WO2020006835A1

WO2020006835A1 - Customer service method, apparatus, and device for engaging in multiple rounds of question and answer, and storage medium

Info

Publication number: WO2020006835A1
Application number: PCT/CN2018/102053
Authority: WO
Inventors: 于凤英; 王健宗; 肖京
Original assignee: 平安科技（深圳）有限公司
Priority date: 2018-07-03
Filing date: 2018-08-24
Publication date: 2020-01-09
Also published as: CN109727041A; CN109727041B

Abstract

A customer service method, apparatus, and device for engaging in multiple rounds of question and answer, and a storage medium. The method comprises: obtaining a target question from a user in a current round (S10); searching a knowledge base for a standard intent matching the target question, and obtaining a degree of matching between the standard intent and the target question (S20); determining whether the degree of matching between the standard intent and the target question exceeds a preset matching threshold (S30); if not, performing named entity recognition on the target question and obtaining a recognition result (S40); detecting whether there is a signal inherited from previous rounds of questions at least one round before the current round, and obtaining a detection result (S50); and determining a target answer on the basis of the recognition result, the detection result, the target question, and the standard intent (S60).

Description

Multi-round question and answer method, equipment, storage medium and device for intelligent customer service

This application claims the priority of a Chinese patent application submitted to the Chinese Patent Office on July 3, 2018, with application number 201810722735.3, and the invention name is "Multi-round question answering method, equipment, storage medium and device for intelligent customer service". Citations are incorporated in the application.

Technical field

The present application relates to the technical field of intelligent customer service, and in particular, to a method, device, storage medium and device for multiple rounds of question and answer for intelligent customer service.

Background technique

With the development of science and technology, the intelligent customer service system is getting more and more attention. When there are multiple rounds of questions and answers between users and intelligent customer service, the next few questions may omit the named entity or signal, which is the signal from the previous rounds of questions The inherited user intentions cause the problem to be ambiguous due to the lack of a named entity or signal. When the intelligent customer service system conducts multiple rounds of questions and answers with the user, it is prone to fail to determine the substance of the problem and fail to respond accordingly. , Resulting in poor user experience. Therefore, how to identify the substance of user questions in multiple rounds of Q & A to improve the accuracy of responses in multiple rounds of Q & A is a technical issue that needs to be solved urgently.

Summary of the invention

The main purpose of this application is to provide a multi-round question and answer method, device, storage medium and device for intelligent customer service, which aims to solve the technical problem of low accuracy in the multi-round question and answer in the prior art.

In order to achieve the above purpose, the present application provides a multi-round question answering method for intelligent customer service. The multi-round question answering method for intelligent customer service includes the following steps: obtaining a target question of the current round of the user; and searching from the knowledge base for a match with the target question. Standard intent, and obtain the degree of matching between the standard intent and the target problem; determine whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold; if it does not exceed, then The target problem is identified by a named entity to obtain a recognition result; detecting whether there is a signal inherited in at least one round of user problems before the current round, and obtaining a detection result; according to the recognition result, the detection result, and the target problem And the criteria are intended to determine the target response.

In addition, in order to achieve the above purpose, this application also proposes a multi-round question answering device for intelligent customer service. The multi-round question answering device for intelligent customer service includes a memory, a processor, and a processor that is stored on the memory and can run on the processor. The intelligent customer service multi-round question and answer readable instruction is configured as the steps of the intelligent customer service multi-round question and answer method described above.

In addition, in order to achieve the above-mentioned object, the present application also proposes a storage medium that stores intelligent customer service multiple rounds of question and answer readable instructions stored on the storage medium. The intelligent customer service multiple rounds of question and answer readable instructions are implemented by the processor as described above. The steps of the multi-round question answering method for intelligent customer service.

In addition, in order to achieve the above purpose, this application also proposes a multi-round question answering device for intelligent customer service, including: an acquisition module for acquiring a target question of a user's current round; a search module for searching for the target from a knowledge base A standard intent for question matching, and obtaining a degree of matching between the standard intent and the target problem; a judgment module, configured to determine whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold; An identification module for identifying a named entity for the target problem to obtain an identification result if it is not exceeded, and a detection module for detecting whether there is a signal inherited from at least one user problem before the current round, Obtaining a detection result; a determination module, configured to determine a target response according to the recognition result, the detection result, the target question, and the standard intent.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic structural diagram of a multi-round question answering device for intelligent customer service in a hardware operating environment according to an embodiment of the present application;

FIG. 2 is a schematic flowchart of a first embodiment of a multi-round question answering method for intelligent customer service in this application;

3 is a schematic flowchart of a second embodiment of a multi-round question answering method for intelligent customer service in this application;

4 is a schematic flowchart of a third embodiment of a multi-round question answering method for intelligent customer service in this application;

FIG. 5 is a structural block diagram of a first embodiment of a multi-round question answering device for intelligent customer service in this application.

detailed description

The specific embodiments described here are only used to explain the application, and are not used to limit the application.

