  • the system can integrate FAQ, KBQA, multiple rounds of Q&A and other technologies. According to the actual classification results corresponding to the current target question, the target technology type that is most suitable for answering the target question is selected.
  • the system can choose to use the common problem answering method based on a preset clustering algorithm for processing; if the system recognizes that the current target problem is a fact-based problem, it can use the combination Question template generation and multi-level semantic analysis KBQA technology for processing; if the system identifies the current target problem as a task line problem, it can use slot filling-based multi-round question and answer technology for processing to obtain the response information for the current target problem And output to the user side.
  • the target question when a target question is received, the target question is classified based on a preset deep network model to determine the question type of the target question; when the target question is detected as the first question Type, perform entity recognition and multi-level semantic analysis on the first target question to obtain the question template and multi-level semantics of the first target question, and generate the first target according to the question template and multi-level semantics
  • the response information of the question wherein the first target question is a target question of the first question type; when it is detected that the target question is a second question type, determine the associated question set of the second target question, and based on multiple rounds
  • the question answering technology receives the multiple-associated replies of the client to the set of associated questions to generate answer information for the second target question based on the multiple-associated replies, where the second target question is a target question of the second question type .
  • this application uses the deep network model to determine the problem type of the target problem, which improves the accuracy of problem positioning; through entity recognition and multi-level semantic analysis for the first target problem, the complexity of the model is effectively reduced, and the In-depth understanding of the semantics of the first target question, so that the first target question with complex semantics can be answered efficiently and accurately; through multiple rounds of question and answer with the client for the second target question, until the client's response to the related question set is obtained
  • Relevant replies enable accurate responses to second target questions that require multiple related information, so that the system can respond to questions that are more complex in other semantics and more diverse in answers, or that require multiple related information to be answered. Respond accurately, which solves the technical problem of low accuracy of the existing intelligent question answering system.
  • step S10 specifically includes:
  • the pre-training language representation model is the above-mentioned BERT.
  • BERT is a natural language pre-training model based on two-way Transformer, combined with natural language processing (NLP, Natural Language Processing) specific downstream tasks combined with other network layers to form suitable for specific The model of the task.
  • NLP Natural Language Processing
  • [CLS] in the BERT model can represent the entire sentence in a vector.
  • the BERT model adds a special mark [CLS] in front of the word sequence of each sentence.
  • the global information of the entire sentence can be encoded to each position.
  • the BERT model can learn the entire sentence by deep characterization of [CLS] The upper-level features.
  • Bi-LSTM can learn the semantic information before and after the corresponding words well, and can integrate deeper semantic information into the model. Then the hidden vectors h 1 , h 2 ..., h n obtained in the previous layer and the representation vector e [CLS ] of the BERT layer [CLS] are used for Attention (attention mechanism) operation, and an operation similar to a weighted sum is performed. Get the target characterization vector Vector. The learned target characterization vector Vector is subjected to a fully connected layer and Softmax operation to obtain the classification of the input sentence (target problem), which is the above-mentioned problem type.
  • step S20 includes:
  • a preset entity recognition model is used to identify the entity information of the first target question, and the question template is obtained based on the entity information;
  • the KBQA module in the system mainly includes two parts: offline processing and online output.
  • the main function of the offline processing part is to use QA corpus, just the information of the map, and extract templates from it to construct a problem template library.
  • the offline processing part of the KBQA module in the system uses long and short-term memory (LSTM, Long Short-Term Memory)-The entity recognition model of conditional random field (CRF) can identify the entity part from the natural sentence.
  • LSTM Long Short-Term Memory
  • CRF conditional random field
  • the entity recognition model in the natural sentence can be Entities are mapped to corresponding concepts, thereby reducing a large number of repetitive problems, and conceptualizing problems can assist the model in learning more precise semantic information.
  • Expand the attributes through the attributes (relation types) in the knowledge graph to get more relationship expressions.
  • the first target question is converted into a structured query of the knowledge graph to obtain the answer information of the first target question, and the answer information of the first target question is stored in In the blockchain.
  • the main function of the online processing part is to parse the first target question to obtain the entity and the corresponding attribute type, convert it into a structured query of the knowledge graph, and finally return the answer to the first target question To the user.
  • the online processing part of the KBQA module in the system first performs multi-level semantic analysis on the first target problem, and then uses the preset probability graph model to comprehensively use the semantic analysis results of the problem, and the problem template obtained through the offline processing part to predict
  • the first target question corresponds to the attribute type in the knowledge graph, that is, to realize the semantic analysis of the question.
  • Model prediction and knowledge graph query function function.
  • the online processing part finally converts the first target question into a structured query of the knowledge graph according to the obtained attribute type and the entity in the first target question identified by the model, and returns the answer to the first target question to the user end.
  • the reply information to the first goal question can also be stored in a node of a blockchain.
  • the step of performing multi-level semantic analysis on the first target question to obtain the multi-level semantics of the first target question includes:
  • Conceptual mapping is performed on the entity information of the first target question based on the question template, and the question-level semantics of the first target question is obtained, and the entity-level semantics, phrase-level semantics, and question-level semantics are used as the multiple Hierarchical semantics.
  • the multi-level semantic analysis includes entity-level semantic analysis, phrase-level semantic analysis, and question-level semantic analysis.
  • the online processing part of the KBQA module uses a preset semantic search model, based on the entity semantic understanding of the knowledge graph, to obtain entity-level semantic information for upper-level semantic computing; mainly uses verb templates for fine-grained semantic representation and comprehensive use of context
  • the information is conceptualized as an entity to obtain semantic information at the phrase level; and based on the expression method of the question template, the entity in the first target question is mapped to the concept to further express the semantics of the first target question to obtain the semantic information at the question level.
  • the first objective question is an opinion question.
  • Figure 3 is a schematic diagram of the KBQA technical framework integrating question templates and multi-level semantic analysis.
  • the KBQA module in the system can be divided into offline processing and online processing.
  • the offline processing module mainly uses QA corpus and knowledge map information to extract templates to construct a problem template library; while the online processing part includes a problem analysis sub-module, a model prediction sub-module and a knowledge map query sub-module, based on these
  • the sub-module converts the opinion question into a structured query of the knowledge graph to obtain the answer to the target question.
  • the complexity of the model can be effectively reduced, and the problem can be understood at a deeper level, and the accuracy of the model can be improved; by analyzing the natural sentences in the QA corpus After the entity recognition operation, the entities in the natural sentence can be mapped to the corresponding concept, thereby reducing a large number of repetitive problems, and the conceptualization of the problem can assist the model to learn more accurate semantic information; attributes are performed through the attribute types in the knowledge graph Extension can get more attribute expressions; through question and answer (QA, Question Answering) corpus of entity concept mapping and attribute expansion of the knowledge graph to obtain high-quality question templates.
  • QA Question Answering
  • the step S30 includes:
  • the target question When it is detected that the target question is of the second question type, determine the related question set from a preset related question library, and conduct multiple rounds of question and answer with the client around the related question set;
  • the multiple association responses are converted into usable information, so as to use the usable information to generate answer information for the second target question, and the second target
  • the answer to the question is stored in the blockchain.
  • the preset associated question library is a preset question library for splitting the second target question into multiple sub-questions.
  • the system When the system detects that the target question belongs to the second question type, it can perform semantic recognition on the second target question, and according to the semantic recognition result, determine from the preset associated question library the multiple sub-subjects used to answer the second target question.
  • the problem is the set of related problems mentioned above. Multi-round Q&A technology can be used.
  • the multi-round Q&A module in the system for answering the second target question conducts multiple rounds of Q&A with the client around the related question set, and receives multiple related replies for the related question set sent by the client At the time, the slot is filled according to the client's multiple association replies; the multi-round question and answer module uses a preset depth network sequence labeling model to convert the client's multiple association replies into usable information, so as to generate reply information for the second target question based on the available information.
  • the preset deep network sequence annotation model is the LSTM-CRF model. It should be emphasized that, in order to further ensure the privacy and security of the reply information to the second target question, the reply information to the second target question can also be stored in a node of a blockchain.
  • the step of receiving the multiple-association reply sent by the client in multiple rounds of question and answering, and performing slot filling according to the multiple-association reply includes:
  • the initial slot information is based on the key information in the user's response to the associated question set.
