WO2019114512A1 - 用于客户服务的方法、装置、电子设备、计算机可读存储介质 - Google Patents

用于客户服务的方法、装置、电子设备、计算机可读存储介质 Download PDF

Info

Publication number
WO2019114512A1
WO2019114512A1 PCT/CN2018/116820 CN2018116820W WO2019114512A1 WO 2019114512 A1 WO2019114512 A1 WO 2019114512A1 CN 2018116820 W CN2018116820 W CN 2018116820W WO 2019114512 A1 WO2019114512 A1 WO 2019114512A1
Authority
WO
WIPO (PCT)
Prior art keywords
answer
reply
auxiliary information
question
customer
Prior art date
Application number
PCT/CN2018/116820
Other languages
English (en)
French (fr)
Inventor
李岚
Original Assignee
株式会社日立制作所
李岚
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 株式会社日立制作所, 李岚 filed Critical 株式会社日立制作所
Publication of WO2019114512A1 publication Critical patent/WO2019114512A1/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
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/01—Customer relationship services
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L51/00—User-to-user messaging in packet-switching networks, transmitted according to store-and-forward or real-time protocols, e.g. e-mail
    • H04L51/02—User-to-user messaging in packet-switching networks, transmitted according to store-and-forward or real-time protocols, e.g. e-mail using automatic reactions or user delegation, e.g. automatic replies or chatbot-generated messages