Referring to FIG. 1, FIG. 1 is a schematic structural diagram of a multi-round question answering device for intelligent customer service in a hardware operating environment according to an embodiment of the present application.

As shown in FIG. 1, the intelligent customer service multi-round question answering device may include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen. The optional user interface 1003 may further include a standard wired interface and a wireless interface. The wired interface of the user interface 1003 may be a USB interface in this application. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WIreless-FIdelity (WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM) memory, or may be a stable memory (Non-volatile memory (NVM), such as a disk memory. The memory 1005 may optionally be a storage device independent of the foregoing processor 1001.

Those skilled in the art can understand that the structure shown in FIG. 1 does not constitute a limitation on the multiple rounds of question answering equipment for intelligent customer service, and may include more or fewer components than shown in the figure, or combine some components, or different components. Layout.

As shown in FIG. 1, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and intelligent customer service multiple rounds of question-and-answer instructions.

In the intelligent customer service multi-round question answering device shown in FIG. 1, the network interface 1004 is mainly used to connect to a background server and perform data communication with the background server; the user interface 1003 is mainly used to connect to intelligent customer service devices; the intelligent customer service multiple rounds The question answering device calls the intelligent customer service multi-round question and answer readable instructions stored in the memory 1005 through the processor 1001, and executes the intelligent customer service multi-round question and answer method provided in the embodiment of the present application.

Based on the above hardware structure, an embodiment of a multi-round question answering method for intelligent customer service in this application is proposed.

Referring to FIG. 2, FIG. 2 is a schematic flowchart of a first embodiment of a multi-round question and answer method for intelligent customer service in this application. In the first embodiment, the multi-round question and answer method for intelligent customer service includes the following steps:

Step S10: Obtain the target question of the user's current round.

It should be understood that the executing subject of this embodiment is an intelligent customer service multi-round question answering device, wherein the intelligent customer service multi-round question answering device may be an electronic device such as a personal computer or a server. The user usually asks the target question by voice, then the intelligent customer service multi-round question answering device can perform voice detection, and when the voice information sent by the user is detected, the voice information is obtained as the target question; or through the The display interface of the multi-round question answering device of the intelligent customer service enters the target question, and obtains text information input by the user in the display interface as the target question.

Step S20: Find a standard intent matching the target problem from the knowledge base, and obtain a degree of matching between the standard intent and the target problem.

In specific implementation, the knowledge base of the intelligent customer service multi-round question answering device includes a variety of standard intents and corresponding standard responses. The standard intents in the knowledge base can be compared with the target questions to find and The standard intent of the target question is the standard intent that matches the target question. The degree of matching between the standard intent and the target question is 100%, that is, a complete match; if it cannot be obtained from the knowledge base, When a standard intent that is consistent with the target problem is found, the similarity between the target problem and the standard intent in the knowledge base can be calculated, and the standard intent with the highest similarity can be matched with the target problem. Standard intent, and use the highest similarity as the degree of matching between the standard intent and the target problem.

Step S30: Determine whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold.

It should be noted that in order to improve the accuracy of the intelligent customer service multi-round question answering device to respond to the target question, the degree of matching between the standard intent and the target question needs to be analyzed. If the degree of matching is low , Indicating that the deviation between the standard intent and the target question is large. At this time, if the response corresponding to the standard intent is used as the answer to the target question, the answer may not be answered. Intent recognition to further determine the intent of the target question, thereby finding a more accurate answer. Therefore, a preset matching threshold is usually set in advance. If the degree of matching between the standard intention and the target problem does not exceed the preset matching threshold, it means that the standard intention cannot reflect the intention of the target problem; if The degree of matching between the standard intent and the target question exceeds the preset matching threshold, it indicates that the standard intent can reflect the intent of the target question, and the response corresponding to the standard intent is used as the target question Response, you can accurately reply to the user ’s target question. In this embodiment, after step S30, the method further includes: if the degree of matching between the standard intent and the target question exceeds a preset matching threshold, obtaining a standard response corresponding to the standard intent as the target response. After obtaining the standard response corresponding to the standard intent as the target response, the method further includes: displaying the target response.

It should be understood that the preset matching threshold may be based on historical data to calculate the degree of matching between historical questions with high response accuracy and standard intent, so as to determine a suitable preset matching threshold, for example, the response is accurate The degree of matching between the highly historical issues and the standard intent is mostly above 80%, and the preset matching threshold may be set to 80%. If the degree of matching between the standard intent and the target question exceeds 80%, the response corresponding to the standard intent may be displayed as a response to the target question; if the standard intent and the target question are If the degree of matching between them does not exceed 80%, the response corresponding to the standard intent cannot be used as the response to the target question. Intent recognition of the target question is required to further determine the intention of the target question, so as to find more Accurate responses.