  • “scheduled vaccination” needs to know the name of the vaccine, the type of vaccine (category 1 and category 2), the vaccination location, and the appointment time for vaccination.
  • the preset validity standard is to determine whether the standard conditions for filling the slot can be completed according to the slot information extracted from the question or reply message sent by the client. It can be flexibly set according to actual needs, for example, for scheduled vaccination.
  • a task-based question is to determine whether the standard conditions for filling the slot can be completed according to the slot information extracted from the question or reply message sent by the client. It can be flexibly set according to actual needs, for example, for scheduled vaccination.
  • the system can determine that the validity standard is still not met, and it is necessary to continue the question and answer to obtain necessary information.
  • the revised response is the user's response to the clarification sent by the system. For example, when the user conducts multiple rounds of question and answer with the system about the task-type question of scheduled vaccination, there are many system terminals in order to explain the information about the slot of vaccine type.
  • the question-and-answer module will send a clarification expression similar to "what kind of vaccines need to be scheduled for vaccination" to the client, so that the client user can give a clear answer.
  • the multi-round question answering module in the system extracts the initial slot information from the multiple associated replies, and judges whether the initial slot information meets the preset validity standard.
  • the initial slot information meets the preset validity standard, it can be determined to complete the question and answer about the second target question; if one or more of the initial slot information does not meet the preset validity standard, use clarification Mechanical energy clarifies the expression.
  • the multi-round Q&A module receives the correction response to the clarification speech sent by the client, it corrects the initial slot information according to the correction response, and standardizes the corrected initial slot information, so that the standardized initial slot information is processed.
  • the slot information is used as the target slot information. At this time, it can be determined that the slot filling is completed, and there is no need to ask questions.
  • the second target problem is set as a task-type problem.
  • “Scheduled vaccination” needs to know the name of the vaccine, the type of vaccine (category 1 and category 2), the vaccination location, and the appointment time for vaccination, etc.
  • To make an appointment for vaccination you need to know four pieces of information. These four pieces of information are the four slots.
  • the information obtained from the dialogue with the user is the filling of the slots (slot filling).
  • slot filling the filling of the slots
  • the clarification words for the vaccination point are "You Which vaccination point do you want to make an appointment for vaccination?", and provide candidates for users to choose, users can also directly use natural language to reply. Whether there is a dependency between slots and slots can be divided into dependent slots and level slots.
  • the type of vaccine depends on the name of the vaccine, and the scheduled vaccination time depends on the vaccination site.
  • the reliance slot needs to have a sequence for clarification, for example, it is necessary to obtain the information of the vaccination point in order to clarify the time information that can be reserved.
  • the system uses the LSTM-CRF deep network sequence annotation model to obtain the available information.
  • the machine asks "Which vaccination spot do you want to make an appointment for vaccination?" and provides candidates to the user. If the user chooses from the candidates, fill in the slot directly; if the user responds in natural language, such as "book a vaccination site”, you need to use the sequence annotation model to get “book a vaccination site” (vaccine Vaccination point)", and then need to verify the legality of the slot information obtained from the dialogue, here it is necessary to verify whether the vaccination point is within the range of options. Then continue to clarify other slot information that needs to be filled in.
  • the user "make an appointment to go to XXX vaccination site tomorrow morning at 10 o'clock to receive hepatitis B vaccine.”
  • using the sequence annotation model can get "book an appointment [tomorrow 10 o'clock in the morning] (appointment vaccination time) to [XXX vaccination site] (vaccine Vaccination point) Vaccination [hepatitis B vaccine] (vaccine name)", and then standardize the information obtained from different slots in the natural sentence and verify whether the information is legal. For example, it is necessary to map 10 o'clock tomorrow morning to standard time information, and it is necessary to verify whether the vaccination point and scheduled vaccination time mentioned by the user are within the range of options. After obtaining all the slot information, you need to reconfirm all the slot information with the user. After the confirmation is correct, the information that needs to be used can be completed.
  • step S10 the method further includes:
  • the target question of the third question type is taken as the third target question, and the third target question and the standard question set constructed based on the preset clustering algorithm are obtained. Similarity between problems;
  • the third target question is mapped to similar standard questions in the standard question set, and standard responses to the similar standard questions are obtained, so as to generate response information for the third target question based on the standard responses.
  • the similarity standard problem has the highest degree of similarity between the standard problem set and the target problem.
  • the third question type may specifically be fact type, opinion type, task type or other types, preferably fact type.
  • the preset clustering algorithm can use the noisy density-based clustering method (DBSCAN, Density-Based Spatial Clustering of Applications with Noise), expectation maximization (EM, expectation maxmization) clustering algorithm, K-means (K-means) clustering algorithm, etc., preferably K-means clustering algorithm.
  • DBSCAN Density-Based Spatial Clustering of Applications with Noise
  • EM expectation maximization
  • K-means K-means clustering algorithm
  • K-means K-means clustering algorithm
  • FAQ technology uses a model to calculate the similarity between user questions and existing standard questions, and maps user questions to existing standard questions to achieve question and answer.
  • the model used by the system in the FAQ module is a deep learning model that reuses the fusion of BERT and BiLSTM in the intention classification module.
  • the input of the model is the user's problem, and the output of the model corresponds to the standard problem with the highest similarity to the third target problem. answer.
  • the number of types of questions that FAQ technology can cover depends on the number of standard questions designed. At present, most standard questions are manually sorted out, and the number of types of questions that can be covered is relatively small.
  • the system in this application is based on the machine learning clustering algorithm K-means algorithm to automatically construct a large number of high-quality standard questions from a large amount of domain question and answer corpus.
  • Automatically constructing a high-quality standard problem mainly includes three steps: the first step is to obtain the initial problem from the data source of the high-quality vaccine vertical website; the second step is to cluster similar problems into a cluster through the clustering algorithm; the third step, Filter standard questions and get answers. In practical applications, a collection of about 5,000 high-quality standard questions in the vaccine field can be constructed through this step.
  • the multi-round question-and-answer module in the system is based on slot filling and multi-round dialogue technology to efficiently solve task-based problems, especially for vaccine question-and-answer-related task-based problems; through the FAQ module based on The K-means clustering algorithm automatically constructs a large-scale and high-quality standard problem set, thereby reducing the cost of manually sorting and labeling the standard problem set.
  • the blockchain referred to in this application is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • Blockchain essentially a decentralized database, is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information for verification. The validity of the information (anti-counterfeiting) and the generation of the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.
  • the embodiment of the present application also provides an intelligent question answering system for multiple types of questions.
  • the intelligent question answering system for multiple types of questions includes:
  • the intention classification module is configured to, when a target question is received, perform intent classification on the target question based on a preset deep network model, so as to determine the question type of the target question;
  • the first question answering module is used to perform entity recognition and multi-level semantic analysis on the first target question when it is detected that the target question is the first question type, so as to obtain the question template and multi-level of the first target question Semantics, and based on the question template and multi-level semantics, generate response information to the first target question, where the first target question is a target question of the first question type
  • the second question answering module is used to determine the associated question set of the second target question when it is detected that the target question is of the second question type, and receive multiple associations of the client to the associated question set based on multiple rounds of question and answer technology Reply, to generate reply information for the second target question based on the multiple associated replies, and the second target question is a target question of a second question type.
  • Each module in the above-mentioned multi-type question intelligent question answering system corresponds to each step in the embodiment of the above-mentioned multi-type question intelligent question answering method, and its functions and implementation processes will not be repeated here.
  • This application also provides an intelligent question answering device for multiple types of questions.
  • the multi-type question intelligent question answering device includes a processor, a memory, and a multi-type question intelligent question answering program stored on the memory and running on the processor, wherein the multi-type question intelligent question answering program is processed by the processor.
  • the steps of the intelligent question answering method for multiple types of questions as described above are implemented.
  • the method implemented when the multi-type question intelligent question answering program is executed can refer to the various embodiments of the multi-type question intelligent question answering method of this application, which will not be repeated here.
  • the embodiments of the present application also provide a computer-readable storage medium.
  • the computer-readable storage medium may be non-volatile or volatile.
  • the computer-readable storage medium of the present application stores a multi-type question intelligent question answering program, wherein when the multi-type question intelligent question answering program is executed by the processor, the steps of the above-mentioned multi-type question intelligent question answering method are realized.
  • the method implemented when the multi-type question intelligent question answering program is executed can refer to the various embodiments of the multi-type question intelligent question answering method of this application, which will not be repeated here.

