CN107799116A

CN107799116A - More wheel interacting parallel semantic understanding method and apparatus

Info

Publication number: CN107799116A
Application number: CN201610793380.8A
Authority: CN
Inventors: 黄鑫; 陈志刚; 王智国
Original assignee: iFlytek Co Ltd
Current assignee: iFlytek Co Ltd
Priority date: 2016-08-31
Filing date: 2016-08-31
Publication date: 2018-03-13

Abstract

The application proposes a kind of more wheel interacting parallel semantic understanding method and apparatus, and more wheel interacting parallel semantic understanding methods include：Receive speech data；Cutting is carried out to the speech data, obtains speech data segment, and, speech recognition is carried out to the speech data segment, obtains current cutting text data；Independent semantic understanding is carried out to current cutting text data, obtains the first semantic understanding result, and, semantic understanding is carried out to current cutting text data according to the historical data of user mutual, obtains the second semantic understanding result；The first semantic understanding result or the second semantic understanding result are selected, as final semantic understanding result.This method can improve the degree of accuracy of semantic understanding result, and, to user, Consumer's Experience is lifted so as to feed back more accurately interaction results.

Description

More wheel interacting parallel semantic understanding method and apparatus

Technical field

The application is related to natural language understanding technology field, more particularly to a kind of more wheel interacting parallel semantic understanding methods and Device.

Background technology

With developing rapidly for intelligent terminal and network technology, people are more and more habitually completed using intelligent terminal various Demand, when such as using entrance of the intelligent sound box as man-machine interaction, user can carry out interactive voice to meet user's with it Different demands, weather is such as looked into, stock is looked into, listens music, or user to be used as man-machine interaction using intelligent vehicle device when driving Entrance, user can carry out interactive voice to complete the application demands such as navigation, radio station inquiry, music query with intelligent vehicle device.Make When completing the various demands of user with intelligent terminal, the general mode for using interactive voice, intelligent terminal for reception needs comprising user The speech data asked, corresponding identification text is obtained after speech recognition is carried out to the speech data, then the identification text is entered After row semantic understanding, system feeds back to user mutual result according to the semantic understanding result, so as to complete once to interact；When with When family there are one or more demands, more wheels can be often carried out with intelligent terminal and be interacted, system is according to each demand of user, no It is disconnected to feed back to user mutual result, so as to give a kind of natural and tripping interactive experience of user, if the demand of user is to navigate to section Greatly, after user can interact with the more wheels of system progress, this demand is met, when specifically interacting, user first says with system：" go to section Greatly ", system puts question to " being University of Science and Technology South, North, middle area or thing area ", and user says：" University of Science and Technology South ", system starts to plan Path, a plurality of candidate route planned is fed back into user, after user selectes guidance path, system starts to navigate.

In correlation technique, when user carries out more wheel interactive voices with intelligent terminal, system usual only root in semantic understanding Semantic understanding is carried out according to the text of current cutting.But if user occurs pausing in a request process, drags phenomena such as sound Or other people speak interference tones when, cutting mistake generally occurs in system, can cause semantic understanding mistake accordingly, and then instead Feed the interaction results of user error, seriously reduce Consumer's Experience.

The content of the invention

The application is intended to one of technical problem at least solving in correlation technique to a certain extent.

Therefore, the purpose of the application is to propose much a kind of wheel interacting parallel semantic understanding methods, this method can be with The degree of accuracy of semantic understanding result is improved, to user, Consumer's Experience is lifted so as to feed back more accurately interaction results.

Further object is to propose a kind of more wheel interacting parallel semantic understanding devices.

To reach above-mentioned purpose, more wheel interacting parallel semantic understanding methods that the application first aspect embodiment proposes, bag Include：Receive speech data；Cutting is carried out to the speech data, obtains speech data segment, and, to the speech data piece It is disconnected to carry out speech recognition, obtain current cutting text data；Independent semantic understanding is carried out to current cutting text data, obtains the One semantic understanding result, and, semantic understanding is carried out to current cutting text data according to the historical data of user mutual, obtained Second semantic understanding result；The first semantic understanding result or the second semantic understanding result are selected, as final Semantic understanding result.

More wheel interacting parallel semantic understanding methods that the application first aspect embodiment proposes, by according to historical data pair Cutting text data carries out semantic understanding, and more information can be combined in semantic understanding, can obtain more accurately semantic Understand result, in addition, selecting one kind in two kinds of semantic understanding results, more suitably semantic manage can be selected according to actual conditions Result is solved, to user, Consumer's Experience is lifted so as to feed back more accurately interaction results.

To reach above-mentioned purpose, more wheel interacting parallel semantic understanding devices that the application second aspect embodiment proposes, bag Include：Receiving module, for receiving speech data；Sound identification module, for carrying out cutting to the speech data, obtain voice Data fragments, and, speech recognition is carried out to the speech data segment, obtains current cutting text data；Semantic understanding mould Block, for carrying out independent semantic understanding to current cutting text data, the first semantic understanding result is obtained, and, according to user Interactive historical data carries out semantic understanding to current cutting text data, obtains the second semantic understanding result；Selecting module, use In selecting the first semantic understanding result or the second semantic understanding result, as final semantic understanding result.

More wheel interacting parallel semantic understanding devices that the application second aspect embodiment proposes, by according to historical data pair Cutting text data carries out semantic understanding, and more information can be combined in semantic understanding, can obtain more accurately semantic Understand result, in addition, selecting one kind in two kinds of semantic understanding results, more suitably semantic manage can be selected according to actual conditions Result is solved, to user, Consumer's Experience is lifted so as to feed back more accurately interaction results.

The aspect and advantage that the application adds will be set forth in part in the description, and will partly become from the following description Obtain substantially, or recognized by the practice of the application.

Brief description of the drawings

The above-mentioned and/or additional aspect of the application and advantage will become from the following description of the accompanying drawings of embodiments Substantially and it is readily appreciated that, wherein：

Fig. 1 is the schematic flow sheet for more wheel interacting parallel semantic understanding methods that the application one embodiment proposes；

Fig. 2 is the schematic flow sheet for the method for being ranked up amendment in the embodiment of the present application to initial semantic understanding result；

Fig. 3 is the flow that the method for the first semantic understanding result or the second semantic understanding result is selected in the embodiment of the present application Schematic diagram；

Fig. 4 is the flow for the method for carrying out semantic understanding in the embodiment of the present application to cutting text data according to historical data Schematic diagram；

Fig. 5 is the schematic flow sheet for the method for being ranked up amendment in the embodiment of the present application to secondary semantic understanding result；

Fig. 6 is the structural representation for more wheel interacting parallel semantic understanding devices that the application one embodiment proposes；

Fig. 7 is the structural representation of more wheel interacting parallel semantic understanding devices of the application another embodiment proposition.

Embodiment

Embodiments herein is described below in detail, the example of the embodiment is shown in the drawings, wherein from beginning to end Same or similar label represents same or similar module or the module with same or like function.Below with reference to attached The embodiment of figure description is exemplary, is only used for explaining the application, and it is not intended that limitation to the application.On the contrary, this All changes that the embodiment of application includes falling into the range of the spirit and intension of attached claims, modification and equivalent Thing.

It is further illustrated in order to be best understood from the application, the problem of first with a specific example in correlation technique.

For example, user inputs speech data as " I wants the snowman for listening model to know tawny daylily ... ", wherein, ellipsis (...) represents user Because think deeply or speak custom and caused by pause or drag sound, due to pausing or dragging the presence of sound, system semantics Understanding Module receives Interaction request twice, i.e. " I wants to listen model to know tawny daylily " and " snowman ", for the interaction request of first time, system progress semantic understanding Afterwards, the song (but being not necessarily " snowman " this song) of model dawn tawny daylily is directly played；For secondary interaction request, system is carried out After semantic understanding, song " snowman " (being probably that model dawn tawny daylily is sung, it is also possible to the song of the same name that other singers sing) is played, also may be used The feedback for chatting result " winter arrives, and I also likes making a snowman " can be provided, no matter which kind of feeds back, and basically can not all meet The real demand of user, user is allowed to feel the not smooth of interaction, Consumer's Experience effect is poor.