Definitions

  • the invention relates to the field of intelligent customer service technology, in particular to a method, device, electronic device and computer readable storage medium for customer service.
  • the existing customer service system has the following problems to be solved: first, how to reduce the workload of manual customer service; second, how to improve the quality of problem response.
  • the technical problem to be solved by the present invention is to provide a method, device, electronic device, and computer readable storage medium for customer service, which can reduce the workload of the manual customer service and improve the quality of the problem response.
  • an embodiment of the present invention provides the following technical solutions:
  • a method for customer service including:
  • the candidate reply answer with the highest confidence is returned to the customer as the final reply text; if the confidence answer of the highest answer is the highest If the degree is less than the first threshold, the service request for the manual customer service access is sent to the chat server, and the plurality of alternative reply answers corresponding to the original question and the auxiliary information database auxiliary information established in advance are presented to the manual customer service.
  • the auxiliary information includes at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer history behavior.
  • the obtaining, from the pre-established question and answer database, a plurality of alternative reply answers corresponding to the original question and corresponding confidence levels include:
  • N is an integer greater than 1, and the question and answer database stores typical questions and alternative responses corresponding to typical problems. answer;
  • the confidence level for calculating each of the alternative reply answers includes:
  • the method further includes:
  • the final reply text of the manual customer service reply is obtained; or the customer has the highest confidence answer answer with the highest confidence as the final reply text, and the final reply text is obtained.
  • the method further includes:
  • the updating the Q&A training cache sample library and the auxiliary information training cache sample library according to the final reply text include:
  • the auxiliary information is trained in the cache sample library.
  • generating the question and answer training cache sample according to the candidate reply answer adopted by the final reply text and storing the sample to the question and answer training cache sample library includes:
  • the final reply text does not adopt the alternative reply answer in the question and answer database, generate the adopted alternative reply answer according to the final reply text, and generate a question and answer training cache sample according to the generated adopted alternative reply answer. And storing in the Q&A training cache sample library;
  • the Q&A training cache sample includes: an original question corresponding to the final reply text, a problem classification, a typical question of matching, and an actual reply content.
  • the method further includes:
  • the updating the Q&A database by using the Q&A training cache sample in the Q&A training cache sample library includes at least one of the following:
  • the retraining the original problem classifier of the question and answer database includes:
  • the original problem and the problem classification of each Q&A training cache sample stored in the Q&A training cache sample library are respectively re-trained as the input and output of the original problem classifier of the Q&A database.
  • the retraining the original problem similarity algorithm of the question and answer database includes:
  • the pairing of the "original problem-matching typical problem” is taken as an input, and the frequency is used as an output to retrain the parameters of the original problem similarity algorithm of the question and answer database.
  • the updating the recommendation degree of the alternative reply answer in the question and answer database includes:
  • the proportion of the responses of the corresponding actual reply contents is counted as the recommendation degree to update the question and answer database.
  • the generating the auxiliary information training sample according to the auxiliary information adopted by the final reply text and storing the auxiliary information training sample into the auxiliary information training cache sample library includes:
  • the final reply text does not adopt the auxiliary information in the auxiliary information database, generate the adopted auxiliary information according to the final reply text, generate an auxiliary information training sample according to the generated adopted auxiliary information, and store the Auxiliary information training cache sample library;
  • the auxiliary information training cache sample includes: an original question corresponding to the final reply text, a problem classification, a typical problem of matching, an auxiliary information category, and an auxiliary information content.
  • the method further includes:
  • auxiliary information training cache sample library satisfies a preset update condition
  • the using the auxiliary information to train the auxiliary information training samples in the cache sample library to update the auxiliary information database includes:
  • the proportion of the reply of each corresponding auxiliary information content is counted as the recommendation degree to update the auxiliary information database.
  • the update condition is that the saved sample number is greater than a third threshold or reaches a preset update time point.
  • An embodiment of the present invention further provides an apparatus for customer service, including:
  • a receiving module for receiving an original question input by the customer
  • a judging module configured to obtain, from a pre-established question and answer database, a plurality of alternative reply answers corresponding to the original question and corresponding confidence levels, and determine whether the confidence of the candidate reply answer with the highest confidence is not less than the first threshold ;
  • a processing module configured to: if the confidence level of the most reliable alternative reply answer is not less than the first threshold, return the candidate with the highest confidence answer as the final reply text; if the confidence is the highest If the confidence level of the reply answer is less than the first threshold, the service request for the manual customer service access is sent to the chat server, and the plurality of alternative reply answers corresponding to the original question are presented to the manual customer service and the pre-established Auxiliary information in the auxiliary information database, the auxiliary information comprising at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer historical behavior.
  • An embodiment of the present invention further provides an electronic device for customer service, including:
  • the candidate reply answer with the highest confidence is returned to the customer as the final reply text; if the confidence answer of the highest answer is the highest If the degree is less than the first threshold, sending a service request for the manual customer service access to the chat server, and presenting, to the manual customer service, the multiple candidate reply answers corresponding to the original question and the pre-established auxiliary information database
  • auxiliary information comprising at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer history behavior.
  • the embodiment of the present invention further provides a computer readable storage medium storing a computer program, when the computer program is executed by a processor, causing the processor to perform the following steps:
  • the candidate reply answer with the highest confidence is returned to the customer as the final reply text; if the confidence answer of the highest answer is the highest If the degree is less than the first threshold, sending a service request for the manual customer service access to the chat server, and presenting, to the manual customer service, the multiple candidate reply answers corresponding to the original question and the pre-established auxiliary information database
  • auxiliary information comprising at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer history behavior.
  • the original question input by the customer is first received, and multiple candidate reply answers corresponding to the original question and corresponding confidence degrees are obtained from the pre-established question and answer database, and the confidence of the candidate reply answer with the highest confidence is determined. Whether the degree is not less than the first threshold, and if the confidence of the highest-reciprocity alternative reply answer is not less than the first threshold, replying the customer with the highest-reliable alternative reply answer as the final reply text; if the confidence level Sending a service request requesting manual customer service access to the chat server, and presenting the plurality of alternative reply answers corresponding to the original question to the manual customer service, and the confidence of the highest candidate reply answer is less than the first threshold Supplementary information.
  • the question and answer database can still be used to provide a quick reply; on the other hand, even in the field that is not familiar to the current manual customer service, It can still provide a variety of auxiliary information for the manual customer service, and assist the manual customer service to respond quickly, which can reduce the workload of the manual customer service and improve the quality of the problem response.
  • FIG. 1 is a schematic flow chart of a method for customer service according to an embodiment of the present invention
  • FIG. 2 is a structural block diagram of an apparatus for customer service according to an embodiment of the present invention.
  • FIG. 3 is a structural block diagram of an electronic device for customer service according to an embodiment of the present invention.
  • FIG. 4 is a system block diagram of a customer service according to an embodiment of the present invention.
  • FIG. 5 is a schematic flowchart of a method for customer service according to an embodiment of the present invention.
  • FIG. 6 is a schematic flowchart of obtaining a confidence level of an alternate reply answer and sorting the alternate reply answers according to an embodiment of the present invention
  • FIG. 7 is a schematic diagram of generating a Q&A training cache sample and an auxiliary information training sample according to an embodiment of the present invention
  • FIG. 8 is a schematic diagram of obtaining an alternate reply answer and auxiliary information adopted according to an embodiment of the present invention.
  • the problem can be assigned to the response robot (ie, the customer service robot) and the manual customer service.
  • the response robot ie, the customer service robot
  • the manual customer service First, it is determined whether the answering robot knows the answer to the question entered in the chat window. If it is known, the responding robot responds; if not, the feedback question is sent to the answering server, and the answering server sends the request to the chat server to request the manual customer service. Incoming request.
  • the chat server then connects to the customer through a series of screening, finding, and determining the most appropriate manual customer service.
  • the conditions for screening and judging the appropriate manual customer service include: whether there is a skill to reply to this category problem, whether it is in an online non-suspended state, whether the connection has been established recently with the answering server, and the manual customer service is currently receiving Whether the number of users is less than the threshold or the like.
  • the customer service robot may participate in answering the customer's question only at the beginning of a session. Once the customer service robot cannot answer the question and transfer to the manual customer service, the customer service robot can no longer be used to help reduce the workload of the manual customer service.
  • the existing solution does not have the function of self-learning, and it is impossible to update its own knowledge base and algorithm according to the operation and feedback of the customer/human customer service.
  • Embodiments of the present invention provide a method, apparatus, electronic device, and computer readable storage medium for customer service, which can reduce the workload of manual customer service and improve the quality of problem recovery.
  • An embodiment of the present invention provides a method for customer service, as shown in FIG. 1, including:
  • Step 101 Receive an original question input by a customer
  • Step 102 Obtain a plurality of alternative reply answers corresponding to the original question and corresponding confidence levels from the pre-established question and answer database, and determine whether the confidence of the candidate reply answers with the highest confidence is not less than the first threshold;
  • Step 103 If the confidence level of the most reliable alternative reply answer is not less than the first threshold, reply to the customer the candidate reply answer with the highest confidence as the final reply text;
  • Step 104 If the confidence level of the most reliable alternative reply answer is less than the first threshold, send a service request requesting manual customer service access to the chat server, and present the plurality of corresponding original questions to the manual customer service.
  • the original question input by the client is first received, and multiple candidate reply answers corresponding to the original question and corresponding confidence levels are obtained from the pre-established question and answer database, and the candidate reply answer with the highest confidence is determined. Whether the confidence is not less than the first threshold, and if the confidence of the most reliable alternative reply answer is not less than the first threshold, replying to the customer the answer with the highest confidence is the final reply text; If the confidence of the highest-level alternative reply answer is less than the first threshold, the service request for the manual customer service access is sent to the chat server, and the plurality of alternative reply answers corresponding to the original question are presented to the manual customer service. And auxiliary information.
  • the question and answer database can still be used to provide a quick reply; on the other hand, even in the field that is not familiar to the current manual customer service, It can still provide a variety of auxiliary information for the manual customer service, and assist the manual customer service to respond quickly, which can reduce the workload of the manual customer service and improve the quality of the problem response.
  • the obtaining, from the pre-established question and answer database, multiple candidate reply answers corresponding to the original question and corresponding confidence levels include:
  • N is an integer greater than 1, and the question and answer database stores typical questions and alternative responses corresponding to typical problems. answer;
  • the confidence to calculate each alternate reply answer includes:
  • the method further includes:
  • the final reply text of the manual customer service reply is obtained; or the customer has the highest confidence answer answer with the highest confidence as the final reply text, and the final reply text is obtained.
  • the method further includes:
  • the updating the Q&A training cache sample library and the auxiliary information training cache sample library according to the final reply text includes:
  • the auxiliary information is trained in the cache sample library.
  • the generating a question and answer training cache sample according to the candidate reply answer adopted by the final reply text and storing the sample to the question and answer training cache sample library includes:
  • the final reply text does not adopt the alternative reply answer in the question and answer database, generate the adopted alternative reply answer according to the final reply text, and generate a question and answer training cache sample according to the generated adopted alternative reply answer. And storing in the Q&A training cache sample library;
  • the Q&A training cache sample includes: an original question corresponding to the final reply text, a problem classification, a typical question of matching, and an actual reply content.
  • the method further includes:
  • the question and answer training cache sample library and the auxiliary information training are discarded. Caches the samples cached in the sample library.
  • the updating the Q&A training cache sample library and the auxiliary information training cache sample library according to the final reply text include:
  • the auxiliary information is trained in the cache sample library.
  • the generating a question and answer training cache sample according to the candidate reply answer adopted by the final reply text and storing the sample to the question and answer training cache sample library includes:
  • the final reply text does not adopt the alternative reply answer in the question and answer database, generate the adopted alternative reply answer according to the final reply text, and generate a question and answer training cache sample according to the generated adopted alternative reply answer. And storing in the Q&A training cache sample library;
  • the Q&A training cache sample includes: an original question corresponding to the final reply text, a problem classification, a typical question of matching, and an actual reply content.
  • the method further includes:
  • the updating the Q&A database by using the Q&A training cache sample in the Q&A training cache sample library includes at least one of the following:
  • the degree of recommendation of the alternate reply answer in the question and answer database is updated.
  • the retraining the original problem classifier of the question and answer database includes:
  • the original problem and the problem classification of each Q&A training cache sample stored in the Q&A training cache sample library are respectively re-trained as the input and output of the original problem classifier of the Q&A database.
  • the retraining the original problem similarity algorithm of the question and answer database includes:
  • the pairing of the "original problem-matching typical problem” is taken as an input, and the frequency is used as an output to retrain the parameters of the original problem similarity algorithm of the question and answer database.
  • the updating the recommendation degree of the alternative reply answer in the question and answer database includes:
  • the proportion of the responses of the corresponding actual reply contents is counted as the recommendation degree to update the question and answer database.
  • the generating the auxiliary information training sample according to the auxiliary information adopted by the final reply text and storing the auxiliary information training sample into the auxiliary information training cache sample library includes:
  • the final reply text does not adopt the auxiliary information in the auxiliary information database, generate the adopted auxiliary information according to the final reply text, generate an auxiliary information training sample according to the generated adopted auxiliary information, and store the Auxiliary information training cache sample library;
  • the auxiliary information training cache sample includes: an original question corresponding to the final reply text, a problem classification, a typical problem of matching, an auxiliary information category, and an auxiliary information content.
  • the method further includes:
  • auxiliary information training cache sample library satisfies a preset update condition
  • the using the auxiliary information to train the auxiliary information training samples in the cache sample library to update the auxiliary information database includes:
  • the proportion of the reply of each corresponding auxiliary information content is counted as the recommendation degree to update the auxiliary information database.
  • the update condition may be that the saved sample quantity is greater than a third threshold or reaches a preset update time point.
  • the value of the third threshold may be set as needed, such as setting the third threshold to 1000, 2000, and the like.
  • the preset update time point can be, for example, a certain day of the week, or a certain number of each month, and the like.
  • the embodiment of the invention further provides a device for customer service, as shown in FIG. 2, comprising:
  • the receiving module 21 is configured to receive an original question input by the client
  • the determining module 22 is configured to obtain, from a pre-established question and answer database, a plurality of alternative reply answers corresponding to the original question and corresponding confidence levels, and determine whether the confidence of the candidate reply answer with the highest confidence is not less than the first Threshold value
  • the processing module 23 is configured to: if the confidence level of the candidate reply answer with the highest confidence is not less than the first threshold, return the candidate reply answer with the highest confidence as the final reply text; if the confidence is the highest And if the confidence of the alternative reply answer is less than the first threshold, sending a service request requesting manual customer service access to the chat server, and presenting the multiple candidate response answers corresponding to the original question to the manual customer service and pre-establishing Auxiliary information in the auxiliary information database, the auxiliary information comprising at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer historical behavior.
  • the original question input by the client is first received, and multiple candidate reply answers corresponding to the original question and corresponding confidence levels are obtained from the pre-established question and answer database, and the candidate reply answer with the highest confidence is determined. Whether the confidence is not less than the first threshold, and if the confidence of the most reliable alternative reply answer is not less than the first threshold, replying to the customer the answer with the highest confidence is the final reply text; If the confidence of the highest-level alternative reply answer is less than the first threshold, the service request for the manual customer service access is sent to the chat server, and the plurality of alternative reply answers corresponding to the original question are presented to the manual customer service. And auxiliary information.
  • the question and answer database can still be used to provide a quick reply; on the other hand, even in the field that is not familiar to the current manual customer service, It can still provide a variety of auxiliary information for the manual customer service, and assist the manual customer service to respond quickly, which can reduce the workload of the manual customer service and improve the quality of the problem response.