Abstract

一种人工智能技术领域的多类型问题智能问答方法、系统、设备及计算机可读存储介质。该方法包括：在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息；在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题。

Description

多类型问题智能问答方法、系统、设备及可读存储介质
本申请要求于2020年6月17日提交中国专利局、申请号为202010558326.1，发明名称为“多类型问题智能问答方法、系统、设备及可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，尤其涉及一种多类型问题智能问答方法、系统、设备及计算机可读存储介质。
 背景技术
现有智能问答系统的核心是可以准确理解问题语义，返回高质量的答案。发明人意识到，目前疫苗等医学领域的问答系统大多是两类，一类是基于常见问题解答（FQA，Frequently Asked Questions）技术的问答系统，只能覆盖有确定性答案的问题；另一类是基于知识库问答（KB-QA，knowledge base question answering）技术的问答系统，只能覆盖部分简单的不存在唯一答案的问题，而现有智能问答系统在对其他语义更为复杂，答案更为多样，或是需要在获取多重相关信息后才可进行回答的问题进行回复时，往往难以为问题提出方提供准确的答复，因而导致了现有智能问答系统的问题回复准确性不高的技术问题。
 技术解决方案
本申请的主要目的在于提供一种多类型问题智能问答方法、系统、设备及计算机可读存储介质，旨在解决现有智能问答系统的问题回复准确性不高的技术问题。
为实现上述目的，本申请提供一种多类型问题智能问答方法，所述多类型问题智能问答方法包括以下步骤：
在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
此外，为实现上述目的，本申请还提供一种多类型问题智能问答系统，所述多类型问题智能问答系统包括：
意图分类模块，用于在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
第一问题答复模块，用于在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题
第二问题答复模块，用于在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
此外，为实现上述目的，本申请还提供一种多类型问题智能问答设备，所述多类型问题智能问答设备包括处理器、存储器、以及存储在所述存储器上并可被所述处理器执行的多类型问题智能问答程序，其中所述多类型问题智能问答程序被所述处理器执行时，实现如下步骤：
在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
此外，为实现上述目的，本申请还提供一种计算机可读存储介质，所述计算机可读存储介质上存储有多类型问题智能问答程序，其中所述多类型问题智能问答程序被处理器执行时，实现如下步骤：
在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
本申请使得能够准确回复需要获知多重相关信息的第二目标问题，从而使得系统能够对其他语义更为复杂，答案更为多样，或是需要在获取多重相关信息后才可进行回答的问题进行准确回复，解决了现有智能问答系统的问题回复准确性不高的技术问题。
 附图说明
图1为本申请实施例方案中涉及的多类型问题智能问答设备的硬件结构示意图；
图2为本申请多类型问题智能问答方法第一实施例的流程示意图；
图3为本申请多类型问题智能问答方法一具体实施例中融合问题模板与多层次语义解析的KBQA技术框架示意图。
 本发明的实施方式
应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
本申请实施例涉及的多类型问题智能问答方法主要应用于多类型问题智能问答设备，该多类型问题智能问答设备可以是PC、便携计算机、移动终端等具有显示和处理功能的设备。
参照图1，图1为本申请实施例方案中涉及的多类型问题智能问答设备的硬件结构示意图。本申请实施例中，多类型问题智能问答设备可以包括处理器1001（例如CPU），通信总线1002，用户接口1003，网络接口1004，存储器1005。其中，通信总线1002用于实现这些组件之间的连接通信；用户接口1003可以包括显示屏（Display）、输入单元比如键盘（Keyboard）；网络接口1004可选的可以包括标准的有线接口、无线接口（如WI-FI接口）；存储器1005可以是高速RAM存储器，也可以是稳定的存储器（non-volatile memory），例如磁盘存储器，存储器1005可选的还可以是独立于前述处理器1001的存储装置。
本领域技术人员可以理解，图1中示出的硬件结构并不构成对多类型问题智能问答设备的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件布置。
继续参照图1，图1中作为一种计算机可读存储介质的存储器1005可以包括操作系统、网络通信模块以及多类型问题智能问答程序。
在图1中，网络通信模块主要用于连接服务器，与服务器进行数据通信；而处理器1001可以调用存储器1005中存储的多类型问题智能问答程序，并执行以下操作：
在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
进一步地，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤包括：
在接收到目标问题时，将所述目标问题输入预训练语言表征模型，得到所述目标问题对应词序列的序列表征向量；
将所述序列表征向量输入预设双向长短时记忆网络中，得到所述词序列的隐藏向量；
结合所述序列表征向量与所述隐藏向量，确定所述目标问题的问题类型。
进一步地，所述在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息的步骤包括：
在检测到所述目标问题为第一问题类型时，基于离线处理方式，使用预设实体识别模型识别出所述第一目标问题的实体信息，并基于所述实体信息得到所述问题模板；
基于线上处理方式，对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义；
使用预设概率图模型，并结合所述问题模板与所述多层次语义，预测所述第一目标问题对应到知识图谱中的属性类别；
根据所述属性类别与所述实体信息，将所述第一目标问题转换为知识图谱的结构化查询，以得到所述第一目标问题的答复信息，所述第一目标问题的答复信息存储于区块链中。
进一步地，所述对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义的步骤包括：
使用预设语义搜索模型对所述第一目标问题进行实体层语义解析，获取实体层语义；
使用预设动词模板对所述第一目标问题进行细粒度语义表示，获取所述第一目标问题的短语层语义；
基于所述问题模板将所述第一目标问题的实体信息进行概念映射，获取所述第一目标问题的问题层语义，以将所述实体层语义、短语层语义与问题层语义作为所述多层次语义。
进一步地，所述在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息的步骤包括：
在检测到所述目标问题为第二问题类型时，从预设关联问题库中确定所述关联问题集，并围绕所述关联问题集与客户端进行多轮问答；
接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充；
在检测到槽填充完成时，基于预设深度网络序列标注模型，将所述多重关联答复转化为可用信息，以使用所述可用信息生成所述第二目标问题的答复信息，所述第二目标问题的答复信息存储于区块链中。
进一步地，所述接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充的步骤包括：
从所述多重关联答复中提取出初始槽位信息；
判断所述初始槽位信息是否符合预设有效性标准，并在所述初始槽位信息不符合预设有效性标准时，使用澄清话术进行澄清表达；
在接收到客户端发送的对于所述澄清话术的修正回复时，根据所述修正回复修正所述初始槽位信息，直至修正后的初始槽位信息符合所述有效性标准；
对修正后的初始槽位信息进行标准化处理，并将经标准化处理后的初始槽位信息作为目标槽位信息，以基于所述目标槽位信息进行槽填充。
进一步地，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤之后，还包括：
在检测到所述目标问题为第三问题类型时，将第三问题类型的目标问题作为第三目标问题，并获取所述第三目标问题与基于预设聚类算法所构建的标准问题集之间的问题相似度；
将所述第三目标问题映射在所述标准问题集中的相似标准问题上，并获取所述相似标准问题的标准回复，以基于所述标准回复生成所述第三目标问题的答复信息，所述相似标准问题在所述标准问题集中与所述目标问题之间的问题相似度最高。
基于上述硬件结构，提出本申请多类型问题智能问答方法的各个实施例。
对于传统的问答系统，主要是以机器自动回答用户所提出来的自然语言问题。目前问答系统主要覆盖的问题主要可以分为三类：事实型问题、观点型问题与任务型问题。事实型问题主要是有确定性的答案的知识型的问题，比如“麻腮风疫苗可以预防哪些疾病？”；观点型问题一般是不存在唯一标准答案的问题，比如“疫苗免费的好还是自费的好？”；而任务型问题一般需要从用户连续的对话中获取任务所需的信息，比如“预约接种疫苗”需要知道接种疫苗种类、预约接种时间、接种地点等信息才可以完成相关问题。
