WO2020258654A1 - 一种答案获取方法及装置 - Google Patents

一种答案获取方法及装置 Download PDF

Info

Publication number
WO2020258654A1
WO2020258654A1 PCT/CN2019/117517 CN2019117517W WO2020258654A1 WO 2020258654 A1 WO2020258654 A1 WO 2020258654A1 CN 2019117517 W CN2019117517 W CN 2019117517W WO 2020258654 A1 WO2020258654 A1 WO 2020258654A1
Authority
WO
WIPO (PCT)
Prior art keywords
answer
question text
preset
question
type
Prior art date
Application number
PCT/CN2019/117517
Other languages
English (en)
French (fr)
Inventor
杨海军
徐倩
杨强
Original Assignee
深圳前海微众银行股份有限公司
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 深圳前海微众银行股份有限公司 filed Critical 深圳前海微众银行股份有限公司
Publication of WO2020258654A1 publication Critical patent/WO2020258654A1/zh

Links

Images

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
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3343—Query execution using phonetics
    • 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
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis

Definitions

  • the present invention relates to the field of financial technology (Fintech) and the field of intelligent communication, and in particular to an answer obtaining method and device.
  • the way to obtain the answers to the user's question in the call-in system is to search in the answer database collected manually according to the keywords of the question text, and feedback the hit answer to the user.
  • the manually collected answer library is limited and inefficient, and the answers obtained by searching based on keywords can hardly guarantee that the semantics of the obtained answers are expected, and the obtained user answers are also not accurate enough.
  • the embodiments of the present application provide an answer obtaining method and device, which solves the problem of insufficient user answers obtained in the prior art.
  • an embodiment of the present application provides an answer obtaining method, including: obtaining a question text; determining whether the question type of the question text matches the answer type in the preset answer type library through a pre-trained question classification model; If the question type of the question text does not match the answer type in the preset answer type library, then according to the first depth semantic vector of each answer in the preset answer library and the second depth of the question text At least one semantic similarity of the semantic vector is used to obtain the answer to the question text.
  • the pre-trained question classification model through the pre-trained question classification model, it is first determined whether the question type of the question text matches the answer type in the preset answer type library. When there is no match, because the at least one semantic similarity characterizes the question text and The semantic consistency of each answer in the answer library is preset, thereby improving the accuracy of obtaining the answer.
  • each key answer element of the at least one key answer element is for the sub-answer of the question text, and the at least one key answer element is The answer element is a combination of all sub-answers of the question text as the answer to the question text.
  • each key answer element disassembled from at least one key answer element the sub-answer of each key answer element to the question text is determined, so as to obtain the answer to the question text in a more fine-grained manner.
  • the accuracy rate of the answers obtained is determined, so as to obtain the answer to the question text in a more fine-grained manner.
  • the question text is determined according to at least one semantic similarity between the first depth semantic vector of each answer in the preset answer library and the second depth semantic vector of the question text
  • the answers include:
  • the answer corresponding to the first depth semantic vector of the first semantic similarity in the preset answer database is used as the question text
  • the answer; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • the answer corresponding to the first depth semantic vector of the first semantic similarity greater than or equal to the preset semantic similarity threshold in the preset answer database is used as the answer to the question text
  • the answer with the greatest semantic similarity of the question text in the preset answer library is obtained.
  • the question text matches the triples in the pre-built knowledge graph; if so, the The matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text.
  • the first semantic similarity when the first semantic similarity is less than the preset semantic similarity threshold, it is determined whether the question text matches the triples in the pre-built knowledge graph, so that the pre-built knowledge graph plays a backup role.
  • the matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text, thereby further improving the acquisition To the accuracy of the answer.
  • the question text does not match the triples in the pre-built knowledge graph, it is determined whether the question text matches the chat sentence in the preset chat library, and if so, then Use the question text and the chat sentence matched in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the answer to the question text.
  • the question text and the chat sentence matching the preset chat library are used as the answer to the question text, or the preset default answer is used as the answer to the question text, so that the question text is When it does not match the triples in the pre-built knowledge graph, there are also backup chat sentences and preset default answers, thereby improving user experience and improving answer accuracy.
  • this application provides an answer obtaining device, including: an obtaining module for obtaining a question text; a processing module for determining whether the question type of the question text is consistent with a preset answer through a pre-trained question classification model The answer types in the type library match; if the question type of the question text does not match the answer type in the preset answer type library, then the first depth semantic vector of each answer in the preset answer library is matched with At least one semantic similarity of the second depth semantic vector of the question text is used to obtain the answer of the question text.
  • the processing module is further configured to: if the question type of the question text matches the answer type in the preset answer type library, obtain the question text in the preset answer At least one key answer element corresponding to the matching answer type in the type library; according to the question text and the at least one key answer element, determine the child of each key answer element in the at least one key answer element to the question text
  • the answer is a combination of all sub-answers of the at least one key answer element to the question text as the answer to the question text.
  • the processing module is specifically configured to: if the first semantic similarity is greater than or equal to a preset semantic similarity threshold, set the first depth semantic vector of the first semantic similarity in The corresponding answer in the preset answer library is used as the answer to the question text; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • the processing module is further configured to: if the first semantic similarity is less than the preset semantic similarity threshold, determine whether the question text is consistent with three of the pre-built knowledge graphs. Tuple matching; if so, the matching triplet in the question text and the pre-built knowledge graph is used as the answer to the question text.
  • the processing module is further configured to: if the question text does not match the triples in the pre-built knowledge graph, determine whether the question text is in the preset chat library If yes, use the question text and the chat sentence in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the question text s answer.
  • an embodiment of the present application provides a computer device, including a program or instruction.
  • the program or instruction When the program or instruction is executed, the following steps are implemented: obtaining a question text; determining the question text through a pre-trained question classification model Whether the question type of the question matches the answer type in the preset answer type library; if the question type of the question text does not match the answer type in the preset answer type library, then according to each of the preset answer types At least one semantic similarity between the first depth semantic vector of the answer and the second depth semantic vector of the question text is used to obtain the answer of the question text.
  • the method further includes: if the question type of the question text matches the answer type in the preset answer type library, obtaining the information that the question text matches in the preset answer type library At least one key answer element corresponding to the answer type; according to the question text and the at least one key answer element, it is determined that each key answer element in the at least one key answer element is a sub-answer of the question text, and the The combination of at least one key answer element to all sub-answers of the question text is used as the answer to the question text.
  • the question text is determined according to at least one semantic similarity between the first depth semantic vector of each answer in the preset answer library and the second depth semantic vector of the question text
  • the answer of includes: if the first semantic similarity is greater than or equal to the preset semantic similarity threshold, then the first depth semantic vector of the first semantic similarity is corresponding to the answer in the preset answer database, As an answer to the question text; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • it further includes: if the first semantic similarity is less than the preset semantic similarity threshold, determining whether the question text matches the triples in the pre-built knowledge graph; if so , The matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text.
