CN108090174A

CN108090174A - A kind of robot answer method and device based on system function syntax

Info

Publication number: CN108090174A
Application number: CN201711341968.0A
Authority: CN
Inventors: 张钫炜; 韩道岐; 孙明哲; 郭雪梅; 陆月明
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2017-12-14
Filing date: 2017-12-14
Publication date: 2018-05-29
Anticipated expiration: 2037-12-14
Also published as: CN108090174B

Abstract

An embodiment of the present invention provides a kind of robot answer methods based on system function syntax and device, this method to include：The target problem received from default knowledge base using the retrieval of problem search method, if retrieving target problem, there are in default knowledge base, then from knowledge base choose with target problem similarity highest the problem of as matching problem, response corresponding with matching problem is searched for from knowledge base again, is denoted as the first response；According to the humanized corresponding system function syntax feature of the first machine, the first response is optimized, the response after being optimized is denoted as target response；Target response is exported to current sessions corresponding with target problem and exports the first response to current sessions corresponding with target problem, current sessions refer to the first response for answering the session of target problem.The accuracy of response problem can be improved using the embodiment of the present invention.

Description

A kind of robot answer method and device based on system function syntax

Technical field

The present invention relates to Systemic Functional Linguistic, natural language processing and intelligent answer technical field, more particularly to one Robot answer method and device of the kind based on system function syntax.

Background technology

Explosive growth trend is presented in Internet era, information.In face of the data of such magnanimity, how to help people more preferable Searching information, identification information, become the most important thing of information retrieval field.

Based on the above problem, in the case where current internet environment unstructured knowledge increases on a large scale, intelligent answer Robot is also more widely applied in the information service system in actual each field and social life.It is used as and asks One of important realization method of system is answered, except being required to provide accurately answer, it is also necessary to which there are the question and answer of remarkable fluency Dialogue ability, accurately semantic analysis ability, abundant context and powerful ability of self-teaching understand that user needs It asks, promotes the development of man-machine interaction mode.

However, in existing intelligent answer, search engine can return to a string of links according to the inquiry request of user, each to link It is directed toward a document and equipped with one section of summary info, user needs to browse the document returned corresponding to lists of links one by one, to seek Look for oneself desired information.But the process of screening document often consumes user's many times, because that user usually requires It is specific answer rather than entire chapter document.The bulk information returned by search engine is only partially what user needed, this is just The problem of bringing information overload.

Intelligent answer robot is mostly the knowledge base that pre-stored target problem and answer pair are relied on based on matching way, After obtaining target problem input by user, characterization expression is carried out to the target problem, and the target characterized after representing is asked Topic is input in advance trained machine learning model, is exported response immediate with the target problem and is opened up Show, wherein, for calculating the similarity of input target problem and target problem in database, prediction most connects the machine learning model Near target problem.But it is this based entirely on matched mode due to not accounting for the basic speech work(of target problem in itself Can, therefore the response for meeting Functional Grammar can not be made, no matter for what target problem, as long as the phase of key feature extraction Together, machine-made response will be exported, therefore causes the accuracy of response target problem relatively low, influence user uses body It tests.

The content of the invention

The embodiment of the present invention is designed to provide a kind of robot answer method and device based on system function syntax, To improve response problem accuracy.Specific technical solution is as follows：

A kind of robot answer method based on system function syntax, applied to intelligent Answer System, the described method includes：

The target problem received from default knowledge base using the retrieval of problem search method, wherein, the target is asked The problem of entitled active user's input；

If retrieving the target problem there are in default knowledge base, chosen and the mesh from the knowledge base The problem of mark problem similarity highest, is as matching problem, then answer corresponding with the matching problem is searched for from the knowledge base It answers, is denoted as the first response；

According to the humanized corresponding system function syntax feature of the first machine, first response is optimized, is obtained Response after optimization, is denoted as target response, so that response is more in line with question and answer context semanteme, first machine is humanized to be The intelligent Answer System is not initialised, and corresponding machine is humanized, and the machine is humanized belongs to including interactive attribute, cognition Property and formal attribute.

Further, it is described the target problem received is forwarded to default robot synthesis responder module before, The method further includes：

Obtain the target corpus data after being marked according to robot attributive classification to target corpus data, the target language material Data are the corpus datas obtained from default field；

Target corpus data after mark is input to default convolutional neural networks to be trained, by deep learning, Robot attributive classification model after being trained；The robot attributive classification model is by the target language material number after mark According to as training set, convolutional neural networks are trained using the training set, by deep learning, the mould after being trained Type；

Using the robot attributive classification model at the beginning of the humanized progress of the first machine in the intelligent Answer System Beginningization.

It further, should to described first described according to the humanized corresponding system function syntax feature of the first machine It answers and optimizes, the response after being optimized is denoted as target response, so that response is more in line with after question and answer context semanteme, The method further includes：

The target response is exported to current sessions corresponding with the target problem.

It further, should to described first described according to the humanized corresponding system function syntax feature of the first machine It answers and optimizes, the response after being optimized is denoted as target response, so that response is more in line with before question and answer context semanteme, The method further includes：

Based on system function syntax principle, the speech function implied in the target problem is refined, and provides the target The corresponding conversational mode of speech function implied in problem, the conversational mode include receiving or retract, perform or refuse, approve Or it rejects, answer or refuse to answer.

Further, it is described from default knowledge base using the target problem that receives of problem search method retrieval it Afterwards, the method further includes：

If not retrieving the target problem in the knowledge base, by deep learning, trained using object module The target problem not in the knowledge base, the depth intelligent answer model after being trained；The object module is volume In product neutral net CNN, shot and long term memory network model LSTM, Recognition with Recurrent Neural Network model RNN and variant or memory network At least one combination；

The target problem is generated into response using the depth intelligent answer model, is denoted as the second response.