  • the determining module 22 is specifically configured to determine a problem classification corresponding to the original problem; and find a N typical problem with the highest similarity to the original problem in the corresponding problem classification in the question and answer database, where N is An integer greater than 1, the candidate question database stores an alternative answer answer corresponding to a typical question and a typical question; obtains an alternate reply answer corresponding to each typical question, and calculates a confidence level for each candidate reply answer.
  • the determining module 22 includes:
  • An alternative reply answer calculation unit configured to obtain, when the candidate reply answer is a plain text answer, obtain a similarity S i of the typical question corresponding to each candidate reply answer and the original question, and each candidate reply answer
  • the device further includes:
  • the obtaining module 24 is configured to obtain a final reply text of the manual customer service reply after the manual customer service replies to the original question; or reply the customer with the most reliable alternative reply answer as the final reply text, and obtain the final Reply to the text.
  • the device further includes:
  • the self-learning training sample buffering unit 25 is configured to determine whether the customer continues to input the original question after obtaining the final reply text of the manual customer service reply; if the customer no longer inputs the original question, and obtains the customer satisfaction with the final reply text, When the satisfaction of the client with the final reply text is greater than the second threshold, the Q&A training cache sample library and the auxiliary information training cache sample library are updated according to the final reply text.
  • the self-learning training sample buffering unit 25 is specifically configured to perform word segmentation on the final reply text; determine an alternative reply answer and auxiliary information adopted by the final reply text according to the word segmentation result; according to the final reply text
  • the adopted alternative reply answer generates a Q&A training cache sample and stores it in the Q&A training cache sample library, generates an auxiliary information training sample according to the auxiliary information adopted by the final reply text, and stores the sample to the auxiliary information training cache sample In the library.
  • the self-learning training sample buffer unit includes:
  • a question and answer training cache sample buffer unit configured to determine, according to the word segmentation result, whether the final reply text adopts an alternate reply answer in the question and answer database; if the final reply text adopts an alternate reply answer in the question and answer database, Generating a Q&A training cache sample according to the adopted alternative reply answer and storing it in the Q&A training cache sample library; if the final reply text does not adopt an alternative reply answer in the Q&A database, according to the final reply Generating an alternative reply answer adopted by the text generation, generating a question and answer training cache sample according to the generated adopted alternative reply answer and storing it in the Q&A training cache sample library; wherein the Q&A training cache sample includes: the final Reply to the original question, the problem classification, the typical question of the match, and the actual response content.
  • the self-learning training sample buffering unit 25 is configured to: after obtaining the final reply text of the manual customer service reply, update the question and answer training cache sample library and the auxiliary information training cache sample library according to the final reply text, and determine Whether the customer continues to input the original question; if the customer no longer inputs the original question, obtains the customer's satisfaction with the final reply text, and discards the cache when the customer's satisfaction with the final reply text is not greater than the second threshold sample.
  • the self-learning training sample buffering unit 25 is specifically configured to perform word segmentation on the final reply text; determine an alternative reply answer and auxiliary information adopted by the final reply text according to the word segmentation result; according to the final reply text
  • the adopted alternative reply answer generates a Q&A training cache sample and stores it in the Q&A training cache sample library, generates an auxiliary information training sample according to the auxiliary information adopted by the final reply text, and stores the sample to the auxiliary information training cache sample In the library.
  • the self-learning training sample buffer unit includes:
  • a question and answer training cache sample buffer unit configured to determine, according to the word segmentation result, whether the final reply text adopts an alternate reply answer in the question and answer database; if the final reply text adopts an alternate reply answer in the question and answer database, Generating a Q&A training cache sample according to the adopted alternative reply answer and storing it in the Q&A training cache sample library; if the final reply text does not adopt an alternative reply answer in the Q&A database, according to the final reply Generating an alternative reply answer adopted by the text generation, generating a question and answer training cache sample according to the generated adopted alternative reply answer and storing it in the Q&A training cache sample library; wherein the Q&A training cache sample includes: the final Reply to the original question, the problem classification, the typical question of the match, and the actual response content.
  • the device further includes:
  • the question and answer database update module 26 is configured to determine, after the question and answer training cache sample is stored in the Q&A training cache sample library, whether the Q&A training cache sample library satisfies a preset update condition, and the QC training cache sample library When the preset update condition is met, the question and answer database is updated by using the Q&A training cache sample in the Q&A training cache sample library.
  • the question and answer database update module is specifically configured to update the recommendation degree of the alternative reply answer in the question and answer database; and/or retrain the original question classifier of the question and answer database; and/or The original problem similarity algorithm of the question and answer database is retrained.
  • question and answer database update module includes:
  • the problem classifier training unit is configured to re-train the original problem classifier of the question and answer database as the input and output, respectively, by using the original problem and the problem classification of each Q&A training cache sample stored in the Q&A training cache sample library.
  • question and answer database update module includes:
  • a problem similarity algorithm training unit configured to select the original problem of the Q&A training cache sample and the matching typical problem in the Q&A training cache sample library, for the same “original problem-matching typical problem”
  • the pairing combination is used to count the frequency of occurrence; the pairing of the "original problem-matching typical problem” is taken as an input, and the frequency is used as an output to retrain the parameters of the original problem similarity algorithm of the question and answer database.
  • question and answer database update module includes:
  • a recommendation degree update unit configured to select, in the Q&A training cache sample library, a typical question of matching the Q&A training cache sample and the actual reply content; for each typical problem of matching, statistics corresponding to each actual The proportion of responses to the reply content is updated as the recommendation level to the question and answer database.
  • the self-learning training sample buffer unit includes:
  • the auxiliary information training buffer sample buffer unit determines whether the final reply text adopts the auxiliary information in the auxiliary information database according to the word segmentation result; if the final reply text adopts the auxiliary information in the auxiliary information database, according to the adopted
  • the auxiliary information generates the auxiliary information training sample and stores it in the auxiliary information training cache sample library; if the final reply text does not adopt the auxiliary information in the auxiliary information database, the adopted final text is generated according to the final reply text.
  • the auxiliary information is generated according to the generated auxiliary information, and is stored in the auxiliary information training cache sample database; wherein the auxiliary information training cache sample includes: an original question corresponding to the final reply text, Problem classification, typical problems of matching, auxiliary information categories and auxiliary information content.
  • the device further includes:
  • the auxiliary information database updating module 27 is configured to determine, after the auxiliary information training sample is stored in the auxiliary information training cache sample library, whether the auxiliary information training cache sample library satisfies a preset update condition, and the auxiliary information training When the cache sample library satisfies the preset update condition, the auxiliary information training sample in the cache sample library is trained by the auxiliary information to update the auxiliary information database.
  • the auxiliary information database update module includes:
  • a recommendation degree update unit configured to select, in the auxiliary information training cache sample library, a typical problem of matching of the auxiliary information training samples and the auxiliary information content; for each typical problem of matching, statistics corresponding to each The proportion of the reply of the auxiliary information content is updated as the degree of recommendation to the auxiliary information database.
  • the update condition may be that the saved sample quantity is greater than a third threshold or reaches a preset update time point.
  • the value of the third threshold may be set as needed, such as setting the third threshold to 1000, 2000, and the like.
  • the preset update time point can be, for example, a certain day of the week, or a certain number of each month, and the like.
  • An embodiment of the present invention further provides an electronic device 30 for customer service, as shown in FIG. 3, including:
  • the processor 32 when the computer program instructions are executed by the processor, the processor 32 is caused to perform the following steps:
  • the candidate reply answer with the highest confidence is returned to the customer as the final reply text; if the confidence answer of the highest answer is the highest If the degree is less than the first threshold, sending a service request for the manual customer service access to the chat server, and presenting, to the manual customer service, the multiple candidate reply answers corresponding to the original question and the pre-established auxiliary information database
  • auxiliary information comprising at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer history behavior.
  • the original question input by the client is first received, and multiple candidate reply answers corresponding to the original question and corresponding confidence levels are obtained from the pre-established question and answer database, and the candidate reply answer with the highest confidence is determined. Whether the confidence is not less than the first threshold, and if the confidence of the most reliable alternative reply answer is not less than the first threshold, replying to the customer the answer with the highest confidence is the final reply text; If the confidence of the highest-level alternative reply answer is less than the first threshold, the service request for the manual customer service access is sent to the chat server, and the plurality of alternative reply answers corresponding to the original question are presented to the manual customer service. And auxiliary information.
  • the question and answer database can still be used to provide a quick reply; on the other hand, even in the field that is not familiar to the current manual customer service, It can still provide a variety of auxiliary information for the manual customer service, and assist the manual customer service to respond quickly, which can reduce the workload of the manual customer service and improve the quality of the problem response.
  • the electronic device 30 for customer service further includes a network interface 31, an input device 33, a hard disk 35, and a display device 36.
  • the bus architecture can be a bus and bridge that can include any number of interconnects.
  • One or more central processing units (CPUs), specifically represented by processor 32, and various circuits of one or more memories represented by memory 34 are coupled together.
  • the bus architecture can also connect various other circuits such as peripherals, voltage regulators, and power management circuits. It will be appreciated that the bus architecture is used to implement connection communication between these components.
  • the bus architecture includes, in addition to the data bus, a power bus, a control bus, and a status signal bus, all of which are well known in the art and therefore will not be described in detail herein.
  • the network interface 31 can be connected to a network (such as the Internet, a local area network, etc.) to obtain relevant data from the network, such as an original question input by the customer, and can be saved in the hard disk 35.
  • a network such as the Internet, a local area network, etc.
  • the input device 33 can receive various instructions input by an operator and send it to the processor 32 for execution.
  • the input device 33 may include a keyboard or a pointing device (eg, a mouse, a trackball, a touch pad or a touch screen, etc.).
  • the display device 36 can display the results obtained by the processor 32 executing the instructions.
  • the memory 34 is configured to store programs and data necessary for the operation of the operating system, and intermediate results such as the calculation process of the processor 32.
  • the memory 34 in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory.
  • the non-volatile memory may be a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM), or Flash memory.
  • Volatile memory can be random access memory (RAM), which acts as an external cache.
  • RAM random access memory
  • memory 34 stores elements, executable modules or data structures, or a subset thereof, or their extended set: operating system 341 and application 342.
  • the operating system 341 includes various system programs, such as a framework layer, a core library layer, a driver layer, and the like, for implementing various basic services and processing hardware-based tasks.
  • the application 342 includes various applications, such as a browser, for implementing various application services.
  • a program implementing the method of the embodiment of the present invention may be included in the application 342.
  • the processor 32 when calling and executing the application and data stored in the memory 34, specifically, may receive an original question input by the client; and obtain a plurality of corresponding original questions from the pre-established question and answer database.
  • the candidate reply answer and the corresponding confidence degree determine whether the confidence of the candidate reply answer with the highest confidence is not less than the first threshold; if the confidence of the highest answer candidate reply answer is not less than the first threshold, then The customer replies with the most reliable alternative reply answer as the final reply text; if the confidence of the highest trusted alternative reply answer is less than the first threshold, the service request for the manual customer service access is sent to the chat server, And presenting, to the manual customer service, the plurality of alternative reply answers corresponding to the original question and auxiliary information in a pre-established auxiliary information database, the auxiliary information including at least one of the following information: the original question Keyword reference information, customer data, customer preferences, customer historical behavior.
  • Processor 32 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method may be completed by an integrated logic circuit of hardware in the processor 32 or an instruction in the form of software.
  • the processor 32 described above may be a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, or discrete hardware.
  • DSP digital signal processor
  • ASIC application specific integrated circuit
  • FPGA off-the-shelf programmable gate array
  • the components, steps, and logic blocks disclosed in the embodiments of the present invention may be implemented or executed.
  • the general purpose processor may be a microprocessor or the processor or any conventional processor or the like.
  • the steps of the method disclosed in the embodiments of the present invention may be directly implemented by the hardware decoding processor, or may be performed by a combination of hardware and software modules in the decoding processor.
  • the software module can be located in a conventional storage medium such as random access memory, flash memory, read only memory, programmable read only memory or electrically erasable programmable memory, registers, and the like.
  • the storage medium is located in memory 34, and processor 32 reads the information in memory 34 and, in conjunction with its hardware, performs the steps of the above method.
  • the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof.
  • the processing unit can be implemented in one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs). ), a general purpose processor, a controller, a microcontroller, a microprocessor, other electronic units for performing the functions described herein, or a combination thereof.
  • ASICs application specific integrated circuits
  • DSPs digital signal processors
  • DSPDs digital signal processing devices
  • PLDs programmable logic devices
  • FPGAs field programmable gate arrays
  • a general purpose processor a controller, a microcontroller, a microprocessor, other electronic units for performing the functions described herein, or a combination thereof.
  • the techniques described herein can be implemented by modules (eg, procedures, functions, and so on) that perform the functions described herein.
  • the software code can be stored in memory and executed by the processor.
  • the memory can be implemented in the processor or external to the processor.
  • the processor 32 determines a problem classification corresponding to the original problem; and finds N typical problems with the highest similarity to the original problem in the question classification database, where N is an integer greater than 1.
  • the question and answer database stores candidate answers corresponding to typical problems and typical questions; obtains alternate reply answers corresponding to each typical question, and calculates the confidence of each candidate reply answer.
  • the processor 32 obtains the similarity S i of the typical question corresponding to each candidate reply answer and the recommendation degree of each candidate reply answer.
  • the candidate reply answer is a template class answer, obtain k auxiliary in the template class answer
  • the processor 32 obtains the final reply text of the manual customer service reply after the manual customer service replies to the original question; or replies to the customer with the most reliable alternative reply answer as the final reply text, and obtains the final Reply to the text.
  • the processor 32 determines whether the client continues to input the original question; if the client continues to input the original question, the process proceeds to the step of receiving the original question input by the client; if the client no longer inputs the original question, the client obtains the final reply text. Satisfaction, when the customer's satisfaction with the final reply text is greater than a second threshold, updating the question and answer training cache sample library and the auxiliary information training cache sample library according to the final reply text.
  • the processor 32 performs word segmentation on the final reply text; determines an alternative reply answer and auxiliary information adopted by the final reply text according to the word segmentation result; and generates a question and answer according to the alternative reply answer adopted by the final reply text.
  • the cache sample is trained and stored in the Q&A training cache sample library, and the auxiliary information training sample is generated according to the auxiliary information adopted by the final reply text and stored in the auxiliary information training cache sample library.
  • the processor 32 determines, according to the word segmentation result, whether the final reply text adopts an alternative reply answer in the question and answer database; if the final reply text adopts an alternative reply answer in the question and answer database, according to the adopted The alternative reply answer generates a question and answer training cache sample and stores it in the Q&A training cache sample library; if the final reply text does not adopt an alternate reply answer in the question and answer database, the final reply text is generated according to the The candidate reply answer is generated, and the Q&A training cache sample is generated according to the generated adopted alternative reply answer and stored in the Q&A training cache sample library; wherein the Q&A training cache sample includes: the final reply text corresponding to The original problem, the problem classification, the typical question of the match, and the actual response.
  • the processor 32 determines whether the Q&A training cache sample library satisfies a preset update condition, and uses the quiz training in the question and answer training cache sample database when the Q&A training cache sample library satisfies a preset update condition.
  • the cache sample updates the question and answer database.
  • the processor 32 updates the recommendation degree of the alternative reply answer in the question and answer database; and/or retrains the original question classifier of the question and answer database; and/or the original of the question and answer database.
  • the problem similarity algorithm is retrained.
  • the processor 32 re-trains the original problem classifier of the Q&A database as the input and output by respectively classifying the original problem and the problem classification of each Q&A training cache sample stored in the Q&A training cache sample library.