Pass through the analysis to above-mentioned example, it is found that a major reason for causing above-mentioned interaction results inaccurate is related Semantic understanding in technology has only carried out semantics recognition to current cutting result, i.e. " I wants to listen model to know tawny daylily " is carried out respectively Semantics recognition and semantics recognition carried out to " snowman ", without will the historical data such as " I wants to listen model to know tawny daylily " and " snow People " is combined identification.

It is inaccurate in order to solve above-mentioned semantics recognition result present in correlation technique, the not smooth and poor user experience of interaction The problems such as, the application will be given below embodiment.

Fig. 1 is the schematic flow sheet for more wheel interacting parallel semantic understanding methods that the application one embodiment proposes.

As shown in figure 1, the method for the present embodiment includes：

S11：Receive speech data.

When user needs to carry out interactive voice with intelligent terminal, intelligent terminal can receive the voice number that user says According to.

S12：Cutting is carried out to the speech data, obtains speech data segment, and, the speech data segment is entered Row speech recognition, obtain current cutting text data.

Under the interactive mode of full duplex, user continuously naturally can carry out more wheels with system and interact, and system is receiving To after the speech data of user's input, it is necessary first to carry out cutting, each voice that cutting is obtained to the speech data of reception As currently interactive speech data, system needs to make feedback to the speech data data fragments.Specific cutting method It is unlimited, prior art or the technology occurred in the future can be used to realize, such as extract the acoustic feature of speech data, such as mel-frequency Cepstrum coefficient (Mel Frequency Cepstrum Coefficient, MFCC) or filterbank, use training in advance Segmentation model carries out cutting to the speech data of reception according to acoustic feature, and the segmentation model is common classification in pattern-recognition Model, such as deep neural network model or supporting vector machine model.

Each speech data piece after cutting is obtained is had no progeny, and is carried out speech recognition to each speech data segment, is obtained Corresponding text data, text data corresponding to each speech data segment are properly termed as cutting text data.Specific voice Recognizer is unlimited, and technology that is existing or occurring in the future can be used to realize.

S13：Independent semantic understanding is carried out to current cutting text data, obtains the first semantic understanding result, and, according to The historical data of user mutual carries out semantic understanding to current cutting text data, obtains the second semantic understanding result.

Wherein, independent semantic understanding refers to directly carry out semantic understanding to current cutting text data, without reference to history number According to, it is corresponding, semantic understanding is carried out to current cutting text data according to historical data and referred to current cutting text data Carry out referring to historical data during semantic understanding.

The flow that semantic understanding is specifically carried out according to historical data may refer to subsequent embodiment.

It is understood that independent semantic understanding and according to historical data carry out semantic understanding sequential do not limit.Enter one Step, in order to improve treatment effeciency, parallel processing can be carried out to both semantic understanding processes.

The flow that independent semantic understanding is carried out to current cutting text data can be as follows：

The independent semantic understanding directly carries out semantic understanding to current cutting text data, not with reference to user mutual Historical data, it is specific when carrying out semantic understanding, using the semantic scene of systemic presupposition, current cutting text data is carried out just Beginning semantic understanding, obtain the initial semantic understanding result of semantic understanding.

The semantic scene of the systemic presupposition is preset according to systematic difference demand, the semantic scene such as music, Film, ticket booking, food and drink etc..

When carrying out initial semantic understanding to current cutting text data, such as semantic analysis based on syntax, base can be used The initial semantic understanding result under each semantic scene is obtained in the semantic analysis of grammar rule network.

, can so that the semantic analysis based on grammar rule network carries out initial semantic understanding to current cutting text data as an example To write in advance under each semantic scene based on extension Backus normal form (Augmented Backus-Naur Form, ABNF) text Method rule, the weighted finite state machine under each semantic scene is then generated according to the grammar rule compiling under each semantic scene (Weighted Finite State Transducer, WFST) network, exists respectively according to the WFST networks under each semantic scene Semantic understanding is carried out to current cutting text data under each semantic scene, obtains the initial semantic understanding knot under each semantic scene Fruit.It is understood that the particular content of the semantic understanding process under each semantic scene may refer to various existing or go out in the future Existing technology, will not be described in detail herein.

Corresponding each semantic scene, the initial semantic understanding result under corresponding semantic scene is 0, one or more.Example Such as, confidence level (static state setting is dynamically determined) can be predefined under every kind of semantic scene, the probable value of respective path is big In confidence level initial semantic understanding result as the alternative initial semantic understanding result under the semantic scene, getting candidate , can be according to predetermined rule, using all alternative initial semantic understanding results as phase after initial semantic understanding result Answer the initial semantic understanding result under semantic scene；Or to all alternative initial semantic understanding results according to respective path Probable value be ranked up, then sequentially select predetermined number alternative initial semantic understanding result as under corresponding semantic scene Initial semantic understanding result, i.e. nbest results, n can determine according to application demand.

The information that each initial semantic understanding result includes includes：Semantic scene name, semantic scene value, semantic groove and language Adopted groove value, can be expressed as { semantic scene name:Semantic scene value, semantic groove 1：The semantic value of groove 1, semantic groove 2：It is semantic The value ... of groove 2, semantic groove n：Semantic groove n values }, wherein semantic groove is preset according to semantic scene and application demand.

1best semantic understandings result when such as current cutting text data is " concert version " is { " service ":” music”,“version”:" concert version ", wherein " service " represents semantic scene name, " music " represents semantic scene Value (music scenario), " version " represent semantic groove (version), and " concert version " represents semantic groove value.

Semantic understanding is carried out to current cutting text data, after obtaining the initial semantic understanding result under each semantic scene, First semantic understanding result can be obtained according to the initial semantic understanding result under each semantic scene.

In some embodiments, probability that can be according to the initial semantic understanding result under each semantic scene according to respective paths The order of value from high to low is ranked up, and sequentially selects the initial semantic understanding result of predetermined number to make from the result after sequence For the first final semantic understanding result.Or

, can be to these initial languages after the initial semantic understanding result under obtaining each semantic scene in some embodiments Reason and good sense solution result is ranked up amendment, in the revised initial semantic understanding result that sorts, sequentially selects the first of predetermined number Beginning semantic understanding result is as the first final semantic understanding result.Particular content can be as will be shown later.

S14：The first semantic understanding result or the second semantic understanding result are selected, as final semanteme reason Solve result.

, can be in both semantic understanding results after the first semantic understanding result and the second semantic understanding result is obtained Middle selection is a kind of as final semantic understanding result, that is, selects the first semantic understanding result as final semantic understanding knot Fruit, or, the second semantic understanding result is selected as final semantic understanding result.

Further, as shown in Fig. 2 the flow for the method for being ranked up amendment to initial semantic understanding result includes：

S21：Obtain the related data of initial semantic understanding result.

Under each semantic scene, the higher number of the correlation of the initial semantic understanding result of vertical search engine search is utilized According to.

During specific search, each initial semantic understanding result is inputted as the search string of search engine, search engine is certainly The dynamic search condition that the search string is switched to inside, the related document of each initial semantic understanding result is searched for, and will sequence Relevant documentation afterwards returns, and each initial semantic understanding result can select TopN relevant documentation, as each initial semantic Understand the related data of result；Detailed process is same as the prior art, will not be described in detail herein.

S22：It is special that the degree of correlation is extracted according to initial semantic understanding result, current cutting text data and the related data Sign.

Wherein it is possible to corresponding each semantic scene extracts the degree of correlation feature under corresponding semantic scene.

The degree of correlation feature is used to describe the degree of correlation between initial semantic understanding result and current cutting text data And the accuracy of initial semantic understanding sort result.