Further, the target problem is generated into response using the depth intelligent answer model described, is denoted as the After two responses, the method further includes：

Based on system function syntax principle, the speech function implied in the target problem is refined, and provides the target The corresponding conversational mode of speech function implied in problem；

Based on the conversational mode, according to the humanized corresponding system function syntax feature of first machine, to described Second response optimizes, and the response after being optimized is denoted as the 3rd response, using the 3rd response as target response, so that should It answers and is more in line with question and answer context semanteme.

The target problem of reception is added to the session that active user prestores to concentrate, and the session collection is carried out more Newly.

Further, the target problem of reception is added to the session that active user prestores concentrates described, and it is right After the session collection is updated, the method further includes：；

According to updated session collection, the target problem is monitored in real time using the robot attributive classification model, really Fixed second machine is humanized；

Using the core emotion of default core sentiment classification model extraction active user, and assess the core emotion Emotion is differential, wherein, the core sentiment classification model is based in sentiment dictionary, weak mark, machine learning, deep learning It is at least one to combine the model for being trained acquisition；

The humanized configuration humanized with first machine of second machine is matched, it, will if mismatch Second machine is humanized humanized as first machine, returns to execution and utilizes the robot attributive classification model pair The humanized carry out initialization step of the first machine in default intelligent Answer System；

It is differential according to the core emotion of active user and the corresponding emotion of active user, judge whether active user's session Content is transferred into artificial objective service platform；The current sessions include the 3rd response and the target problem；

If it has, then active user's session content is accessed into manually objective service platform, the artificial objective service platform It is the platform for manually conversating with active user.

Further, it is described differential according to the core emotion of active user and the corresponding emotion of active user, judge current Whether session transfers into artificial objective service platform, including：

The 3rd response and the target problem are obtained, and the 3rd response and the target problem are generated currently The session of user；

According to the core sentiment classification model, the emotion potentiality of output active user's session；

Monitor the emotion potentiality of active user's session of output in real time；

Whether with pre-set emotion attribute value match to determine by the emotion potentiality for checking the current user conversation Whether need to transfer active user's session content into artificial objective service platform.

A kind of robot answering device based on system function syntax, applied to intelligent Answer System, described device：

Problem search module, the target for the retrieval of problem search method to be used to receive from default knowledge base are asked Topic, wherein, the target problem is the problem of active user inputs；

If first responder module for retrieving the target problem there are in default knowledge base, is known from described Know storehouse in choose with the target problem similarity highest the problem of as matching problem, then from the knowledge base search for and institute The corresponding response of matching problem is stated, is denoted as the first response；

First optimization module, for according to the humanized corresponding system function syntax feature of the first machine, to described first Response optimizes, and the response after being optimized is denoted as target response, so that response is more in line with question and answer context semanteme, institute It is that the intelligent Answer System is not initialised corresponding machine at the another aspect of the invention implemented that it is humanized, which to state the first machine, and also A kind of computer readable storage medium is provided, instruction is stored in the computer readable storage medium, when it is in computer During upper operation so that computer performs any of the above-described robot answer method based on system function syntax.

At the another aspect that the present invention is implemented, the embodiment of the present invention additionally provides a kind of computer program production comprising instruction Product, when run on a computer so that computer performs any of the above-described robot based on system function syntax Answer method.

A kind of robot answer method and device based on system function syntax provided in an embodiment of the present invention, can be from pre- If knowledge base in using the target problem that receives of problem search method retrieval, if retrieving target problem, there are default In knowledge base, then chosen from knowledge base with the problem of target problem similarity highest as matching problem, then from knowledge base Search response corresponding with matching problem, is denoted as the first response；It is special according to the humanized corresponding system function syntax of the first machine Sign, optimizes the first response, the response after being optimized is denoted as target response；Target response is exported to target and is asked Inscribe corresponding current sessions.This method is according to the humanized corresponding system function syntax feature of the first machine, to from knowledge base The problem of choosing with target problem similarity highest, and the problem is carried out in response i.e. the first response of knowledge base Corresponding matching Optimization, improves the accuracy of response problem.Certainly, implement any of the products of the present invention or method must be not necessarily required to reach simultaneously To all the above advantage.

Description of the drawings

It in order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, below will be to embodiment or existing There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this Some embodiments of application, for those of ordinary skill in the art, without creative efforts, can be with Other attached drawings are obtained according to these attached drawings.

Fig. 1 is the first flow of the robot answer method provided by the embodiments of the present application based on system function syntax Figure；

Fig. 2 is second of flow of the robot answer method provided by the embodiments of the present application based on system function syntax Figure；

Fig. 3 is the third flow of the robot answer method provided by the embodiments of the present application based on system function syntax Figure；

Fig. 4 is the structure diagram of the robot answering device provided by the embodiments of the present application based on system function syntax；

Fig. 5 is the structure diagram of electronic equipment provided by the embodiments of the present application.

Specific embodiment

Below in conjunction with the attached drawing in the embodiment of the present application, the technical solution in the embodiment of the present application is carried out clear, complete Site preparation describes, it is clear that described embodiments are only a part of embodiments of the present application, instead of all the embodiments.It is based on Embodiment in the application, those of ordinary skill in the art are obtained every other without creative efforts Embodiment shall fall in the protection scope of this application.

The function body of the present invention is intelligent Answer System, and above-mentioned intelligent Answer System is embedded into page end or mobile terminal Service is provided in APP, the present invention does not limit this.

Fig. 1 is the first flow of the robot answer method provided by the embodiments of the present application based on system function syntax Figure, this method include：

S101, the target problem received from default knowledge base using the retrieval of problem search method, wherein, the mesh The problem of mark problem inputs for active user；

Wherein, above-mentioned target problem can be the key entry of user's keyboard, phonetic entry and candidate item selection etc., and the present invention is right The active user that intelligent Answer System obtains target problem does not limit the embodiment of the target problem.