  • it further includes: if the question text does not match the triples in the pre-built knowledge graph, determining whether the question text matches the chat sentence in the preset chat library, If yes, use the question text and the chat sentence matched in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the answer to the question text.
  • an embodiment of the present application provides a storage medium that includes a program or instruction.
  • the program or instruction When the program or instruction is executed, the following steps are implemented: obtaining a question text; determining the question text through a pre-trained question classification model Whether the question type of the question matches the answer type in the preset answer type library; if the question type of the question text does not match the answer type in the preset answer type library, then according to each of the preset answer types At least one semantic similarity between the first depth semantic vector of the answer and the second depth semantic vector of the question text is used to obtain the answer of the question text.
  • the method further includes: if the question type of the question text matches the answer type in the preset answer type library, obtaining the information that the question text matches in the preset answer type library At least one key answer element corresponding to the answer type; according to the question text and the at least one key answer element, it is determined that each key answer element in the at least one key answer element is a sub-answer of the question text, and the The combination of at least one key answer element to all sub-answers of the question text is used as the answer to the question text.
  • the question text is determined according to at least one semantic similarity between the first depth semantic vector of each answer in the preset answer library and the second depth semantic vector of the question text
  • the answer of includes: if the first semantic similarity is greater than or equal to the preset semantic similarity threshold, then the first depth semantic vector of the first semantic similarity is corresponding to the answer in the preset answer database, As an answer to the question text; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • it further includes: if the first semantic similarity is less than the preset semantic similarity threshold, determining whether the question text matches the triples in the pre-built knowledge graph; if so , The matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text.
  • it further includes: if the question text does not match the triples in the pre-built knowledge graph, determining whether the question text matches the chat sentence in the preset chat library, If yes, use the question text and the chat sentence matched in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the answer to the question text.
  • FIG. 1 is a schematic diagram of the architecture of an incoming call intelligent system to which an answer obtaining method according to an embodiment of the application can be applied;
  • FIG. 2 is a schematic diagram of time sequence interaction of an answer obtaining method provided by an embodiment of this application.
  • FIG. 3 is a schematic flowchart of steps of an answer obtaining method provided by an embodiment of this application.
  • FIG. 4 is a schematic flowchart of specific steps of an answer obtaining method provided by an embodiment of this application.
  • FIG. 5 is a schematic structural diagram of an answer obtaining device provided by an embodiment of this application.
  • the call-in system is a system for users to ask questions and provide answers to users.
  • the call-in system is a system that supports users' call-in and provides guidance for answering questions.
  • an embodiment of the present application provides an answer obtaining method.
  • FIG. 1 it is a schematic diagram of the architecture of a telephone call-in smart system to which an answer obtaining method provided in an embodiment of this application is applicable.
  • the system architecture includes the following parts:
  • Terminal equipment The terminal equipment is the equipment used by the customer to make calls and ask questions.
  • Telecom operators are used to transfer incoming calls from terminals.
  • Company network The company network is used to transfer calls from telecom operators and transfer the calls to smart robots.
  • Agent call service platform includes multiple artificial agents, which are used to manually take over incoming calls.
  • the intelligent robot is used to answer the user's call, convert the call recording into text, or convert the text into call recording, and return the answer to the user's call.
  • S101 The user dials the customer service agent phone through the terminal device.
  • S102 The incoming call is forwarded through the telecommunication operator and company network, and the intelligent robot automatically answers the incoming call.
  • S103 The intelligent robot converts the interactive text into voice through the speech synthesis engine and plays it to the user, and guides the user to ask questions.
  • S104 The intelligent robot converts the user's call recording into a question text through a voice recognition engine.
  • S105 The intelligent customer service robot obtains the answer through the question text, and then converts it into voice through the speech synthesis engine and plays it to the user.
  • S106 Repeat S103-S105, if the user actively hangs up the call, the call ends the library; if the user enters "switch to manual" by voice input, the call will be hosted by a human agent until the user hangs up the call and the call ends.
  • FIG. 2 a schematic diagram of time sequence interaction of an answer obtaining method provided by an embodiment of this application. It includes the following steps:
  • Step 201 Make a call.
  • the user makes a call through the terminal device and generates an incoming call.
  • the call state of the terminal device is the outgoing state.
  • Step 202 Notify the phone to connect.
  • the telecom operator notifies the smart robot of the incoming call.
  • Step 203 Send a call connection command.
  • the smart robot After the smart robot receives the incoming call, it sends a call connection command to the telecom operator.
  • Step 204 Answer the call.
  • the telecom operator connects to the phone, and the call status of the terminal device is connected.
  • Step 205 Record the call voice to obtain the call recording.
  • the telecom operator answers the user's voice and obtains the recording of the call.
  • Step 206 Send the call recording.
  • the telecom operator sends the converted call recording to the speech recognition engine.
  • Step 207 Convert the call recording into a question text.
  • the voice recognition engine converts the call recording into question text.
  • Step 208 Return the answer text in response to the question text.
  • the intelligent robot obtains the answer text of the question text and returns it to the speech synthesis engine.
  • Step 209 Send the answer voice.
  • the answer text of the speech synthesis engine is converted into answer speech, and the answer speech is sent to the telecom operator.
  • Step 2010 Send the answer voice.
  • the telecom operator sends the answer voice to the terminal device.
  • step 205 to step 2010 are executed in a loop.
  • step 2011 to step 2012 are executed.
  • Step 2011 Send a manual transfer command.
  • the intelligent robot sends a manual transfer command to the telecom operator.
  • Step 2012 Access manual services.
  • the telecom operator connects the terminal equipment with the agent call service platform to talk, thereby realizing access to manual services.
  • Step 2013 Hang up the phone.
  • step 2013 may be executed between any two adjacent steps in step 201 to step 2012.
  • intelligent robots are used to automatically answer calls, and can conduct multiple rounds of dialogue interaction for user-specific task-type questions, perform in-depth semantic analysis on non-task-type questions and find answers, and also support users to transfer to manual agent services.
  • This solution can achieve a good balance between intelligent customer service robot service and agent service, solve the above-mentioned shortcomings, and bring huge commercial benefits.
  • Intelligent robot is an automatic question answering system implemented using natural language processing technology, machine learning and deep learning frameworks. It can conduct multiple rounds of conversations with users and deeply semantically understand the user's question intentions, and obtain what users need by performing tasks or searching knowledge bases. s answer.
  • the technologies used include intention understanding, semantic matching, search engines, recommendation engines, task engines, knowledge graphs, etc.
  • the preset answer type and the preset chat library are generated through machine self-learning assistance.
  • the question text cannot match the content of the preset answer type and the preset chat library, it is passed to the business system, and the business The staff confirms that if it belongs to the existing content and does not match accurately, the question text will be classified as existing content. If it does not belong to the existing content, the new content will be edited and added to the corresponding knowledge base. Thereby, it continues to accumulate into preset answer types and preset chat libraries, thereby continuously expanding the content, and constantly classifying similar questions into one category.
  • the question part of the knowledge base of this solution consists of only the text of the user's question, rather than keywords and patterns, which conforms to the natural conversation habits of humans and greatly reduces the difficulty of editing the knowledge base.
  • the answer part includes preset answer types, preset answer libraries, pre-built knowledge graphs and chat libraries, which more powerfully supports the increase in the intelligence of dialogue and the richness of answers.
  • FIG. 3 it is a schematic flow chart of the steps of an answer obtaining method provided by an embodiment of this application.
  • Step 301 Obtain the question text.