  • the processor 32 selects, in the Q&A training cache sample library, the original problem of the Q&A training cache sample and the paired data of the typical problem, for the pairing combination of the same “original problem-matching typical problem”, The frequency of occurrence is counted; the pairing of "original problem-matching typical problem” is taken as input, and the frequency is used as an output to retrain the parameters of the original problem similarity algorithm of the question and answer database.
  • the processor 32 selects, in the Q&A training cache sample library, a typical question of the matching of the Q&A training cache sample and the actual reply content; for each typical question of the match, the corresponding actual reply is counted.
  • the proportion of the content is updated as the recommendation level to update the question and answer database.
  • the processor 32 determines, according to the word segmentation result, whether the final reply text adopts the auxiliary information in the auxiliary information database; if the final reply text adopts the auxiliary information in the auxiliary information database, according to the adopted auxiliary
  • the information generation auxiliary information training sample is stored in the auxiliary information training cache sample library; if the final reply text does not adopt the auxiliary information in the auxiliary information database, the adopted auxiliary information is generated according to the final reply text.
  • the processor 32 determines whether the auxiliary information training cache sample library satisfies a preset update condition, and uses the auxiliary information to train the cache sample database when the auxiliary information training cache sample library satisfies a preset update condition.
  • the auxiliary information training sample updates the auxiliary information database.
  • the processor 32 selects, in the auxiliary information training cache sample library, a typical problem of matching of the auxiliary information training samples and the two sets of auxiliary information content; for each typical problem of matching, statistics corresponding to each auxiliary The proportion of the reply of the information content is updated as the degree of recommendation to the auxiliary information database.
  • the embodiment of the present invention further provides a computer readable storage medium storing a computer program, when the computer program is executed by a processor, causing the processor to perform the following steps:
  • the candidate reply answer with the highest confidence is returned to the customer as the final reply text; if the confidence answer of the highest answer is the highest If the degree is less than the first threshold, sending a service request for the manual customer service access to the chat server, and presenting, to the manual customer service, the multiple candidate reply answers corresponding to the original question and the pre-established auxiliary information database
  • auxiliary information comprising at least one of the following: keyword reference information in the original question, customer data, customer preferences, customer history behavior.
  • the present invention provides a customer service assistance system, as shown in FIG. 4, including:
  • the question and answer database 110 is used to store a list of known questions and their alternate reply answers.
  • the auxiliary information database 120 is configured to store various reference information that may be used when performing customer service, including but not limited to brand, product information, product reputation, customer basic data, customer preferences, customer historical behavior records, and the like.
  • the alternative reply answer calculation unit 150 is configured to calculate one or more alternative reply answers and their confidence based on the question and answer database 110 according to the customer's question.
  • the manual customer service auxiliary panel 100 which assists the manual customer service to reply to the customer problem interface, can be divided into a plurality of auxiliary sub-panels according to the auxiliary content.
  • auxiliary content E.g:
  • Alternate reply answer sub-panel used to provide multiple alternative reply answers or reply templates according to priority.
  • Auxiliary information sub-panel used to provide auxiliary information that may be used when responding to customer questions. Can include but is not limited to:
  • Reference information based on keywords in the original question extract keywords from the customer's original question and list relevant information. For example, if a product name is mentioned in the customer question, the text description, specifications, price, preferential activities, product word of mouth, similar product list and other information of the product can be given.
  • Customer's basic data the customer's name or nickname, age, gender, geographical location, height, weight, skin quality, etc. can describe the customer's basic attributes.
  • Customer preference data Customer preference information extracted from the customer's historical purchase record and product evaluation, such as the style of the preferred product, the product attributes of interest, and so on.
  • Manual customer service reply input field It is used for the manual customer service to refer to the above alternative reply answer sub-panel and auxiliary information sub-panel, and by clicking, selecting, editing, finally generating an input column that needs to reply to the customer's answer.
  • the self-learning training sample buffer unit 160 updates the question and answer training cache sample library 130 and the auxiliary information training cache sample library 140 according to the final reply text of the manual customer service.
  • the self-learning unit 170 continuously adjusts the question and answer database 110, the auxiliary information database 120, and retrains the alternative reply answer according to the sample stored in the question and answer training cache sample library 130 and the auxiliary information training cache sample library 140 and the customer satisfaction survey result. Calculate the algorithm in the unit to improve the performance of the system.
  • each candidate reply answer in the question and answer database 110 is as shown in Table 1:
  • the question and answer database 110 can pre-define a plurality of problem classifications, so that when searching for the closest problem, the search scope is narrowed according to the application scenario and the question and answer context, and the matching speed and precision are improved.
  • the types of questions that can be referred to are: product introduction, product recommendation, quality complaints, etc.
  • a typical question it refers to the text of the question. Since many natural language processing techniques can judge the similarity between problems (such as grammatical similarity, editing distance, statement vector similarity, etc., we will not repeat them here), so there are many similar questions for the same typical problem. In the case, only one typical question can be saved in the question and answer database 110.
  • the response type refers to the type of alternative reply answer, and the alternative reply answer can be divided into two types: "plain text answer” and "template class answer".
  • the plain text answer is only text;
  • the template type answer is that some keywords in the reply are replaceable words, for example, the reply content "to the [provincial name] postage is [price]” belongs to a template type answer, where "[province The words in the square brackets of the city name "" and “[price]” are all replaceable words in the reply, and also correspond to the information in the column of "auxiliary information category” in the auxiliary information database 120.
  • the words that can be replaced by the [] position correspond to the information in the column of "auxiliary information content” in the auxiliary information database 120.
  • the timestamp was recently used to record the timestamp that the alternate reply answer was last selected.
  • the method for customer service in this embodiment specifically includes the following steps:
  • Step 401 The customer raises an original question
  • Step 402 Determine whether the confidence of the alternative reply answer is sufficiently high, if it is high enough, go to step 403; if not high enough, go to step 404;
  • the alternative reply answer calculation unit 150 obtains the candidate answer of the highest confidence by the calculation of the problem similarity.
  • the process of obtaining an alternate answer with the highest confidence includes the following steps:
  • Step 601 Determine a problem classification corresponding to the original problem.
  • the classification of the problem can be judged simply by some rules, templates, or by a "problem classifier" that trains through a large number of pre-marked texts through machine learning.
  • Step 502 Find N typical problems with the highest similarity with the original problem in the corresponding problem classification in the question and answer database;
  • the N typical problems with the highest similarity with the original problem (assumed to be Q 1 , Q 2 , ..., Q N ) are found in the question and answer database 110, and the problem similarity values are respectively obtained.
  • the calculation method of the problem similarity may use one of a plurality of existing algorithms such as edit distance and eigenvector cosine similarity, or may use multiple similarity algorithms at the same time, and take a set of matching results with the largest similarity value. (hereinafter referred to as "integrated similarity algorithm").
  • Step 503 Acquire an alternative reply answer corresponding to each typical question
  • each typical problem corresponds to multiple alternative reply answers, and has their own similarity.
  • M alternative reply answers assumed to be A 1 , A 2 , ..., A M
  • their recommended degrees are respectively Is R 1 , R 2 , ..., R M .
  • Step 504 Calculate a confidence level of each candidate reply answer
  • the calculator confidence C ij For the jth alternative reply answer to the i-th typical question, the calculator confidence C ij .
  • Step 505 Sort all candidate reply answers according to the high to low confidence level.
  • Step 403 Display an alternate reply answer to the client
  • Step 404 Transfer to the answer by the manual customer service
  • the auxiliary information provided by the system to the manual customer service includes an alternative reply answer list, and auxiliary information related to the question, etc., which are displayed on the alternate reply answer sub-panel and the auxiliary information sub-panel.
  • Alternate Reply Answers Sub-panel Displays a list of alternate reply answers sorted by confidence level from highest to lowest.
  • the auxiliary information sub-panel can be composed of multiple classification panels, such as reference information given according to keywords in customer questions, customer basic data, customer preferences, customer historical behavior, and the like.
  • Each category panel can be either a "self-learning update panel” and a "other category panel”.
  • the auxiliary information given in the "self-learning update classification panel” can continuously learn and update the recommendation according to the final reply text of the manual customer service, and the "other classification panel” can be assisted by other existing recommendation algorithms.
  • Information such as customer basic data, customer preferences, customer historical behavior, personalized recommendations for customers, and more.
  • the auxiliary information displayed in the "self-study update classification panel" is the focus of the present invention.
  • auxiliary information sub-panel includes only the "self-learning update classification panel”.
  • the auxiliary information displayed in the "Other Classification Panel” can be generated, sorted, and updated in a variety of ways, and is not within the scope of the present invention, but the present invention provides "training cache sample library” and "question/answer information”
  • the updated method of the database can also be used as a reference for the "Other Classification Panel” update.
  • Each classification panel represents an "auxiliary information category” (corresponding to the "auxiliary information category” in the auxiliary information database 120, and if the auxiliary information category has a multi-level classification, it can represent the first-level classification therein), and the display is in accordance with the confidence.
  • a list of alternate response answers sorted from high to low. (The confidence level of the auxiliary information is similar to the confidence calculation method for the alternative reply answer.)
  • the manual customer service finally edits and completes the reply to the customer in the manual customer service reply input column (hereinafter referred to as "final reply text").
  • Step 405 The self-learning training sample buffer unit 160 updates the Q&A training cache sample library and the auxiliary information training cache sample library according to the final reply text.
  • the update question and answer training cache sample library and the auxiliary information training cache sample library include the following steps:
  • Step 601 Acquire an alternative reply answer and auxiliary information adopted by the final reply text
  • the self-learning training sample buffer unit 160 first determines which alternative reply answer text/template and auxiliary information are specifically used in the final reply text. That is, what reply texts/templates and auxiliary information need to be updated in the knowledge base.
  • the alternate answer and auxiliary information adopted to obtain the final reply text includes the following steps:
  • Step 701 Perform word segmentation on the final reply text
  • the wording of the final reply text is divided into words.
  • Step 702 Initially select an alternative reply answer and auxiliary information that may be adopted according to the word segmentation result
  • auxiliary information that has appeared on the panel when answering the customer's original question and the click history of the customer service, which alternative reply answers and auxiliary information (words) are initially screened may be used in the final reply text.
  • Step 703 Compare the preliminary screening result with the final reply text, and determine an alternative reply answer and auxiliary information that are adopted therein;
  • Step 704 Determine whether there is other auxiliary information in the final reply text, if yes, go to step 705; if no, go to step 706;
  • auxiliary information database 120 In addition to the alternative reply answers and auxiliary information identified in step 703, it is further determined whether there is other content belonging to the auxiliary information database 120 in the final reply text. Although it does not appear as auxiliary information on the panel, it is finally applied to the customer service. In the reply - if the auxiliary information is already in the auxiliary information database 120 corresponding to the problem, the recommendation degree needs to be increased; if it is not in the auxiliary information database 120 corresponding to the problem, it needs to be added.
  • Step 705 Appending to the adopted auxiliary information
  • Step 706 whether to find the alternative reply answer corresponding to the final reply text, if yes, then end; if not, go to step 707;
  • Step 707 Generate an alternate answer answer that is adopted.
  • an "adopted" alternative reply answer is generated based on the existing information.
  • the generating method is: in the "final reply text", the "auxiliary information content” identified in steps 703 and 704 is replaced with the corresponding "auxiliary information category", that is, a "received reply template” is obtained; If the final reply text does not contain any known "auxiliary information content", the final reply text will be treated directly as a "received reply text”.
  • a determination is made as to whether the accepted reply text or reply template already exists in the question and answer database 110. If it already exists, you need to update its recommendation. If it does not exist, it means that it may be necessary to add a reply or reply template in the question and answer database 110.
  • the information is stored in the question and answer training cache sample library and the auxiliary information training cache sample library.
  • the "Typical question of matching” column fills in the "typical problem” corresponding to the alternative reply answer used in the final reply text. If the final reply text does not use an existing alternate reply answer, add a new "matching typical question” with the same content as the "original question.”
  • the "Question Classification” column is filled with the "question classification” corresponding to the "typical problem of matching” in the question and answer database 110. If the "typical problem of matching" is a new item, the "problem classification” can be uniformly set. It is classified as “pending classification” or classified according to "original problem” and “problem classifier”.
  • the “actual reply content” in the Q&A training cache sample library is filled in “final reply text” or “final reply template”.
  • the "auxiliary information category" of the auxiliary information training cache sample library is filled in the "auxiliary information category” corresponding to the adopted auxiliary information.
  • the specific filling method is as follows: if in the auxiliary information database 120, under the "typical problem” matching the original question in the dialogue, there is exactly the same item as the auxiliary information adopted this time, and the corresponding "direct problem” is directly adopted.
  • the auxiliary information category is filled in the "auxiliary information category” of the auxiliary information training cache sample library; if it is in the auxiliary information database 120, the "typical problem” that matches the original question in the dialogue is not adopted this time. The same information for the auxiliary information.
  • auxiliary information category is filled in the auxiliary information training cache sample library; if there is no entry in the auxiliary information database 120 that is the same as the auxiliary information that is adopted this time, the "auxiliary information category” is set to "pending classification".
  • the Q&A training cache sample library and the auxiliary information training cache sample library are filled in as “to be classified”, which can be additionally labeled by manual or machine learning, and modified into classification results.
  • Step 602 Generate a Q&A training cache sample according to the candidate reply answer adopted by the final reply text and store it in the Q&A training cache sample library, and generate auxiliary information training cache samples according to the auxiliary information adopted by the final reply text and store To the auxiliary information training cache sample library.
  • Step 406 determine whether the customer still has the next question, if there is, then go to step 401; if there is no next question, then go to step 407;
  • Step 407 determine whether the customer is satisfied with the final reply text; if satisfied, go to step 408; if not, go to step 409;
  • Step 408 It is judged whether a sufficient number of training samples have been accumulated, and if so, the process proceeds to step 410; if not, it ends.
  • Step 409 If the customer is not satisfied with the final reply text, discard all the training cache samples in the session.
  • Step 410 The self-learning unit performs update of the database and re-training of the alternative reply answer calculation unit.
  • the self-learning unit 170 utilizes the incremental training samples for the question and answer database 110 and every predetermined period of time, or whenever the number of new samples in the question and answer training cache sample library and the auxiliary information training cache sample pool is accumulated to a threshold.
  • the auxiliary information database 120 performs update of the recommendation degree and retraining of the alternative reply answer calculation unit.
  • the retraining of the alternative reply answer calculation unit 150 includes:
  • Retraining the "problem classifier” In the cumulative question and answer training cache sample library, select the original problem, the problem classification two columns of data, respectively as input and output, retrain the existing "problem classifier". For a sample whose problem is classified as “to be classified”, it can be marked by manual method before participating in the training.
  • the original problem, the matching typical problem is taken as the input feature, and the matching frequency is taken as the output result.
  • the trained "statistical similarity regression algorithm” can be used as one of the “integrated similarity algorithms” to participate in the calculation of problem similarity.
  • the two types of data matching the typical problem and the actual reply content are selected.
  • the proportion of responses of each "actual reply content” is counted as the "customer overall recommendation degree” to update the question and answer database 110.
  • the specific method is: assuming that in the accumulated question and answer training cache sample library, for a certain "typical question of matching", there are a total of P "actual reply contents” corresponding thereto, and the P “matching typical problems” - " The actual reply content "the number of times the combination appears is M 1 , M 2 , ..., M P .
  • the response ratio is:
  • the value of the column of "Customer Overall Recommendation Degree” corresponding to the corresponding "Typical Question” - “Reply Content” is updated as its response ratio R j .
  • the update method of the recommendation degree of the auxiliary information database 120 is similar to the update of the recommendation degree in the question and answer database, and details are not described herein again.
  • the customer service robot and the question and answer database capable of improving the alternative reply answer are not only useful at the beginning of the session, but have an opportunity to be used for each question of the client.
  • the question of high confidence in the alternative reply answer it can be directly answered by the customer service robot; and the question of low confidence in the alternative reply answer can still be listed as the manual customer service according to the confidence level, as a reference template for the reply. .
  • the alternative reply answer sub-panel and the auxiliary information sub-panel can provide a multi-faceted reply material for the manual customer service. Even if the customer raises the business problem that the human customer service is not familiar with, the question and answer database and the auxiliary information database can be used to help the manual customer service. Give an answer effectively. Through the final manual customer service operations (such as alternative response answers, auxiliary information selected when generating the final answer) and after-the-fact customer satisfaction surveys, the system can continuously update and learn the existing database to improve system performance.