It is noted that default knowledge base is the knowledge base created in advance, which is the row based on this field The knowledge base that industry experience and Practical Statistic create；Technical staff can be according to the situation of industry development and user's general concern to knowing Know storehouse and carry out timely update extension.

Before S101, this method further includes：

Target corpus data after mark is input to default convolutional neural networks to be trained, by deep learning, Robot attributive classification model after being trained；

Using the robot attributive classification model to the humanized progress of the first machine in default intelligent Answer System Initialization.

As it can be seen that the present embodiment is instructed by using the robot attributive classification model using the target corpus data after mark Robot attributive classification model after white silk is initialized so that the first machine after initialization is humanized more to be helped to improve The accuracy of response.

After S101, this method further includes：

If not retrieving the target problem in the knowledge base, by deep learning, trained using object module The target problem not in the knowledge base, the depth intelligent answer model after being trained；The object module is volume Product neutral net (CNN, convolutional neural network), shot and long term memory network model (LSTM, Convolutional neural network), Recognition with Recurrent Neural Network model (RNN, Recurrent neural network) And at least one of variant or memory network (Memory Network) combination；

As it can be seen that present embodiment to knowledge base scope outside the problem of mainly by trained depth intelligent answer model into Row response can be used as depth intelligent answer model described in training data back feeding with stylish question and answer, enhance model adaptability. It has taken into full account the source of problem and the scalability of system, has utmostly ensured the validity of response, be conducive to promote user The impression used system.

It is described using the depth intelligent answer model by the target problem generate response, be denoted as the second response it Afterwards, this method further includes：

Based on system function syntax principle, the speech function implied in the target problem is refined, and provides the target In problem imply the corresponding conversational mode of speech function, turn-taking to be promoted to carry out, the conversational mode include receive or It retracts, perform or refuses, approve or reject, answer or refuse to answer；

As it can be seen that present embodiment, by using object module training objective corpus data, the depth after being trained is intelligent Question-Answering Model has taken into full account session speech function and its conversion in question answering process, improves specific aim and the dialogue of response Flexibility avoids the answer that open problem is brought and obscures with exchanging situations such as having some setbacks so that the depth intelligent answer mould Type further improves the accuracy of response.

It is described using the depth intelligent answer model by the target problem generate response, be denoted as the second response it Afterwards, the method further includes：

It the target problem of reception is added to the session that active user prestores concentrates described, and to the session collection After being updated, the method further includes：

As it can be seen that core emotion of the present embodiment using core sentiment classification model extraction active user, and described in assessment The emotion of core emotion is differential, has taken into full account that user feeling is fluctuated with emotion in question answering process, by cutting artificial visitor in time Service platform reduces the not pleasant mood that user in question answering process may accumulate, and is conducive to improve user to intelligent answer system The use wish of system.

It is described differential according to the core emotion of active user and the corresponding emotion of active user, judge whether current sessions turn The artificial objective service platform of access；The current sessions include a kind of embodiment of the 3rd response and the target problem For：

Whether with pre-set emotion attribute value match to determine by the emotion potentiality for checking the current user conversation Whether need to transfer active user's session content into artificial objective service platform；

Emotion potentiality is as shown in table 1 in present embodiment, and table 1 is classified and the differential signal table of emotion for emotion core attitude, Emotion core attitude is divided into three classes：Satisfied, happy and safety.Satisfied to be divided into interest and be discontented with, happiness, which is divided into, to be liked and dislikes, peace It is divided into trust and distrust entirely.It is very poor per the emotion of one kind emotional attitude that there are three grade is respectively high, medium and low.Such as to " emerging Interest " is a kind of, and it is " height " that user conversation illustrates user's attitude " interest " very poor value for " please be handled to me " at once, and " I determines substantially Will this " reaction " interest " attitude it is very poor value for " in ", " I will also consider further that " reaction " interest " attitude it is very poor Value is compared with " low ".Specific emotion core attitude, which is classified, and emotion is very poor is shown in Table 1.

Table 1 is the classification of emotion core attitude and the differential signal table of emotion

As it can be seen that emotion potentiality of the present embodiment by active user's session of real time monitoring output, and should by checking Whether the emotion potentiality of active user's session matches with pre-set emotion attribute value determines so that dialogue is more smooth, more Add intelligence, improve the experience effect of user.

Above-mentioned target problem search method is：

Utilize vector space model COS distance algorithm, smallest edit distance algorithm or longest common subsequence algorithm, pin To the target problem of each active user input, the first object problem similarity with problem in knowledge base respectively is calculated, The first object problem similarity between all problems in the knowledge base respectively is obtained, the first object problem is Any one target problem in the target problem of each active user's input；

It is similar to being selected in similarity of all the problems in the knowledge base respectively from the obtained first object problem The problem of spending highest is as matching problem.

As it can be seen that the problem search method can quick obtaining the problem of being matched with target problem, it is and corresponding with matching problem Response so that intelligent response system response is fast.

It is described from the obtained first object problem respectively with being selected in the knowledge base in similarity of all the problems The problem of similarity highest as matching problem, including：

Utilize vector space model COS distance algorithm, smallest edit distance algorithm or longest common subsequence algorithm, pin To the target problem of each active user input, it is of all the problems similar to knowledge base respectively to calculate the first object problem Degree, obtain the first object problem respectively with similarity of all the problems in the knowledge base；

To similarity of all the problems in the obtained first object problem and the knowledge base according to similarity height It is ranked up, the similarity sequence after being sorted；

The problem of similarity highest corresponds to is chosen from the similarity sequence as matching target problem.

As it can be seen that present embodiment using similarity according to being just ranked up so that it is quick and obtain similarity in an orderly manner The problem of highest.