Step S40: if not, perform named entity recognition on the target problem to obtain a recognition result.

If the degree of matching between the standard intent and the target problem does not exceed a preset matching threshold, it means that the standard intent and the target problem have a large deviation, and the target problem may have a missing named entity or a missing manifestation. The situation of the words intended by the user causes the intent presented by the target question to be unclear. At this time, if the response corresponding to the standard intent is used as the answer to the target question, the answer will not be asked, and you need to pass The target question is intentionally identified, and the intention of the target question is further determined, so as to find a more accurate answer. The intent recognition includes naming entity recognition of the target question and detecting whether there is a signal inherited in at least one round of user questions before the current round. The Named Entity Recognition (NER), also known as "proper name recognition", refers to entities with a specific meaning in the recognition text, mainly including person names, place names, agency names, and proper nouns. The named entity recognition of the target problem may be implemented by a Long-Short Short-Term Memory Network (LSTM) + Conditional Random Field Algorithm (CRF) model.

Step S50: Detect whether there are signals inherited from at least one round of user questions before the current round, and obtain a detection result.

It should be noted that the signal refers to the user intention inherited from the questions raised by the user in the previous round or previous rounds, and the signal includes the name inherited from the questions raised by the users in the previous round or previous rounds For physical or other information, the signal will become weaker as the number of rounds passed increases. Of course, the user's signal can also be set to specify the number of rounds passed. For example, the signal is set to 2, which means that the current user's intention can only go backwards. Pass two rounds of dialogue. When the third round of dialogue starts, the signal will be invalid. Therefore, the signal has a life cycle. This life cycle can be set by the user. The user can pass the round of signals in the history of multiple rounds of dialogue. The number of statistics is used to calculate the number of dialogue rounds with the largest proportion of signal transmission in the history of multiple rounds of dialogue as the life cycle of the signal, and the signal will gradually decay during the transmission. Detect whether there is a signal inherited from at least one round of user issues before the current round, by detecting whether the life cycle of the signal is greater than zero, and if the life cycle of the signal is greater than zero, it can be obtained from the previous round Or the signals inherited from the questions raised by the users in the previous rounds are used as the signals of the target question, and the signals can reflect the user's intention.

Step S60: Determine a target response according to the recognition result, the detection result, the target question, and the standard intent.

It can be understood that if the recognition result is the existence of a named entity and the detection result is the presence of a signal, the signal intent can be determined from the named entity and the signal in the target problem, and the knowledge can be obtained from the knowledge according to the signal intent The library searches for a standard intent that matches the signal intent, and uses the response corresponding to the found standard intent as the target response. If the detection result is the presence of a signal, but the recognition result is the absence of a named entity, determine whether the signal includes a named entity, and if the signal does not include the named entity, issue a named entity inquiry and receive A named entity response made by the user according to the named entity inquiry, a signal intent is determined based on the named entity response and a signal in the target question, and the signal intent can be found in the knowledge base and the signal according to the signal intent For a standard intent that matches an intent, the response corresponding to the found standard intent is used as the target response.

If the identified result is the existence of a named entity, but the detection result is the absence of a signal, a preset integrity query is sent, and an integrity reply made by the user according to the preset integrity query is received, according to the The completeness reply and the named entity in the target problem determine the signal intent, and according to the signal intent, a standard intent that matches the signal intent can be found from the knowledge base, and the response corresponding to the found standard intent is taken as The target responds. The detection result is that there is no signal, indicating that the target problem has not been inherited from the user's intention from the previous round or previous rounds of user questions, then the user's intention can be directly asked, and the preset integrity reply It can be to inform the user that the target question is incomplete, and ask the user to reply to the complete question, so that according to the complete response made by the user, a clear-cut question can be obtained, so that a clear response can be found from the knowledge base as the target. Answer. Usually, the user ’s problem is composed of a named entity and an intent. When the user interacts with the intelligent customer service multi-round question and answer device, the user will omit the named entity or intent that has appeared in the previous round or rounds because of multiple rounds of question and answer , But it is usually not the case that both the named entity and the intent are omitted.

In the first embodiment, the target problem of the user's current round is obtained, the standard intent matching the target problem is found from the knowledge base, and the degree of matching between the standard intent and the target problem is obtained, and Whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold; if not, the target problem is identified by a named entity, a recognition result is obtained, and it is detected whether there is at least one before the current round Signals inherited from a round of user questions to obtain detection results, so as to determine the substantial content of the target problem according to the recognition results and the detection results, and to perform corresponding evaluations based on the determined substantial content and the standard intent. Answer, can intelligently answer questions with users, improve the accuracy of responses in multiple rounds of questions and answers, and improve user experience.