  • the question text is the text converted by the user's recording.
  • Step 302 Determine whether the question type of the question text matches the answer type in the preset answer type library through the pre-trained question classification model.
  • the question type that matches the answer type in the preset answer type library is a question type that requires an intelligent robot to calculate a specific answer.
  • the specific answer is not directly given in the preset answer library, but first through the preset answer type Database, determine the answer type, and then determine the answer to the question text based on the specific information in the question text.
  • Step 303 If the question type of the question text does not match the answer type in the preset answer type library, then according to the difference between the first depth semantic vector of each answer in the preset answer library and the question text At least one semantic similarity of the second depth semantic vector is used to obtain the answer to the question text.
  • the first depth semantic vector is a vector used to describe the semantics of the corresponding answer in the preset answer library, and each dimension can describe an attribute of the answer.
  • the second depth semantic vector is a vector used to describe the semantics of the question text.
  • the pre-trained question classification model through the pre-trained question classification model, it is first determined whether the question type of the question text matches the answer type in the preset answer type library. When there is no match, because the at least one semantic similarity characterizes the question text and The semantic consistency of each answer in the answer library is preset, thereby improving the accuracy of obtaining the answer.
  • step 303 determining the answer to the question text according to the at least one semantic similarity can be performed in the following manner:
  • the answer corresponding to the first depth semantic vector of the first semantic similarity in the preset answer database is used as the question text
  • the answer; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • each semantic similarity in at least one semantic similarity is not limited.
  • the first semantic similarity can be represented by the cosine value of the angle between the first depth semantic vector and the second depth semantic vector.
  • the answer corresponding to the first depth semantic vector of the first semantic similarity greater than or equal to the preset semantic similarity threshold in the preset answer database is used as the answer to the question text
  • the answer with the greatest semantic similarity of the question text in the preset answer library is obtained.
  • the question text matches the triples in the pre-built knowledge graph; if so, the The matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text.
  • the question text includes the keyword A1.
  • A1A2A3 the triple matched by A1 in the pre-built knowledge graph
  • A1A2A3 the triple matched by A1 in the pre-built knowledge graph
  • the first semantic similarity when the first semantic similarity is less than the preset semantic similarity threshold, it is determined whether the question text matches the triples in the pre-built knowledge graph, so that the pre-built knowledge graph plays a backup role.
  • the matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text, thereby further improving the acquisition To the accuracy of the answer.
  • the question text does not match the triples in the pre-built knowledge graph, it is determined whether the question text matches the chat sentence in the preset chat library, and if so, then Use the question text and the chat sentence matched in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the answer to the question text.
  • the question text B does not match the triples in the pre-built knowledge graph, it is determined whether B matches the chat sentence in the preset chat library, and if B matches the chat sentence B in the preset chat library If it matches, the chat sentence B is used as the answer to the question text B.
  • the question text and the chat sentence matching the preset chat library are used as the answer to the question text, or the preset default answer is used as the answer to the question text, so that the question text is When it does not match the triples in the pre-built knowledge graph, there are also backup chat sentences and preset default answers, thereby improving user experience and improving answer accuracy.
  • step 303 if the question type of the question text matches the answer type in the preset answer type library, an optional implementation manner is:
  • the key answer element is a combination of all sub-answers of the at least one key answer element to the question text as the answer to the question text.
  • Answer type 1 contains 4 key answer elements: key answer element 1, key answer element 2, key answer element 3, and key answer element 4.
  • Question text C matches key answer element 1, key answer element 2, and key answer element 3 in answer type 1, then sub-answer 1 that matches key answer element 1, key answer element 2, and key answer element 3 respectively according to question text C ,
  • Sub-answer 2, sub-answer 3, the combination of sub-answer 1, sub-answer 2, and sub-answer 3 is used as the answer to question text C.
  • the intelligent robot will also ask whether it needs to obtain the sub-answer of key answer element 4.
  • each key answer element disassembled from at least one key answer element the sub-answer of each key answer element to the question text is determined, so as to obtain the answer to the question text in a more fine-grained manner.
  • the accuracy rate of the answers obtained is determined, so as to obtain the answer to the question text in a more fine-grained manner.
  • the intelligent robot matches the preset answer type library, the intelligent robot uses the corresponding answer type to perform a single or multiple rounds of dialogue to obtain all necessary slot information, then use the slot information to perform the task, and return the result of the task execution to the user ; If it matches the question in the preset answer library, it will be the answer; if the corresponding triple is found in the knowledge graph, the triple will be returned as the answer; finally if the answer is found in the chat library, the chat sentence will be returned as answer.
  • the user asks "what is the interest rate for borrowing 10,000 yuan for 10 days?".
  • the intelligent robot first uses the question classification model to classify the question type of the current question. If the question type can match the answer type in the preset answer type library, the corresponding answer type is used for multiple rounds of conversations to obtain the required answer type All key answer elements are calculated and the result is returned to the user.
  • the element analysis of the question text finds that the amount is 10,000 yuan and the borrowing time is 10 days, then the key answer element "daily interest rate" is missing, and the intelligent robot will ask the user "what is your daily interest rate?" After getting the user's answer and analyzing the value of the daily interest rate, the task will perform the final calculation to obtain the interest amount and return it to the user.
  • the intelligent robot will choose to search in the preset answer library, and the intelligent robot will calculate the depth semantic vector of the user's question and the depth of the candidate answer question in the preset answer library
  • the semantic similarity between the semantic vectors, and the answer with the highest semantic similarity is taken as the result and returned to the user. From the actual effect, the accuracy and recall rate of the answer are greatly improved.
  • the question in the prior art must contain the word “loan” to recall the answer, while the deep semantic vector does not need to be the same as the keyword, as long as the semantics are the same, such as "borrow money” , "Borrowing”, etc., the recall rate has been improved; on the other hand, keywords generally occupy a relatively small proportion of words in a sentence.
  • keywords When using keywords to find answers, the influence of other words in the sentence is not considered, which may lead to large deviations in the searched answers.
  • the deep semantic vector considers the semantic impact of all words in the sentence, so the accuracy of the answers found is higher.
  • the robot will choose to find the answer in the knowledge graph.
  • the triple in the knowledge graph is composed of subject-predication-object (subject-predication-object, SPO) triples can also be other types of triples, which are not limited here.
  • the intelligent robot will extract the key SPO in the text, in order to find one of the missing SPO (subject, predicate and object) three elements in the knowledge graph, or just look up and reason about the possible relationship between SPOs in the question on the way.
  • the chat library If none of the above three steps can find a reasonable answer, enter the chat library to find the chat sentence and return it to the user as the answer. Finally, if no chat sentence is matched in the chat library, the preset default answer is returned to the user. For example, the default answer is: "I'm sorry, but I can't answer your question at this time? I can answer the following questions.".
  • the knowledge base of this application has strong self-learning ability. It is mainly reflected as follows: 1. Perform clustering, classification, filtering and other processing on the questions that cannot be found, and regularly push the obtained question set to the business staff to associate existing knowledge points or add corresponding answers. 2. Cluster all user questions in history, find out the questions similar to the existing questions in the knowledge base and submit them to the business staff for review to supplement the similar questions of the knowledge points, thereby improving the accuracy and recall rate of the answers. 3. Analyze the conversation data after the transfer of labor, mine new knowledge points and push them to the business staff for review, modification, and add to the knowledge base. As the completion rate of the knowledge base increases, the rate of user transfers will be further reduced. 4. Organize existing knowledge points, such as arranging repetitive and similar knowledge points into one place, splitting knowledge points containing multiple topics into multiple knowledge points, etc.