S102 if retrieving the target problem there are in the knowledge base, chooses and institute from the knowledge base The problem of stating target problem similarity highest is as matching problem, then search is corresponding with the matching problem from the knowledge base Response, be denoted as the first response；

Even if it should be noted that checked from the knowledge base with the highest matching problem of target problem matching degree, but Be due to not accounting for the basic speech function of target problem in itself, the response that meets Functional Grammar can not be made, it is necessary to Further first response problem is optimized, to obtain the response for meeting current user context semanteme.

S103 according to the humanized corresponding system function syntax feature of the first machine, optimizes first response, Response after being optimized, is denoted as target response, so that response is more in line with question and answer context semanteme, the first machine Genus Homo Property be that the intelligent Answer System corresponding machine that is not initialised is humanized, the machine it is humanized including interactive attribute, recognize Know attribute and formal attribute；

It should be noted that based on system function syntax interpersonal metaphor can determine the humanized i.e. interactive attribute of machine, Attribute, formal attribute are recognized, specific as shown in table 2, table 2 is the attribute list of intelligent Answer System robot classification, based on system The interpersonal metaphor of Functional Grammar determines the humanized i.e. interactive attribute of machine, cognition attribute, formal attribute.Interactivity defines robot Interactive interpersonal functions label is respectively：Statement puts question to, provides and order；Awareness is divided into certainty and uncertainty, such as： " you have a sore throat be likely to be flu caused by." reaction is to recognize the uncertainty of attribute, it is opposite that " it must be sense that you, which have a sore throat, Caused by emitting." there is apparent certainty；Formal property be divided into three aspects it is professional, apart from property, serious property, it is corresponding very poor It is respectively high, medium and low.

The attribute list of 2 intelligent Answer System robot of table classification

It is interactive：The system function syntax that linguist Halliday is proposed points out interpersonal metaphor metaphor containing the tone and feelings State metaphor.A variety of tone embody interpersonal functions, and in verbal Communication, speaker provides information with statement, and information is asked for enquirement. Generally, speech tool gives information, asks for information there are four types of basic function, give article and service and ask for article and Service.In conclusion the interpersonal functions label that the embodiment of the present invention can define robot interactive is respectively：Statement is putd question to, carried For and order.

Formal property：From linguistics angle, formal property is a kind of semantic domain of interpersonal property, is in particular in " serious Property ", these three Attribute transpositions can be high, medium and low three by " apart from property " and " professional " three aspects, the embodiment of the present invention Attribute is very poor, and enumerates corresponding example such as word, sentence to illustrate this method.This three aspect is not what is be independently distributed, and It is to correspond to conjunction relation；

Awareness：For the modality metaphor of system function syntax, Givon proposes a kind of tendency of the awareness as mood. Awareness can be abstracted as two attribute by the embodiment of the present invention：Certainty and uncertainty.

Before S103, including：

Based on system function syntax principle, the speech function implied in the target problem is refined, and provides the target The corresponding conversational mode of speech function implied in problem, the conversational mode include receiving or retract, perform or refuse, approve Or it rejects, answer or refuse to answer；

This step is in order to which turn-taking is promoted to carry out, and as shown in table 3, table 3 is embodied in words rotation for system function syntax The signal table changed.

During turn-taking speech tool gives information, asks for information, give article and service there are four types of basic function And article and service are asked for, the present invention, which defines speech function, to be included：It provides, order, statement is with puing question to.It is with " giving information " Example, completely taking turns for one to be：

- user：I has 500,000 dollar, it is contemplated that need not all be used in 30 days.

- robot：Alright, parent.Next will recommend for you.

User can also carry out unrestricted choice and the functions such as be rejected, refuses to answer, retract or refuse.

The signal table of 3 turn-taking of table

It should be noted that the speech function implied in above-mentioned target problem can include：It provides, order, state and carries It asks.

As it can be seen that present embodiment has taken into full account session speech function and its conversion in question answering process, response is improved Specific aim and the flexibility of dialogue, avoid the answer that open problem is brought and obscure with exchanging situations such as having some setbacks, so that The intelligent Answer System is more in line with session context, and more hommization.

After S103, this method further includes：

It should be noted that by further being optimized to the first response problem, obtain and be more in line in question and answer Semanteme hereafter.

It can be seen that method provided in an embodiment of the present invention is special according to the humanized corresponding system function syntax of the first machine Sign to the problem of selection is with target problem similarity highest from knowledge base, and answers the problem in knowledge base Corresponding matching It answers i.e. the first response to be optimized, improves the accuracy of response problem.

Fig. 2 is the first flow of the robot answer method provided by the embodiments of the present application based on system function syntax Figure, this method include：

S201 obtains the target corpus data after being marked according to robot attributive classification to target corpus data, the mesh It is the corpus data obtained from default field to mark corpus data；

It should be noted that language material source can be existing question and answer language material, be asked towards common problem collection, towards specific area Inscribe the synthesis of collection, existing question and answer language material such as wikipedia data set, Baidu's question and answer data set etc., towards common problem collection such as Beijing University Chinese corpus, internet corpus etc., towards specific area problem set such as insurance industry language material, financial industry language material and medical treatment Industry language material etc..The language material of acquisition is labeled language material in conjunction with above-mentioned robot categorical attribute table, as convolutional neural networks Input data.By the continuous study of convolutional neural networks, trained robot disaggregated model is exported, as intelligent answer The basic model of system.

Target corpus data after mark is input to default convolutional neural networks and is trained, passes through depth by S202 Study, the robot attributive classification model after being trained；

It should be noted that the robot attributive classification model is to be used as training by the target corpus data after mark Collection, is trained convolutional neural networks using the training set, by deep learning, the model after being trained；

S203, it is humanized to the first machine in default intelligent Answer System using the robot attributive classification model It is initialized；

It should be noted that this step is the basic mould using the robot attributive classification model as intelligent Answer System Type, in case the target problem subsequently inputted directly uses the intelligent Answer System of the basis intelligence model.