Referring to FIG. 3, FIG. 3 is a schematic flowchart of a second embodiment of a multi-round question answering method for intelligent customer service in this application. Based on the first embodiment shown in FIG. 2 above, a second embodiment of a multi-round question answering method for intelligent customer service in this application is proposed.

In a second embodiment, the step S60 includes:

Step S601: If the recognition result is the existence of a named entity and the detection result is the presence of a signal, determine the signal intention according to the named entity and the signal in the target problem.

It can be understood that the detection result is a presence signal, indicating that the target problem has inherited the user intention from the user problems of the previous round or previous rounds of the current round, and that there is a named entity in the target problem , The signal intent may be determined according to the named entity in the target question and a signal inherited from a user question from a previous round or previous rounds of the current round. Example: Multiple rounds of Q & A scenario 1, Question 1: When does the event start? Answer 1: Which activity do you ask? Question 2: Activity A. Answer 2: The start time of activity A is: May 4, 2018.

In the above-mentioned multiple rounds of Q & A scenario 1, there is a named entity in question 2; the named entity is activity A; there is also a signal, and the signal is a signal inherited from question 1: when does it start, according to the name The signal determined by the entity and said signal is: When does activity A begin?

Step S602: Calculate a first similarity between the signal intent and the standard intent.

It should be noted that the standard intent is an intent found from the knowledge base that matches the target problem. In order to further determine whether the standard intent is consistent with the signal intent, the standard intent can be corresponding The standard response is used as the target response. Whether the standard intent is consistent with the signal intent can be determined by calculating the semantic similarity between the standard intent and the inherited signal intent. If the semantic similarity exceeds a preset similarity threshold, it is considered to be consistent , Otherwise it is considered inconsistent.

It should be understood that the semantic similarity calculation mainly uses a deep semantic model based on LSTM to obtain the cosine similarity between the standard intent semantic vector and the signal intent semantic vector as a measure of the consistency of the signal intent and the standard intent standard. Through the LSTM network, the standard intent and the inherited signal intent are represented as semantic vectors, and then the cosine similarity is calculated for the two semantic vectors. The cosine similarity is used to determine whether they are consistent. If the cosine similarity exceeds a preset similarity threshold , The signal intent and the standard intent are considered to be consistent, otherwise the signal intent and the standard intent are considered to be inconsistent. Therefore, in this embodiment, the step S602 includes: performing semantic feature extraction of the signal intent and the standard intent through a long-term short-term memory network model, obtaining a signal semantic vector and a standard semantic vector, and calculating the signal semantic vector. The cosine similarity with the standard semantic vector, and the cosine similarity is used as the first similarity between the signal intent and the standard intent.

Step S603: Determine whether the first similarity exceeds a preset similarity threshold.

In a specific implementation, the preset similarity threshold may be based on historical data to calculate the degree of matching between the signal intent corresponding to each historical question with a high accuracy of recovery and the standard question, so as to determine an appropriate preset matching threshold. For example, if the degree of matching between the signal intent and the standard intent corresponding to each historical question with high accuracy is more than 90%, the preset similarity threshold may be set to 90%. If the first similarity between the signal intent and the standard intent exceeds 90%, the response corresponding to the standard intent may be displayed as a response to the target question; if the signal intent and the standard intent If the first similarity between the intents does not exceed 90%, the response corresponding to the standard intent cannot be used as the response to the target question.

Step S604: if it exceeds, determine the target response according to the standard intent.

It can be understood that if the first similarity exceeds the preset similarity threshold, it indicates that the signal intent is consistent with the standard intent, and a response corresponding to the standard intent can be obtained from the knowledge base as The target response, and the target response is displayed, which can be displayed by voice, and can also be displayed through the display interface of the intelligent customer service multi-round question answering device, or by sending the target response to a user device. The display may also be performed in other ways, which is not limited in this embodiment.

It should be noted that if the first similarity does not exceed the preset similarity threshold, indicating that the signal intent is inconsistent with the standard intent, it may be directly asked whether the customer's intention is the signal intent. The user response determines the user's true intention, and then finds a response corresponding to the user's true intention from the knowledge base as the target response.

In the second embodiment, before step S602, the method further includes: if the detection result is a signal, but the recognition result is that no named entity exists, determining whether the signal includes a named entity; if the signal If the named entity is not included, a named entity query is issued, and a named entity response made by the user according to the named entity query is received; a signal intent is determined according to the named entity response and a signal in the target question.