  • FIG. 4 a schematic diagram of a specific step flow provided by an embodiment of the present application.
  • Step 401 Obtain the user's question text.
  • Step 402 Determine whether the call state of the user is a task state.
  • step 406 If yes, go to step 406; otherwise, go to step 403.
  • Step 403 Perform slot analysis to determine whether the slot is full.
  • the slot is the position of questions that can be continuously asked in the current round of conversation. If the slot is full, it means that no additional questions can be asked in this round of conversation.
  • step 404 If yes, go to step 404; otherwise, go to step 405.
  • Step 404 Re-initiate a round of conversation.
  • step 404 After step 404, return to step 401.
  • Step 405 Perform a task to find at least one key answer element corresponding to the answer type.
  • step 405 the sub-answers of the at least one key answer element are combined as the answer to the question text, and step 411 is executed.
  • Step 406 Determine whether the question type of the question text has a matching answer type in the preset answer type library.
  • step 407 If yes, go to step 407; otherwise, go to step 408.
  • Step 407 Set the call state of the user to the task state.
  • step 403 is executed.
  • Step 408 Determine whether the question type has a matching answer in the preset answer library.
  • Step 409 Determine whether the question type has a matching answer in the chat library.
  • step 411 If yes, use the answer as the answer to the question text, and execute step 411; otherwise, execute step 410.
  • Step 410 Use the preset default answer as the answer to the question text, and execute step 411.
  • Step 411 Return the answer to the question text to the user.
  • Step 412 End the call.
  • steps 401 to 411 the user may hang up the call at any time, that is, jump directly to step 412.
  • FIG. 5 a schematic structural diagram of an answer obtaining device provided by an embodiment of this application.
  • This application proposes an answer obtaining device, which includes: an obtaining module 501 for obtaining question text; a processing module 502 for passing A pre-trained question classification model to determine whether the question type of the question text matches the answer type in the preset answer type library; if the question type of the question text does not match the answer type in the preset answer type library , The answer to the question text is obtained according to at least one semantic similarity between the first depth semantic vector of each answer in the preset answer library and the second depth semantic vector of the question text.
  • the processing module 502 is further configured to: if the question type of the question text matches the answer type in the preset answer type library, obtain the question text in the preset answer type library. At least one key answer element corresponding to the matching answer type in the answer type library; according to the question text and the at least one key answer element, it is determined that each key answer element in the at least one key answer element is relative to the question text Sub-answer, a combination of all sub-answers of the at least one key answer element to the question text as the answer to the question text.
  • the processing module 502 is specifically configured to: if the first semantic similarity is greater than or equal to a preset semantic similarity threshold, convert the first depth semantic vector of the first semantic similarity The corresponding answer in the preset answer library is used as the answer to the question text; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • the processing module 502 is further configured to: if the first semantic similarity is less than the preset semantic similarity threshold, determine whether the question text is consistent with that in the pre-built knowledge graph Matching triples; if so, use the matching triples in the question text and the pre-built knowledge graph as the answer to the question text.
  • the processing module 502 is further configured to: if the question text does not match the triples in the pre-built knowledge graph, determine whether the question text matches a preset chat library If yes, the question text and the chat sentence matched in the preset chat library are used as the answer to the question text; or, if not, the preset default answer is used as the question Text answer.
  • the embodiment of the present application provides a computer device, including a program or instruction.
  • the program or instruction When the program or instruction is executed, the following steps are implemented: obtaining a question text; using a pre-trained question classification model to determine whether the question type of the question text is Matches with the answer type in the preset answer type library; if the question type of the question text does not match the answer type in the preset answer type library, then according to the first answer of each answer in the preset answer library At least one semantic similarity between the deep semantic vector and the second deep semantic vector of the question text is used to obtain an answer to the question text.
  • the method further includes: if the question type of the question text matches the answer type in the preset answer type library, obtaining the information that the question text matches in the preset answer type library At least one key answer element corresponding to the answer type; according to the question text and the at least one key answer element, it is determined that each key answer element in the at least one key answer element is a sub-answer of the question text, and the The combination of at least one key answer element to all sub-answers of the question text is used as the answer to the question text.
  • the question text is determined according to at least one semantic similarity between the first depth semantic vector of each answer in the preset answer library and the second depth semantic vector of the question text
  • the answer of includes: if the first semantic similarity is greater than or equal to the preset semantic similarity threshold, then the first depth semantic vector of the first semantic similarity is corresponding to the answer in the preset answer database, As an answer to the question text; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • it further includes: if the first semantic similarity is less than the preset semantic similarity threshold, determining whether the question text matches the triples in the pre-built knowledge graph; if so , The matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text.
  • it further includes: if the question text does not match the triples in the pre-built knowledge graph, determining whether the question text matches the chat sentence in the preset chat library, If yes, use the question text and the chat sentence matched in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the answer to the question text.
  • the embodiment of the present application provides a storage medium, including a program or instruction.
  • the program or instruction When the program or instruction is executed, the following steps are implemented: obtaining a question text; using a pre-trained question classification model to determine whether the question type of the question text is Matches with the answer type in the preset answer type library; if the question type of the question text does not match the answer type in the preset answer type library, then according to the first answer of each answer in the preset answer library At least one semantic similarity between the deep semantic vector and the second deep semantic vector of the question text is used to obtain an answer to the question text.
  • the method further includes: if the question type of the question text matches the answer type in the preset answer type library, obtaining the information that the question text matches in the preset answer type library At least one key answer element corresponding to the answer type; according to the question text and the at least one key answer element, it is determined that each key answer element in the at least one key answer element is a sub-answer of the question text, and the The combination of at least one key answer element to all sub-answers of the question text is used as the answer to the question text.
  • the question text is determined according to at least one semantic similarity between the first depth semantic vector of each answer in the preset answer library and the second depth semantic vector of the question text
  • the answer of includes: if the first semantic similarity is greater than or equal to the preset semantic similarity threshold, then the first depth semantic vector of the first semantic similarity is corresponding to the answer in the preset answer database, As an answer to the question text; the first semantic similarity is the largest semantic similarity among the at least one semantic similarity.
  • it further includes: if the first semantic similarity is less than the preset semantic similarity threshold, determining whether the question text matches the triples in the pre-built knowledge graph; if so , The matching triples in the question text and the pre-built knowledge graph are used as the answer to the question text.
  • it further includes: if the question text does not match the triples in the pre-built knowledge graph, determining whether the question text matches the chat sentence in the preset chat library, If yes, use the question text and the chat sentence matched in the preset chat library as the answer to the question text; or, if not, use the preset default answer as the answer to the question text.
  • the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program codes.
  • a computer-usable storage media including but not limited to disk storage, optical storage, etc.
  • These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