S204, the target problem received from default knowledge base using the retrieval of problem search method, wherein, the mesh The problem of mark problem inputs for active user；

Wherein, S204~S205 is identical with the method that step S102~S103 in Fig. 1 embodiments is performed respectively.Therefore, All embodiments in Fig. 1 are suitable for Fig. 2, and can reach the same or similar advantageous effect, and details are not described herein.

S205, if retrieving the target problem there are in default knowledge base, chosen from the knowledge base with The problem of target problem similarity highest, is as matching problem, then search and the matching problem pair from the knowledge base The response answered is denoted as the first response；

S206 based on system function syntax principle, refines the speech function implied in the target problem, and provides described In target problem imply the corresponding conversational mode of speech function, the conversational mode include receive or retract, perform or refuse, Accreditation is rejected, answers or refuse to answer；

The speech function implied wherein in target problem can include：It provides, order, state and puts question to.

S207, based on the conversational mode, according to the humanized corresponding system function syntax feature of the first machine, to described First response optimizes, and the response after being optimized is denoted as target response, so that response is more in line with question and answer context language Justice, it is that the intelligent Answer System corresponding machine that is not initialised is humanized that first machine is humanized, the robot Attribute includes interactive attribute, cognition attribute and formal attribute；

S208, if not retrieving the target problem in the knowledge base, by deep learning, using object module The target problem of the training not in the knowledge base, the depth intelligent answer model after being trained；The object module For convolutional neural networks CNN, shot and long term memory network model LSTM, Recognition with Recurrent Neural Network model RNN and variant or memory net The combination of at least one of network；

It should be noted that before S208 or S205, this method further includes：Statistics dialogue wheel number, with default experience Upper limit threshold is compared, and is then transferred artificial customer service platform more than the experience upper limit threshold；

Wherein, the experience upper limit threshold be available with statistical method from substantial amounts of dialogue data talk about wheel number desired value obtain .The dialogue wheel number refers to the current session content that target problem is located at, and a target problem corresponds to a target response, is denoted as one Dialogue wheel.

S209 based on system function syntax principle, refines the speech function implied in the target problem, and provides described The corresponding conversational mode of speech function implied in target problem；

S210 is right according to the humanized corresponding system function syntax feature of first machine based on the conversational mode Second response optimizes, and the response after being optimized is denoted as the 3rd response, using the 3rd response as target response, with Response is made to be more in line with question and answer context semanteme；

This step obtains being more in line with question and answer context semanteme, makes user and intelligence by being optimized to the second response The exchange of problem system is more smooth, improves the experience effect of user.

S211 exports the target response to current sessions corresponding with the target problem.

It can be seen that method provided in an embodiment of the present invention by using robot attributive classification model to intelligent answer system The first machine in system is humanized to be initialized so that the first machine after initialization is humanized more to help to improve response Accuracy, there are problems that for target problem in knowledge base, from knowledge base choose it is highest with target problem similarity Problem according to the humanized corresponding system function syntax feature of the first machine, and answers the problem in knowledge base Corresponding matching It answers i.e. the first response to be optimized, improves the accuracy of response problem, there is no asking in knowledge base for target problem Topic, response is carried out by trained depth intelligent answer model, can be used as described in training data back feeding with stylish question and answer Depth intelligent answer model enhances model adaptability.The source of problem and the scalability of system, maximum journey are taken into full account Degree ensures the validity of response, is conducive to promote the impression that user uses system.

Fig. 3 is the third flow of the robot answer method provided by the embodiments of the present application based on system function syntax Figure, this method include：

S301 obtains the target corpus data after being marked according to robot attributive classification to target corpus data, the mesh It is the corpus data obtained from default field to mark corpus data；

Wherein, S301~S310 is identical with the method that step S201~S210 in Fig. 2 embodiments is performed respectively.Therefore, All embodiments in Fig. 2 are suitable for Fig. 3, and can reach the same or similar advantageous effect, and details are not described herein.

Target corpus data after mark is input to default convolutional neural networks and is trained, passes through depth by S302 Study, the robot attributive classification model after being trained；

S303, it is humanized to the first machine in default intelligent Answer System using the robot attributive classification model It is initialized；

S304, the target problem received from default knowledge base using the retrieval of problem search method, wherein, the mesh The problem of mark problem inputs for active user；

S305, if retrieving the target problem there are in default knowledge base, chosen from the knowledge base with The problem of target problem similarity highest, is as matching problem, then search and the matching problem pair from the knowledge base The response answered is denoted as the first response；

S306 based on system function syntax principle, refines the speech function implied in the target problem, and provides described The corresponding conversational mode of speech function implied in target problem, turn-taking to be promoted to carry out, the conversational mode includes connecing By or retract, perform or refuse, approve or reject, answer or refuse to answer；

S307, based on the conversational mode, according to the humanized corresponding system function syntax feature of the first machine, to described First response optimizes, and the response after being optimized is denoted as target response, so that response is more in line with question and answer context language Justice, it is that the intelligent Answer System corresponding machine that is not initialised is humanized that first machine is humanized, the robot Attribute includes interactive attribute, cognition attribute and formal attribute；

S308, if not retrieving the target problem in the knowledge base, by deep learning, using object module The target problem of the training not in the knowledge base, the depth intelligent answer model after being trained；The object module For convolutional neural networks CNN, shot and long term memory network model LSTM, Recognition with Recurrent Neural Network model RNN and variant or memory net The combination of at least one of network；

S309 based on system function syntax principle, refines the speech function implied in the target problem, and provides described The corresponding conversational mode of speech function implied in target problem；

S310 is right according to the humanized corresponding system function syntax feature of first machine based on the conversational mode Second response optimizes, and the response after being optimized is denoted as the 3rd response, using the 3rd response as target response, with Response is made to be more in line with question and answer context semanteme；

The target problem of reception is added to the session that active user prestores and concentrated by S311, and to the session collection It is updated；

Wherein, session collection refers to the session that active user once stored in the intelligent Answer System.