It should be understood that the detection result is a signal indicating that the target problem has inherited the user intention from the user problems of the previous round or previous rounds of the current round, and the signal may be a named entity or other Information, if the signal includes a named entity, the user's signal intention can be determined based on the signal and the target problem. Example: Multi-round Q & A scenario two, Question 3: When does activity A start? Answer 3: The start time of activity A is: April 5, 2018. Question 4: When does it end? Answer 4: The end time of activity A is: April 8, 2018.

In the second round of Q & A scenario two, there is no named entity in question 4 but there is a signal inherited from the previous round of question 3 as a named entity: activity A, it can be determined according to the signal and the question 4 The signal of the user is: When does activity A end?

It should be noted that if the signal does not include the named entity, it means that the named entity does not exist in the target problem, and the named entity cannot be known from the signal, so the target problem is missing a named entity , A named entity query may be directly issued to determine the named entity of the target question through the user's reply, thereby more accurately identifying the intention of the target question.

In specific implementation, according to the named entity response, the named entity that the user wants to know can be clearly known, and then the signal in the target problem is combined to determine the complete signal intention.

In a second embodiment, the step S40 includes: performing sequence feature extraction of the target problem through the long-term short-term memory network model; performing entity probability calculation of the extracted sequence features through a conditional random field algorithm, and determining the entity Whether the maximum probability exceeds a preset probability threshold; if the maximum probability of the entity exceeds the preset probability threshold, the feature corresponding to the maximum probability of the entity is identified as the named entity of the target problem, and the recognition result is present A named entity; if the maximum value of the entity probability does not exceed the preset probability threshold, a recognition result is obtained that there is no named entity.

It should be understood that the named entity recognition is realized by the LSTM + CRF model, and the processing by the LSTM network is equivalent to obtaining a better representation method of the input target problem. The final output vector of the LSTM unit can be regarded as It is a representation of the input target problem. Finally, in the labeling phase, LSTM is combined with CRF, and LSTM is used to solve the problem of extracting sequence features. Using CRF effectively uses sentence-level tagging information.

Under the LSTM + CRF model, the output will no longer be independent labels, but the best label sequence. For the input: | X = (x1, x2, ..., xn) |, we can define the output probability matrix of the LSTM | Pn * k | where k is the number of output labels. | Pi, j | refers to the probability that the i-th word is marked as the j-th label. For the label sequence to be predicted: | y = (y1, y2, ..., yn) |, we can have the following definition:

Where A is the state transition matrix, | Ai, j | represents the probability of transition from the i-th tag to the j-th tag. By finding the largest | s (X, y) |, the best output label sequence can be obtained. The CRF introduced here is actually just modeling the output label two-tuple, and then using dynamic programming to calculate it, and finally labeling according to the obtained optimal path.

The target problem is input into an LSTM + CRF model, and sequence feature extraction is performed on the target problem through LSTM. Each feature in the extracted feature sequence is subjected to entity probability calculation through CRF, and the feature with the highest entity probability is identified as the feature. The target problem entity. However, when the output entity recognition result indicates that there is no named entity, or the probability of the named entity is too small, it is considered that there is no named entity in the target problem.

In a second embodiment, a signal intention is determined by a named entity in the target problem and a signal inherited from a previous round or previous rounds, and if the standard intention is consistent with the signal intention, the The standard response corresponding to the standard intent is used as the target response, which combines the information transmitted in the previous round or previous rounds, thereby improving the accuracy of the responses in multiple rounds of question and answer.

Referring to FIG. 4, FIG. 4 is a schematic flowchart of a third embodiment of a multi-round question and answer method for intelligent customer service in this application. Based on the second embodiment shown in FIG. 3 described above, a third embodiment of a multi-round question and answer method for intelligent customer service in this application is proposed.

In a third embodiment, the step S20 includes:

Step S201: Find a candidate intent set matching the target problem from a knowledge base through ES retrieval.

It is understandable that ES is short for Elaticsearch, which is mainly based on an inverted index to quickly filter the candidate intent set. First, use the word segmentation system to automatically segment each candidate intent in the knowledge base into word sequences, so that each candidate intent is converted into a data stream composed of word sequences. For the convenience of subsequent processing, it is necessary to assign unique words to each different word. Number, and record which candidate intent in the knowledge base contains this word, so as to get the simplest inverted index. For example, the knowledge base includes 5 candidate intents. In the inverted index, the "word ID" column records the word number of each word, the second column can record the corresponding word, and the third column can record the corresponding inversion of each word. Ranking list. For example, the word "activity A" has a word number of 1 and an inverted list of {1,2,3,4,5}, indicating that each candidate question in the knowledge base contains the word. Then, the target problem can be quickly divided into target word sequences by cutting the target problem into target word sequences, and the inverse index method retrieved by the ES from the knowledge base, the candidate intents and corresponding responses constitute the Candidate intent set.