现有智能问答系统的核心是可以准确理解问题语义，返回高质量的答案。目前疫苗等医学领域的问答系统大多是两类，一类是基于常见问题解答（FQA，Frequently Asked Questions）技术的问答系统，只能覆盖有确定性答案的问题；另一类是基于知识库问答（KB-QA，knowledge base question answering）技术的问答系统，只能覆盖部分简单的不存在唯一答案的问题，而现有智能问答系统在对其他语义更为复杂，答案更为多样，或是需要在获取多重相关信息后才可进行回答的问题进行回复时，往往难以为问题提出方提供准确的答复，因而导致了现有智能问答系统的问题回复准确性不高的技术问题。
为解决上述问题，本申请提供一种多类型问题智能问答方法，即通过深度网络模型确定目标问题的问题类型，提升了问题定位的准确性；通过针对第一目标问题进行实体识别与多层次语义解析，有效降低了模型复杂度，实现了对于第一目标问题的语义的深层次理解，从而能够高效精准地回复具有复杂语义的第一目标问题；通过针对第二目标问题与客户端展开多轮问答，直至获取到客户端对于关联问题集的关联答复，使得能够准确回复需要获知多重相关信息的第二目标问题，从而使得系统能够对其他语义更为复杂，答案更为多样，或是需要在获取多重相关信息后才可进行回答的问题进行准确回复，从而解决了现有的智能问答系统所支持回答的问题类型不全面的技术问题。本申请应用于多类型问题智能问答系统。
参照图2，图2为本申请多类型问题智能问答方法第一实施例的流程示意图。
本申请第一实施例提供一种多类型问题智能问答方法，所述多类型问题智能问答方法包括以下步骤：
步骤S10，在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
在本实施例中，目标问题通常为用户通过客户端以自然语言输入或语音输入等方式向多类型问题智能问答系统输入的需要获得回复的问题。预设深度网络模型融合了Transformer的双向编码器表征(BERT，Bidirectional Encoder Representations from Transformers)与双向长短时记忆网络(Bi-LSTM,Bi-directional Long Short-Term Memory)。意图分类旨在将用户所提出的目标问题通过深度网络模型进行分类。问题类型具体可包括常见型问题、事实性问题、任务型问题与其他无关问题。多类型问题智能问答系统（以下简称系统）在接收到用户所提出的当前需要答复的目标问题时，将目标问题作为模型输入BERT层，BERT层得到所述用户问题对应词序列的序列表征向量。系统在获取到词序列的序列表征向量后，再将序列表征向量输入Bi-LSTM层，得到目标问题的词序列的隐藏向量。系统结合序列表征向量和隐藏向量学习得到目标表征向量，并经过全连接层以及Softmax（输出层的激励函数）操作，即可得到目标问题的所属分类，也即是上述分类结果。
步骤S20，在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，其中，所述第一目标问题为第一问题类型的目标问题；
在本实施例中，第一问题类型具体可为事实型、观点型、任务型或其他类型，优选为观点型。当系统识别出当前的目标问题属于观点型问题时，采用KBQA技术进行答复。KBQA技术即采用知识库问答（KB-QA，knowledge base question answering）技术，通过给定自然语言问题，对问题进行语义理解和解析，进而利用知识库进行查询、推理得出答案。具体的，从应用领域的角度划分，知识库问答可以分为开放域的知识问答，如百科知识问答，和特定域的知识问答，如金融领域，医疗领域，宗教领域等。系统先对第一目标问题的实体进行识别，以根据第一目标问题的实体信息得到与第一目标问题相关的问题模板，并对第一目标问题进行多层次语义识别，以深度理解第一目标问题的语义。最后结合得到的问题模板与解析出的多层次语义，进行知识图谱的结构化查询，以生成第一目标问题的答复信息。
步骤S30，在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，其中，所述第二目标问题为第二问题类型的目标问题。
在本实施例中，第二问题类型具体可为事实型、观点型、任务型或其他类型，优选为任务型。多轮问答技术为，先获取用户端输入的初始信息也即是第二目标问题，并对初始信息进行识别，确定出在不同条件组合下对应不同答案的预定问题，也即是关联问题集，并开始进行对于所确定的预定问题的第一轮问答。在每一轮问答中，输出对于所确定的预定问题的一个未知条件的提问，以供用户回答。在获取到用户所答复的条件，也即是多重关联答复时，根据预定的条件组合和答案之间的对应关系，判断已获取的条件的组合是否存在对应的答案。若存在，则输出已获取的条件的组合所对应的答案；若不存在，则进行下一轮问答。系统可融合了FAQ、KBQA、多轮问答等技术，按照当前目标问题所对应的实际分类结果，选择最适用于答复目标问题的目标技术类型。具体地，若系统识别出当前的目标问题为常见型问题，则可选用基于预设聚类算法的常见问题解答方式进行处理；若系统识别出当前的目标问题为事实型问题，则可采用结合问题模板生成与多层次语义解析的KBQA技术进行处理；若系统识别出当前的目标问题为任务线问题，则可选用基于槽填充的多轮问答技术进行处理，以得到针对当前目标问题的答复信息并向用户端进行输出。
在本实施例中，通过在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，其中，所述第一目标问题为第一问题类型的目标问题；在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，其中，所述第二目标问题为第二问题类型的目标问题。通过上述方式，本申请通过深度网络模型确定目标问题的问题类型，提升了问题定位的准确性；通过针对第一目标问题进行实体识别与多层次语义解析，有效降低了模型复杂度，实现了对于第一目标问题的语义的深层次理解，从而能够高效精准地回复具有复杂语义的第一目标问题；通过针对第二目标问题与客户端展开多轮问答，直至获取到客户端对于关联问题集的关联答复，使得能够准确回复需要获知多重相关信息的第二目标问题，从而使得系统能够对其他语义更为复杂，答案更为多样，或是需要在获取多重相关信息后才可进行回答的问题进行准确回复，解决了现有智能问答系统的问题回复准确性不高的技术问题。
进一步地，图中未示的，基于上述图2所示的第一实施例，提出本申请多类型问题智能问答方法的第二实施例，本实施例中，步骤S10具体包括：
在接收到目标问题时，将所述目标问题输入预训练语言表征模型，得到所述目标问题对应词序列的序列表征向量；
将所述序列表征向量输入预设双向长短时记忆网络中，得到所述词序列的隐藏向量；
结合所述序列表征向量与所述隐藏向量，确定所述目标问题的问题类型。
本实施例中，预训练语言表征模型即为上述BERT，BERT是基于双向Transformer的自然语言预训练模型，再结合自然语言处理（NLP，Natural Language Processing)具体下游任务组合其他网络层形成适应于特定任务的模型。需要说明BERT模型中[CLS]可以对整个句子进行向量表示。BERT模型在每个句子的词序列前面加一个特殊的记号[CLS]，基于Transformer网络结构可以将整个句子的全局信息编码到每个位置，BERT模型对[CLS]进行深度表征可以学到整个句子的上层特征。
具体地，对于一句话先将句子分词得到句子的词序列 w 1 ， w 2 ……， w n ，并在整个句子的输入序列开头增加[CLS]得到模型的输入序列 w [CLS ， w 1 ，…… w n 。通过BERT层获取的得到词序列的表征向量 e [CLS ， e 1 ， e 2 ……， e n ，其中 e [CLS 可以学习到整个句子的上层特征。将一句话的词向量 e 1 ， e 2 ……， e n 输入Bi-LSTM层，到隐藏向量 h 1 ， h 2 ……， h n 。Bi-LSTM可以很好的学习到对应词前后的语义信息，可以将更深层次的语义信息融入到模型中。然后将上一层得到的隐藏向量 h 1 ， h 2 ……， h n 与BERT层[CLS]的表征向量 e [CLS 进行Attention（注意力机制）操作，做一个近似于加权求和的操作，得到目标表征向量Vector。将学习到的目标表征向量Vector经过全连接层以及Softmax操作，得到输入句子（目标问题）的所属分类，也即是上述问题类型。
进一步地，在本实施例中，步骤S20包括：
在检测到所述目标问题为第一问题类型时，基于离线处理方式，使用预设实体识别模型识别出所述第一目标问题的实体信息，并基于所述实体信息得到所述问题模板；
在本实施例中，需要说明的是系统中的KBQA模块主要包含离线处理和线上输出两个部分。离线处理部分的主要功能是利用QA语料、只是图谱的信息，从中抽取模板构建问题模板库。具体地，系统中KBQA模块的离线处理部分使用基于长短时记忆（LSTM，Long Short-Term Memory）-条件随机场（CRF，conditional random field）的实体识别模型可以从自然语句中识别出实体部分，通过对QA语料中的自然语句进行实体识别操作后，可以将自然语句中的实体映射到对应的概念上，从而减少大量重复的问题，而且将问题概念化可以辅助模型学习到更精准的语义信息。通过知识图谱中的属性（关系类型）进行属性扩展获取得到更多的关系表达。通过对QA语料中的实体概念映射以及对知识图谱的属性扩展从而获取得到高质量的问题模板。