Abstract

一种答案获取方法及装置，其中方法为：获取提问文本(301)；通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配(302)；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案(303)。上述方法应用于金融科技时，通过预训练的提问分类模型，首先确定提问文本的提问类型是否与预设答案类型库中的答案类型匹配，在不匹配时，由于所述至少一个语义相似度表征了提问文本与预设答案库中每个答案的语义吻合程度，从而提升了获取到答案的准确率。

Description

一种答案获取方法及装置
相关申请的交叉引用
本申请要求在2019年06月27日提交中国专利局、申请号为201910570444.1、申请名称为“一种答案获取方法及装置”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
本发明涉及金融科技(Fintech)领域和智能通信领域，尤其涉及一种答案获取方法及装置。
背景技术
随着计算机技术的发展，越来越多的技术(大数据、分布式、区块链(Blockchain)、人工智能等)应用在金融领域，传统金融业正在逐步向金融科技(Fintech)转变。目前，金融科技领域中的智能通信过程中，经常需要根据用户的提问，向用户返回问题的答案。
目前电话呼入系统中获取用户问题答案的方式为，根据提问文本的关键词在人工收集的答案库中进行搜索，并将命中的答案反馈给用户。显然，人工收集的答案库有限，且效率低下，而且根据关键词进行搜索获取到的答案，难以保证获取到的答案语义为预期语义，从而导致获取到的用户答案也不够准确。
发明内容
本申请实施例提供一种答案获取方法及装置，解决了现有技术中获取到的用户答案不够准确的问题。
第一方面，本申请实施例提供一种答案获取方法，包括：获取提问文本；通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案 类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
上述方法中，通过预训练的提问分类模型，首先确定提问文本的提问类型是否与预设答案类型库中的答案类型匹配，在不匹配时，由于所述至少一个语义相似度表征了提问文本与预设答案库中每个答案的语义吻合程度，从而提升了获取到答案的准确率。
一种可选实施方式中，若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
上述方法中，根据至少一个关键答案要素拆解出来的每个关键答案要素，确定每个关键答案要素对于所述提问文本的子答案，从而更细粒度地获取所述提问文本的答案，提升了获取到答案的准确率。
一种可选实施方式中，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：
若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
通过设置预设语义相似度阈值，将大于或等于预设语义相似度阈值的第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案，在保证语义相似度的前提下，获取提问文本在所述预设 答案库中语义相似度最大的答案。
一种可选实施方式中，若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
上述方法中，在第一语义相似度小于所述预设语义相似度阈值时，确定提问文本是否与预构建的知识图谱中的三元组匹配，从而预构建的知识图谱起到了备份作用，用于第一语义相似度小于所述预设语义相似度阈值时，将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案，从而进一步提升了获取到答案的准确率。
一种可选实施方式中，若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
上述方法中，将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案，或者将预设默认答案作为所述提问文本的答案，从而在所述提问文本不与所述预构建的知识图谱中的三元组匹配时，还有备份的聊天语句以及预设默认答案，从而提高了用户感受，并提升了答案准确率。
第二方面，本申请提供一种答案获取装置，包括：获取模块，用于获取提问文本；处理模块，用于通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
一种可选实施方式中，所述处理模块还用于：若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提 问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
一种可选实施方式中，所述处理模块具体用于：若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
一种可选实施方式中，所述处理模块还用于：若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
一种可选实施方式中，所述处理模块还用于：若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
上述第二方面及第二方面各个实施例的有益效果，可以参考上述第一方面及第一方面各个实施例的有益效果，这里不再赘述。
第三方面，本申请实施例提供一种计算机设备，包括程序或指令，当所述程序或指令被执行时，实现如下步骤：获取提问文本；通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述 至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
一种可选实施方式中，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
一种可选实施方式中，还包括：若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
第四方面，本申请实施例提供一种存储介质，包括程序或指令，当所述程序或指令被执行时，实现如下步骤：获取提问文本；通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述 至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
一种可选实施方式中，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
一种可选实施方式中，还包括：若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
附图说明
图1为本申请实施例提供的一种答案获取方法可应用的电话呼入智能系统的架构示意图；
图2为本申请实施例提供的一种答案获取方法的时序交互示意图；
图3为本申请实施例提供的一种答案获取方法的步骤流程示意图；
图4为本申请实施例提供的一种答案获取方法的具体步骤流程示意图；
图5为本申请实施例提供的一种答案获取装置的结构示意图。
具体实施方式
为了更好的理解上述技术方案，下面将结合说明书附图及具体的实施方式对上述技术方案进行详细的说明，应当理解本申请实施例以及实施例中的具体特征是对本申请技术方案的详细的说明，而不是对本申请技术方案的限定，在不冲突的情况下，本申请实施例以及实施例中的技术特征可以相互结合。
在电信运营业务的运转过程中，用户经常会碰到各种各样的问题，需要解答。电话呼入系统是一个用于用户提问并给用户解答的系统，电话呼入系统即支持用户电话呼入并提供问题答疑指引的系统。
目前电话呼入系统中获取用户问题答案的方式会导致获取到的用户答案不够准确，为此本申请实施例提供一种答案获取方法。
如图1所示，为本申请实施例提供的一种答案获取方法可应用的电话呼入智能系统的架构示意图。
该系统架构包括以下部分：
终端设备：终端设备为客户用于拨打电话提问的设备，发起来电。
电信运营商：电信运营商用于转接终端的来电。
公司网络：公司网络用于将电信运营商转接的来电，并将来电再转接给智能机器人。
坐席话务服务平台：坐席话务服务平台包括多个人工座席，用于人工接管来电。
智能机器人：智能机器人用于接听用户来电，将通话录音转化成文本，或将文本转化为通话录音，返回用户来电提问的答案。
基于上述架构，可以通过以下流程完成答案获取。
S101：用户通过终端设备拨打客服坐席电话。
S102：来电经过电信运营商和公司网络转接，并由智能机器人自动接听来电。
S103：智能机器人将交互文本通过语音合成引擎转换成语音播放给用户， 引导用户提问。
S104：智能机器人将用户通话录音通过语音识别引擎转换成提问文本。
S105：智能客服机器人通过提问文本，获取答案，然后通过语音合成引擎转换成语音播放给用户。
S106：重复S103-S105，若用户主动挂掉电话则通话结束库，若用户语音输入“转人工”则由人工坐席托管通话，直到用户挂掉电话，通话结束。
下面结合图2详细介绍上述系统架构，如图2所示，为本申请实施例提供的一种答案获取方法的时序交互示意图。包括以下步骤：
步骤201：拨打电话。
用户通过终端设备拨打电话，产生来电，此时终端设备的通话状态为呼出状态。
步骤202：通知电话接入。
电信运营商通知智能机器人来电。
步骤203：发送接通电话命令。
智能机器人接收到来电后，向电信运营商发送接通电话命令。
步骤204：接听电话。
电信运营商接通电话，此时终端设备的通话状态为接通状态。
步骤205：录制通话语音，获得通话录音。
电信运营商接听用户的语音，并获得通话录音。
步骤206：发送通话录音。
电信运营商将转换的通话录音发送至语音识别引擎。
步骤207：将通话录音转化为提问文本。
语音识别引擎将通话录音转化为提问文本。
步骤208：返回响应提问文本的答案文本。
智能机器人获取提问文本的答案文本，并返回至语音合成引擎。
步骤209：发送答案语音。
语音合成引擎答案文本转化为答案语音，并将答案语音发送至电信运营 商。
步骤2010：发送答案语音。
电信运营商发送答案语音至终端设备。
若智能机器人未接收到转人工命令且用户未挂断电话，则循环执行步骤205～步骤2010。
若智能机器人在步骤207时接收到的问题文本为“转人工”，则执行步骤2011～步骤2012。
步骤2011：发送转人工命令。
智能机器人向电信运营商发送转人工命令。
步骤2012：接入人工服务。
电信运营商将终端设备与坐席话务服务平台连接通话，从而实现接入人工服务。
步骤2013：挂断电话。
需要说明的是，步骤2013可能在步骤201～步骤2012中任何两个相邻步骤之间执行。
本申请中使用智能机器人自动接听电话，并可以对用户特定任务型提问进行多轮对话交互，对非任务型提问进行深度语义分析并查找答案，同时也支持用户转人工坐席服务。此方案可在智能客服机器人服务和坐席服务之间取得良好的平衡，解决了上面提到的缺陷，带来了巨大的商业收益。
智能机器人是使用自然语言处理技术、机器学习与深度学习框架等来实现的自动问答系统，可与用户进行多轮会话并深度语义理解用户的提问意图，通过执行任务或查找知识库获得用户所需要的答案。使用的技术包括意图理解、语义匹配、搜索引擎、推荐引擎、任务引擎、知识图谱等。
本申请的知识库组成如表1所示：
问题 答案 知识点产生方式
提问文本 预设答案类型 机器自学习辅助
提问文本 预设聊天库 机器自学习辅助
提问文本 知识图谱 定期编辑，内容相对稳定
提问文本 预设聊天库 定期编辑，内容相对稳定
表1
需要说明的是，预设答案类型和预设聊天库是通过机器自学习辅助产生的，当提问文本在预设答案类型和预设聊天库匹配不到内容时，则传递到业务系统，由业务人员确认若属于已有的内容，未匹配准确，则将该提问文本归为已有内容一类，若不属于已有的内容，则将编辑新内容，加入到相应知识库中。从而不断积累成预设答案类型和预设聊天库，从而不断扩充内容，并不断将相似问题归为一类。
本方案知识库的问题部分只有用户提问文本组成，而不是关键词和模式，符合人类自然对话习惯，极大降低知识库的编辑难度。答案部分则包含预设答案类型、预设答案库、预构建的知识图谱和聊天库，更有力的支持了对话的智能化程度提升、答案的丰富度提升。