Step S202: Calculate a second similarity between the target problem and a candidate intent in the candidate intent set.

It should be understood that, in order to determine a standard intent matching the target problem from the candidate intent set, a second similarity between the target problem and the candidate intent in the candidate intent set is generally calculated, Taking the standard intent with the highest similarity as the standard intent that matches the target problem best reflects the true intention of the target problem.

Further, the step S202 includes: calculating a semantic feature, a text feature, a syntactic feature, and a topic feature between the target question and the candidate intents in the candidate intent set; The text features, the syntactic features, and the topic features are aggregated to obtain a second similarity between the target problem and candidate intents in the candidate intent set.

It should be noted that the semantic features are mainly obtained based on the deep semantic model of LSTM; the text features are mainly based on TF-IDF (TF * IDF, TF word frequency, Term Frequency, IDF reverse file frequency, Inverse Document Frequency) value, editing distance, Features such as the longest common substring and / or shared words; syntactic features based on Harbin Institute of Technology's Language Technology Platform (LTP) model for similarity calculation; topic similarity based on document theme generation (Latent, Dirichlet, Allocation, abbreviated LDA) ) Model. Finally, a Logistic Regression (LR) model is trained, and the above features are combined and aggregated to obtain a second similarity between the target problem and each candidate intent.

Step S203: Use the candidate intent with the second highest similarity as the standard intent matching the target problem, and obtain the highest second similarity as the degree of matching between the standard intent and the target problem.

It should be noted that if the candidate intent with the highest similarity is closest to the true intention of the target problem, the candidate intent with the second highest similarity may be used as the standard intent that matches the target problem and obtained The highest second similarity is taken as the degree of matching between the standard intent and the target question, and whether the response corresponding to the candidate intent is taken as the target response is determined by the degree of matching. The candidate intent with the highest similarity is closest to the true intent of the target question, thereby improving the accuracy of the responses in multiple rounds of question and answer.

In addition, an embodiment of the present application further provides a storage medium, where the intelligent customer service multi-round question-and-answer readable instructions are stored, and the intelligent customer service multi-round question-and-answer readable instructions are implemented as described above when executed by a processor. Steps of multiple rounds of question and answer methods for smart customer service. The storage medium may be a non-volatile readable storage medium.

In addition, referring to FIG. 5, an embodiment of the present application also proposes a multi-round question answering device for an intelligent customer service. The multi-round question answering device for an intelligent customer service includes: an obtaining module 10 for obtaining a target question of a user's current round; a finding module 20, It is used to find the standard intent that matches the target problem from the knowledge base, and to obtain the degree of matching between the standard intent and the target problem; the determination module 30 is used to determine the standard intent and the target problem Whether the degree of matching between them exceeds a preset matching threshold; an identification module 40 configured to identify a named entity for the target problem to obtain a recognition result if it does not exceed; a detection module 50 configured to detect whether the current round exists Signals inherited from at least one previous round of user questions to obtain detection results; a determination module 60 is configured to determine a target response based on the recognition results, the detection results, the target question, and the standard intent.

For other embodiments or specific implementations of the intelligent customer service multi-round question answering device described in this application, reference may be made to the foregoing method embodiments, and details are not described herein again.

The technical solution of the present application is essentially in the form of a software product that contributes to the existing technology. The computer software product is stored in a storage medium (such as a read-only memory image (Read Only Memory image, ROM)). / Random Access Memory (Random Access Memory, RAM), magnetic disks, compact discs, including a number of instructions to enable a terminal device (can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute this application Methods described in various embodiments. Any equivalent structure or equivalent process transformation made by using the contents of the specification and drawings of this application, or directly or indirectly used in other related technical fields, is equally included in the scope of patent protection of this application.

Claims

An intelligent customer service multi-round question and answer method is characterized in that the intelligent customer service multi-round question and answer method includes the following steps:

Get the target question for the user's current round;

Searching for a standard intent matching the target problem from a knowledge base, and obtaining a degree of matching between the standard intent and the target problem;

Determining whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold;

If not, perform named entity recognition on the target problem to obtain a recognition result;

Detecting whether there are signals inherited from at least one round of user problems before the current round, and obtaining a detection result;

A target response is determined based on the recognition result, the detection result, the target question, and the standard intent.
The multi-round question answering method for intelligent customer service according to claim 1, wherein the determining a target response based on the recognition result, the detection result, the target question, and the standard intention comprises:

If the identified result is the existence of a named entity and the detection result is the presence of a signal, determining the signal intention according to the named entity and the signal in the target problem;

Calculating a first similarity between the signal intent and the standard intent;