基于线上处理方式，对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义；
使用预设概率图模型，并结合所述问题模板与所述多层次语义，预测所述第一目标问题对应到知识图谱中的属性类别；
根据所述属性类别与所述实体信息，将所述第一目标问题转换为知识图谱的结构化查询，以得到所述第一目标问题的答复信息，所述第一目标问题的答复信息存储于区块链中。
在本实施例中，线上处理部分的主要功能是对第一目标问题进行解析获取得到实体以及对应的属性类型，并将其转换成知识图谱结构化查询，最终将第一目标问题的答案返回给用户。系统中KBQA模块的线上处理部分先对第一目标问题进行多层次语义解析，再使用预设的概率图模型综合使用问题的语义解析结果，以及通过离线处理部分获取到的问题模板，预测出第一目标问题对应到知识图谱中的属性类型，也即是实现问题语义解析。模型预测与知识图谱查询的功能。线上处理部分最后根据得到的属性类型以及模型识别出的第一目标问题中的实体等信息，将第一目标问题转换成知识图谱的结构化查询，并将第一目标问题的答案返回到用户端。需要强调的是，为进一步保证上述第一目标问题的答复信息的私密和安全性，上述第一目标问题的答复信息还可以存储于一区块链的节点中。
进一步地，在本实施例中，所述对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义的步骤包括：
使用预设语义搜索模型对所述第一目标问题进行实体层语义解析，获取实体层语义；
使用预设动词模板对所述第一目标问题进行细粒度语义表示，获取所述第一目标问题的短语层语义；
基于所述问题模板将所述第一目标问题的实体信息进行概念映射，获取所述第一目标问题的问题层语义，以将所述实体层语义、短语层语义与问题层语义作为所述多层次语义。
在本实施例中，多层次语义解析包括实体层语义解析、短语层语义解析与问题层语义解析。KBQA模块中的线上处理部分使用预设语义搜索模型，基于知识图谱的实体语义理解，为上层语义计算获取实体层语义信息；主要使用动词模板用来进行细粒度的语义表示，并综合使用上下文信息进行实体的概念化，以获取短语层语义信息；还基于问题模板的表示方法，将第一目标问题中的实体映射到概念上，进一步表示第一目标问题的语义，以获取问题层语义信息。
作为一具体实施例，第一目标问题为观点型问题。如图3所示，图3为融合问题模板与多层次语义解析的KBQA技术框架示意图。系统中的KBQA模块具体可分为离线处理与线上处理两部分。离线处理部分中离线处理模块主要是利用QA语料、知识图谱的信息从中抽取模板构建出问题模板库；而线上处理部分包括问题解析子模块、模型预测子模块与知识图谱查询子模块，基于这些子模块将观点型问题转换成知识图谱的结构化查询，以获取目标问题的答案。
在本实施例中，进一步通过融合问题模板以及多层次的语义解析技术，可以有效的降低模型的复杂度，并且可以深层次理解问题，提高了模型的精度；通过对QA语料中的自然语句进行实体识别操作后，可以将自然语句中的实体映射到对应的概念上，从而减少大量重复的问题，而且将问题概念化可以辅助模型学习到更精准的语义信息；通过知识图谱中的属性类型进行属性扩展可以获取得到更多的属性表达；通过对问答（QA，Question Answering）语料的中实体概念映射以及对知识图谱的属性扩展从而获取得到高质量的问题模板。
进一步地，图中未示的，基于上述图2所示的第一实施例，提出本申请多类型问题智能问答方法的第三实施例。本实施例中，所述步骤S30包括：
在检测到所述目标问题为第二问题类型时，从预设关联问题库中确定所述关联问题集，并围绕所述关联问题集与客户端进行多轮问答；
接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充；
在检测到槽填充完成时，基于预设深度网络序列标注模型，将所述多重关联答复转化为可用信息，以使用所述可用信息生成所述第二目标问题的答复信息，所述第二目标问题的答复信息存储于区块链中。
在本实施例中，预设关联问题库为预先设定的用于将第二目标问题拆分为多个子问题的问题库。
系统在检测到目标问题属于第二问题类型时，则可对第二目标问题进行语义识别，并根据语义识别结果从预设的关联问题库中确定出与用于回答第二目标问题的多个子问题，也即是上述关联问题集。可采用多轮问答技术，系统中的用于答复第二目标问题的多轮问答模块围绕关联问题集与客户端进行多轮问答，并在接收到客户端发送的针对关联问题集的多重关联回复时，根据客户端的多重关联回复进行槽填充；多轮问答模块使用预设深度网络序列标注模型，将客户端的多重关联回复转化为可用信息，以根据可用信息生成第二目标问题的答复信息。其中，预设深度网络序列标注模型即为LSTM-CRF模型。需要强调的是，为进一步保证上述第二目标问题的答复信息的私密和安全性，上述第二目标问题的答复信息还可以存储于一区块链的节点中。
进一步地，在本实施例中，所述接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充的步骤包括：
从所述多重关联答复中提取出初始槽位信息；
判断所述初始槽位信息是否符合预设有效性标准，并在所述初始槽位信息不符合预设有效性标准时，使用澄清话术进行澄清表达；
在接收到客户端发送的对于所述澄清话术的修正回复时，根据所述修正回复修正所述初始槽位信息，直至修正后的初始槽位信息符合所述有效性标准；
对修正后的初始槽位信息进行标准化处理，并将经标准化处理后的初始槽位信息作为目标槽位信息，以基于所述目标槽位信息进行槽填充。
在本实施例中，初始槽位信息为根据用户针对关联问题集的回复中的关键信息。例如，关于“预约疫苗接种”这一任务型问题，“预约疫苗接种”需要知道疫苗名称、疫苗种类（一类、二类）、疫苗接种点、预约接种时间等。预约疫苗接种需要知道四个信息，这四个信息即是四个槽位。预设有效性标准为，判断能否根据从客户端所发送的问题或回复消息中提取出的槽位信息来完成填槽的标准条件，可根据实际需求灵活设置，例如，对于预约疫苗接种这一任务型问题，客户端发送的回复语句当中没有关于疫苗种类的信息，则系统可判定当前仍不满足有效性标准，需要继续进行问答以获取必要信息。修正回复为用户对于系统发出的澄清表达的回复信息，例如，用户在关于预约疫苗接种这一任务型问题与系统进行多轮问答时，为说明关于疫苗种类这一槽位的信息，系统终端多轮问答模块则会向客户端发送类似于“需要预约接种哪一种类的疫苗”等的澄清表达，以便客户端用户进行明确答复。系统中的多轮问答模块从所述多重关联答复中提取出初始槽位信息，并判断初始槽位信息是否符合预设有效性标准。若初始槽位信息符合预设有效性标准，则可判定当前以完成关于第二目标问题的问答；若初始槽位信息中有一个或多个不符合预设有效性标准，则使用澄清话术机械能澄清表达。多轮问答模块在接收到客户端发送的对于澄清话术的修正回复时，根据修正回复修正初始槽位信息，并对修正后的初始槽位信息进行标准化处理，以将经标准化处理后的初始槽位信息作为目标槽位信息，此时则可判定完成槽填充，无需再进行问答。
作为一具体实施例，设定第二目标问题为任务型问题。以下以“预约疫苗接种”为例说明多轮问答的流程。“预约疫苗接种”需要知道疫苗名称、疫苗种类（一类、二类）、疫苗接种点、预约接种时间等。预约疫苗接种需要知道四个信息，这四个信息即是四个槽位，从与用户的对话中获取四个槽位的信息即是填槽（槽填充）。首先需要从用户的初始问题中获取四个槽位的信息，当检测到当前存在槽位的信息没有填写时，需要使用澄清话术进行澄清表达，比如针对疫苗接种点的澄清话术是“您想预约接种疫苗的疫苗接种点是哪个？”，并提供候选项给用户进行选择，用户也可以直接使用自然语言进行回复。槽与槽之间的是否有依赖关系可以分为依赖槽和平级槽，疫苗种类这个槽是依赖于疫苗名称的，预约的接种时间是依赖于疫苗接种点。依赖槽在澄清时需要有先后顺序，比如需要获取疫苗接种点的信息，才能澄清可预约的时间信息。当用户使用自然语言进行回复时，系统使用基于LSTM-CRF深度网络序列标注模型来获取可用信息。比如当疫苗接种点信息需要澄清时，机器提问“您想预约接种疫苗的疫苗接种点是哪个？”，并提供候选项给用户。如果用户是从候选项进行选择，则直接填槽；如果用户是使用自然语言进行回复，比如“预约某疫苗接种点”，需要使用序列标注模型可以获取到“预约【某疫苗接种点】(疫苗接种点)”，然后需要验证从对话中获取的槽位信息的合法性，这里需要验证疫苗接种点是否在可供选择的范围内。然后继续澄清其他需要填写的槽位信息。比如用户“预约明天上午10点钟去XXX疫苗接种点接种乙肝疫苗。”，使用序列标注模型可以获取到“预约【明天上午10点钟】（预约接种时间）去【XXX疫苗接种点】（疫苗接种点）接种【乙肝疫苗】（疫苗名称）”，然后将从自然语句中获取的不同槽位的信息进行标准化处理和验证信息是否合法。比如需要将明天上午10点钟映射成标准时间信息，需要验证用户提及的疫苗接种点和预约接种时间是否在可供选择的范围内等等。当获取了所有的槽位信息后，需要将所有的槽位信息与用户进行再次确认，确认无误后即可完成需要使用的信息。
进一步地，在本实施例中，在步骤S10之后，还包括：
在检测到所述目标问题为第三问题类型时，将第三问题类型的目标问题作为第三目标问题，并获取所述第三目标问题与基于预设聚类算法所构建的标准问题集之间的问题相似度；
将所述第三目标问题映射在所述标准问题集中的相似标准问题上，并获取所述相似标准问题的标准回复，以基于所述标准回复生成所述第三目标问题的答复信息，所述相似标准问题在所述标准问题集中与所述目标问题之间的问题相似度最高。