Degree of correlation feature includes at least one in following item：

Initial semantic understanding result is with including key in current cutting text data same words number, initial semantic understanding result Semantic slot number, with crucial semantic groove value temperature contained by the initial semantic understanding result of word identical in current cutting text data, Current cutting text data includes the quantity of the keyword related to semantic scene.

Specific extracting method is as described below：

(1) initial semantic understanding result and current cutting text data same words number

The initial semantic understanding result and current cutting text data same words number, refer to initial semantic understanding result institute Word in the current cutting text data understood, specifically can contained semantic groove value in matching initial semantic understanding result successively Each word after being segmented with current cutting text data, it is determined that semantic groove value and current cutting text in initial semantic understanding result Notebook data identical word number.

(2) crucial semantic slot number is included in initial semantic understanding result

Count the key included under each semantic scene in each initial semantic understanding result in its affiliated semantic understanding scene Semantic slot number, more comprising crucial semantic slot number, the correlation of semantic understanding result and its affiliated semantic scene is higher, so that In each semantic scene, the initial semantic understanding result higher with each semantic scene correlation is come before；In different languages Between adopted scene, the crucial semantic groove of scene that initial semantic understanding result includes is more, and the initial semantic understanding sort result is got over It is forward.

(3) and in current cutting text data crucial semantic groove value is hot contained by the initial semantic understanding result of word identical Degree.

The crucial semantic groove in each initial semantic understanding result is found, each crucial semantic groove value of matching is with working as successively Whether word is identical in preceding cutting text data, if identical, calculates the temperature of current key semanteme groove value；Crucial semantic groove The related data that the temperature of value can search according to the initial semantic understanding result of the affiliated semantic scene of the semantic groove of key It is calculated；The data that vertical page in the related data such as search result includes, it is corresponding to hang down such as in music scenario The straight page is search dog music, cruel my music, Tengxun's music etc., the use of crucial semantic groove value according to the vertical page Family click volume, user's playback volume, user's volumes of searches, crucial semantic groove value are in information such as the distributions of multiple vertical pages, to working as The temperature information of the preceding semantic groove of key carries out comprehensive analysis, obtains the temperature of crucial semantic groove, concrete analysis computational methods with it is existing There is technology identical, will not be described in detail herein.

(4) current cutting text data includes the quantity of the keyword related to semantic scene

The keyword of each semantic scene can be entered by collecting the mass text data under each semantic scene in advance After row participle, the word frequency of each word is counted, word frequency is more than to keyword of the word as each semantic scene of threshold value；Such as music field Keyword in scape is " listening ", " song ", " first ", " broadcasting " etc.；The semantic scene keyword that current cutting text data includes More, the semantic understanding sort result under the semantic scene is more forward.

S23：According to the degree of correlation feature and the order models built in advance, initial semantic understanding result is ranked up Amendment.

Wherein, order models generate after being trained in the training stage to training data.For example, collect a large amount of Text data, semantic understanding is carried out to text data and obtains the initial semantic understanding result under each semantic scene, and extraction Degree of correlation feature, the sequence to the initial semantic understanding result under each semantic scene is manually marked, afterwards according to the degree of correlation Feature and the sequence manually marked are trained, and generate order models.The input of order models is degree of correlation feature, and output is row Initial semantic understanding result under the revised each semantic scene of sequence.

Therefore, after extraction obtains degree of correlation feature, the revised initial semanteme that sorts can be obtained according to order models Understand result.

After the revised initial semantic understanding result that obtains sorting, TopN can be selected semantic to be managed as final first Result is solved, the N can determine according to application demand, if N values are 10.

, can be in semanteme by carrying out semantic understanding to current cutting text data according to historical data in the present embodiment More information are combined when understanding, more accurately semantic understanding result can be obtained, in addition, being selected in two kinds of semantic understanding results One kind is selected, more suitably semantic understanding result can be selected according to actual conditions, so as to feed back more accurately interaction results To user, Consumer's Experience is lifted.

Fig. 3 is the flow that the method for the first semantic understanding result or the second semantic understanding result is selected in the embodiment of the present application Schematic diagram.

It is that first extraction selection is special when a kind of semantic understanding result is selected in two kinds of semantic understanding results in the present embodiment Sign, selected further according to selection feature and semantic results preference pattern.

Selection feature describes the first semantic understanding result and the second semantic understanding result, and current cutting text from different perspectives Relation between notebook data and user interaction history, for the semantic understanding result for selecting system finally to use；Generally, If when current cutting text data and the smaller correlation of user interaction history data, made using the first semantic understanding result It is appropriate for the final semantic understanding result of system, if current cutting text data and the correlation of user interaction history data When larger, then appropriate using the second semantic understanding result, specifically chosen process is as described below.

As shown in figure 3, this method includes：

S31：According to initial the semantic understanding result and user mutual of current cutting text data, current cutting text data Historical data extraction selection feature.

Wherein, the historical data includes at least one in following item：

The feedback text data of last round of interaction, the history semantic understanding result of last round of interaction, last round of interaction are gone through History cutting text data.

Select feature can be including at least one in following item：

Current cutting text data length, current cutting text data replace the feedback text data middle finger of last round of interaction For the semantic matching degree before and after body, the semantic integrity degree of current cutting text data, current cutting text data initial semanteme Understand semantic groove set contained by result with the registration of semantic groove set contained by the last round of history semantic understanding result interacted, when Semantic groove value contained by the initial semantic understanding result of preceding cutting text data and the last round of history semantic understanding result interacted Semantic groove contained by the matching degree of contained semantic groove value, the initial semantic understanding result of current cutting text data and last round of friendship The joint degree of semantic groove, current cutting text data and the last round of history cutting interacted contained by mutual history semantic understanding result Semantic integrity degree after text data connection.

The specific extraction content of each selection feature is as follows：

(1) current cutting text data length

The number of words that the current cutting text data length is typically included using current cutting text data represents.Statistics is worked as The number of words that preceding cutting text data includes, you can obtain the length of current cutting text data.General current cutting text data It is longer, independent user's request is more likely to be, final semantic understanding result directly uses the first semantic understanding result, such as The current cutting text data of fruit is shorter, and current cutting text data content is more possible to the historical data phase with user mutual Close, final semantic understanding result is more suitable using the second semantic understanding result.

(2) current cutting text data replaces the feedback text data middle finger of last round of interaction for the semantic matches before and after body Degree

The current cutting text data replaces the feedback text data middle finger of last round of interaction for semantic before and after body After the feedback text data middle finger for the system for referring to utilize current cutting text data to replace last round of interaction with degree is for body, generation New feedback text data, the semantic matches before calculating the new feedback text data and replacing between feedback text data Degree.

The semantic groove that the reference body typically includes according to feedback text data obtains, and system generation is anti-comprising reference body When presenting text data, two kinds of clause, i.e. general question, alternative question are usually contained；The general question such as feeds back text data For " you think when set out", by interactive system according to interaction data and corresponding template " you think $ when" generation, this In $ when i.e. refer to body " when "；The alternative question is as fed back text data for " you want first block, coach seat Or sleeping berth" by interactive system, according to interaction data and corresponding template, " you want $ seat [0], $ seat [1], or $ Seat [2]" generation, " $ seat [0], $ seat [1], or $ seat [2] " here namely to refer to body " first block, second-class Seat, sleeping berth ", wherein, seat [0], seat [1], seat [2] represent three kinds of modes of train order of seats respectively, with specific reference to upper During one wheel interaction, system generation is fed back in the data searched during text data, is searched according to semantic groove seat time (seat) Value dynamically obtains after being clustered.

During feature extraction, the semantic reason of text data progress after body is replaced is referred to the feedback text data of last round of interaction Solution, specific semantic analytic method is identical with independent semantic understanding method, refers to so as to obtain the feedback text data of last round of interaction Semantic understanding result after being replaced for body；The matching degree that the reference body replaces front and rear feedback text semantic understanding is calculated, specifically Calculate same as the prior art, will not be described in detail herein.