如图3所示，为本申请实施例提供的一种答案获取方法的步骤流程示意图。
步骤301：获取提问文本。
其中，提问文本是用户录音转化的文本。
步骤302：通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配。
与预设答案类型库中的答案类型匹配的提问类型，是需要智能机器人运算得出具体答案的提问类型，一般不直接在预设答案库中给出具体答案，而是先通过预设答案类型库，确定答案类型，再根据提问文本中的具体信息，确定提问文本的答案。
步骤303：若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答 案。
第一深度语义向量为用于描述预设答案库中相应答案语义的一个向量，每个维度可描述该答案的一项属性。相应地，第二深度语义向量为用于描述提问文本语义的一个向量。
上述方法中，通过预训练的提问分类模型，首先确定提问文本的提问类型是否与预设答案类型库中的答案类型匹配，在不匹配时，由于所述至少一个语义相似度表征了提问文本与预设答案库中每个答案的语义吻合程度，从而提升了获取到答案的准确率。
步骤303中，根据所述至少一个语义相似度，确定所述提问文本的答案，可以按照以下方式进行：
若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
需要说明的是，至少一个语义相似度中每个语义相似度的描述方式不做限定。举例来说，用第一语义相似度可用第一深度语义向量与第二深度语义向量的夹角余弦值来表示。
通过设置预设语义相似度阈值，将大于或等于预设语义相似度阈值的第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案，在保证语义相似度的前提下，获取提问文本在所述预设答案库中语义相似度最大的答案。
在上述可选实施方式中，若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
举例来说，提问文本中包括关键词A1。其中，A1在预构建的知识图谱中匹配的三元组为A1A2A3，则返回A1A2A3。
上述方法中，在第一语义相似度小于所述预设语义相似度阈值时，确定提问文本是否与预构建的知识图谱中的三元组匹配，从而预构建的知识图谱起到了备份作用，用于第一语义相似度小于所述预设语义相似度阈值时，将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案，从而进一步提升了获取到答案的准确率。
在上述可选实施方式中，若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
举例来说，提问文本B不与所述预构建的知识图谱中的三元组匹配，则确定B是否与预设聊天库中的聊天语句匹配，如果B与预设聊天库中的聊天语句B匹配，则将聊天语句B作为提问文本B的答案。
上述方法中，将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案，或者将预设默认答案作为所述提问文本的答案，从而在所述提问文本不与所述预构建的知识图谱中的三元组匹配时，还有备份的聊天语句以及预设默认答案，从而提高了用户感受，并提升了答案准确率。
步骤303中，若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则一种可选实施方式为：
获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
举例来说，若提问文本C与答案类型“答案类型1”匹配。答案类型1包含关键答案要素1、关键答案要素2、关键答案要素3、关键答案要素4这4个关键答案要素。提问文本C与答案类型1中关键答案要素1、关键答案要素2、关键答案要素3匹配，则根据提问文本C分别与关键答案要素1、关键 答案要素2、关键答案要素3匹配的子答案1、子答案2、子答案3，将子答案1、子答案2、子答案3的组合作为提问文本C的答案。而且，在返回提问文本C的答案后，智能机器人还会反问是否需要获取关键答案要素4的子答案。
上述方法中，根据至少一个关键答案要素拆解出来的每个关键答案要素，确定每个关键答案要素对于所述提问文本的子答案，从而更细粒度地获取所述提问文本的答案，提升了获取到答案的准确率。
下面以一个具体实施例来说明。
智能机器人如果匹配到预设答案类型库则智能机器人使用对应的答案类型则执行单轮或多轮对话获得所有必要槽位信息，然后使用槽位信息执行任务，并将任务执行的结果返回给用户；如果匹配到预设答案库的问题则，作为答案；如果在知识图谱中查找到对应三元组则返回三元组，作为答案；最后如果在聊天库中查找到答案则返回聊天语句，作为答案。
用户提问“借款10000元10天的话利息是多少？”。
智能机器人首先使用提问分类模型对当前问题进行提问类型分类，如果该提问类型能在预设答案类型库中匹配到答案类型，则使用对应的答案类型进行多轮会话，获得该答案类型所需的所有关键答案要素，然后进行计算得到结果返回给用户。在此例中，假如对提问文本进行要素分析发现了金额是10000元，借款时间是10天，那么还缺少关键答案要素“日利率”，智能机器人会反问用户“您的日利率是多少？”，在得到用户回答并解析出日利率的值后，该任务将执行最后的计算得到利息数额，并返回给用户。
如果预设答案类型库中并没有对应的答案类型，则智能机器人将选择在预设答案库中进行查找，智能机器人将计算用户提问的深度语义向量与预设答案库中候选答案的问题的深度语义向量之间的语义相似度，并取语义相似度最高的答案作为结果返回给用户，从实际效果上看答案的准确率和召回率大幅提升。举例来说如果关键词“贷款”作为，那么现有技术中提问必须包含“贷款”两字才能召回答案，而深度语义向量则不需要关键词一样，只要 语义一样即可，比如“借钱”、“借款”等，召回率得到提升；另一方面关键词一般在句子中字的占比较小，使用关键词查找答案的时候由于没有考虑句子间其他词影响，可能导致查找的答案偏差较大，而深度语义向量则是考虑了句子中所有词的语义影响，因此查找到的答案准确率也更高。
如果预设答案库中没有此提问文本的答案，则机器人将选择在知识图谱中查找答案，举例来说，知识图谱中的三元组是，由主语-谓语-宾语(subject-predication-object,SPO)三元组，还可以是其他类型三元组，在此不做限定。智能机器人将提取文本中关键的SPO，以期能在知识图谱中查找到缺失的SPO(主谓宾)三要素之一，或在只是途中查找、推理提问中SPO间可能存在的关系。作为对预设答案类型库返回的关键答案要素组合，或者预设答案库中返回的答案补充。
如果上述三个步骤都不能找到合理答案，则进入聊天库中查找聊天语句返回给用户，作为答案。最后，如果聊天库中没有匹配到聊天语句的话，则返回预设默认答案给用户。比如预设默认答案是：”很抱歉，目前无法回答您提出的问题？我可以回答以下问题...”。
通过上述多层次的知识库、多轮交互机制、深度语义分析模型可很好的解决现有技术存在的问题，极大提升回答问题的准确率和召回率，极大提升对话交互中用户体验。
除此之外，本申请的知识库拥有较强的自学习能力。主要体现为：1、对查找不到答案的提问进行聚类、分类、过滤等处理，并将得到的问题集定期推送给业务人员去关联已有知识点或添加对应答案。2、对历史上所有的用户提问进行聚类，查找到与知识库中已有问题相似的问法并提交给业务人员审核以便补充知识点的相似问题，从而提升答案的准确率和召回率。3、对转人工后的对话数据进行分析，挖掘新的知识点并推送给业务人员审核、修改并添加到知识库中，随着知识库的完备率提升，用户转人工率将进一步降低。4、对存在的知识点进行规整，比如将重复和相似的知识点规整到一处，将包含多个主题的知识点拆分成多个知识点等。
下面结合图4，详细介绍本申请实施例提供的一种答案获取方法，如图4所示，为本申请实施例提供的具体步骤流程示意图。
步骤401：获取用户的提问文本。
步骤402：确定用户的通话状态是否为任务状态。
若用户已预设答案类型库中有对应的答案类型，且正在针对该答案类型通话，则确定用户在任务状态。
若是，则执行步骤406；否则执行步骤403。
步骤403：进行槽位分析，确定槽位是否已满。
槽位即当前一轮会话可连续提问的问题的位置，若槽位已满则说明在该轮会话中已不能额外提问。
若是，则执行步骤404；否则执行步骤405。
步骤404：重新发起一轮会话。
步骤404之后，返回步骤401。
步骤405：执行任务，查找该答案类型中对应的至少一个关键答案要素。
步骤405之后，将该至少一个关键答案要素的子答案组合，作为提问文本的答案，并执行步骤411。
步骤406：确定提问文本的提问类型是否在预设答案类型库中有匹配的答案类型。
若是，则执行步骤407；否则，执行步骤408。
步骤407：将用户的通话状态设置为任务状态。
步骤407之后，执行步骤403。
步骤408：确定提问类型是否在预设答案库中有匹配的答案。
若有，则将该答案作为提问文本的答案，并执行步骤411；否则，执行步骤409。
步骤409：确定提问类型是否在聊天库中有匹配的答案。
若是，则将该答案作为提问文本的答案，并执行步骤411；否则，执行步骤410。
步骤410：将预设默认答案作为提问文本的答案，并执行步骤411。
步骤411：将提问文本的答案返回用户。
步骤412：结束通话。
需要说明的是，步骤401～步骤411中，用户可能随时挂掉电话，即直接跳转到步骤412。
从中可以看出，在智能机器人答案查找方面本申请使用的技术更先进、层次更丰富、流程更合理，极大提高了答案返回的准确率和召回率，降低了转人工的概率，提升用户体验。
如图5所示，为本申请实施例提供的一种答案获取装置的结构示意图，本申请提出一种答案获取装置，包括：获取模块501，用于获取提问文本；处理模块502，用于通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
一种可选实施方式中，所述处理模块502还用于：若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
一种可选实施方式中，所述处理模块502具体用于：若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
一种可选实施方式中，所述处理模块502还用于：若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识 图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
一种可选实施方式中，所述处理模块502还用于：若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
本申请实施例提供一种计算机设备，包括程序或指令，当所述程序或指令被执行时，实现如下步骤：获取提问文本；通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
一种可选实施方式中，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
一种可选实施方式中，还包括：若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹 配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
本申请实施例提供一种存储介质，包括程序或指令，当所述程序或指令被执行时，实现如下步骤：获取提问文本；通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
一种可选实施方式中，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
一种可选实施方式中，还包括：若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹 配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
一种可选实施方式中，还包括：若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，若否，则将预设默认答案作为所述提问文本的答案。
最后应说明的是：本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。