在本实施例中，第三问题类型具体可为事实型、观点型、任务型或其他类型，优选为事实型。预设聚类算法可选用具有噪声的基于密度的聚类方法（DBSCAN，Density-Based Spatial Clustering of Applications with Noise）、期望最大化（EM，expectation maxmization）聚类算法、K均值（K-means）聚类算法等，优选为K-means聚类算法。常见问题回答方式即采用常见问题解答（FQA，Frequently Asked Questions）技术，FAQ是当前网络上提供在线帮助的主要手段，通过事先组织好一些可能的常问问答对，发布在网页上为用户提供咨询服务。
FAQ技术通过模型计算用户问题和已有的标准问题之间的相似度，将用户的问题映射到已有的标准问题上从而实现问答。系统在FAQ模块使用的模型是复用了意图分类模块中的融合BERT与BiLSTM的深度学习模型，模型的输入是用户的问题，模型的输出是与第三目标问题相似度最高的标准问题对应的答案。FAQ技术能覆盖的问题的种类数取决于设计的标准问题的多少，目前标准问题多是通过人工整理得到，能覆盖的问题的种类数比较少。而本申请中的系统基于机器学习聚类算法K-means算法从海量的领域问答语料中自动构建大量高质量的标准问题。自动构建高质量的标准问题主要包括三步：第一步，从高质量疫苗垂直型网站数据源获取初始的问题；第二步，通过聚类算法将相似的问题聚成一簇；第三步，筛选标准问题并获取答案。在实际应用中，通过该步骤可构建五千左右疫苗领域的高质量标准问题集合。
在本实施例中，进一步通过在系统中的多轮问答模块中基于槽填充、多轮对话技术，高效解决了任务型问题，尤其是针对疫苗问答相关的任务型问题；通过在FAQ模块中基于K-means聚类算法自动构建大规模高质量的标准问题集，从而减少了人工整理标注标准问题集的成本。
本申请所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链(Blockchain)，本质上是一个去中心化的数据库，是一串使用密码学方法相关联产生的数据块，每一个数据块中包含了一批次网络交易的信息，用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
此外，本申请实施例还提供一种多类型问题智能问答系统。本实施例中，所述多类型问题智能问答系统包括：
意图分类模块，用于在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
第一问题答复模块，用于在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题
第二问题答复模块，用于在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
上述多类型问题智能问答系统中各个模块与上述多类型问题智能问答方法实施例中各步骤相对应，其功能和实现过程在此处不再一一赘述。
本申请还提供一种多类型问题智能问答设备。
所述多类型问题智能问答设备包括处理器、存储器及存储在所述存储器上并可在所述处理器上运行的多类型问题智能问答程序，其中所述多类型问题智能问答程序被所述处理器执行时，实现如上所述的多类型问题智能问答方法的步骤。
所述多类型问题智能问答程序被执行时所实现的方法可参照本申请多类型问题智能问答方法的各个实施例，此处不再赘述。
此外，本申请实施例还提供一种计算机可读存储介质，所述计算机可读存储介质可以是非易失性的，也可以是易失性的。
本申请计算机可读存储介质上存储有多类型问题智能问答程序，其中所述多类型问题智能问答程序被处理器执行时，实现如上述的多类型问题智能问答方法的步骤。
多类型问题智能问答程序被执行时所实现的方法可参照本申请多类型问题智能问答方法的各个实施例，此处不再赘述。
需要说明的是，在本文中，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者系统不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者系统所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括该要素的过程、方法、物品或者系统中还存在另外的相同要素。
上述本申请实施例序号仅仅为了描述，不代表实施例的优劣。
通过以上的实施方式的描述，本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件，但很多情况下前者是更佳的实施方式。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，空调器，或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种多类型问题智能问答方法，其中，所述多类型问题智能问答方法包括以下步骤：
    在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
    在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
    在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
  2. 如权利要求1所述的多类型问题智能问答方法，其中，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤包括：
    在接收到目标问题时，将所述目标问题输入预训练语言表征模型，得到所述目标问题对应词序列的序列表征向量；
    将所述序列表征向量输入预设双向长短时记忆网络中，得到所述词序列的隐藏向量；
    结合所述序列表征向量与所述隐藏向量，确定所述目标问题的问题类型。
  3. 如权利要求1所述的多类型问题智能问答方法，其中，所述在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息的步骤包括：
    在检测到所述目标问题为第一问题类型时，基于离线处理方式，使用预设实体识别模型识别出所述第一目标问题的实体信息，并基于所述实体信息得到所述问题模板；
    基于线上处理方式，对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义；
    使用预设概率图模型，并结合所述问题模板与所述多层次语义，预测所述第一目标问题对应到知识图谱中的属性类别；
    根据所述属性类别与所述实体信息，将所述第一目标问题转换为知识图谱的结构化查询，以得到所述第一目标问题的答复信息，所述第一目标问题的答复信息存储于区块链中。
  4. 如权利要求3所述的多类型问题智能问答方法，其中，所述对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义的步骤包括：
    使用预设语义搜索模型对所述第一目标问题进行实体层语义解析，获取实体层语义；
    使用预设动词模板对所述第一目标问题进行细粒度语义表示，获取所述第一目标问题的短语层语义；
    基于所述问题模板将所述第一目标问题的实体信息进行概念映射，获取所述第一目标问题的问题层语义，以将所述实体层语义、短语层语义与问题层语义作为所述多层次语义。
  5. 如权利要求1所述的多类型问题智能问答方法，其中，所述在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息的步骤包括：
    在检测到所述目标问题为第二问题类型时，从预设关联问题库中确定所述关联问题集，并围绕所述关联问题集与客户端进行多轮问答；
    接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充；
    在检测到槽填充完成时，基于预设深度网络序列标注模型，将所述多重关联答复转化为可用信息，以使用所述可用信息生成所述第二目标问题的答复信息，所述第二目标问题的答复信息存储于区块链中，所述第二目标问题的答复信息存储于区块链中。
  6. 如权利要求5所述的多类型问题智能问答方法，其中，所述接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充的步骤包括：
    从所述多重关联答复中提取出初始槽位信息；
    判断所述初始槽位信息是否符合预设有效性标准，并在所述初始槽位信息不符合预设有效性标准时，使用澄清话术进行澄清表达；
    在接收到客户端发送的对于所述澄清话术的修正回复时，根据所述修正回复修正所述初始槽位信息，直至修正后的初始槽位信息符合所述有效性标准；
    对修正后的初始槽位信息进行标准化处理，并将经标准化处理后的初始槽位信息作为目标槽位信息，以基于所述目标槽位信息进行槽填充。
  7. 如权利要求1所述的多类型问题智能问答方法，其中，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤之后，还包括：
    在检测到所述目标问题为第三问题类型时，将第三问题类型的目标问题作为第三目标问题，并获取所述第三目标问题与基于预设聚类算法所构建的标准问题集之间的问题相似度；
    将所述第三目标问题映射在所述标准问题集中的相似标准问题上，并获取所述相似标准问题的标准回复，以基于所述标准回复生成所述第三目标问题的答复信息，所述相似标准问题在所述标准问题集中与所述目标问题之间的问题相似度最高。