If without body is referred in the feedback text data of last round of interaction, this feature value is 0；

Cutting text data such as last round of interaction is " helping me to look into lower Hefei to Beijing ", the feedback textual data that system provides According to for " you think by air still train", the current cutting text data of user's currently interaction is " train of tomorrow ", ought After the reference body in feedback text data when preceding cutting text data replaces last round of interactive, new feedback text data is obtained " you want to sit the train of tomorrow ", during feature extraction, semantic understanding is carried out to referring to the feedback text data after body is replaced, is referred to Semantic understanding result after being replaced for body, the matching degree for calculating feedback text data semantic understanding result before and after reference body is replaced are Can.

(3) the semantic integrity degree of current cutting text data

Whether the semanteme that the semantic integrity degree of the current cutting text data is used to describe current cutting text data is complete Whole, such as " Liu Dehua lustily water ", semantic integrity degree is 1.0, and " Liu Dehua's " semantic integrity degree is 0.3, the language of " " Adopted integrity degree is 0.0；Specific extracting method can be by extracting the semantic feature of current cutting text data, and training in advance text Notebook data semanteme integrity degree forecast model carries out semantic integrity degree to current cutting text data and predicts to obtain；The semantic feature Can be obtained according to the independent semantic understanding result of current cutting text data, such as whether comprising semantic groove, semantic groove sum, Whether the length (number of words included using semantic groove value is represented) and semantic understanding result of each semantic groove value are comprising semanteme The features such as scene keyword, the semantic integrity degree forecast model is that regression model is commonly used in pattern-recognition, such as logistic regression mould Type, after specifically can carrying out independent semantic understanding by collecting a large amount of text datas in advance, semantic feature is extracted, and mark text The semantic integrity degree of data, train to obtain according to the semantic integrity degree of the semantic feature and mark.

(4) semantic groove set and the last round of history interacted contained by the initial semantic understanding result of current cutting text data The registration of semantic groove set contained by semantic understanding result

Wherein, the history semantic understanding result of last round of interaction refers to the final language of the cutting text data of last round of interaction Reason and good sense solution result.

Language contained by the initial semantic understanding result (referred to as initial semantic understanding result) of the current cutting text data Adopted groove set describes with the registration of semantic groove set contained by the last round of history semantic understanding result interacted from the semantic angle of the v-groove Current cutting text data and the degree of correlation of last round of interactive history data.

During specific extraction, semantic groove set contained by initial semantic understanding result and history semantic understanding result are obtained respectively Contained semantic groove set, wherein, semantic groove set contained by initial semantic understanding result is specifically referred to by initial semantic understanding result The set of contained semantic groove composition；Semantic groove set is specifically referred to by history semantic understanding result contained by history semantic understanding result The set of contained semantic groove composition.Further, above-mentioned history semantic understanding result is specifically the history language of last round of interaction Reason and good sense solution result.

Generally, the registration is higher, and current cutting text data is more likely to be to last round of interactive history Data are modified, and at this moment should use the first semantic understanding result, and therefore, system uses the probability of the first semantic understanding result Higher, conversely, registration is lower, it is two to be more likely to be same sentence interactive voice data by the cutting of system mistake, at this moment The second semantic understanding result should be used, therefore, system is finally higher using the probability of the second semantic understanding result.

(5) semantic groove value and the last round of history interacted contained by the initial semantic understanding result of current cutting text data The matching degree of semantic groove value contained by semantic understanding result

During specific extraction, semantic groove value and history semantic understanding knot contained in initial semantic understanding result is extracted respectively Semantic groove value contained by fruit, this two groups of semanteme groove values can separately constitute a character string, and above-mentioned matching degree feature can To specifically refer to the editing distance between the two character strings.Editing distance (Edit Distance), also known as Levenshtein Distance, between referring to two character strings, as the minimum edit operation number needed for one changes into another.The edit operation of license Including a character is substituted for into another character, a character is inserted, deletes a character.In general, editing distance is got over Small, the similarity of two character strings is bigger.The specific calculation of editing distance can use existing or occurring in the future various Technology realizes that such as redirecting cost matrix based on pronunciation phonemes calculates editing distance.

(6) semantic groove contained by the initial semantic understanding result of current cutting text data and the last round of history interacted are semantic Understand the joint degree of semantic groove contained by result

The joint degree of the semantic groove is calculated according to a large amount of interaction datas collected in advance, and the interaction data can be with Obtained according to user interaction history data, when specifically calculating, according to collected interaction data, calculate current cutting textual data According to the initial semantic understanding result of semantic groove contained by the history semantic understanding result of last round of interaction and current cutting text data The number and the last round of history language interacted of current cutting text data that contained semantic groove occurs simultaneously in the interaction data The ratio of semantic groove occurrence number in the interaction data of collection contained by reason and good sense solution result, it is described while occur referring in interaction data In the interaction of adjacent two-wheeled in, order and current cutting text that last round of interaction data and next round user interactive data occur The last round of interaction data of data is as the order that current cutting text data occurs, shown in circular such as formula (1)：

Wherein, T_cRepresent that semantic groove contained by the initial semantic understanding result of current cutting text data interacts with last round of The joint degree of semantic groove contained by history semantic understanding result；count(slot_c|slot_hc) represent the upper of current cutting text data Contained by the initial semantic understanding result of semantic groove and current cutting text data contained by the history semantic understanding result of one wheel interaction The number that semantic groove occurs simultaneously in the interaction data of collection, count (slot_hc) represent the last round of of current cutting text data The number that semantic groove contained by interactive history semantic understanding result occurs in interaction data.

(7) the semantic integrity degree after current cutting text data is connected with the history cutting text data of last round of interaction

The current cutting text data is connected with the history cutting text data of last round of interaction to be referred to currently directly Cutting text data is put into behind last round of interactive history cutting text data, so as to form new text data；Calculating should The semantic integrity degree of new text data, circular and the method phase for calculating current cutting text data semanteme integrity degree Together, will not be described in detail herein.

S32：According to the selection feature and the semantic results preference pattern built in advance, the first semantic understanding result is selected Or the second semantic understanding result.

Input using the selection feature as the preference pattern, mould is selected using the semantic results of the training in advance Type is selected parallel semantic understanding result, is exported as every kind of language in the first semantic understanding result and the second semantic understanding result Reason and good sense solution result is the select probability of correct semantic understanding result, and the larger semantic understanding result of select probability is as final language Reason and good sense solution result.

The semantic results preference pattern by collecting a large number of users interaction data in advance, after carrying out parallel semantic understanding, Extraction selection feature, and mark correct semantic understanding result in the parallel semantic understanding result；Utilize the selection feature And the annotation results of semantic understanding result train the semantic results preference pattern, using the selection feature as semantic understanding knot The input of fruit preference pattern, make every kind of semantic understanding in parallel semantic understanding result as the probability of correct semantic understanding result For output, the model parameter is updated according to the mark feature of the parallel semantic understanding, after parameter renewal terminates, obtained To semantic results preference pattern；Detailed process is same as the prior art, will not be described in detail herein；The semantic results preference pattern is Common classification model in pattern-recognition, such as decision-tree model, supporting vector machine model, deep neural network model.

In the present embodiment, pass through the semantic understanding that can select to be more suitable for according to selection feature and semantic results preference pattern As a result.

It is above-mentioned to relate to obtain the second semantic understanding to current cutting text data progress semantic understanding according to historical data As a result content, the flow for carrying out semantic understanding to current cutting text data according to historical data is carried out specifically below It is bright.

Fig. 4 is the flow for the method for carrying out semantic understanding in the embodiment of the present application to cutting text data according to historical data Schematic diagram.

As shown in figure 4, the method for the present embodiment includes：

S41：Independent semantic understanding is carried out to current cutting text data, obtains initial semantic understanding result.

In independent semantic understanding, each semantic scene can be corresponded to, under each semantic scene, to current cutting text Notebook data carries out independent semantic understanding, obtains the initial semantic understanding result under each semantic scene.

The semantic scene can be preset according to systematic difference demand, the semantic scene such as music, film, be ordered Ticket, food and drink etc..