Claims (20)

  1. 一种答案获取方法，其特征在于，包括：
    获取提问文本；
    通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；
    若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
  2. 如权利要求1所述的方法，其特征在于，还包括：
    若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；
    根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
  3. 如权利要求1或2所述的方法，其特征在于，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：
    若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
  4. 如权利要求3所述的方法，其特征在于，还包括：
    若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与 所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
  5. 如权利要求4所述的方法，其特征在于，还包括：
    若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，
    若否，则将预设默认答案作为所述提问文本的答案。
  6. 一种答案获取装置，其特征在于，包括：
    获取模块，用于获取提问文本；
    处理模块，用于通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
  7. 如权利要求6所述的装置，其特征在于，所述处理模块还用于：
    若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；
    根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
  8. 如权利要求6或7所述的装置，其特征在于，所述处理模块具体用于：
    若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
  9. 如权利要求8所述的装置，其特征在于，所述处理模块还用于：
    若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
  10. 如权利要求9所述的装置，其特征在于，所述处理模块还用于：
    若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，
    若否，则将预设默认答案作为所述提问文本的答案。
  11. 一种计算机设备，其特征在于，包括程序或指令，当所述程序或指令被执行时，实现如下步骤：
    获取提问文本；
    通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；
    若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
  12. 如权利要求11所述的计算机设备，其特征在于，还包括：
    若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；
    根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
  13. 如权利要求11或12所述的计算机设备，其特征在于，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：
    若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
  14. 如权利要求13所述的计算机设备，其特征在于，还包括：
    若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
  15. 如权利要求14所述的计算机设备，其特征在于，还包括：
    若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，
    若否，则将预设默认答案作为所述提问文本的答案。
  16. 一种存储介质，其特征在于，包括程序或指令，当所述程序或指令被执行时，实现如下步骤：
    获取提问文本；
    通过预训练的提问分类模型，确定所述提问文本的提问类型是否与预设答案类型库中的答案类型匹配；
    若所述提问文本的提问类型不与所述预设答案类型库中的答案类型匹配，则根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，获取所述提问文本的答案。
  17. 如权利要求16所述的存储介质，其特征在于，还包括：
    若所述提问文本的提问类型与所述预设答案类型库中的答案类型匹配，则获取所述提问文本在所述预设答案类型库中匹配的答案类型对应的至少一个关键答案要素；
    根据所述提问文本和所述至少一个关键答案要素，确定所述至少一个关键答案要素中每个关键答案要素对于所述提问文本的子答案，将所述至少一 个关键答案要素对于所述提问文本的所有子答案的组合，作为所述提问文本的答案。
  18. 如权利要求16或17所述的存储介质，其特征在于，所述根据所述预设答案库中每个答案的第一深度语义向量与所述提问文本的第二深度语义向量的至少一个语义相似度，确定所述提问文本的答案，包括：
    若所述第一语义相似度大于或等于预设语义相似度阈值，则将所述第一语义相似度的第一深度语义向量在所述预设答案库中对应的答案，作为所述提问文本的答案；所述第一语义相似度为所述至少一个语义相似度中最大的语义相似度。
  19. 如权利要求18所述的存储介质，其特征在于，还包括：
    若所述第一语义相似度小于所述预设语义相似度阈值，则确定所述提问文本是否与预构建的知识图谱中的三元组匹配；若是，则将所述提问文本与所述预构建的知识图谱中匹配的三元组，作为所述提问文本的答案。
  20. 如权利要求19所述的存储介质，其特征在于，还包括：
    若所述提问文本不与所述预构建的知识图谱中的三元组匹配，则确定所述提问文本是否与预设聊天库中的聊天语句匹配，若是，则将所述提问文本与所述预设聊天库中匹配的聊天语句，作为所述提问文本的答案；或者，
    若否，则将预设默认答案作为所述提问文本的答案。
PCT/CN2019/117517 2019-06-27 2019-11-12 一种答案获取方法及装置 WO2020258654A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910570444.1A CN110263144A (zh) 2019-06-27 2019-06-27 一种答案获取方法及装置
CN201910570444.1 2019-06-27

Publications (1)

Publication Number Publication Date
WO2020258654A1 true WO2020258654A1 (zh) 2020-12-30

Family

ID=67922619

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/117517 WO2020258654A1 (zh) 2019-06-27 2019-11-12 一种答案获取方法及装置

Country Status (2)

Country Link
CN (1) CN110263144A (zh)
WO (1) WO2020258654A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112818222A (zh) * 2021-01-26 2021-05-18 吾征智能技术(北京)有限公司 一种基于知识图谱的个性化饮食推荐方法及系统
CN113032540A (zh) * 2021-03-19 2021-06-25 北京百度网讯科技有限公司 人机交互方法、装置、设备和存储介质
CN113688252A (zh) * 2021-08-09 2021-11-23 广西师范大学 基于多特征协同知识图谱和区块链的安全跨域推荐方法
CN113792153A (zh) * 2021-08-25 2021-12-14 北京百度网讯科技有限公司 问答推荐方法及其装置
TWI790069B (zh) * 2021-12-28 2023-01-11 高雄榮民總醫院 文獻評讀學習系統