  8. 一种多类型问题智能问答系统，其中，所述多类型问题智能问答系统包括：
    意图分类模块，用于在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
    第一问题答复模块，用于在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
    第二问题答复模块，用于在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
  9. 一种多类型问题智能问答设备，其中，所述多类型问题智能问答设备包括处理器、存储器、以及存储在所述存储器上并可被所述处理器执行的多类型问题智能问答程序，其中所述多类型问题智能问答程序被所述处理器执行时，实现如下步骤：
    在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
    在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
    在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
  10. 如权利要求9所述的多类型问题智能问答设备，其中，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤包括：
    在接收到目标问题时，将所述目标问题输入预训练语言表征模型，得到所述目标问题对应词序列的序列表征向量；
    将所述序列表征向量输入预设双向长短时记忆网络中，得到所述词序列的隐藏向量；
    结合所述序列表征向量与所述隐藏向量，确定所述目标问题的问题类型。
  11. 如权利要求9所述的多类型问题智能问答设备，其中，所述在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息的步骤包括：
    在检测到所述目标问题为第一问题类型时，基于离线处理方式，使用预设实体识别模型识别出所述第一目标问题的实体信息，并基于所述实体信息得到所述问题模板；
    基于线上处理方式，对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义；
    使用预设概率图模型，并结合所述问题模板与所述多层次语义，预测所述第一目标问题对应到知识图谱中的属性类别；
    根据所述属性类别与所述实体信息，将所述第一目标问题转换为知识图谱的结构化查询，以得到所述第一目标问题的答复信息，所述第一目标问题的答复信息存储于区块链中。
  12. 如权利要求11所述的多类型问题智能问答设备，其中，所述对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义的步骤包括：
    使用预设语义搜索模型对所述第一目标问题进行实体层语义解析，获取实体层语义；
    使用预设动词模板对所述第一目标问题进行细粒度语义表示，获取所述第一目标问题的短语层语义；
    基于所述问题模板将所述第一目标问题的实体信息进行概念映射，获取所述第一目标问题的问题层语义，以将所述实体层语义、短语层语义与问题层语义作为所述多层次语义。
  13. 如权利要求9所述的多类型问题智能问答设备，其中，所述在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息的步骤包括：
    在检测到所述目标问题为第二问题类型时，从预设关联问题库中确定所述关联问题集，并围绕所述关联问题集与客户端进行多轮问答；
    接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充；
    在检测到槽填充完成时，基于预设深度网络序列标注模型，将所述多重关联答复转化为可用信息，以使用所述可用信息生成所述第二目标问题的答复信息，所述第二目标问题的答复信息存储于区块链中，所述第二目标问题的答复信息存储于区块链中。
  14. 如权利要求13所述的多类型问题智能问答设备，其中，所述接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充的步骤包括：
    从所述多重关联答复中提取出初始槽位信息；
    判断所述初始槽位信息是否符合预设有效性标准，并在所述初始槽位信息不符合预设有效性标准时，使用澄清话术进行澄清表达；
    在接收到客户端发送的对于所述澄清话术的修正回复时，根据所述修正回复修正所述初始槽位信息，直至修正后的初始槽位信息符合所述有效性标准；
    对修正后的初始槽位信息进行标准化处理，并将经标准化处理后的初始槽位信息作为目标槽位信息，以基于所述目标槽位信息进行槽填充。
  15. 如权利要求9所述的多类型问题智能问答设备，其中，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤之后，还包括：
    在检测到所述目标问题为第三问题类型时，将第三问题类型的目标问题作为第三目标问题，并获取所述第三目标问题与基于预设聚类算法所构建的标准问题集之间的问题相似度；
    将所述第三目标问题映射在所述标准问题集中的相似标准问题上，并获取所述相似标准问题的标准回复，以基于所述标准回复生成所述第三目标问题的答复信息，所述相似标准问题在所述标准问题集中与所述目标问题之间的问题相似度最高。
  16. 一种计算机可读存储介质，其中，所述计算机可读存储介质上存储有多类型问题智能问答程序，其中所述多类型问题智能问答程序被处理器执行时，实现如下步骤：
    在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型；
    在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息，所述第一目标问题为第一问题类型的目标问题；
    在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息，所述第二目标问题为第二问题类型的目标问题。
  17. 如权利要求16所述的计算机可读存储介质，其中，所述在接收到目标问题时，基于预设深度网络模型对所述目标问题进行意图分类，以确定所述目标问题的问题类型的步骤包括：
    在接收到目标问题时，将所述目标问题输入预训练语言表征模型，得到所述目标问题对应词序列的序列表征向量；
    将所述序列表征向量输入预设双向长短时记忆网络中，得到所述词序列的隐藏向量；
    结合所述序列表征向量与所述隐藏向量，确定所述目标问题的问题类型。
  18. 如权利要求16所述的计算机可读存储介质，其中，所述在检测到所述目标问题为第一问题类型时，对第一目标问题进行实体识别与多层次语义解析，以获取所述第一目标问题的问题模板与多层次语义，并根据所述问题模板与多层次语义，生成所述第一目标问题的答复信息的步骤包括：
    在检测到所述目标问题为第一问题类型时，基于离线处理方式，使用预设实体识别模型识别出所述第一目标问题的实体信息，并基于所述实体信息得到所述问题模板；
    基于线上处理方式，对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义；
    使用预设概率图模型，并结合所述问题模板与所述多层次语义，预测所述第一目标问题对应到知识图谱中的属性类别；
    根据所述属性类别与所述实体信息，将所述第一目标问题转换为知识图谱的结构化查询，以得到所述第一目标问题的答复信息，所述第一目标问题的答复信息存储于区块链中。
  19. 如权利要求18所述的计算机可读存储介质，其中，所述对所述第一目标问题进行多层次语义解析，得到所述第一目标问题的多层次语义的步骤包括：
    使用预设语义搜索模型对所述第一目标问题进行实体层语义解析，获取实体层语义；
    使用预设动词模板对所述第一目标问题进行细粒度语义表示，获取所述第一目标问题的短语层语义；
    基于所述问题模板将所述第一目标问题的实体信息进行概念映射，获取所述第一目标问题的问题层语义，以将所述实体层语义、短语层语义与问题层语义作为所述多层次语义。
  20. 如权利要求16所述的计算机可读存储介质，其中，所述在检测到所述目标问题为第二问题类型时，确定第二目标问题的关联问题集，并基于多轮问答技术接收客户端对于所述关联问题集的多重关联答复，以基于所述多重关联答复生成所述第二目标问题的答复信息的步骤包括：
    在检测到所述目标问题为第二问题类型时，从预设关联问题库中确定所述关联问题集，并围绕所述关联问题集与客户端进行多轮问答；
    接收多轮问答中客户端发送的所述多重关联答复，并根据所述多重关联答复进行槽填充；
    在检测到槽填充完成时，基于预设深度网络序列标注模型，将所述多重关联答复转化为可用信息，以使用所述可用信息生成所述第二目标问题的答复信息，所述第二目标问题的答复信息存储于区块链中，所述第二目标问题的答复信息存储于区块链中。
PCT/CN2020/117816 2020-06-17 2020-09-25 多类型问题智能问答方法、系统、设备及可读存储介质 WO2021109690A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010558326.1 2020-06-17
CN202010558326.1A CN111858877A (zh) 2020-06-17 2020-06-17 多类型问题智能问答方法、系统、设备及可读存储介质