When carrying out independent semantic understanding to current cutting text data, the semantic analysis based on syntax, base can be used The initial semantic understanding result under each semantic scene is obtained in semantic analysis of grammar rule network etc..

So that the semantic analysis based on grammar rule network carries out independent semantic understanding to current cutting text data as an example, It can write in advance under each semantic scene based on extension Backus normal form (Augmented Backus-Naur Form, ABNF) Grammar rule, the weighted finite state machine under each semantic scene is then generated according to the grammar rule compiling under each semantic scene (Weighted Finite State Transducer, WFST) network, exists respectively according to the WFST networks under each semantic scene Semantic understanding is carried out to current cutting text data under each semantic scene, obtains the initial semantic understanding knot under each semantic scene Fruit.It is understood that the particular content of the semantic understanding process under each semantic scene may refer to various existing or go out in the future Existing technology, will not be described in detail herein.

S42：According to the initial semantic understanding result and the historical data of user mutual, current cutting text data is extracted Semantic feature.

Wherein, historical data includes：History semantic understanding result, and/or, feed back the semantic understanding result of text data.

The semantic feature is used to carry out secondary semantic understanding to current cutting text data.Semantic feature includes following item In it is one or more：

Initial semantic understanding result feature, history semantic understanding result feature, semantic groove contained by initial semantic understanding result Set takes with the semantic groove of key contained by the registration of semantic groove set contained by history semantic understanding result, initial semantic understanding result Value and semantic groove set contained by the matching degree of crucial semantic groove value contained by history semantic understanding result, initial semantic understanding result With the semantic understanding result of last round of feedback text data contained by semantic groove set registration.

The specific extraction flow of above-mentioned each semantic feature is as follows：

(1) initial semantic understanding result feature

The initial semantic understanding result feature obtains according to initial semantic understanding result.

During specific extraction, feature name and feature value can be extracted according to initial semantic understanding result, by feature name and spy The information after valued combinations is levied as initial semantic understanding result feature.

Wherein, the determination mode of feature name and feature value can be set.For example, semantic scene name and semantic scene are taken Value connected by the use of connector after as a feature name, feature value corresponding to this feature name show be in initial semantic understanding result It is no this feature name occur；Using semantic groove as another feature name, feature value shows initial semanteme corresponding to this feature name Understand the semantic groove of this in result whether there is value.

Above-mentioned connector be it is settable, for example with " " either "-" or other symbols as connector.Feature Value can be 0 or 1.For example, corresponding to above-mentioned previous feature name, it is right feature name occur in 1 expression semantic understanding result The semantic scene name and semantic scene value answered, otherwise, feature value are 0.Or corresponding above-mentioned later feature name, 1 table Show that semantic groove has value in initial semantic understanding result, otherwise, feature value is 0.

For example, initial semantic understanding result is { " service ":”music”,“version”:" concert version ", then carry The initial semantic understanding result taken is characterized as (service.music:1,version:1), wherein " service.music " is language Feature name after the connection of adopted scene and semantic scene value, is the value of this feature after colon, " version " semantic cavity feature , it is the value of this feature, because in semantic understanding result, " version " has corresponding value, the then feature extracted after colon The value of " version " is 1.

(2) history semantic understanding result feature

The history semantic understanding result feature can obtain according to history semantic understanding result.Specifically it is referred to basis The mode that initial semantic understanding result obtains initial semantic understanding result feature obtains.

History semantic understanding result refers to the final semantic understanding result to historical data, and further, historical data can To specifically refer to history cutting text data, i.e., text data corresponding to the speech data of user's input.

Further, history semantic understanding result can be the semantic understanding result of more wheel historical datas, now, extraction History semantic understanding result is characterized as multiple, forms semantic understanding result characteristic sequence, specifically considers that a few wheel historical datas can be with Determined according to application demand, such as consider 5 wheel historical datas, then including 5 history semantic understanding result features.

After often wheel interaction terminates, system, which can preserve, often takes turns user mutual text data and its final semantic understanding result, institute Final semantic understanding result is stated as the form of expression of initial semantic understanding result, i.e., is all by semantic scene, semantic scene Value, semantic groove and semantic groove value composition, therefore, feature extracting method and the extracting method phase to initial semantic understanding result Together, will not be described in detail herein.

For example, current cutting text data is " concert version ", it is assumed that the historical data of user mutual is included " TV pass ", " faith for carrying out first Zhang Xinzhe ", user interaction history includes two-wheeled altogether, then the history semantic understanding characteristic sequence extracted is total to Comprising two features, represented using bracket, i.e. [(service.smartHome:1,object.tv:1,action:1), (service.music:1,artist:1,song:1)]；Wherein, the information in each round bracket is a history semantic understanding As a result feature, " service.smartHome " and " object.tv " is that semantic scene name takes with semantic scene in first feature Feature name after value connection, " action " is semantic cavity feature name, and " service.music " is semantic scene in second feature Feature name after being connected with semantic scene value, " artist " and " song " are semantic cavity feature name；Each the value of feature is Numeral after colon.

(3) weight of semantic groove set and semantic groove set contained by history semantic understanding result contained by initial semantic understanding result It is right

The registration describes the degree of correlation of current cutting text data and user interaction history from the semantic angle of the v-groove.

During specific extraction, semantic groove set contained by initial semantic understanding result and history semantic understanding result are obtained respectively Contained semantic groove set, wherein, semantic groove set contained by initial semantic understanding result is specifically referred to by initial semantic understanding result The set of contained semantic groove composition；Semantic groove set is specifically referred to by history semantic understanding result contained by history semantic understanding result The set of contained semantic groove composition.Further, when multiple history semantic understanding results be present, a history language can be selected Reason and good sense solution result, then semantic groove set corresponding to determining.One history semantic understanding result of selection can be specifically last round of Interactive history semantic understanding result.

The calculation of above-mentioned registration is：The quantity with identical semantic slot name is removed in above-mentioned two semanteme groove set With the element sum in semantic groove set contained by initial semantic understanding result.

For example, cutting text data is " concert version ", initial semantic understanding result is { " service ":”music”, “version”:" concert version ", then the element sum in semantic groove set contained by initial semantic understanding result is 1, and this yuan Element is specially：Semantic groove is " version ", it is assumed that last round of historical data is " come first Zhang Xinzhe faith ", the last round of history The song of various versions corresponding to the faith sung in the semantic understanding result of data comprising Zhang Xinzhe, therefore, what is obtained goes through The semantic groove that semantic groove set contained by history semantic understanding result includes is " song title ", " singer ", " version ", " affiliated special edition "； Then identical semanteme groove, only " version " in two semantic groove set, identical semantic slot number amount are 1, use 1 divided by initial semanteme Understand the number of the element in semantic groove set contained by result, only 1 element, the registration finally given is 1.

(4) crucial semantic groove value contained by initial semantic understanding result and crucial semantic groove contained by history semantic understanding result The matching degree of value

The semantic groove of key refers under each semantic scene according to application demand semantic groove set in advance, using music scenario as Example, such as semantic groove of key set in advance, " song title ", " singer ", " special edition ", " version ", " source (such as video display interlude, variety Program) ", " language ", " region " etc..

After crucial semantic groove is determined, it can enter respectively in initial semantic understanding result and history semantic understanding result Row extraction, obtains the semantic groove of key therein and the semantic groove value of corresponding key, so as to get initial semanteme respectively Understand crucial semantic groove value contained by crucial semantic groove value and history semantic understanding result contained by result.Further, depositing In multiple history semantic understanding results, specifically chosen one or more history semantic understanding results can be true with application demand It is fixed.

Getting respectively contained by crucial semantic groove value and history semantic understanding result contained by initial semantic understanding result After crucial semantic groove value, this two groups crucial semantic groove values can separately constitute a character string, above-mentioned matching degree feature The editing distance between the two character strings can be specifically referred to.Editing distance (Edit Distance), also known as Levenshtein distances, between referring to two character strings, as the minimum edit operation number needed for one changes into another.Perhaps Can edit operation include a character being substituted for another character, insert a character, delete a character.It is general next Say, editing distance is smaller, and the similarity of two character strings is bigger.The specific calculation of editing distance can be used existing or incited somebody to action Various technologies to occur realize that such as redirecting cost matrix based on pronunciation phonemes calculates editing distance.