Abstract

本发明提供一种用于客户服务的方法、装置、电子设备、计算机可读存储介质，属于智能客服技术领域。用于客户服务的方法包括：接收客户输入的原始问题；从预先建立的问答数据库中获取多个对应原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；若置信度最高的备选回复答案的置信度不小于第一阈值，则向客户回复置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向人工客服展现多个对应原始问题的备选回复答案以及辅助信息。本发明能够减轻人工客服的工作量，提高问题回复的质量。

Description

用于客户服务的方法、装置、电子设备、计算机可读存储介质
相关申请的交叉引用
本申请主张在2017年12月14日在中国提交的中国专利申请号No.201711339712.6的优先权，其全部内容通过引用包含于此。
技术领域
本发明涉及智能客服技术领域，特别是指一种用于客户服务的方法、装置、电子设备、计算机可读存储介质。
背景技术
在日常生活中，客户通常需要与人工客服或者客服机器人通过网络平台、电话或者面对面的交流以获得问题咨询、产品推荐、投诉等服务。但是受制于目前语义分析、多轮对话等自然语言处理领域的专业技术发展所限，目前客服机器人能回答的问题数量及满意程度还有一定限制；而人工客服通常较为繁忙，却又要处理很多重复的问题或流程。此外，由于客服人员各自熟悉的领域不同、客服人员流动性大、客服人员对前来咨询的客户背景信息不了解等问题，都会造成人工客服的服务水平也存在一定差距。
综上，现有客户服务系统存在如下问题需要解决：第一、如何减轻人工客服的工作量；第二、如何提高问题回复的质量。
发明内容
本发明要解决的技术问题是提供一种用于客户服务的方法、装置、电子设备、计算机可读存储介质，能够减轻人工客服的工作量，并能够提高问题回复的质量。
为解决上述技术问题，本发明的实施例提供技术方案如下：
一方面，提供一种用于客户服务的方法，包括：
接收客户输入的原始问题；
从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
进一步地，所述从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度包括：
判断所述原始问题所对应的问题分类；
在所述问答数据库中对应的问题分类下查找与所述原始问题相似度最高的N个典型问题，N为大于1的整数，所述问答数据库中存储有典型问题及典型问题对应的备选回复答案；
获取每一典型问题对应的备选回复答案，并计算每一备选回复答案的置信度。
进一步地，所述计算每一备选回复答案的置信度包括：
在所述备选回复答案为纯文本答案时，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j计算得到每一备选回复答案的置信度C ij；
在所述备选回复答案为模板类答案时，获取所述模板类答案中k个辅助信息类别中每一辅助信息类别的最高推荐度RAI k，其中k个最高推荐度中的最小值为min(RAI k)，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j*min(RAI k)计算得到每一备选回复答案的置信度C ij。
进一步地，所述方法还包括：
在人工客服回复所述原始问题后，获取人工客服回复的最终回复文本；或向所述客户回复所述置信度最高的备选回复答案作为最终回复文本，获取所述最终回复文本。
进一步地，获取所述最终回复文本之后，所述方法还包括：
判断客户是否继续输入原始问题；
若客户继续输入原始问题，转向所述接收客户输入的原始问题的步骤；
若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度大于第二阈值时，根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库。
进一步地，所述根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库包括：
对所述最终回复文本进行分词；
根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；
根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
进一步地，所述根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中包括：
根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；
若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
若所述最终回复文本未采纳所述问答数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、 问题分类、匹配的典型问题和实际回复内容。
进一步地，将问答训练缓存样本存储至所述问答训练缓存样本库中之后，所述方法还包括：
判断所述问答训练缓存样本库是否满足预设的更新条件，在所述问答训练缓存样本库满足预设的更新条件时，利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新。
进一步地，所述利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新包括以下至少一种：
对所述问答数据库中的备选回复答案的推荐度进行更新；
对所述问答数据库的原问题分类器进行再训练；
对所述问答数据库的原问题相似度算法进行再训练。
进一步地，所述对所述问答数据库的原问题分类器进行再训练包括：
将所述问答训练缓存样本库中存储的每一问答训练缓存样本的原始问题和问题分类分别作为输入和输出对所述问答数据库的原问题分类器进行再训练。
进一步地，所述对所述问答数据库的原问题相似度算法进行再训练包括：
在所述问答训练缓存样本库中，选择问答训练缓存样本的原始问题以及匹配的典型问题这两组数据，对于相同的“原始问题-匹配的典型问题”的配对组合，统计其出现的频度；
以“原始问题-匹配的典型问题”的配对作为输入，以其频度作为输出，对所述问答数据库的原问题相似度算法的参数进行再训练。
进一步地，所述对所述问答数据库中的备选回复答案的推荐度进行更新包括：
在所述问答训练缓存样本库中，选择问答训练缓存样本的匹配的典型问题以及实际回复内容这两组数据；
对于每一个匹配的典型问题，统计其对应的各个实际回复内容的回复比例作为推荐度对所述问答数据库进行更新。
进一步地，所述根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中包括：
根据分词结果判断所述最终回复文本是否采纳所述辅助信息数据库中的辅助信息；
若所述最终回复文本采纳所述辅助信息数据库中的辅助信息，则根据所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；
若所述最终回复文本未采纳所述辅助信息数据库中的辅助信息，则根据所述最终回复文本生成所采纳的辅助信息，根据生成的所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；
其中，所述辅助信息训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题，辅助信息类别和辅助信息内容。
进一步地，将辅助信息训练样本存储至所述辅助信息训练缓存样本库之后，所述方法还包括：
判断所述辅助信息训练缓存样本库是否满足预设的更新条件，在所述辅助信息训练缓存样本库满足预设的更新条件时，利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新。
进一步地，所述利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新包括：
在所述辅助信息训练缓存样本库中，选择辅助信息训练样本的匹配的典型问题以及辅助信息内容这两组数据；
对于每一个匹配的典型问题，统计其对应的各个辅助信息内容的回复比例作为推荐度对所述辅助信息数据库进行更新。
进一步地，所述更新条件为所保存的样本数量大于第三阈值或达到预设的更新时间点。
本发明实施例还提供了一种用于客户服务的装置，包括：
接收模块，用于接收客户输入的原始问题；
判断模块，用于从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
处理模块，用于若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本发明实施例还提供了一种用于客户服务的电子设备，包括：
处理器；和
存储器，在所述存储器中存储有计算机程序指令，
其中，在所述计算机程序指令被所述处理器运行时，使得所述处理器执行以下步骤：
接收客户输入的原始问题；
从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本发明实施例还提供了一种计算机可读存储介质，所述计算机可读存储介 质存储有计算机程序，所述计算机程序被处理器运行时，使得所述处理器执行以下步骤：
接收客户输入的原始问题；
从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本发明的实施例具有以下有益效果：
上述方案中，首先接收客户输入的原始问题，从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值，若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及辅助信息。这样不仅在会话开始时可以使用问答数据库中的备选回复答案，而且在转接入人工客服后，仍能利用问答数据库提供快速答复；另一方面，即使是当前人工客服所不熟悉的领域，仍能为人工客服提供多种辅助信息，协助人工客服进行快速回复，从而能够减轻人工客服的工作量，并能够提高问题回复的质量。
附图说明
图1为本发明实施例用于客户服务的方法的流程示意图；
图2为本发明实施例用于客户服务的装置的结构框图；
图3为本发明实施例用于客户服务的电子设备的结构框图；
图4为本发明实施例用于客户服务的系统框图；
图5为本发明具体实施例用于客户服务的方法的流程示意图；
图6为本发明实施例获取备选回复答案的置信度并对备选回复答案进行排序的流程示意图；
图7为本发明实施例生成问答训练缓存样本和辅助信息训练样本的示意图；
图8为本发明实施例获取被采纳的备选回复答案和辅助信息的示意图。
具体实施方式
为使本发明的实施例要解决的技术问题、技术方案和优点更加清楚，下面将结合附图及具体实施例进行详细描述。
为了减轻人工客服的工作量，可以对应答机器人(即客服机器人)和人工客服进行问题的分配。首先，判断应答机器人是否知道聊天窗口中输入的问题的答案，若知道，则由应答机器人进行回复；若不知道，则反馈问题至应答服务端，由应答服务端向聊天服务器发送要求人工客服接入的请求。然后，聊天服务器通过一系列筛选、查找和判断分配最合适的人工客服与客户进行连接。其中，用于筛选和判断合适的人工客服人选的条件包括：是否具备回复此类别问题的技能、处于在线非挂起的状态、是否与该应答服务端最近一次建立过连接、人工客服当前正在接待的用户数是否小于阈值等。
但是上述方案仍然存在以下缺点：
客服机器人仅在一段会话的开始阶段有可能参与回答客户的问题，一旦由于客服机器人无法回答问题而转接至人工客服后，就不能再利用客服机器人辅助减轻人工客服的工作量。
由于在筛选人工客服时，判断是基于客户当前所提的问题类别，若客户后 续又问到了其他类别的问题，仍可能超出所选人工客服的熟悉领域。虽然可以通过转接其他人工客服的方式解决此问题，但对于客户而言，等待转接的时间过长、需要重新描述问题等因素还是会影响用户体验，客户通常更倾向于在一个人工客服处解决所有问题。
现有方案并没有自我学习的功能，无法根据客户/人工客服的操作及反馈，对自身的知识库以及算法进行更新。
本发明的实施例提供一种用于客户服务的方法、装置、电子设备、计算机可读存储介质，能够减轻人工客服的工作量，并能够提高问题回复的质量。
实施例一
本发明的实施例提供一种用于客户服务的方法，如图1所示，包括：
步骤101：接收客户输入的原始问题；
步骤102：从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
步骤103：若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；
步骤104：若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本实施例中，首先接收客户输入的原始问题，从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值，若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所 述多个对应所述原始问题的备选回复答案以及辅助信息。这样不仅在会话开始时可以使用问答数据库中的备选回复答案，而且在转接入人工客服后，仍能利用问答数据库提供快速答复；另一方面，即使是当前人工客服所不熟悉的领域，仍能为人工客服提供多种辅助信息，协助人工客服进行快速回复，从而能够减轻人工客服的工作量，并能够提高问题回复的质量。
作为一个示例，所述从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度包括：
判断所述原始问题所对应的问题分类；
在所述问答数据库中对应的问题分类下查找与所述原始问题相似度最高的N个典型问题，N为大于1的整数，所述问答数据库中存储有典型问题及典型问题对应的备选回复答案；
获取每一典型问题对应的备选回复答案，并计算每一备选回复答案的置信度。
作为一个示例，所述计算每一备选回复答案的置信度包括：
在所述备选回复答案为纯文本答案时，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j计算得到每一备选回复答案的置信度C ij；
在所述备选回复答案为模板类答案时，获取所述模板类答案中k个辅助信息类别中每一辅助信息类别的最高推荐度RAI k，其中k个最高推荐度中的最小值为min(RAI k)，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j*min(RAI k)计算得到每一备选回复答案的置信度C ij。
作为一个示例，所述方法还包括：
在人工客服回复所述原始问题后，获取人工客服回复的最终回复文本；或向所述客户回复所述置信度最高的备选回复答案作为最终回复文本，获取所述最终回复文本。
作为一个示例，获取所述最终回复文本之后，所述方法还包括：
判断客户是否继续输入原始问题；
若客户继续输入原始问题，转向所述接收客户输入的原始问题的步骤；
若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度大于第二阈值时，根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库。
作为一个示例，所述根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库包括：
对所述最终回复文本进行分词；
根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；
根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
作为一个示例，所述根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中包括：
根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；
若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
若所述最终回复文本未采纳所述问答数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题和实际回复内容。
作为另外一个示例，获取所述最终回复文本之后，所述方法还包括：
根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库；
判断客户是否继续输入原始问题；
若客户继续输入原始问题，转向所述接收客户输入的原始问题的步骤；
若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度不大于第二阈值时，丢弃问答训练缓存样本库以及辅助信息训练缓存样本库中缓存的样本。
其中，所述根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库包括：
对所述最终回复文本进行分词；
根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；
根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
作为一个示例，所述根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中包括：
根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；
若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
若所述最终回复文本未采纳所述问答数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题和实际回复内容。
进一步地，将问答训练缓存样本存储至所述问答训练缓存样本库中之后，所述方法还包括：
判断所述问答训练缓存样本库是否满足预设的更新条件，在所述问答训练 缓存样本库满足预设的更新条件时，利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新。
进一步地，所述利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新包括以下至少一种：
对所述问答数据库的原问题分类器进行再训练；
对所述问答数据库的原问题相似度算法进行再训练；
对所述问答数据库中的备选回复答案的推荐度进行更新。
进一步地，所述对所述问答数据库的原问题分类器进行再训练包括：
将所述问答训练缓存样本库中存储的每一问答训练缓存样本的原始问题和问题分类分别作为输入和输出对所述问答数据库的原问题分类器进行再训练。
进一步地，所述对所述问答数据库的原问题相似度算法进行再训练包括：
在所述问答训练缓存样本库中，选择问答训练缓存样本的原始问题以及匹配的典型问题这两组数据，对于相同的“原始问题-匹配的典型问题”的配对组合，统计其出现的频度；