Publications (1)

Publication Number Publication Date
WO2021109690A1 true WO2021109690A1 (zh) 2021-06-10

Family

ID=72986186

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/117816 WO2021109690A1 (zh) 2020-06-17 2020-09-25 多类型问题智能问答方法、系统、设备及可读存储介质

Country Status (2)

Country Link
CN (1) CN111858877A (zh)
WO (1) WO2021109690A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113420111A (zh) * 2021-06-17 2021-09-21 中国科学院声学研究所 一种用于多跳推理问题的智能问答方法及装置
CN113449117A (zh) * 2021-06-24 2021-09-28 武汉工程大学 一种基于Bi-LSTM和中文知识图谱的复合问答方法
CN113590776A (zh) * 2021-06-23 2021-11-02 北京百度网讯科技有限公司 基于知识图谱的文本处理方法、装置、电子设备及介质
CN113689851A (zh) * 2021-07-27 2021-11-23 国家电网有限公司 调度专业语言理解系统及方法
CN113887621A (zh) * 2021-09-30 2022-01-04 中国平安财产保险股份有限公司 问答资源调整方法、装置、设备及存储介质
CN116821312A (zh) * 2023-08-30 2023-09-29 中国石油大学(华东) 一种基于学科领域知识图谱的复杂问答方法
CN117171332A (zh) * 2023-11-02 2023-12-05 江西拓世智能科技股份有限公司 基于ai的智能问答方法及系统

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112380343A (zh) * 2020-11-19 2021-02-19 平安科技（深圳）有限公司 问题解析方法、装置、电子设备及存储介质
CN112565663B (zh) * 2020-11-26 2022-11-18 平安普惠企业管理有限公司 需求问题回复方法、装置、终端设备及存储介质
CN113704432A (zh) * 2021-08-31 2021-11-26 广州方舟信息科技有限公司 一种基于互联网医院的人工智能客服系统构建方法及装置
CN115062628A (zh) * 2022-06-15 2022-09-16 北京信息科技大学 一种基于知识图谱的医患交流对话自动模拟方法
CN115238101B (zh) * 2022-09-23 2023-01-03 中国电子科技集团公司第十研究所 一种面向多类型知识库的多引擎智能问答系统
CN115470338B (zh) * 2022-10-27 2023-04-14 之江实验室 一种基于多路召回的多场景智能问答方法和系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104598445A (zh) * 2013-11-01 2015-05-06 腾讯科技（深圳）有限公司 自动问答系统和方法
CN105159996A (zh) * 2015-09-07 2015-12-16 百度在线网络技术（北京）有限公司 基于人工智能的深度问答服务提供方法和装置
CN108491394A (zh) * 2018-06-27 2018-09-04 杭州贝店科技有限公司 一种语义分析方法、装置、计算机设备及存储介质
CN109284363A (zh) * 2018-12-03 2019-01-29 北京羽扇智信息科技有限公司 一种问答方法、装置、电子设备及存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105868313B (zh) * 2016-03-25 2019-02-12 浙江大学 一种基于模板匹配技术的知识图谱问答系统及方法
JP6986978B2 (ja) * 2018-01-16 2021-12-22 ヤフー株式会社 情報処理装置、情報処理方法、及び情報処理プログラム
CN109492077B (zh) * 2018-09-29 2020-09-29 北京智通云联科技有限公司 基于知识图谱的石化领域问答方法及系统
CN110245240A (zh) * 2019-05-31 2019-09-17 国网电子商务有限公司 一种问题数据答案的确定方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104598445A (zh) * 2013-11-01 2015-05-06 腾讯科技（深圳）有限公司 自动问答系统和方法
CN105159996A (zh) * 2015-09-07 2015-12-16 百度在线网络技术（北京）有限公司 基于人工智能的深度问答服务提供方法和装置
CN108491394A (zh) * 2018-06-27 2018-09-04 杭州贝店科技有限公司 一种语义分析方法、装置、计算机设备及存储介质
CN109284363A (zh) * 2018-12-03 2019-01-29 北京羽扇智信息科技有限公司 一种问答方法、装置、电子设备及存储介质

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113420111A (zh) * 2021-06-17 2021-09-21 中国科学院声学研究所 一种用于多跳推理问题的智能问答方法及装置
CN113420111B (zh) * 2021-06-17 2023-08-11 中国科学院声学研究所 一种用于多跳推理问题的智能问答方法及装置
CN113590776A (zh) * 2021-06-23 2021-11-02 北京百度网讯科技有限公司 基于知识图谱的文本处理方法、装置、电子设备及介质
CN113590776B (zh) * 2021-06-23 2023-12-12 北京百度网讯科技有限公司 基于知识图谱的文本处理方法、装置、电子设备及介质
CN113449117B (zh) * 2021-06-24 2023-09-26 武汉工程大学 一种基于Bi-LSTM和中文知识图谱的复合问答方法
CN113449117A (zh) * 2021-06-24 2021-09-28 武汉工程大学 一种基于Bi-LSTM和中文知识图谱的复合问答方法
CN113689851A (zh) * 2021-07-27 2021-11-23 国家电网有限公司 调度专业语言理解系统及方法
CN113689851B (zh) * 2021-07-27 2024-02-02 国家电网有限公司 调度专业语言理解系统及方法
CN113887621A (zh) * 2021-09-30 2022-01-04 中国平安财产保险股份有限公司 问答资源调整方法、装置、设备及存储介质
CN113887621B (zh) * 2021-09-30 2024-04-30 中国平安财产保险股份有限公司 问答资源调整方法、装置、设备及存储介质
CN116821312A (zh) * 2023-08-30 2023-09-29 中国石油大学(华东) 一种基于学科领域知识图谱的复杂问答方法
CN116821312B (zh) * 2023-08-30 2023-11-14 中国石油大学(华东) 一种基于学科领域知识图谱的复杂问答方法
CN117171332A (zh) * 2023-11-02 2023-12-05 江西拓世智能科技股份有限公司 基于ai的智能问答方法及系统

Also Published As

Publication number Publication date
CN111858877A (zh) 2020-10-30

Similar Documents

Publication Publication Date Title
WO2021109690A1 (zh) 多类型问题智能问答方法、系统、设备及可读存储介质
US11816435B1 (en) Applied artificial intelligence technology for contextualizing words to a knowledge base using natural language processing
CN111159385B (zh) 一种基于动态知识图谱的无模板通用智能问答方法
CN107846350B (zh) 一种语境感知网络聊天的方法、计算机可读介质和系统
CN109800306B (zh) 意图分析方法、装置、显示终端及计算机可读存储介质
CN115238101B (zh) 一种面向多类型知识库的多引擎智能问答系统
CN110727779A (zh) 基于多模型融合的问答方法及系统
CN111708869B (zh) 人机对话的处理方法及装置
CN109857846B (zh) 用户问句与知识点的匹配方法和装置
CN110162675B (zh) 应答语句的生成方法、装置、计算机可读介质及电子设备
US10326863B2 (en) Speed and accuracy of computers when resolving client queries by using graph database model
CN111680144A (zh) 多轮对话语音交互的方法及系统、存储介质、电子设备
CN116737910B (zh) 智能对话处理方法、装置、设备及存储介质
CN116244412A (zh) 多意图识别方法及装置
CN116974626B (zh) 分析序列图生成方法、装置、设备和计算机可读存储介质
CN113821587A (zh) 文本相关性确定方法、模型训练方法、装置及存储介质
CN113157887A (zh) 知识问答意图识别方法、装置、及计算机设备
CN115714030A (zh) 一种基于疼痛感知和主动交互的医疗问答系统及方法
CN111930919B (zh) 一种面向企业在线教育app语音交互的实现方法
CN115129859A (zh) 意图识别方法、装置、电子装置及存储介质
CN115408500A (zh) 问答一致性的评估方法、装置、电子设备及介质
CN112860873A (zh) 智能应答方法、装置及存储介质
CN117725185B (zh) 智能对话生成方法及系统
RU2818036C1 (ru) Способ и система управления диалоговым агентом в канале взаимодействия с пользователем
CN111428018B (zh) 智能问答方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20895415

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20895415

Country of ref document: EP

Kind code of ref document: A1