S312 according to updated session collection, monitors the target using the robot attributive classification model and asks in real time Topic, determines that the second machine is humanized；

S313 using the core emotion of default core sentiment classification model extraction active user, and assesses the core The emotion of emotion is differential, wherein, the core sentiment classification model is based on sentiment dictionary, weak mark, machine learning, depth Practise at least one of combination be trained the model of acquisition；

This step is analyzed and handled to the emotion in session content based on core sentiment classification model, that is to say, that The core emotion of active user can be extracted by using default core sentiment classification model；

And session content feelings are presented in the differential visualization that can be understood as analysis result of emotion for assessing the core emotion Feel situation.

Wherein, the sentiment analysis result of session content will combine meeting with the data of core emotion-very poor bi-values of emotion Words emotion attribute table is matched, such as：" the quality too bad of your this product " will export discontented-high as a result, working as Analysis result is negative for core emotion, and such as discontented, dislike is distrusted；Emotion it is very poor for it is middle and high when, system will send session The notice of transfer, and session content will be given to artificial customer service processing.

S314 matches the humanized configuration humanized with first machine of second machine, if not Match somebody with somebody, second machine is humanized humanized as first machine, it returns and performs S303 steps；

Wherein, execution S303 is returned to can be understood as performing S303~S314.

S315, it is differential according to the core emotion of active user and the corresponding emotion of active user, judge whether to use current Family session content is transferred into artificial objective service platform；The current sessions include the 3rd response and the target problem；If It is yes, performs S316；

Active user's session content is accessed manually objective service platform by S316, and the artificial objective service platform is to use In the platform manually to conversate with active user.

This step provides manually monitoring interface and access artificial visitor's service platform, by active user's session content or by feelings The visualization achievement of sense analysis is transferred to artificial customer service, and permission is manually checked or chased after to the situation curve or situation map of presentation It the operations such as traces back.

It can be seen that method provided in an embodiment of the present invention by using robot attributive classification model to intelligent answer system The first machine in system is humanized to be initialized so that the first machine after initialization is humanized more to help to improve response Accuracy, there are problems that for target problem in knowledge base, from knowledge base choose it is highest with target problem similarity Problem according to the humanized corresponding system function syntax feature of the first machine, and answers the problem in knowledge base Corresponding matching It answers i.e. the first response to be optimized, improves the accuracy of response problem, there is no asking in knowledge base for target problem Topic, response is carried out by trained depth intelligent answer model, can be used as described in training data back feeding with stylish question and answer Depth intelligent answer model enhances model adaptability.The source of problem and the scalability of system, maximum journey are taken into full account Degree ensures the validity of response, is conducive to promote the impression that user uses system；Worked as using the extraction of core sentiment classification model The core emotion of preceding user, and the emotion for assessing the core emotion is differential, taken into full account in question answering process user feeling with Emotion fluctuates, and by cutting manually objective service platform in time, reduces the not pleasant mood that user in question answering process may accumulate, Be conducive to improve use wish of the user to intelligent Answer System.

Corresponding with the above-mentioned robot answer method based on system function syntax, the embodiment of the present application additionally provides one kind Robot answering device based on system function syntax.

Fig. 4 is the structure diagram of the robot answering device provided by the embodiments of the present application based on system function syntax, The device includes：

Problem search module 401, for using the target that the retrieval of problem search method receives from default knowledge base Problem, wherein, the target problem is the problem of active user inputs；

First responder module 402, if for retrieving the target problem there are in default knowledge base, from described In knowledge base choose with the target problem similarity highest the problem of as matching problem, then from the knowledge base search and The corresponding response of the matching problem, is denoted as the first response；

First optimization module 403, for according to the humanized corresponding system function syntax feature of the first machine, to described One response optimizes, and the response after being optimized is denoted as target response, so that response is more in line with question and answer context semanteme, It is that the intelligent Answer System is not initialised that corresponding machine is humanized, and the machine is humanized that first machine is humanized Including interactive attribute, cognition attribute and formal attribute.

Wherein, which further includes：

Response output module, for exporting the target response to current sessions corresponding with the target problem.

The device further includes：

Target language material data acquisition module, after being marked for acquisition according to robot attributive classification to target corpus data Target corpus data, the target corpus data are the corpus datas obtained from default field；

First training module is instructed for the target corpus data after marking to be input to default convolutional neural networks Practice, by deep learning, the robot attributive classification model after being trained；The robot attributive classification model is to pass through mark Target corpus data after note is trained convolutional neural networks using the training set, as training set by depth It practises, the model after being trained；

First initialization module, for using the robot attributive classification model to the in the intelligent Answer System One machine is humanized to be initialized.

The device further includes：

First conversational mode provides module, for being based on system function syntax principle, refines in the target problem and implies Speech function, and provide the corresponding conversational mode of speech function implied in the target problem, the conversational mode includes Receive or retract, perform or refuse, approve or reject, answer or refuse to answer.

The device further includes：

Second training module, if for not retrieving the target problem in the knowledge base, by deep learning, The target problem not in the knowledge base, the depth intelligent answer model after being trained are trained using object module； The object module for convolutional neural networks CNN, shot and long term memory network model LSTM, Recognition with Recurrent Neural Network model RNN and The combination of at least one of variant or memory network；

Second responder module for the target problem to be generated response using the depth intelligent answer model, is denoted as Second response.

The device can also include：

Second conversational mode provides module, for being based on system function syntax principle, refines in the target problem and implies Speech function, and provide in the target problem imply the corresponding conversational mode of speech function；

Second optimization module, for being based on the conversational mode, according to the humanized corresponding system work(of first machine Energy grammar property, optimizes second response, and the response after being optimized is denoted as the 3rd response, and the 3rd response is made For target response, so that response is more in line with question and answer context semanteme.