以“原始问题-匹配的典型问题”的配对作为输入，以其频度作为输出，对所述问答数据库的原问题相似度算法的参数进行再训练。
进一步地，所述对所述问答数据库中的备选回复答案的推荐度进行更新包括：
在所述问答训练缓存样本库中，选择问答训练缓存样本的匹配的典型问题以及实际回复内容这两组数据；
对于每一个匹配的典型问题，统计其对应的各个实际回复内容的回复比例作为推荐度对所述问答数据库进行更新。
作为一个示例，所述根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中包括：
根据分词结果判断所述最终回复文本是否采纳所述辅助信息数据库中的辅助信息；
若所述最终回复文本采纳所述辅助信息数据库中的辅助信息，则根据所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；
若所述最终回复文本未采纳所述辅助信息数据库中的辅助信息，则根据所述最终回复文本生成所采纳的辅助信息，根据生成的所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；
其中，所述辅助信息训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题，辅助信息类别和辅助信息内容。
进一步地，将辅助信息训练样本存储至所述辅助信息训练缓存样本库之后，所述方法还包括：
判断所述辅助信息训练缓存样本库是否满足预设的更新条件，在所述辅助信息训练缓存样本库满足预设的更新条件时，利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新。
进一步地，所述利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新包括：
在所述辅助信息训练缓存样本库中，选择辅助信息训练样本的匹配的典型问题以及辅助信息内容这两组数据；
对于每一个匹配的典型问题，统计其对应的各个辅助信息内容的回复比例作为推荐度对所述辅助信息数据库进行更新。
其中，上述更新条件具体可以为所保存的样本数量大于第三阈值或达到预设的更新时间点。可以根据需要设置第三阈值的值，比如将第三阈值设置为1000、2000等等。预设的更新时间点比如可以为每周的某一天，或者每月的某一号等等。
实施例二
本发明实施例还提供了一种用于客户服务的装置，如图2所示，包括：
接收模块21，用于接收客户输入的原始问题；
判断模块22，用于从预先建立的问答数据库中获取多个对应所述原始问 题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
处理模块23，用于若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本实施例中，首先接收客户输入的原始问题，从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值，若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及辅助信息。这样不仅在会话开始时可以使用问答数据库中的备选回复答案，而且在转接入人工客服后，仍能利用问答数据库提供快速答复；另一方面，即使是当前人工客服所不熟悉的领域，仍能为人工客服提供多种辅助信息，协助人工客服进行快速回复，从而能够减轻人工客服的工作量，并能够提高问题回复的质量。
进一步地，所述判断模块22具体用于判断所述原始问题所对应的问题分类；在所述问答数据库中对应的问题分类下查找与所述原始问题相似度最高的N个典型问题，N为大于1的整数，所述问答数据库中存储有典型问题及典型问题对应的备选回复答案；获取每一典型问题对应的备选回复答案，并计算每一备选回复答案的置信度。
进一步地，所述判断模块22包括：
备选回复答案计算单元，用于在所述备选回复答案为纯文本答案时，获取 每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j计算得到每一备选回复答案的置信度C ij；在所述备选回复答案为模板类答案时，获取所述模板类答案中k个辅助信息类别中每一辅助信息类别的最高推荐度RAI k，其中k个最高推荐度中的最小值为min(RAI k)，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j*min(RAI k)计算得到每一备选回复答案的置信度C ij。
进一步地，如图2所示，所述装置还包括：
获取模块24，用于在人工客服回复所述原始问题后，获取人工客服回复的最终回复文本；或向所述客户回复所述置信度最高的备选回复答案作为最终回复文本，获取所述最终回复文本。
进一步地，所述装置还包括：
自学习训练样本缓存单元25，用于在获取人工客服回复的最终回复文本之后，判断客户是否继续输入原始问题；若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度大于第二阈值时，根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库。
进一步地，所述自学习训练样本缓存单元25具体用于对所述最终回复文本进行分词；根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
进一步地，所述自学习训练样本缓存单元包括：
问答训练缓存样本缓存单元，用于根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；若所述最终回复文本未采纳所述问答 数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题和实际回复内容。
作为另外一个示例，所述自学习训练样本缓存单元25，用于在获取人工客服回复的最终回复文本之后，根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库，并判断客户是否继续输入原始问题；若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度不大于第二阈值时，丢弃缓存的样本。
进一步地，所述自学习训练样本缓存单元25具体用于对所述最终回复文本进行分词；根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
进一步地，所述自学习训练样本缓存单元包括：
问答训练缓存样本缓存单元，用于根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；若所述最终回复文本未采纳所述问答数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题和实际回复内容。
进一步地，所述装置还包括：
问答数据库更新模块26，用于在将问答训练缓存样本存储至所述问答训练缓存样本库中之后，判断所述问答训练缓存样本库是否满足预设的更新条件，在所述问答训练缓存样本库满足预设的更新条件时，利用所述问答训练缓存样 本库中的问答训练缓存样本对所述问答数据库进行更新。
进一步地，所述问答数据库更新模块具体用于对所述问答数据库中的备选回复答案的推荐度进行更新；和/或对所述问答数据库的原问题分类器进行再训练；和/或对所述问答数据库的原问题相似度算法进行再训练。
进一步地，所述问答数据库更新模块包括：
问题分类器训练单元，用于将所述问答训练缓存样本库中存储的每一问答训练缓存样本的原始问题和问题分类分别作为输入和输出对所述问答数据库的原问题分类器进行再训练。
进一步地，所述问答数据库更新模块包括：
问题相似度算法训练单元，用于在所述问答训练缓存样本库中，选择问答训练缓存样本的原始问题以及匹配的典型问题这两组数据，对于相同的“原始问题-匹配的典型问题”的配对组合，统计其出现的频度；以“原始问题-匹配的典型问题”的配对作为输入，以其频度作为输出，对所述问答数据库的原问题相似度算法的参数进行再训练。
进一步地，所述问答数据库更新模块包括：
推荐度更新单元，用于在所述问答训练缓存样本库中，选择问答训练缓存样本的匹配的典型问题以及实际回复内容这两组数据；对于每一个匹配的典型问题，统计其对应的各个实际回复内容的回复比例作为推荐度对所述问答数据库进行更新。
进一步地，所述自学习训练样本缓存单元包括：
辅助信息训练缓存样本缓存单元，根据分词结果判断所述最终回复文本是否采纳所述辅助信息数据库中的辅助信息；若所述最终回复文本采纳所述辅助信息数据库中的辅助信息，则根据所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；若所述最终回复文本未采纳所述辅助信息数据库中的辅助信息，则根据所述最终回复文本生成所采纳的辅助信息，根据生成的所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；其中，所述辅助信息训练缓存样本包括：所述最终回复文 本对应的原始问题、问题分类、匹配的典型问题，辅助信息类别和辅助信息内容。
进一步地，所述装置还包括：
辅助信息数据库更新模块27，用于在将辅助信息训练样本存储至所述辅助信息训练缓存样本库之后，判断所述辅助信息训练缓存样本库是否满足预设的更新条件，在所述辅助信息训练缓存样本库满足预设的更新条件时，利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新。
所述辅助信息数据库更新模块包括：
推荐度更新单元，用于在所述辅助信息训练缓存样本库中，选择辅助信息训练样本的匹配的典型问题以及辅助信息内容这两组数据；对于每一个匹配的典型问题，统计其对应的各个辅助信息内容的回复比例作为推荐度对所述辅助信息数据库进行更新。
其中，上述更新条件具体可以为所保存的样本数量大于第三阈值或达到预设的更新时间点。可以根据需要设置第三阈值的值，比如将第三阈值设置为1000、2000等等。预设的更新时间点比如可以为每周的某一天，或者每月的某一号等等。
实施例三
本发明实施例还提供了一种用于客户服务的电子设备30，如图3所示，包括：
处理器32；和
存储器34，在所述存储器34中存储有计算机程序指令，
其中，在所述计算机程序指令被所述处理器运行时，使得所述处理器32执行以下步骤：
接收客户输入的原始问题；
从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第 一阈值；
若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本实施例中，首先接收客户输入的原始问题，从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值，若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及辅助信息。这样不仅在会话开始时可以使用问答数据库中的备选回复答案，而且在转接入人工客服后，仍能利用问答数据库提供快速答复；另一方面，即使是当前人工客服所不熟悉的领域，仍能为人工客服提供多种辅助信息，协助人工客服进行快速回复，从而能够减轻人工客服的工作量，并能够提高问题回复的质量。
进一步地，如图3所示，用于客户服务的电子设备30还包括网络接口31、输入设备33、硬盘35、和显示设备36。
上述各个接口和设备之间可以通过总线架构互连。总线架构可以是可以包括任意数量的互联的总线和桥。具体由处理器32代表的一个或者多个中央处理器(CPU)，以及由存储器34代表的一个或者多个存储器的各种电路连接在一起。总线架构还可以将诸如外围设备、稳压器和功率管理电路等之类的各种其它电路连接在一起。可以理解，总线架构用于实现这些组件之间的连接通信。总线架构除包括数据总线之外，还包括电源总线、控制总线和状态信号总线，这些都是本领域所公知的，因此本文不再对其进行详细描述。
所述网络接口31，可以连接至网络(如因特网、局域网等)，从网络中获取相关数据，例如客户输入的原始问题，并可以保存在硬盘35中。
所述输入设备33，可以接收操作人员输入的各种指令，并发送给处理器32以供执行。所述输入设备33可以包括键盘或者点击设备(例如，鼠标，轨迹球(trackball)、触感板或者触摸屏等。
所述显示设备36，可以将处理器32执行指令获得的结果进行显示。
所述存储器34，用于存储操作系统运行所必须的程序和数据，以及处理器32计算过程中的中间结果等数据。
可以理解，本发明实施例中的存储器34可以是易失性存储器或非易失性存储器，或可包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(ROM)、可编程只读存储器(PROM)、可擦除可编程只读存储器(EPROM)、电可擦除可编程只读存储器(EEPROM)或闪存。易失性存储器可以是随机存取存储器(RAM)，其用作外部高速缓存。本文描述的装置和方法的存储器34旨在包括但不限于这些和任意其它适合类型的存储器。
在一些实施方式中，存储器34存储了如下的元素，可执行模块或者数据结构，或者他们的子集，或者他们的扩展集：操作系统341和应用程序342。
其中，操作系统341，包含各种系统程序，例如框架层、核心库层、驱动层等，用于实现各种基础业务以及处理基于硬件的任务。应用程序342，包含各种应用程序，例如浏览器(Browser)等，用于实现各种应用业务。实现本发明实施例方法的程序可以包含在应用程序342中。
上述处理器32，当调用并执行所述存储器34中所存储的应用程序和数据，具体的，可以是接收客户输入的原始问题；从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅 助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
本发明上述实施例揭示的方法可以应用于处理器32中，或者由处理器32实现。处理器32可能是一种集成电路芯片，具有信号的处理能力。在实现过程中，上述方法的各步骤可以通过处理器32中的硬件的集成逻辑电路或者软件形式的指令完成。上述的处理器32可以是通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件，可以实现或者执行本发明实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。结合本发明实施例所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器34，处理器32读取存储器34中的信息，结合其硬件完成上述方法的步骤。
可以理解的是，本文描述的这些实施例可以用硬件、软件、固件、中间件、微码或其组合来实现。对于硬件实现，处理单元可以实现在一个或多个专用集成电路(ASIC)、数字信号处理器DSP)、数字信号处理设备(DSPD)、可编程逻辑设备(PLD)、现场可编程门阵列(FPGA)、通用处理器、控制器、微控制器、微处理器、用于执行本申请所述功能的其它电子单元或其组合中。
对于软件实现，可通过执行本文所述功能的模块(例如过程、函数等)来实现本文所述的技术。软件代码可存储在存储器中并通过处理器执行。存储器可以在处理器中或在处理器外部实现。
具体地，处理器32判断所述原始问题所对应的问题分类；在所述问答数据库中对应的问题分类下查找与所述原始问题相似度最高的N个典型问题，N为大于1的整数，所述问答数据库中存储有典型问题及典型问题对应的备选回复答案；获取每一典型问题对应的备选回复答案，并计算每一备选回复答案的置信度。
具体地，处理器32在所述备选回复答案为纯文本答案时，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j计算得到每一备选回复答案的置信度C ij；在所述备选回复答案为模板类答案时，获取所述模板类答案中k个辅助信息类别中每一辅助信息类别的最高推荐度RAI k，其中k个最高推荐度中的最小值为min(RAI k)，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j*min(RAI k)计算得到每一备选回复答案的置信度C ij。
具体地，处理器32在人工客服回复所述原始问题后，获取人工客服回复的最终回复文本；或向所述客户回复所述置信度最高的备选回复答案作为最终回复文本，获取所述最终回复文本。
具体地，处理器32判断客户是否继续输入原始问题；若客户继续输入原始问题，转向所述接收客户输入的原始问题的步骤；若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度大于第二阈值时，根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库。
具体地，处理器32对所述最终回复文本进行分词；根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
具体地，处理器32根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；若所述最终回复文本未采纳所述问答数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存 样本库中；其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题和实际回复内容。
具体地，处理器32判断所述问答训练缓存样本库是否满足预设的更新条件，在所述问答训练缓存样本库满足预设的更新条件时，利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新。
具体地，处理器32对所述问答数据库中的备选回复答案的推荐度进行更新；和/或对所述问答数据库的原问题分类器进行再训练；和/或对所述问答数据库的原问题相似度算法进行再训练。
具体地，处理器32将所述问答训练缓存样本库中存储的每一问答训练缓存样本的原始问题和问题分类分别作为输入和输出对所述问答数据库的原问题分类器进行再训练。
具体地，处理器32在所述问答训练缓存样本库中，选择问答训练缓存样本的原始问题以及匹配的典型问题这两组数据，对于相同的“原始问题-匹配的典型问题”的配对组合，统计其出现的频度；以“原始问题-匹配的典型问题”的配对作为输入，以其频度作为输出，对所述问答数据库的原问题相似度算法的参数进行再训练。
具体地，处理器32在所述问答训练缓存样本库中，选择问答训练缓存样本的匹配的典型问题以及实际回复内容这两组数据；对于每一个匹配的典型问题，统计其对应的各个实际回复内容的回复比例作为推荐度对所述问答数据库进行更新。