(5) semantic groove set contained by initial semantic understanding result and the semantic understanding result of last round of feedback text data The registration of contained semantic groove set.

Used before the last round of feedback text data refers to current cutting text data in the last interactive history Obtained system response results are asked at family, and such as user, the last request is " Zhang Xinzhe faith ", and system is to the anti-of user Present as " faith that Zhang Xinzhe please be appreciate ", then last round of feedback text data is " faith that please appreciate Zhang Xinzhe ".It is right afterwards The feedback text data carries out semantic understanding, and specific semantic understanding flow is referred to carry out semantic understanding to cutting text data The flow of initial semantic understanding result is obtained, can obtain feeding back the semantic understanding result of text data.Further, it is similar right The processing of initial semantic understanding result corresponding to cutting text data, it can obtain feeding back the semantic understanding result institute of text data Containing semantic groove set.

Obtain semantic groove set contained by initial semantic understanding result and feed back the semantic understanding result institute of text data , can semantic groove set contained by similar initial semantic understanding result and language contained by history semantic understanding result after semantic groove set The calculation of the registration of adopted groove set, calculate semantic groove set contained by initial semantic understanding result and last round of feedback The registration of semantic groove set contained by the semantic understanding result of text data.

S43：According to the semantic feature and the semantic understanding sort result model built in advance, to current cutting textual data According to secondary semantic understanding is carried out, secondary semantic understanding result is obtained.

Semantic understanding sort result model can generate in the training stage after being trained according to the training data of collection. The input of semantic understanding sort result model is the semantic feature of text data, and output is to multiple first corresponding to this article notebook data The sequencing information of beginning semantic understanding result, the probable value of such as each initial semantic understanding result, or can also be will correctly just The initial semantic understanding sort result information that beginning semantic understanding result makes number one, it is therefore, special in the semanteme for extracting above-mentioned After sign, the input using the semantic feature of extraction as semantic understanding sort result model, it can be obtained to initial language according to output The ranking results of reason and good sense solution result.

Specifically, the flow of training semantic understanding sort result model can include：Collect a large number of users text data and The feedback text data of system, after carrying out initial semantic understanding to these data, obtain initial language corresponding to user version data Initial semantic understanding result corresponding to reason and good sense solution result and feedback text data, further according to features described above extraction flow extraction text The semantic feature of data, the clooating sequence of corresponding initial semantic understanding result is manually marked, or, directly mark just True initial semantic understanding result, using the correctly initial semantic understanding result as first semantic understanding knot after sequence Fruit, other initial semantic understanding results are put into behind correct initial semantic understanding result, and order is indefinite, according to the semanteme of extraction After feature and the clooating sequence manually marked are trained, it is possible to which generative semantics understands sort result model.The sequence mould Type is that order models are commonly used in pattern-recognition, such as Ranking SVM, Ranking CNN.

Further, when sub-semantic field scape carries out semantic understanding to text data, in training semantic understanding sort result During model, each semantic scene training one semantic understanding sort result model of generation can be corresponded to, is generating some semantic field During the semantic understanding sort result model of scape, the training data of use be the semantic scene user version data and system it is anti- Present text data etc..

S44：Second semantic understanding result is obtained according to the secondary semantic understanding result.

In some embodiments, the second semantic understanding result directly can be obtained according to secondary semantic understanding result.

For example, preceding N number of (TopN) the secondary semantic understanding result of selected and sorted is as the second semantic understanding result.It is described N determines according to application demand, forward preceding 10 semantic understanding results such as selected and sorted.Wherein, secondary semantic understanding result Clooating sequence can determine according to the probable value of respective paths, such as be sorted according to the order of probable value from big to small.

In order to further improve the degree of accuracy, in some embodiments, secondary semantic understanding result can also be ranked up and repaiied Just, according to revised secondary semantic understanding result as final semantic understanding result.

As shown in figure 5, the flow for the method for being ranked up amendment to secondary semantic understanding result includes：

S51：Obtain the related data of secondary semantic understanding result.

Under each semantic scene, the higher data of the correlation of the secondary semantic understanding of vertical search engine search are utilized.

During specific search, inputted each secondary semantic understanding result as the search string of search engine, search engine is certainly The dynamic search condition that the search string is switched to inside, the related document of each secondary semantic understanding result of search, and will sequence Relevant documentation afterwards returns, and each secondary semantic understanding result can select TopN relevant documentation, as each secondary semanteme Understand the related data of result；Detailed process is same as the prior art, will not be described in detail herein.

S52：Degree of correlation feature is extracted according to current cutting text data, secondary semantic understanding result and related data.

According to the secondary semantic understanding result data extraction degree of correlation feature associated therewith under each semantic scene.

The degree of correlation feature is used to describe the degree of correlation between secondary semantic understanding result and current cutting text data And the accuracy of secondary semantic understanding sort result.

Degree of correlation feature includes at least one in following item：

Secondary semantic understanding result is with including key in current cutting text data same words number, secondary semantic understanding result Semantic slot number, with crucial semantic groove value temperature contained by the secondary semantic understanding result of word identical in current cutting text data, Current quantity of the cutting text data comprising the keyword related to semantic scene, semantic groove collection contained by initial semantic understanding result The registration of conjunction and semantic groove set contained by related data.

Specific extracting method is as described below：

(1) secondary semantic understanding result and current cutting text data same words number

The secondary semantic understanding result and current cutting text data same words number, refer to secondary semantic understanding result institute Word in the current cutting text data understood, contained semantic groove value in secondary semantic understanding result can be specifically matched successively Each word after being segmented with current cutting text data, determine semantic groove value and current cutting text in secondary semantic understanding result Notebook data identical word number.

(2) crucial semantic slot number is included in secondary semantic understanding result

Count the key included under each semantic scene in each secondary semantic understanding result in its affiliated semantic understanding scene Semantic slot number, more comprising crucial semantic slot number, the correlation of semantic understanding result and its affiliated semantic scene is higher, so that In each semantic scene, the secondary semantic understanding result higher with each semantic scene correlation is come before；In different languages Between adopted scene, the crucial semantic groove of scene that secondary semantic understanding result includes is more, and the secondary semantic understanding sort result is got over It is forward.

(3) and in current cutting text data crucial semantic groove value is hot contained by the secondary semantic understanding result of word identical Degree.

The crucial semantic groove in each secondary semantic understanding result is found, each crucial semantic groove value of matching is with working as successively Whether word is identical in preceding cutting text data, if identical, calculates the temperature of current key semanteme groove value；Crucial semantic groove The related data that the temperature of value can search according to the secondary semantic understanding result of the affiliated semantic scene of the semantic groove of key It is calculated；The data that vertical page in the related data such as search result includes, it is corresponding to hang down such as in music scenario The straight page is search dog music, cruel my music, Tengxun's music etc., the use of crucial semantic groove value according to the vertical page Family click volume, user's playback volume, user's volumes of searches, crucial semantic groove value are in information such as the distributions of multiple vertical pages, to working as The temperature information of the preceding semantic groove of key carries out comprehensive analysis, obtains the temperature of crucial semantic groove, concrete analysis computational methods with it is existing There is technology identical, will not be described in detail herein.

(5) registration of semantic groove set and semantic groove set contained by related data contained by initial semantic understanding result

The related data that semantic groove contained by the related data searches according to secondary semantic understanding result obtains, to described After the related data searched carries out structuring, the semantic groove set that related data includes is obtained, concrete structure method is with showing There is technology identical, will not be described in detail herein；After semantic groove in two of extraction semantic groove set is compared, two languages are determined The quantity of the semantic groove of same names in adopted groove set, reuse language contained by the identical quantity divided by initial semantic understanding result First prime number in adopted groove set, obtain semantic groove set contained by the initial semantic understanding result and semantic groove contained by related data The registration of set.