The device can also include：

Update module is concentrated for the target problem received to be added to the session that active user prestores, and to institute Session collection is stated to be updated.

The device can also include：

Second robot attribute determination module, for according to updated session collection, utilizing the robot attributive classification Model monitors the target problem in real time, determines that the second machine is humanized；

Assessment module for being extracted the core emotion of active user using default core sentiment classification model, and is assessed The emotion of the core emotion is differential, wherein, the core sentiment classification model is based on sentiment dictionary, weak mark, engineering Practise, deep learning at least one of combination be trained the model of acquisition；

Matching module, for the humanized configuration humanized with first machine of second machine to be matched, It is if mismatch, second machine is humanized humanized as first machine, it returns to execution and utilizes the robot Attributive classification model is to the humanized carry out initialization step of the first machine in default intelligent Answer System；

Judgment module for differential according to the core emotion of active user and the corresponding emotion of active user, judges whether Active user's session content is transferred into artificial objective service platform；The current sessions include the 3rd response and the target Problem；If it is, trigger artificial module；

Artificial module, for active user's session content to be accessed manually objective service platform, the artificial customer service business Platform is the platform for manually conversating with active user.

Above-mentioned artificial module includes：

Acquisition submodule, for obtaining the 3rd response and the target problem, and by the 3rd response and described Target problem generates the session of active user；

Output sub-module, for according to the core sentiment classification model, the emotion potentiality of output active user's session；

Monitoring submodule, for monitoring the emotion potentiality of active user's session of output in real time；

Judging module, for pass through check the current user conversation emotion potentiality whether with pre-set emotion attribute Value matches to decide whether to transfer active user's session content into artificial objective service platform.

Above-mentioned target problem search method is：

It can be seen that device provided in an embodiment of the present invention is special according to the humanized corresponding system function syntax of the first machine Sign to the problem of selection is with target problem similarity highest from knowledge base, and answers the problem in knowledge base Corresponding matching It answers i.e. the first response to be optimized, improves the accuracy of response problem.

The embodiment of the present invention additionally provides a kind of electronic equipment, as shown in figure 5, including processor 501, communication interface 502, Memory 503 and communication bus 504, wherein, processor 501, communication interface 502, memory 503 is complete by communication bus 504 Into mutual communication,

Memory 503, for storing computer program；

Processor 501 during for performing the program stored on memory 503, realizes following steps：

According to the humanized corresponding system function syntax feature of the first machine, first response is optimized, is obtained Response after optimization, is denoted as target response, so that response is more in line with question and answer context semanteme, first machine is humanized to be The intelligent Answer System is not initialised, and corresponding machine is humanized, and the machine is humanized belongs to including interactive attribute, cognition Property and formal attribute；

The communication bus that above-mentioned electronic equipment is mentioned can be Peripheral Component Interconnect standard (Peripheral Component Interconnect, PCI) bus or expanding the industrial standard structure (Extended Industry Standard Architecture, EISA) bus etc..The communication bus can be divided into address bus, data/address bus, controlling bus etc..For just It is only represented in expression, figure with a thick line, it is not intended that an only bus or a type of bus.

Communication interface is for the communication between above-mentioned electronic equipment and other equipment.

Memory can include random access memory (Random Access Memory, RAM), can also include non-easy The property lost memory (Non-Volatile Memory, NVM), for example, at least a magnetic disk storage.Optionally, memory may be used also To be at least one storage device for being located remotely from aforementioned processor.

Above-mentioned processor can be general processor, including central processing unit (Central Processing Unit, CPU), network processing unit (Network Processor, NP) etc.；It can also be digital signal processor (Digital Signal Processing, DSP), it is application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), existing It is field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic device, discrete Door or transistor logic, discrete hardware components.

In another embodiment provided by the invention, a kind of computer readable storage medium is additionally provided, which can It reads to be stored with instruction in storage medium, when run on a computer so that computer performs any institute in above-described embodiment The robot answer method based on system function syntax stated.

In another embodiment provided by the invention, a kind of computer program product for including instruction is additionally provided, when it When running on computers so that computer performs any robot based on system function syntax in above-described embodiment Answer method.

In the above-described embodiments, can come wholly or partly by software, hardware, firmware or its any combination real It is existing.When implemented in software, can entirely or partly realize in the form of a computer program product.The computer program Product includes one or more computer instructions.When loading on computers and performing the computer program instructions, all or It partly generates according to the flow or function described in the embodiment of the present invention.The computer can be all-purpose computer, special meter Calculation machine, computer network or other programmable devices.The computer instruction can be stored in computer readable storage medium In or from a computer readable storage medium to another computer readable storage medium transmit, for example, the computer Instruction can pass through wired (such as coaxial cable, optical fiber, number from a web-site, computer, server or data center Active user's line (DSL)) or wireless (such as infrared, wireless, microwave etc.) mode to another web-site, computer, service Device or data center are transmitted.The computer readable storage medium can be any usable medium that computer can access The data storage devices such as server, the data center either integrated comprising one or more usable mediums.The usable medium Can be magnetic medium, (for example, floppy disk, hard disk, tape), optical medium (for example, DVD) or semiconductor medium (such as solid-state Hard disk Solid State Disk (SSD)) etc..

It should be noted that herein, relational terms such as first and second and the like are used merely to a reality Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation In any this actual relation or order.Moreover, term " comprising ", "comprising" or its any other variant are intended to Non-exclusive inclusion, so that process, method, article or equipment including a series of elements not only will including those Element, but also including other elements that are not explicitly listed or further include as this process, method, article or equipment Intrinsic element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that Also there are other identical elements in process, method, article or equipment including the element.