具体地，处理器32根据分词结果判断所述最终回复文本是否采纳所述辅助信息数据库中的辅助信息；若所述最终回复文本采纳所述辅助信息数据库中的辅助信息，则根据所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；若所述最终回复文本未采纳所述辅助信息数据库中的辅助信息，则根据所述最终回复文本生成所采纳的辅助信息，根据生成的所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；其中，所述辅助信息训练缓存样本包括：所述最终回复文本对应的原始 问题、问题分类、匹配的典型问题，辅助信息类别和辅助信息内容。
具体地，处理器32判断所述辅助信息训练缓存样本库是否满足预设的更新条件，在所述辅助信息训练缓存样本库满足预设的更新条件时，利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新。
具体地，处理器32在所述辅助信息训练缓存样本库中，选择辅助信息训练样本的匹配的典型问题以及辅助信息内容这两组数据；对于每一个匹配的典型问题，统计其对应的各个辅助信息内容的回复比例作为推荐度对所述辅助信息数据库进行更新。
实施例四
本发明实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器运行时，使得所述处理器执行以下步骤：
接收客户输入的原始问题；
从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
实施例五
下面结合附图对本发明的用于客户服务的方案进行进一步介绍。为了实现本发明的技术目的，本发明提出了一种客户服务辅助系统，如图4所示，包括：
问答数据库110，用于存储已知的问题及其备选回复答案列表。
辅助信息数据库120，用于存储进行客户服务时可能用到的各种参考信息，包括但不限于品牌、产品信息、产品口碑、客户基本数据、客户偏好、客户历史行为记录等等。
备选回复答案计算单元150，用于根据客户的问题，基于问答数据库110推算出一个或多个备选回复答案及其置信度。
人工客服辅助面板100，辅助人工客服回复客户问题的界面，根据辅助内容，可划分为多个辅助子面板。例如：
(1)备选回复答案子面板：用于按照优先级提供多个备选回复答案或者回复模板。
(2)辅助信息子面板：用于提供回复客户问题时可能用到的辅助信息。可以包括但不限于：
①基于原始问题中关键词的参考信息：从客户的原始问题中提取关键词，并列出相关信息。例如客户问题中提到了某产品名称，则可以给出该产品的文字介绍、规格、价格、优惠活动、产品口碑、相似产品列表等信息。
②客户的基本数据：客户的姓名或昵称、年龄、性别、地理位置、身高、体重、肤质等可以描述客户基础属性的信息。
③客户偏好数据：从客户的历史购买记录、产品评价中提取出的客户偏好信息，例如偏好产品的风格、关注的产品属性等。
④客户历史行为记录：如近期购买过的产品、操作、商品评价、近期服务记录等。
(3)人工客服回复输入栏：用于让人工客服参考以上备选回复答案子面板、辅助信息子面板，通过点击、选择、编辑，最终生成需要回复给客户的答案的输入栏。
自学习训练样本缓存单元160，根据人工客服的最终回复文本，更新问答训练缓存样本库130和辅助信息训练缓存样本库140。
自学习单元170，根据问答训练缓存样本库130和辅助信息训练缓存样本 库140中存储的样本以及客户的满意度调查结果，不断调整问答数据库110、辅助信息数据库120，并重新训练备选回复答案计算单元中算法，使系统的性能不断改善。
其中，问答数据库110中每个备选回复答案的储存格式如表1所示：
表1
Figure PCTCN2018116820-appb-000001
对于问题分类，问答数据库110可以预定义多个问题分类，便于在检索最接近问题时根据应用场景和问答上下文缩小搜索范围，提高匹配的速度和精度。可参考的问题分类有：产品介绍、产品推荐、质量投诉等。
对于典型问题，是指问题的文字内容。由于现有很多自然语言处理技术可以判断问题之间的相似度(例如语法相似度、编辑距离、语句向量相似度等，这里不再赘述)所以对于同一个典型问题可以有多种相似问法的情况，问答数据库110中只保存一个典型问题即可。
回复类型是指备选回复答案的类型，备选回复答案可分为两种：“纯文本答案”和“模板类答案”。纯文本答案即只有文字；模板类答案即回复中有部分关键词是可替换的词语，例如回复内容“到[省市名]的邮费是[价格]”属于一条模板类答案，其中“[省市名]”、“[价格]”这些方括号中的词语均是回复中可替换的词语，也对应着辅助信息数据库120中“辅助信息类别”这一栏里的信息。而可以被替换到[]位置的词语，则对应着辅助信息数据库120中“辅 助信息内容”这一栏里的信息。(此处的[省市名]、[价格]只是示例，在实际应用中可以是更复杂的分类，例如多级分类。)每个典型问题可以对应于多条回复内容，即同一个问题在问答数据库110中可以储存多个对应的备选回复答案。
最近使用时间戳记录了该条备选回复答案最近一次被选择使用的时间戳。
客服总体推荐度：根据所有客服的选择结果计算出的该备选回复答案的总体推荐度。
辅助信息数据库120中每条辅助信息的储存格式如表2所示：
表2
Figure PCTCN2018116820-appb-000002
问答训练缓存样本库中存储的问答训练缓存样本的格式如表3所示：
表3
Figure PCTCN2018116820-appb-000003
辅助信息训练缓存样本库中存储的辅助信息训练样本的格式如表4所示：
表4
Figure PCTCN2018116820-appb-000004
Figure PCTCN2018116820-appb-000005
如图5所示，本实施例的用于客户服务的方法具体包括以下步骤：
步骤401：客户提出原始问题；
每当客户提出一个问题(可以通过文本输入或语音输入)时，如果输入的问题非文本格式，可以通过格式转换工具转换为文本。文本格式的客户问题以下简称为“原始问题”。
步骤402：判断备选回复答案的置信度是否足够高，如果足够高，转向步骤403；如果不足够高，转向步骤404；
备选回复答案计算单元150通过问题相似度的计算，得到最高置信度的备选回复答案。
如图6所示，得到最高置信度的备选回复答案的过程包括以下步骤：
步骤601：判断所述原始问题所对应的问题分类；
首先根据当前原始问题及最近的聊天记录等信息判断问题的分类，以缩小搜索范围。判断问题的分类可以简单地通过一些规则、模板判断，也可以使用通过大量预先标注过的文本经过机器学习训练的“问题分类器”进行判断。
步骤502：在所述问答数据库中对应的问题分类下查找与所述原始问题相似度最高的N个典型问题；
在对应的问题分类下，通过问题匹配，在问答数据库110中找到与原始问题相似度最高的N个典型问题(假设为Q 1，Q 2，…，Q N)，并分别得到问题相似度数值(假设原始问题与它们的相似度分别为S 1，S 2，…，S N)。问题相似度的计算方法可以使用编辑距离、特征向量余弦相似度等多种现有算法中的某一种算法，也可以同时使用多种相似度算法，并取相似度数值最大的一组匹配结果(下文简称“综合相似度算法”)。在筛选相似度最高的N个典型问题时，还需要参考一个最低可接受的相似度阈值，假设为S th，那么对于每一个S i(i＝1，2，…，N)都需要满足S i≥S th。
步骤503：获取每一典型问题对应的备选回复答案；
针对相似度最高的N个典型问题，每一个典型问题又对应于多条备选回复答案，并有各自的相似度。假设第i个相似度最高的典型问题(i＝1，2，…，N)中，有M条备选回复答案(假设为A 1，A 2，…，A M)，它们的推荐度分别为R 1，R 2，…，R M。
步骤504：计算每一备选回复答案的置信度；
那么，对于第i个典型问题的第j个备选回复答案，计算器置信度C ij。计算方法为：如果该备选回复答案是纯文本答案，可以计算其置信度C ij为：C ij＝S i*R j；如果该备选回复答案是模板类答案，假设模板中出现了k个辅助信息类别，且每个辅助信息类别在辅助信息数据库120中相同的典型问题及辅助信息类别下所对应的最高推荐度的辅助信息的客服总体推荐度为RAI k，则可以计算该模板类答案的置信度C ij为：C ij＝S i*R j*min(RAI k)，其中min(RAI k)表示k个辅助信息类别所对应的RAI k中的最小值；
步骤505：按照置信度的从高到低对所有备选回复答案进行排序。
这样就可以得到最高置信度的备选回复答案或备选回复答案列表。
步骤403：向客户展示备选回复答案；
如果最高置信度的一个备选回复答案其置信度C大于等于某一个阈值C th，则认为该备选回复答案足够可信，可以直接向客户回复该备选回复答案。
步骤404：转入由人工客服回答；
如果最高置信度的一个备选回复答案其置信度C小于某一个阈值C th，则需要由人工客服来回答，但系统会给人工客服提示多种辅助信息，帮助其快速的完成对客户的回复。
系统给人工客服提供的辅助信息包括备选回复答案列表，及与问题相关的辅助信息等，这些信息显示在备选回复答案子面板和辅助信息子面板上。
备选回复答案子面板：显示按照置信度从高到低排序的备选回复答案列表。
辅助信息子面板：根据实际需求，辅助信息子面板可以由多个分类面板组成，如依据客户问题中的关键词给出的参考信息、客户基本数据、客户偏好、客户历史行为等等。每个分类面板可以是“可自学更新的分类面板”和“其他 分类面板”两种。其中“可自学更新的分类面板”里给出的辅助信息是可以根据人工客服的最终回复文本不断学习并更新推荐度的，而“其他分类面板”可以是由其他已有推荐算法给出的辅助信息，比如客户基本数据、客户偏好、客户历史行为、针对客户的个性化推荐内容等。其中“可自学更新的分类面板”中显示的辅助信息为本发明重点关注的内容，为简化说明，下文假设“辅助信息子面板”中仅包括“可自学更新的分类面板”。“其他分类面板”中显示的辅助信息可以由多种方式生成、排序及更新，不在本发明重点讨论范围之内，但本发明所给出的关于“训练缓存样本库”与“问答/辅助信息数据库”的更新的方法也可以作为“其他分类面板”更新的参考。每个分类面板代表一种“辅助信息类别”(对应于辅助信息数据库120中的“辅助信息类别”，如果辅助信息类别有多级分类，则可代表其中的第一级分类)，显示按照置信度从高到低排序的备选回复答案列表。(辅助信息的置信度与备选回复答案的置信度计算方法相似。)
人工客服根据上述辅助信息，在人工客服回复输入栏中最终编辑完成给客户的回复(下面简称“最终回复文本”)。
步骤405：自学习训练样本缓存单元160根据最终回复文本，更新问答训练缓存样本库和辅助信息训练缓存样本库；
如图7所示，更新问答训练缓存样本库和辅助信息训练缓存样本库包括以下步骤：
步骤601：获取最终回复文本所采纳的备选回复答案及辅助信息；
初始状态下，同一个典型问题的所有备选回复答案、辅助信息可由人工或特定算法设置不同的初始推荐度，也可以默认设定相同的初始推荐度。
当某人工客服在人工客服输入栏编辑完最终回复文本并确认发出后，首先由自学习训练样本缓存单元160判断此最终回复文本中具体采用了哪些备选回复答案文本/模板、辅助信息——即有哪些回复文本/模板、辅助信息需要更新其在知识库中的推荐度。
如图8所示，获取最终回复文本所采纳的备选回复答案及辅助信息包括以 下步骤：
步骤701：对最终回复文本进行分词；
对最终回复文本进行分词，即将每条语句分割成词语。
步骤702：根据分词结果初步筛选可能采纳的备选回复答案及辅助信息；
根据回答客户的原始问题时面板上曾经出现过的辅助信息以及客服的点击历史记录，初步筛选哪些备选回复答案、辅助信息(词语)可能被用到了最终回复文本中。
步骤703：将初步筛选结果与最终回复文本进行比对，确定其中被采纳的备选回复答案及辅助信息；
将初步筛查的结果，与最终回复文本进行对比，确认其是否最终被使用到了最终回复文本中——这些备选回复答案、辅助信息称为“被采纳的”备选回复答案/辅助信息，它们的推荐度需要被新增或更新到数据库中。
步骤704：判断最终回复文本中是否还有其他辅助信息，如果是，转向步骤705；如果否，转向步骤706；
除了步骤703中已识别的备选回复答案和辅助信息以外，继续判断最终回复文本中是否有其他属于辅助信息数据库120的内容，虽然没在面板上作为辅助信息出现，但最终被应用到了客服的回复中——这些辅助信息如果已经在此问题对应的辅助信息数据库120中，则需要提高其推荐度；如果不在此问题对应的辅助信息数据库120中，则需要新增。
步骤705：追加到被采纳的辅助信息中；
步骤706：是否找到最终回复文本所对应的备选回复答案，如果是，则结束；如果否，转向步骤707；
步骤707：生成被采纳的备选回复答案。
如果经过上述步骤，未能找到最终回复文本所对应的问答数据库110中的备选回复答案，则根据已有信息生成一个“被采纳的”备选回复答案。生成方法是：在“最终回复文本”中，将步骤703和步骤704中识别出来的“辅助信息内容”用相应的“辅助信息类别”做替换，即得到一个“被采纳的回复模板”； 如果最终回复文本中未包含任何已知的“辅助信息内容”，则直接将最终回复文本视为一条“被采纳的回复文本”。判断该被采纳的回复文本或者回复模板是否已存在于问答数据库110中。如果已存在，则需要更新其推荐度。如果不存在，则意味着可能需要在问答数据库110新增一条回复内容或者回复模板。
对于那些本次最终回复文本中“被采纳的”备选回复答案的文本/回复模板、辅助信息，将这些信息存储到问答训练缓存样本库及辅助信息训练缓存样本库中。其中，“匹配的典型问题”这一栏填写最终回复文本所采用的备选回复答案所对应的“典型问题”。如果最终回复文本未使用已有的备选回复答案，则新增一条“匹配的典型问题”，内容与“原始问题”相同。“问题分类”这一栏填写的是“匹配的典型问题”在问答数据库110中所对应的“问题分类”，如果“匹配的典型问题”是新增的条目，则“问题分类”可以统一设定为“待定分类”，或者根据“原始问题”和“问题分类器”对其进行分类。问答训练缓存样本库中的“实际回复内容”填写的是“最终回复文本”或“最终回复模板”。
辅助信息训练缓存样本库的“辅助信息类别”填写的是被采纳的辅助信息所对应的“辅助信息类别”。具体填写方法如下：如果在辅助信息数据库120中，与此次对话中原始问题所匹配的“典型问题”下，恰好有与此次被采纳的辅助信息相同的条目，则直接采用其对应的“辅助信息类别”填写到辅助信息训练缓存样本库的“辅助信息类别”中；如果在辅助信息数据库120中，与此次对话中原始问题所匹配的“典型问题”下，没有与此次被采纳的辅助信息相同的条目。则按照与原始问题的匹配度从高到低的顺序再查找其他“典型问题”中是否储存了与“被采纳的辅助信息”相同的条目，如果找到了，则读取其“辅助信息类别”填写到辅助信息训练缓存样本库的“辅助信息类别”中；如果在辅助信息数据库120中，没有与此次被采纳的辅助信息相同的条目，则“辅助信息类别”设置为“待定分类”。
问答训练缓存样本库及辅助信息训练缓存样本库中，填写为“待定分类”的栏目，可以后期由人工或机器学习的方式来进行追加标注，修改为分类结果。
步骤602：根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练缓存样本并存储至辅助信息训练缓存样本库中。
步骤406：判断客户是否还有下一个问题，如果有，则转向步骤401；如果没有下一个问题，则转向步骤407；
步骤407：判断客户对最终回复文本是否满意；如果满意，转向步骤408；如果不满意，转向步骤409；
如果客户对最终回复文本满意，则保留此次会话中所有的训练缓存样本。
步骤408：判断是否已积累足够多的训练样本，如果是，转向步骤410；如果否，则结束。
步骤409：如果客户对最终回复文本不满意，则丢弃此次会话中所有的训练缓存样本。
步骤410：自学习单元进行数据库的更新和备选回复答案计算单元的再训练。
每隔一段预设的时间，或者每当问答训练缓存样本库和辅助信息训练缓存样本库中的新增样本数量累积到一个阈值，则自学习单元170利用增量的训练样本对问答数据库110和辅助信息数据库120进行推荐度的更新，以及备选回复答案计算单元的再训练。
其中，备选回复答案计算单元150的再训练包括：
对“问题分类器”的再训练：在累积的问答训练缓存样本库中，选择原始问题、问题分类这两列数据，分别作为输入和输出，对已有的“问题分类器”进行再训练。对于问题分类为“待定分类”的样本，可以采用人工的方法对其进行标注后再参加训练。
对“问题相似度计算”的再训练：在计算原始问题与典型问题的相似度时，如果采用的是多种算法结合的“综合相似度算法”，则可在其中追加一种相似度计算方法，该方法通过机器学习训练样本，采用参数回归的方法推算“原始问题”与“典型问题”的相似度(下文简称“统计相似度回归算法”)。具体方 法为：
在累积的问答训练缓存样本库中，选择原始问题、匹配的典型问题这两列数据，对于相同文本内容的“原始问题”-“匹配的典型问题”的配对组合，统计其出现的频度(下文简称“配对频度”)。计算方法为：假设在问答训练缓存样本库中，同一个“原始问题”总共有L个“匹配的典型问题”配对结果，且这L个“原始问题”-“匹配的典型问题”组合出现的次数分别为N 1，N 2，…，N L。那么对于“原始问题”-第k个“匹配的典型问题”组合，其配对频度为：
Figure PCTCN2018116820-appb-000006
以原始问题、匹配的典型问题作为输入特征，以配对频度为输出结果。对“统计相似度回归算法”的参数进行重新训练。经过训练后的“统计相似度回归算法”可以作为“综合相似度算法”中的其中一种算法，参与到问题相似度的计算中。
对于问答数据库中推荐度的更新，在累积的问答训练缓存样本库中，选择匹配的典型问题、实际回复内容这两列数据。针对每一个“匹配的典型问题”，统计各个“实际回复内容”的回复比例作为“客服总体推荐度”对问答数据库110进行更新。具体方法为：假设在累积的问答训练缓存样本库中，针对某个“匹配的典型问题”，总共有P个“实际回复内容”与之对应，且这P个“匹配的典型问题”-“实际回复内容”组合出现的次数分别为M 1，M 2，…，M P。那么对于“原始问题”-第j个“匹配的典型问题”组合，其回复比例为：
Figure PCTCN2018116820-appb-000007
在问答数据库110中，更新相应的“典型问题”-“回复内容”所对应的“客服总体推荐度”这一栏的数值为其回复比例R j。
辅助信息数据库120的推荐度的更新方法与问答数据库中推荐度的更新类似，在此不再赘述。
本实施例中，能够提高备选回复答案的客服机器人及问答数据库不是仅在会话开始阶段有用，而是有机会用于客户的每一个问题。对于备选回复答案置信度高的问题，可以直接由客服机器人回答；而对于备选回复答案置信度不高 的问题，仍可以按照置信度高低为人工客服列出，作为回复的参考模板供选择。大大减轻了人工客服的工作量。
另外，备选回复答案子面板和辅助信息子面板可以给人工客服提供多方面的回复素材，即使客户提了人工客服不太熟悉的业务的问题，仍然可以利用问答数据库、辅助信息数据库帮助人工客服有效地给出回答。通过最终人工客服的操作(例如生成的最终答案时所选择的备选回复答案、辅助信息)以及事后的客户满意度调查，系统可以对已有的数据库进行不断更新、学习，提升系统性能。
以上是本发明的优选实施方式，应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明所述原理的前提下，还可以作出若干改进和润饰，这些改进和润饰也应视为本发明的保护范围。