S53：Amendment is ranked up to secondary semantic understanding result according to degree of correlation feature and the order models pre-established.

Wherein, order models generate after being trained in the training stage to training data.For example, collect a large amount of Text data, semantic understanding is carried out to text data and obtains the secondary semantic understanding result under each semantic scene, and extraction Degree of correlation feature, the sequence to the secondary semantic understanding result under each semantic scene is manually marked, afterwards according to the degree of correlation Feature and the sequence manually marked are trained, and generate order models.The input of order models is degree of correlation feature, and output is row Secondary semantic understanding result under the revised each semantic scene of sequence.

Therefore, after extraction obtains degree of correlation feature, the revised secondary semanteme that sorts can be obtained according to order models Understand result.

After the revised secondary semantic understanding result that obtains sorting, TopN can be selected as the final second semantic reason Result is solved, the N can determine according to application demand, if N values are 10.

In the present embodiment, by being ranked up amendment to secondary semantic understanding result, semantic understanding can be further improved As a result the degree of accuracy, and then the degree of accuracy of feedback result is further improved, lift Consumer's Experience.

Fig. 6 is the structural representation for more wheel interacting parallel semantic understanding devices that the application one embodiment proposes.

As shown in fig. 6, the device 60 of the present embodiment includes：Receiving module 61, sound identification module 62, semantic understanding module 63 and selecting module 64.

Receiving module 61, for receiving speech data；

Sound identification module 62, for carrying out cutting to the speech data, speech data segment is obtained, and, to institute Predicate sound data fragments carry out speech recognition, obtain current cutting text data；

Semantic understanding module 63, for carrying out independent semantic understanding to current cutting text data, obtain the first semantic reason Result is solved, and, semantic understanding is carried out to current cutting text data according to the historical data of user mutual, it is semantic to obtain second Understand result；

Selecting module 64, for selecting the first semantic understanding result or the second semantic understanding result, as Final semantic understanding result.

In some embodiments, referring to Fig. 7, the selecting module 64 includes：

Select feature extraction submodule 641, for according to current cutting text data, current cutting text data it is initial The historical data of semantic understanding result and user mutual extraction selection feature；

Submodule 642 is selected, for according to the selection feature and the semantic results preference pattern built in advance, selecting institute The first semantic understanding result or the second semantic understanding result are stated, as final semantic understanding result.

In some embodiments, the historical data includes at least one in following item：

In some embodiments, the selection feature includes at least one in following item：

In some embodiments, referring to Fig. 7, semantic understanding module 63 includes：

Independent semantic understanding submodule 631, for carrying out independent initial semantic understanding to current cutting text data, obtain To initial semantic understanding result；And the initial semantic understanding result of predetermined number is directly sequentially selected, as the first semantic reason Solve result；Or independent initial semantic understanding is carried out to current cutting text data, obtain initial semantic understanding result；It is right Initial semantic understanding result is ranked up amendment；And sequentially select the revised initial semantic understanding of sequence of predetermined number As a result, as the first semantic understanding result.

In some embodiments, the independent semantic understanding submodule is used to be ranked up initial semantic understanding result to repair Just, including：

Obtain the related data of initial semantic understanding result；

Degree of correlation feature is extracted according to current cutting text data, initial semantic understanding result and the related data；

According to the degree of correlation feature and the order models built in advance, initial semantic understanding result is ranked up and repaiied Just.

In some embodiments, the degree of correlation feature includes at least one in following item：

In some embodiments, referring to Fig. 7, semantic understanding module 63 also includes：

Initial semantic understanding submodule 632, for carrying out initial semantic understanding to the cutting text data, obtain initial Semantic understanding result；

Extracting sub-module 633, for the historical data according to the initial semantic understanding result and user mutual, extract institute State the semantic feature of cutting text data；

Secondary semantic understanding submodule 634, for according to the semantic feature and the semantic understanding result built in advance row Sequence model, secondary semantic understanding is carried out to the cutting text data, obtains secondary semantic understanding result；

Acquisition submodule 635, for obtaining the second semantic understanding result according to the secondary semantic understanding result.

In some embodiments, the acquisition submodule is specifically used for：

The secondary semantic understanding result of predetermined number is directly sequentially selected, as the second semantic understanding result；Or

Amendment is ranked up to the secondary semantic understanding result, sequentially selects the sequence of predetermined number revised secondary Semantic understanding result, as the second semantic understanding result.

In some embodiments, the acquisition submodule is used to be ranked up amendment to the secondary semantic understanding result, wraps Include：

Obtain the related data of secondary semantic understanding result；

Degree of correlation feature is extracted according to current cutting text data, secondary semantic understanding result and related data；

Amendment is ranked up to secondary semantic understanding result according to degree of correlation feature and the order models pre-established.

History semantic understanding result, the semantic understanding result of the feedback text data of last round of interaction.

In some embodiments, the semantic feature includes at least one in following item：

It is understood that the device of the present embodiment is corresponding with above method embodiment, particular content may refer to method The associated description of embodiment, is no longer described in detail herein.

It is understood that same or similar part can mutually refer in the various embodiments described above, in certain embodiments Unspecified content may refer to same or analogous content in other embodiment.

It should be noted that in the description of the present application, term " first ", " second " etc. are only used for describing purpose, without It is understood that to indicate or implying relative importance.In addition, in the description of the present application, unless otherwise indicated, the implication of " multiple " Refer at least two.

Any process or method described otherwise above description in flow chart or herein is construed as, and represents to include Module, fragment or the portion of the code of the executable instruction of one or more the step of being used to realize specific logical function or process Point, and the scope of the preferred embodiment of the application includes other realization, wherein can not press shown or discuss suitable Sequence, including according to involved function by it is basic simultaneously in the way of or in the opposite order, carry out perform function, this should be by the application Embodiment person of ordinary skill in the field understood.

It should be appreciated that each several part of the application can be realized with hardware, software, firmware or combinations thereof.Above-mentioned In embodiment, software that multiple steps or method can be performed in memory and by suitable instruction execution system with storage Or firmware is realized.If, and in another embodiment, can be with well known in the art for example, realized with hardware Any one of row technology or their combination are realized：With the logic gates for realizing logic function to data-signal Discrete logic, have suitable combinational logic gate circuit application specific integrated circuit, programmable gate array (PGA), scene Programmable gate array (FPGA) etc..

Those skilled in the art are appreciated that to realize all or part of step that above-described embodiment method carries Suddenly it is that by program the hardware of correlation can be instructed to complete, described program can be stored in a kind of computer-readable storage medium In matter, the program upon execution, including one or a combination set of the step of embodiment of the method.

In addition, each functional unit in each embodiment of the application can be integrated in a processing module, can also That unit is individually physically present, can also two or more units be integrated in a module.Above-mentioned integrated mould Block can both be realized in the form of hardware, can also be realized in the form of software function module.The integrated module is such as Fruit is realized in the form of software function module and as independent production marketing or in use, can also be stored in a computer In read/write memory medium.

Storage medium mentioned above can be read-only storage, disk or CD etc..

In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show The description of example " or " some examples " etc. means specific features, structure, material or the spy for combining the embodiment or example description Point is contained at least one embodiment or example of the application.In this manual, to the schematic representation of above-mentioned term not Necessarily refer to identical embodiment or example.Moreover, specific features, structure, material or the feature of description can be any One or more embodiments or example in combine in an appropriate manner.

Although embodiments herein has been shown and described above, it is to be understood that above-described embodiment is example Property, it is impossible to the limitation to the application is interpreted as, one of ordinary skill in the art within the scope of application can be to above-mentioned Embodiment is changed, changed, replacing and modification.