Each embodiment in this specification is described using relevant mode, identical similar portion between each embodiment Point just to refer each other, and the highlights of each of the examples are difference from other examples.Especially for device, For electronic equipment, computer readable storage medium or computer product embodiment, since it is substantially similar to embodiment of the method, So description is fairly simple, the relevent part can refer to the partial explaination of embodiments of method.

The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the scope of the present invention.It is all Any modifications, equivalent replacements and improvements are made within the spirit and principles in the present invention, are all contained in protection scope of the present invention It is interior.

Claims

1. a kind of robot answer method based on system function syntax, which is characterized in that described applied to intelligent Answer System Method includes：

The target problem received from default knowledge base using the retrieval of problem search method, wherein, the target problem is The problem of active user inputs；

If retrieving the target problem there are in default knowledge base, choose from the knowledge base and asked with the target The problem of inscribing similarity highest is as matching problem, then response corresponding with the matching problem is searched for from the knowledge base, It is denoted as the first response；

According to the humanized corresponding system function syntax feature of the first machine, first response is optimized, is optimized Response afterwards, is denoted as target response, so that response is more in line with question and answer context semanteme, humanized first machine is described The intelligent Answer System corresponding machine that is not initialised is humanized, the machine it is humanized including interactive attribute, cognition attribute and Formal attribute.

2. the method as described in claim 1, which is characterized in that the target problem received is forwarded to default machine described Before device people integrates responder module, the method further includes：

Obtain the target corpus data after being marked according to robot attributive classification to target corpus data, the target corpus data It is the corpus data obtained from default field；

Target corpus data after mark is input to default convolutional neural networks to be trained, by deep learning, is obtained Robot attributive classification model after training；

It is initialized using the robot attributive classification model is humanized to the first machine in the intelligent Answer System.

3. method as claimed in claim 2, which is characterized in that described according to the humanized corresponding system function of the first machine Grammar property optimizes first response, and the response after being optimized is denoted as target response, so that response more accords with After closing question and answer context semanteme, the method further includes：

4. the method as described in claim 1 or 3, which is characterized in that described according to the humanized corresponding system of the first machine Functional Grammar feature optimizes first response, and the response after being optimized is denoted as target response, so that response is more Add before meeting question and answer context semanteme, the method further includes：

Based on system function syntax principle, the speech function implied in the target problem is refined, and provides the target problem In imply the corresponding conversational mode of speech function, the conversational mode include receive or retract, perform or refuse, approve or refute It returns, answer or refuse to answer.

5. method as claimed in claim 4, which is characterized in that problem search method is used from default knowledge base described After retrieving the target problem received, the method further includes：

If not retrieving the target problem in the knowledge base, by deep learning, do not existed using object module training The target problem in the knowledge base, the depth intelligent answer model after being trained；The object module is convolution god Through in network C NN, shot and long term memory network model LSTM, Recognition with Recurrent Neural Network model RNN and variant or memory network at least A kind of combination；

6. method as claimed in claim 5, which is characterized in that utilize the depth intelligent answer model by the mesh described Mark problem generates response, is denoted as after the second response, the method further includes：

Based on system function syntax principle, the speech function implied in the target problem is refined, and provides the target problem In imply the corresponding conversational mode of speech function；

Based on the conversational mode, according to the humanized corresponding system function syntax feature of first machine, to described second Response optimizes, and the response after being optimized is denoted as the 3rd response, using the 3rd response as target response, so that response is more Add and meet question and answer context semanteme.

7. method as claimed in claim 6, which is characterized in that utilize the depth intelligent answer model by the mesh described Mark problem generates response, is denoted as after the second response, the method further includes：

The target problem of reception is added to the session that active user prestores to concentrate, and the session collection is updated.

8. the method for claim 7, which is characterized in that the target problem of reception is added to current use described Session that family prestores is concentrated, and after being updated to the session collection, the method further includes：

According to updated session collection, the target problem is monitored in real time using the robot attributive classification model, determines Two machines are humanized；

Using the core emotion of default core sentiment classification model extraction active user, and assess the emotion of the core emotion It is differential, wherein, the core sentiment classification model be based in sentiment dictionary, weak mark, machine learning, deep learning at least It is a kind of to combine the model for being trained acquisition；

The humanized configuration humanized with first machine of second machine is matched, if mismatch, by described in Second machine is humanized humanized as first machine, returns and performs using the robot attributive classification model to default Intelligent Answer System in the humanized carry out initialization step of the first machine；

It is differential according to the core emotion of active user and the corresponding emotion of active user, judge whether active user's session content It transfers into artificial objective service platform；The current sessions include the 3rd response and the target problem；

If it has, then active user's session content is accessed manually objective service platform, the artificial objective service platform is to use In the platform manually to conversate with active user.

9. method as claimed in claim 8, which is characterized in that described according to the core emotion of active user and active user couple The emotion answered is differential, judges whether current sessions transfer into artificial objective service platform, including：

The 3rd response and the target problem are obtained, and the 3rd response and the target problem are generated into active user Session；

Whether with pre-set emotion attribute value match to decide whether by the emotion potentiality for checking the current user conversation It needs to transfer active user's session content into artificial objective service platform.

10. a kind of robot answering device based on system function syntax, which is characterized in that applied to intelligent Answer System, institute State device：

Problem search module, for using the target problem that the retrieval of problem search method receives from default knowledge base, In, the target problem is the problem of active user inputs；

First responder module, if for retrieving the target problem there are in default knowledge base, from the knowledge base The problem of middle selection and the target problem similarity highest, searches for and described as matching problem, then from the knowledge base With the corresponding response of problem, the first response is denoted as；

First optimization module, for according to the humanized corresponding system function syntax feature of the first machine, to first response It optimizing, the response after being optimized is denoted as target response, so that response is more in line with question and answer context semanteme, described the It is that the intelligent Answer System is not initialised that corresponding machine is humanized, and the machine is humanized including mutually that one machine is humanized Dynamic attribute, cognition attribute and formal attribute.