Claims (19)

  1. 一种用于客户服务的方法，其特征在于，包括：
    接收客户输入的原始问题；
    从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
    若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
  2. 根据权利要求1所述的用于客户服务的方法，其特征在于，所述从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度包括：
    判断所述原始问题所对应的问题分类；
    在所述问答数据库中对应的问题分类下查找与所述原始问题相似度最高的N个典型问题，N为大于1的整数，所述问答数据库中存储有典型问题及典型问题对应的备选回复答案；
    获取每一典型问题对应的备选回复答案，并计算每一备选回复答案的置信度。
  3. 根据权利要求2所述的用于客户服务的方法，其特征在于，所述计算每一备选回复答案的置信度包括：
    在所述备选回复答案为纯文本答案时，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式 C ij＝S i*R j计算得到每一备选回复答案的置信度C ij；
    在所述备选回复答案为模板类答案时，获取所述模板类答案中k个辅助信息类别中每一辅助信息类别的最高推荐度RAI k，其中k个最高推荐度中的最小值为min(RAI k)，获取每一备选回复答案对应的典型问题与所述原始问题的相似度S i以及每一备选回复答案的推荐度R j，通过公式C ij＝S i*R j*min(RAI k)计算得到每一备选回复答案的置信度C ij。
  4. 根据权利要求1所述的用于客户服务的方法，其特征在于，所述方法还包括：
    在人工客服回复所述原始问题后，获取人工客服回复的最终回复文本；或向所述客户回复所述置信度最高的备选回复答案作为最终回复文本，获取所述最终回复文本；或向所述客户回复所述置信度最高的备选回复答案作为最终回复文本，获取所述最终回复文本。
  5. 根据权利要求4所述的用于客户服务的方法，其特征在于，获取所述最终回复文本之后，所述方法还包括：
    判断客户是否继续输入原始问题；
    若客户继续输入原始问题，转向所述接收客户输入的原始问题的步骤；
    若客户不再输入原始问题，获取客户对所述最终回复文本的满意度，在所述客户对所述最终回复文本的满意度大于第二阈值时，根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库。
  6. 根据权利要求5所述的用于客户服务的方法，其特征在于，所述根据所述最终回复文本更新问答训练缓存样本库以及辅助信息训练缓存样本库包括：
    对所述最终回复文本进行分词；
    根据分词结果确定所述最终回复文本所采纳的备选回复答案和辅助信息；
    根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中，根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中。
  7. 根据权利要求6所述的用于客户服务的方法，其特征在于，所述根据所述最终回复文本所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中包括：
    根据分词结果判断所述最终回复文本是否采纳所述问答数据库中的备选回复答案；
    若所述最终回复文本采纳所述问答数据库中的备选回复答案，则根据所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
    若所述最终回复文本未采纳所述问答数据库中的备选回复答案，则根据所述最终回复文本生成所采纳的备选回复答案，根据生成的所采纳的备选回复答案生成问答训练缓存样本并存储至所述问答训练缓存样本库中；
    其中，所述问答训练缓存样本包括：所述最终回复文本对应的原始问题、问题分类、匹配的典型问题和实际回复内容。
  8. 根据权利要求7所述的用于客户服务的方法，其特征在于，将问答训练缓存样本存储至所述问答训练缓存样本库中之后，所述方法还包括：
    判断所述问答训练缓存样本库是否满足预设的更新条件，在所述问答训练缓存样本库满足预设的更新条件时，利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新。
  9. 根据权利要求8所述的用于客户服务的方法，其特征在于，所述利用所述问答训练缓存样本库中的问答训练缓存样本对所述问答数据库进行更新包括以下至少一种：
    对所述问答数据库中的备选回复答案的推荐度进行更新；
    对所述问答数据库的原问题分类器进行再训练；
    对所述问答数据库的原问题相似度算法进行再训练。
  10. 根据权利要求9所述的用于客户服务的方法，其特征在于，所述对所述问答数据库的原问题分类器进行再训练包括：
    将所述问答训练缓存样本库中存储的每一问答训练缓存样本的原始问题 和问题分类分别作为输入和输出对所述问答数据库的原问题分类器进行再训练。
  11. 根据权利要求9所述的用于客户服务的方法，其特征在于，所述对所述问答数据库的原问题相似度算法进行再训练包括：
    在所述问答训练缓存样本库中，选择问答训练缓存样本的原始问题以及匹配的典型问题这两组数据，对于相同的“原始问题-匹配的典型问题”的配对组合，统计其出现的频度；
    以“原始问题-匹配的典型问题”的配对作为输入，以其频度作为输出，对所述问答数据库的原问题相似度算法的参数进行再训练。
  12. 根据权利要求9所述的用于客户服务的方法，其特征在于，所述对所述问答数据库中的备选回复答案的推荐度进行更新包括：
    在所述问答训练缓存样本库中，选择问答训练缓存样本的匹配的典型问题以及实际回复内容这两组数据；
    对于每一个匹配的典型问题，统计其对应的各个实际回复内容的回复比例作为推荐度对所述问答数据库进行更新。
  13. 根据权利要求6所述的用于客户服务的方法，其特征在于，所述根据所述最终回复文本所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中包括：
    根据分词结果判断所述最终回复文本是否采纳所述辅助信息数据库中的辅助信息；
    若所述最终回复文本采纳所述辅助信息数据库中的辅助信息，则根据所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；
    若所述最终回复文本未采纳所述辅助信息数据库中的辅助信息，则根据所述最终回复文本生成所采纳的辅助信息，根据生成的所采纳的辅助信息生成辅助信息训练样本并存储至所述辅助信息训练缓存样本库中；
    其中，所述辅助信息训练缓存样本包括：所述最终回复文本对应的原始问 题、问题分类、匹配的典型问题，辅助信息类别和辅助信息内容。
  14. 根据权利要求13所述的用于客户服务的方法，其特征在于，将辅助信息训练样本存储至所述辅助信息训练缓存样本库之后，所述方法还包括：
    判断所述辅助信息训练缓存样本库是否满足预设的更新条件，在所述辅助信息训练缓存样本库满足预设的更新条件时，利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新。
  15. 根据权利要求14所述的用于客户服务的方法，其特征在于，所述利用所述辅助信息训练缓存样本库中的辅助信息训练样本对所述辅助信息数据库进行更新包括：
    在所述辅助信息训练缓存样本库中，选择辅助信息训练样本的匹配的典型问题以及辅助信息内容这两组数据；
    对于每一个匹配的典型问题，统计其对应的各个辅助信息内容的回复比例作为推荐度对所述辅助信息数据库进行更新。
  16. 根据权利要求8或14所述的用于客户服务的方法，其特征在于，
    所述更新条件为所保存的样本数量大于第三阈值或达到预设的更新时间点。
  17. 一种用于客户服务的装置，其特征在于，包括：
    接收模块，用于接收客户输入的原始问题；
    判断模块，用于从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
    处理模块，用于若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户 偏好、客户历史行为。
  18. 一种用于客户服务的电子设备，其特征在于，包括：
    处理器；和
    存储器，在所述存储器中存储有计算机程序指令，
    其中，在所述计算机程序指令被所述处理器运行时，使得所述处理器执行以下步骤：
    接收客户输入的原始问题；
    从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
    若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
  19. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器运行时，使得所述处理器执行以下步骤：
    接收客户输入的原始问题；
    从预先建立的问答数据库中获取多个对应所述原始问题的备选回复答案及相应的置信度，判断其中置信度最高的备选回复答案的置信度是否不小于第一阈值；
    若置信度最高的备选回复答案的置信度不小于第一阈值，则向所述客户回复所述置信度最高的备选回复答案作为最终回复文本；若置信度最高的备选回复答案的置信度小于第一阈值，则向聊天服务器发送要求人工客服接入的服务 请求，并向所述人工客服展现所述多个对应所述原始问题的备选回复答案以及预先建立的辅助信息数据库中的辅助信息，所述辅助信息包括以下信息中的至少一种：所述原始问题中的关键词参考信息、客户数据、客户偏好、客户历史行为。
PCT/CN2018/116820 2017-12-14 2018-11-22 用于客户服务的方法、装置、电子设备、计算机可读存储介质 WO2019114512A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711339712.6 2017-12-14
CN201711339712.6A CN110069607B (zh) 2017-12-14 2017-12-14 用于客户服务的方法、装置、电子设备、计算机可读存储介质

Publications (1)

Publication Number Publication Date
WO2019114512A1 true WO2019114512A1 (zh) 2019-06-20

Family

ID=66818924

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/116820 WO2019114512A1 (zh) 2017-12-14 2018-11-22 用于客户服务的方法、装置、电子设备、计算机可读存储介质

Country Status (2)

Country Link
CN (1) CN110069607B (zh)
WO (1) WO2019114512A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113672719A (zh) * 2021-09-08 2021-11-19 中国平安人寿保险股份有限公司 对话辅助信息推送方法、装置、计算机设备及存储介质

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110570215A (zh) * 2019-09-06 2019-12-13 北京金融资产交易所有限公司 智能客服系统
CN110795546A (zh) * 2019-09-12 2020-02-14 易谷网络科技股份有限公司 一种基于智能客服的收敛方法、装置及电子设备
CN113127613B (zh) * 2020-01-10 2024-01-09 北京搜狗科技发展有限公司 聊天信息处理方法及装置
CN111666394A (zh) * 2020-04-30 2020-09-15 北京捷通华声科技股份有限公司 一种文本处理方法及装置
CN112040075A (zh) * 2020-08-28 2020-12-04 厦门羽星科技股份有限公司 基于语音识别的坐席语音辅助系统
CN112527972A (zh) * 2020-12-25 2021-03-19 东云睿连(武汉)计算技术有限公司 基于深度学习的智能客服聊天机器人实现方法和系统
CN112948403B (zh) * 2021-01-25 2023-08-08 北京房江湖科技有限公司 一种用于自动更新房源数据的方法和装置
CN113420137A (zh) * 2021-06-29 2021-09-21 山东新一代信息产业技术研究院有限公司 基于端到端框架的智能问答系统实现方法、设备及介质
CN113282737B (zh) * 2021-07-21 2021-11-12 中信建投证券股份有限公司 人机协作的智能客服对话方法及装置
CN113626569A (zh) * 2021-08-02 2021-11-09 浪潮软件股份有限公司 一种基于政务事项与历史数据的便民办事方法及系统
CN114925192A (zh) * 2022-07-21 2022-08-19 北京聆心智能科技有限公司 一种人机协同对话方法、装置、设备及存储介质
CN115577088B (zh) * 2022-11-18 2023-11-03 易点生活电子商务有限公司 一种对聊天信息进行分配的技术开发咨询系统
CN117041618B (zh) * 2023-10-10 2024-02-06 北京装库创意科技有限公司 一种用于电商的智能语音客服方法和系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105591882A (zh) * 2015-12-10 2016-05-18 北京中科汇联科技股份有限公司 一种智能机器人与人混合客服的方法及系统
CN106683678A (zh) * 2016-11-30 2017-05-17 厦门快商通科技股份有限公司 一种人工电话客服辅助系统及方法
CN106874441A (zh) * 2017-02-07 2017-06-20 腾讯科技（上海）有限公司 智能问答方法和装置
US20170337261A1 (en) * 2014-04-06 2017-11-23 James Qingdong Wang Decision Making and Planning/Prediction System for Human Intention Resolution

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102043793A (zh) * 2009-10-09 2011-05-04 卢健华 一种面向知识服务的推荐方法
US20120084112A1 (en) * 2010-09-24 2012-04-05 International Business Machines Corporation Providing community for customer questions
CN104809197A (zh) * 2015-04-24 2015-07-29 同程网络科技股份有限公司 基于智能机器人的在线问答方法
CN105389358A (zh) * 2015-11-04 2016-03-09 浙江工商大学 基于关联规则的Web服务推荐方法
CN105373940A (zh) * 2015-11-12 2016-03-02 江南大学 智能商场信息系统
US10572516B2 (en) * 2015-12-15 2020-02-25 [24]7.ai, Inc. Method and apparatus for managing natural language queries of customers
CN105512349B (zh) * 2016-02-23 2019-03-26 首都师范大学 一种用于学习者自适应学习的问答方法及装置
CN105808694B (zh) * 2016-03-04 2019-05-31 上海携程商务有限公司 在线客服应答系统及方法
CN106409283B (zh) * 2016-08-31 2020-01-10 上海交通大学 基于音频的人机混合交互系统及方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170337261A1 (en) * 2014-04-06 2017-11-23 James Qingdong Wang Decision Making and Planning/Prediction System for Human Intention Resolution
CN105591882A (zh) * 2015-12-10 2016-05-18 北京中科汇联科技股份有限公司 一种智能机器人与人混合客服的方法及系统
CN106683678A (zh) * 2016-11-30 2017-05-17 厦门快商通科技股份有限公司 一种人工电话客服辅助系统及方法
CN106874441A (zh) * 2017-02-07 2017-06-20 腾讯科技（上海）有限公司 智能问答方法和装置

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113672719A (zh) * 2021-09-08 2021-11-19 中国平安人寿保险股份有限公司 对话辅助信息推送方法、装置、计算机设备及存储介质

Also Published As

Publication number Publication date
CN110069607B (zh) 2024-03-05
CN110069607A (zh) 2019-07-30

Similar Documents

Publication Publication Date Title
WO2019114512A1 (zh) 用于客户服务的方法、装置、电子设备、计算机可读存储介质
US9646079B2 (en) Method and apparatus for identifiying similar questions in a consultation system
CN110134765B (zh) 一种基于情感分析的餐厅用户评论分析系统及方法
CN109299245B (zh) 知识点召回的方法和装置
CN111641514A (zh) 电子会议智能
US20140314311A1 (en) System and method for classification with effective use of manual data input
JP2017504883A (ja) オーディオキューに基づくモデル駆動型の候補者のソート
US11314790B2 (en) Dynamic field value recommendation methods and systems
US11120268B2 (en) Automatically evaluating caption quality of rich media using context learning
US10496751B2 (en) Avoiding sentiment model overfitting in a machine language model
US20160292163A1 (en) Proactive identification of content items for a member of a social network
CN115526590B (zh) 一种结合专家知识和算法的高效人岗匹配与复推方法
US11790894B2 (en) Machine learning based models for automatic conversations in online systems
WO2022218139A1 (zh) 融合注意力机制的个性化搜索方法和搜索系统
CN112417848A (zh) 语料生成方法、装置及计算机设备
WO2020135054A1 (zh) 视频推荐方法、装置、设备及存储介质
JP2023017921A (ja) コンテンツ推薦とソートモデルトレーニング方法、装置、機器、記憶媒体及びコンピュータプログラム
US11550830B2 (en) Systems and methods for multi-source reference class identification, base rate calculation, and prediction
US20210097424A1 (en) Dynamic selection of features for training machine learning models
US20230085599A1 (en) Method and device for training tag recommendation model, and method and device for obtaining tag
US11921754B2 (en) Systems and methods for categorization of ingested database entries to determine topic frequency
CN111046151B (zh) 一种消息处理方法及装置
Imam Curriculum coherence when subject-specific standards are absent: a case study using coursework-based master of finance programs at Australian universities
JP7452801B2 (ja) 評価支援システム及び評価支援プログラム
US20240160642A1 (en) Systems and methods for categorization of ingested database entries to determine topic frequency

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: 18889589

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: 18889589

Country of ref document: EP

Kind code of ref document: A1