Families Citing this family (35)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8977255B2 (en) 2007-04-03 2015-03-10 Apple Inc. Method and system for operating a multi-function portable electronic device using voice-activation
US8676904B2 (en) 2008-10-02 2014-03-18 Apple Inc. Electronic devices with voice command and contextual data processing capabilities
CN104969289B (zh) 2013-02-07 2021-05-28 苹果公司 数字助理的语音触发器
US10170123B2 (en) 2014-05-30 2019-01-01 Apple Inc. Intelligent assistant for home automation
US9338493B2 (en) 2014-06-30 2016-05-10 Apple Inc. Intelligent automated assistant for TV user interactions
US10747498B2 (en) 2015-09-08 2020-08-18 Apple Inc. Zero latency digital assistant
US10691473B2 (en) 2015-11-06 2020-06-23 Apple Inc. Intelligent automated assistant in a messaging environment
DK201670540A1 (en) 2016-06-11 2018-01-08 Apple Inc Application integration with a digital assistant
DK180048B1 (en) 2017-05-11 2020-02-04 Apple Inc. MAINTAINING THE DATA PROTECTION OF PERSONAL INFORMATION
DK201770427A1 (en) 2017-05-12 2018-12-20 Apple Inc. LOW-LATENCY INTELLIGENT AUTOMATED ASSISTANT
DK179496B1 (en) 2017-05-12 2019-01-15 Apple Inc. USER-SPECIFIC Acoustic Models
US11145294B2 (en) 2018-05-07 2021-10-12 Apple Inc. Intelligent automated assistant for delivering content from user experiences
US10928918B2 (en) 2018-05-07 2021-02-23 Apple Inc. Raise to speak
DK180639B1 (en) 2018-06-01 2021-11-04 Apple Inc DISABILITY OF ATTENTION-ATTENTIVE VIRTUAL ASSISTANT
US11462215B2 (en) 2018-09-28 2022-10-04 Apple Inc. Multi-modal inputs for voice commands
US11348573B2 (en) 2019-03-18 2022-05-31 Apple Inc. Multimodality in digital assistant systems
US11307752B2 (en) 2019-05-06 2022-04-19 Apple Inc. User configurable task triggers
DK201970509A1 (en) 2019-05-06 2021-01-15 Apple Inc Spoken notifications
US11140099B2 (en) 2019-05-21 2021-10-05 Apple Inc. Providing message response suggestions
US11227599B2 (en) 2019-06-01 2022-01-18 Apple Inc. Methods and user interfaces for voice-based control of electronic devices
CN110263144A (zh) * 2019-06-27 2019-09-20 深圳前海微众银行股份有限公司 一种答案获取方法及装置
CN110837549B (zh) * 2019-11-06 2023-08-11 腾讯科技（深圳）有限公司 一种信息处理方法、装置及存储介质
CN111444724B (zh) * 2020-03-23 2022-11-15 腾讯科技（深圳）有限公司 医疗问答对质检方法、装置、计算机设备和存储介质
US11061543B1 (en) 2020-05-11 2021-07-13 Apple Inc. Providing relevant data items based on context
US11610065B2 (en) 2020-06-12 2023-03-21 Apple Inc. Providing personalized responses based on semantic context
WO2021252827A1 (en) * 2020-06-12 2021-12-16 Apple Inc. Providing personalized responses based on semantic context
CN113807148A (zh) * 2020-06-16 2021-12-17 阿里巴巴集团控股有限公司 文本识别匹配方法和装置、终端设备
US11490204B2 (en) 2020-07-20 2022-11-01 Apple Inc. Multi-device audio adjustment coordination
US11438683B2 (en) 2020-07-21 2022-09-06 Apple Inc. User identification using headphones
CN111931514B (zh) * 2020-07-28 2021-03-16 链博(成都)科技有限公司 基于深度学习和大数据的信息处理方法及区块链服务平台
CN111966840A (zh) * 2020-08-18 2020-11-20 北京猿力未来科技有限公司 一种语言教学的人机交互管理方法和管理系统
CN112084315A (zh) * 2020-09-07 2020-12-15 腾讯科技（深圳）有限公司 问答互动方法、装置、存储介质及设备
CN112163083B (zh) * 2020-10-20 2023-11-03 珠海格力电器股份有限公司 智能问答方法、装置、电子设备和存储介质
CN112287085B (zh) * 2020-11-06 2023-12-05 中国平安财产保险股份有限公司 语义匹配方法、系统、设备及存储介质
CN116886656B (zh) * 2023-09-06 2023-12-08 北京小糖科技有限责任公司 面向聊天室的舞蹈知识推送方法及其装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107133349A (zh) * 2017-05-24 2017-09-05 北京无忧创新科技有限公司 一种对话机器人系统
CN107180029A (zh) * 2016-03-09 2017-09-19 阿里巴巴集团控股有限公司 一种基于咨询业务的信息处理方法及装置
CN108073576A (zh) * 2016-11-09 2018-05-25 上海诺悦智能科技有限公司 智能搜索方法、搜索装置以及搜索引擎系统
CN108197167A (zh) * 2017-12-18 2018-06-22 深圳前海微众银行股份有限公司 人机对话处理方法、设备及可读存储介质
CN110263144A (zh) * 2019-06-27 2019-09-20 深圳前海微众银行股份有限公司 一种答案获取方法及装置

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107220380A (zh) * 2017-06-27 2017-09-29 北京百度网讯科技有限公司 基于人工智能的问答推荐方法、装置和计算机设备
CN109740077B (zh) * 2018-12-29 2021-02-12 北京百度网讯科技有限公司 基于语义索引的答案搜索方法、装置及其相关设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107180029A (zh) * 2016-03-09 2017-09-19 阿里巴巴集团控股有限公司 一种基于咨询业务的信息处理方法及装置
CN108073576A (zh) * 2016-11-09 2018-05-25 上海诺悦智能科技有限公司 智能搜索方法、搜索装置以及搜索引擎系统
CN107133349A (zh) * 2017-05-24 2017-09-05 北京无忧创新科技有限公司 一种对话机器人系统
CN108197167A (zh) * 2017-12-18 2018-06-22 深圳前海微众银行股份有限公司 人机对话处理方法、设备及可读存储介质
CN110263144A (zh) * 2019-06-27 2019-09-20 深圳前海微众银行股份有限公司 一种答案获取方法及装置

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112818222A (zh) * 2021-01-26 2021-05-18 吾征智能技术(北京)有限公司 一种基于知识图谱的个性化饮食推荐方法及系统
CN112818222B (zh) * 2021-01-26 2024-02-23 吾征智能技术(北京)有限公司 一种基于知识图谱的个性化饮食推荐方法及系统
CN113032540A (zh) * 2021-03-19 2021-06-25 北京百度网讯科技有限公司 人机交互方法、装置、设备和存储介质
CN113032540B (zh) * 2021-03-19 2023-06-23 北京百度网讯科技有限公司 人机交互方法、装置、设备和存储介质
CN113688252A (zh) * 2021-08-09 2021-11-23 广西师范大学 基于多特征协同知识图谱和区块链的安全跨域推荐方法
CN113688252B (zh) * 2021-08-09 2023-09-26 广西师范大学 基于多特征协同知识图谱和区块链的安全跨域推荐方法
CN113792153A (zh) * 2021-08-25 2021-12-14 北京百度网讯科技有限公司 问答推荐方法及其装置
CN113792153B (zh) * 2021-08-25 2023-12-12 北京度商软件技术有限公司 问答推荐方法及其装置
TWI790069B (zh) * 2021-12-28 2023-01-11 高雄榮民總醫院 文獻評讀學習系統

Also Published As

Publication number Publication date
CN110263144A (zh) 2019-09-20

Similar Documents

Publication Publication Date Title
WO2020258654A1 (zh) 一种答案获取方法及装置
US11380327B2 (en) Speech communication system and method with human-machine coordination
CN111866022B (zh) 感知质量比会议中原始听到的更高的后会议回放系统
CN104598445B (zh) 自动问答系统和方法
US20140081643A1 (en) System and method for determining expertise through speech analytics
TW201935273A (zh) 語句的使用者意圖識別方法和裝置
CN110059182A (zh) 面向客服的话术推荐方法和装置
CN106202301A (zh) 一种基于深度学习的智能应答系统
US10956480B2 (en) System and method for generating dialogue graphs
CN110310663A (zh) 违规话术检测方法、装置、设备及计算机可读存储介质
US20170032027A1 (en) Contact Center Virtual Assistant
CN112487810B (zh) 一种智能客服服务方法、装置、设备及存储介质
CN113282737B (zh) 人机协作的智能客服对话方法及装置
CN110008308B (zh) 针对用户问句补充信息的方法和装置
CN117441165A (zh) 减少生成语言模型的偏差
JP2012003704A (ja) Ｆａｑ候補抽出システムおよびｆａｑ候補抽出プログラム
CN112015879B (zh) 基于文本结构化管理的人机交互引擎实现方法及装置
CN113779217A (zh) 一种基于人机交互的智能语音外呼业务方法及系统
CN109726002B (zh) 一种处理流程调整方法及装置
WO2023090380A1 (ja) プログラム、情報処理システム及び情報処理方法
CN111324719B (zh) 用于法律咨询的模糊识别系统
CN111858851A (zh) 一种智能客服知识库多维度训练方法及装置
CN113744741B (zh) 识别电话接听行为的方法及装置
JP7169030B1 (ja) プログラム、情報処理装置、情報処理システム、情報処理方法、情報処理端末
CN113297361B (zh) 一种基于可视化流程图的智能问答交互系统及其方法

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: 19935077

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 22.04.2022)

122 Ep: pct application non-entry in european phase

Ref document number: 19935077

Country of ref document: EP

Kind code of ref document: A1