Determining whether the first similarity exceeds a preset similarity threshold;

If it exceeds, the target response is determined according to the standard intent.
The multi-round question answering method for intelligent customer service according to claim 2, wherein before calculating the first similarity between the signal intention and the standard intent, the multi-round question answering method for intelligent customer service further comprises:

If the detection result is a signal, but the recognition result is that no named entity exists, determining whether the signal includes a named entity;

If the signal does not include the named entity, sending a named entity inquiry, and receiving a named entity response made by the user according to the named entity inquiry;

A signal intent is determined based on the named entity response and a signal in the target problem.
The multi-round question answering method for intelligent customer service according to claim 2, wherein the calculating a first similarity between the signal intent and the standard intent includes:

Performing semantic feature extraction of the signal intent and the standard intent through a long-term short-term memory network model to obtain a signal semantic vector and a standard semantic vector;

A cosine similarity between the signal semantic vector and a standard semantic vector is calculated, and the cosine similarity is used as a first similarity between the signal intent and the standard intent.
The multi-round question answering method for intelligent customer service according to claim 4, wherein said performing named entity recognition on said target question to obtain a recognition result comprises:

Performing sequence feature extraction on the target problem through the long-term short-term memory network model;

Calculate the entity probability using the conditional random field algorithm on the extracted sequence features, and determine whether the maximum value of the entity probability exceeds a preset probability threshold;

If the maximum value of the entity probability exceeds the preset probability threshold, determining the feature corresponding to the maximum value of the entity probability as the named entity of the target problem, and obtaining the recognition result as the existence of the named entity;

If the maximum value of the entity probability does not exceed the preset probability threshold, the recognition result is obtained that there is no named entity.
The intelligent customer service multi-round question answering method according to claim 1, wherein the search for a standard intent matching the target problem from a knowledge base, and obtaining a match between the standard intent and the target problem Degree, including:

Finding a candidate intent set matching the target problem from a knowledge base through ES retrieval;

Calculating a second similarity between the target problem and candidate intents in the candidate intent set;

The candidate intent with the highest second similarity is used as the standard intent matching the target problem, and the highest second similarity is obtained as the degree of matching between the standard intent and the target problem.
The intelligent customer service multi-round question answering method according to claim 6, wherein the calculating a second similarity between the target problem and a candidate intent in the candidate intent set comprises:

Calculating semantic features, text features, syntactic features, and topic features between the target problem and candidate intents in the candidate intent set;

Aggregate the semantic features, the text features, the syntactic features, and the topic features through logistic regression to obtain a second similarity between the target problem and candidate intents in the candidate intent set.
An intelligent customer service multi-round question answering device, characterized in that the intelligent customer service multi-round question answering device includes: a memory, a processor, and an intelligent customer service multi-round question answering device stored on the memory and operable on the processor. Read instructions, the intelligent customer service multiple rounds of question-and-answer readable instructions are executed by the processor to implement the following steps:

Get the target question for the user's current round;

Searching for a standard intent matching the target problem from a knowledge base, and obtaining a degree of matching between the standard intent and the target problem;

Determining whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold;

If not, perform named entity recognition on the target problem to obtain a recognition result;

Detecting whether there are signals inherited from at least one round of user problems before the current round, and obtaining a detection result;

A target response is determined based on the recognition result, the detection result, the target question, and the standard intent.
The intelligent customer service multi-round question answering device according to claim 8, wherein the step of determining a target response based on the recognition result, the detection result, the target question, and the standard intent includes:

If the identified result is the existence of a named entity and the detection result is the presence of a signal, determining the signal intention according to the named entity and the signal in the target problem;

Calculating a first similarity between the signal intent and the standard intent;

Determining whether the first similarity exceeds a preset similarity threshold;

If it exceeds, the target response is determined according to the standard intent.
The intelligent customer service multi-round question answering device according to claim 9, wherein before the step of calculating a first similarity between the signal intent and the standard intent includes:

If the detection result is a signal, but the recognition result is that no named entity exists, determining whether the signal includes a named entity;

If the signal does not include the named entity, sending a named entity inquiry and receiving a named entity response made by the user according to the named entity inquiry;

A signal intent is determined based on the named entity response and a signal in the target problem.
The intelligent customer service multi-round question answering device according to claim 9, wherein the step of calculating a first similarity between the signal intent and the standard intent includes:

Performing semantic feature extraction of the signal intent and the standard intent through a long-term short-term memory network model to obtain a signal semantic vector and a standard semantic vector;

A cosine similarity between the signal semantic vector and a standard semantic vector is calculated, and the cosine similarity is used as a first similarity between the signal intent and the standard intent.
The intelligent customer service multi-round question answering device according to claim 11, wherein the step of performing named entity recognition on the target question and obtaining a recognition result comprises:

Performing sequence feature extraction on the target problem through the long-term short-term memory network model;

Calculate the entity probability using the conditional random field algorithm on the extracted sequence features, and determine whether the maximum value of the entity probability exceeds a preset probability threshold;

If the maximum value of the entity probability exceeds the preset probability threshold, determining the feature corresponding to the maximum value of the entity probability as the named entity of the target problem, and obtaining the recognition result as the existence of the named entity;

If the maximum value of the entity probability does not exceed the preset probability threshold, the recognition result is obtained that there is no named entity.
The intelligent customer service multi-round question answering device according to claim 8, characterized in that, the standard intent matching the target question is found from the knowledge base, and a match between the standard intent and the target question is obtained The steps include:

Finding a candidate intent set matching the target problem from a knowledge base through ES retrieval;

Calculating a second similarity between the target problem and candidate intents in the candidate intent set;

The candidate intent with the highest second similarity is used as the standard intent matching the target problem, and the highest second similarity is obtained as the degree of matching between the standard intent and the target problem.
A storage medium is characterized in that the storage medium stores readable instructions of multiple rounds of questions and answers for intelligent customer service, which are executed by a processor to implement the following steps:

Get the target question for the user's current round;

Searching for a standard intent matching the target problem from a knowledge base, and obtaining a degree of matching between the standard intent and the target problem;

Determining whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold;

If not, perform named entity recognition on the target problem to obtain a recognition result;

Detecting whether there are signals inherited from at least one round of user problems before the current round, and obtaining a detection result;

A target response is determined based on the recognition result, the detection result, the target question, and the standard intent.
The storage medium according to claim 14, wherein the step of determining a target response based on the recognition result, the detection result, the target question, and the standard intent includes:

If the identified result is the existence of a named entity and the detection result is the presence of a signal, determining the signal intention according to the named entity and the signal in the target problem;

Calculating a first similarity between the signal intent and the standard intent;

Determining whether the first similarity exceeds a preset similarity threshold;

If it exceeds, the target response is determined according to the standard intent.
The storage medium according to claim 15, wherein the step of calculating a first similarity between the signal intent and the standard intent includes before:

If the detection result is a signal, but the recognition result is that no named entity exists, determining whether the signal includes a named entity;

If the signal does not include the named entity, sending a named entity inquiry and receiving a named entity response made by the user according to the named entity inquiry;

A signal intent is determined based on the named entity response and a signal in the target problem.
The storage medium according to claim 15, wherein the step of calculating a first similarity between the signal intent and the standard intent includes:

Performing semantic feature extraction of the signal intent and the standard intent through a long-term short-term memory network model to obtain a signal semantic vector and a standard semantic vector;

A cosine similarity between the signal semantic vector and a standard semantic vector is calculated, and the cosine similarity is used as a first similarity between the signal intent and the standard intent.
The storage medium according to claim 17, wherein the step of performing named entity recognition on the target problem and obtaining a recognition result comprises:

Performing sequence feature extraction on the target problem through the long-term short-term memory network model;

Calculate the entity probability using the conditional random field algorithm on the extracted sequence features, and determine whether the maximum value of the entity probability exceeds a preset probability threshold;

If the maximum value of the entity probability exceeds the preset probability threshold, determining the feature corresponding to the maximum value of the entity probability as the named entity of the target problem, and obtaining the recognition result as the existence of the named entity;

If the maximum value of the entity probability does not exceed the preset probability threshold, the recognition result is obtained that there is no named entity.
The storage medium according to claim 14, wherein the steps of finding a standard intent matching the target question from a knowledge base, and obtaining a degree of matching between the standard intent and the target question include :

Finding a candidate intent set matching the target problem from a knowledge base through ES retrieval;

Calculating a second similarity between the target problem and candidate intents in the candidate intent set;

The candidate intent with the highest second similarity is used as the standard intent matching the target problem, and the highest second similarity is obtained as the degree of matching between the standard intent and the target problem.
An intelligent customer service multi-round question answering device, characterized in that the intelligent customer service multi-round question answering device includes: an acquisition module for acquiring a target question of a current round of a user;

A search module, configured to find a standard intent matching the target problem from a knowledge base, and obtain a degree of matching between the standard intent and the target problem;

A judging module, configured to judge whether the degree of matching between the standard intent and the target problem exceeds a preset matching threshold;

A recognition module, configured to identify a named entity for the target problem if the number is not exceeded, and obtain a recognition result;

A detection module, configured to detect whether there are signals inherited from at least one round of user questions before the current round, and obtain a detection result;

A determination module, configured to determine a target response according to the recognition result, the detection result, the target question, and the standard intention.