Claims

A kind of 1. more wheel interacting parallel semantic understanding methods, it is characterised in that including：

Receive speech data；

Cutting is carried out to the speech data, obtains speech data segment, and, voice knowledge is carried out to the speech data segment Not, current cutting text data is obtained；

Independent semantic understanding is carried out to current cutting text data, obtains the first semantic understanding result, and, according to user mutual Historical data to current cutting text data carry out semantic understanding, obtain the second semantic understanding result；

The first semantic understanding result or the second semantic understanding result are selected, as final semantic understanding result.
2. according to the method for claim 1, it is characterised in that selection the first semantic understanding result or described Second semantic understanding result, as final semantic understanding result, including：

According to current cutting text data, the history number of initial the semantic understanding result and user mutual of current cutting text data Feature is selected according to extraction；

According to the selection feature and the semantic results preference pattern that builds in advance, selection the first semantic understanding result or The second semantic understanding result, as final semantic understanding result.
3. according to the method for claim 2, it is characterised in that the historical data includes at least one in following item：

The feedback text data of last round of interaction, history semantic understanding result, the history of last round of interaction of last round of interaction are cut Single cent notebook data.
4. according to the method for claim 3, it is characterised in that the selection feature includes at least one in following item：

Current cutting text data length, current cutting text data replace the feedback text data middle finger of last round of interaction for body Front and rear semantic matching degree, the semantic integrity degree of current cutting text data, the initial semantic understanding of current cutting text data As a result contained semantic groove set and the registration of semantic groove set contained by the last round of history semantic understanding result interacted, currently cut Contained by semantic groove value contained by the initial semantic understanding result of single cent notebook data and the last round of history semantic understanding result interacted Semantic groove contained by the matching degree of semantic groove value, the initial semantic understanding result of current cutting text data interacts with last round of The joint degree of semantic groove, current cutting text data and the last round of history cutting text interacted contained by history semantic understanding result Semantic integrity degree after data connection.
5. according to the method for claim 1, it is characterised in that described that independent semantic reason is carried out to current cutting text data Solution, obtains the first semantic understanding result, including：

Independent initial semantic understanding is carried out to current cutting text data, obtains initial semantic understanding result；

The initial semantic understanding result of predetermined number is directly sequentially selected, as the first semantic understanding result.
6. according to the method for claim 1, it is characterised in that described that independent semantic reason is carried out to current cutting text data Solution, obtains the first semantic understanding result, including：

Independent initial semantic understanding is carried out to current cutting text data, obtains initial semantic understanding result；

Amendment is ranked up to initial semantic understanding result；

The revised initial semantic understanding result of sequence of predetermined number is sequentially selected, as the first semantic understanding result.
7. according to the method for claim 6, it is characterised in that it is described that amendment is ranked up to initial semantic understanding result, Including：

Obtain the related data of initial semantic understanding result；

Degree of correlation feature is extracted according to current cutting text data, initial semantic understanding result and the related data；

According to the degree of correlation feature and the order models built in advance, amendment is ranked up to initial semantic understanding result.
8. according to the method for claim 7, it is characterised in that the degree of correlation feature includes at least one in following item ：

Initial semantic understanding result is with including crucial semanteme in current cutting text data same words number, initial semantic understanding result Slot number, with crucial semantic groove value temperature contained by the initial semantic understanding result of word identical in current cutting text data, current Cutting text data includes the quantity of the keyword related to semantic scene.
9. according to the method for claim 1, it is characterised in that the historical data according to user mutual is to current cutting Text data carries out semantic understanding, obtains the second semantic understanding result, including：

Independent semantic understanding is carried out to current cutting text data, obtains initial semantic understanding result；

According to the initial semantic understanding result and the historical data of user mutual, the semanteme for extracting current cutting text data is special Sign；

According to the semantic feature and the semantic understanding sort result model built in advance, two are carried out to current cutting text data Secondary semantic understanding, obtain secondary semantic understanding result；

Second semantic understanding result is obtained according to the secondary semantic understanding result.
10. according to the method for claim 9, it is characterised in that the result according to secondary semantic understanding obtains second Semantic understanding result, including：

The secondary semantic understanding result of predetermined number is directly sequentially selected, as the second semantic understanding result；Or

Amendment is ranked up to the secondary semantic understanding result, sequentially selects the revised secondary semanteme of sequence of predetermined number Result is understood, as the second semantic understanding result.
11. according to the method for claim 10, it is characterised in that described that the secondary semantic understanding result is ranked up Amendment, including：

Obtain the related data of secondary semantic understanding result；

Degree of correlation feature is extracted according to current cutting text data, secondary semantic understanding result and related data；

Amendment is ranked up to secondary semantic understanding result according to degree of correlation feature and the order models pre-established.
12. according to the method for claim 9, it is characterised in that the historical data includes at least one in following item：

History semantic understanding result, the semantic understanding result of the feedback text data of last round of interaction.
13. according to the method for claim 12, it is characterised in that the semantic feature includes at least one in following item ：

Initial semantic understanding result feature, history semantic understanding result feature, semantic groove set contained by initial semantic understanding result With history semantic understanding result contained by the registration of semantic groove set, crucial semantic groove value contained by initial semantic understanding result with The matching degree of crucial semantic groove value contained by history semantic understanding result, semantic groove set contained by initial semantic understanding result with it is upper The registration of semantic groove set contained by the semantic understanding result of the feedback text data of one wheel.
14. according to the method for claim 11, it is characterised in that the degree of correlation feature includes at least one in following item ：

Secondary semantic understanding result is with including crucial semanteme in current cutting text data same words number, secondary semantic understanding result Slot number, with crucial semantic groove value temperature contained by the secondary semantic understanding result of word identical in current cutting text data, current Cutting text data includes the quantity of the keyword related to semantic scene, semantic groove set contained by initial semantic understanding result and The registration of semantic groove set contained by related data.
A kind of 15. more wheel interacting parallel semantic understanding devices, it is characterised in that including：

Receiving module, for receiving speech data；

Sound identification module, for carrying out cutting to the speech data, speech data segment is obtained, and, to the voice Data fragments carry out speech recognition, obtain current cutting text data；

Semantic understanding module, for carrying out independent semantic understanding to current cutting text data, the first semantic understanding result is obtained, And semantic understanding is carried out to current cutting text data according to the historical data of user mutual, obtain the second semantic understanding knot Fruit；

Selecting module, for selecting the first semantic understanding result or the second semantic understanding result, as final Semantic understanding result.
16. device according to claim 15, it is characterised in that the selecting module includes：

Feature extraction submodule is selected, for the initial semantic reason according to current cutting text data, current cutting text data Solve the historical data extraction selection feature of result and user mutual；

Submodule is selected, for according to the selection feature and the semantic results preference pattern built in advance, selecting described first Semantic understanding result or the second semantic understanding result, as final semantic understanding result.
17. device according to claim 16, it is characterised in that the historical data includes at least one in following item ：

The feedback text data of last round of interaction, history semantic understanding result, the history of last round of interaction of last round of interaction are cut Single cent notebook data.
18. device according to claim 17, it is characterised in that the selection feature includes at least one in following item ：

Current cutting text data length, current cutting text data replace the feedback text data middle finger of last round of interaction for body Front and rear semantic matching degree, the semantic integrity degree of current cutting text data, the initial semantic understanding of current cutting text data As a result contained semantic groove set and the registration of semantic groove set contained by the last round of history semantic understanding result interacted, currently cut Contained by semantic groove value contained by the initial semantic understanding result of single cent notebook data and the last round of history semantic understanding result interacted Semantic groove contained by the matching degree of semantic groove value, the initial semantic understanding result of current cutting text data interacts with last round of The joint degree of semantic groove, current cutting text data and the last round of history cutting text interacted contained by history semantic understanding result Semantic integrity degree after data connection.
19. device according to claim 15, it is characterised in that the semantic understanding module includes：Independent semantic understanding Submodule, the independent semantic understanding submodule are used for：

Independent initial semantic understanding is carried out to current cutting text data, obtains initial semantic understanding result；And directly press Sequence selects the initial semantic understanding result of predetermined number, as the first semantic understanding result；

Or

Independent initial semantic understanding is carried out to current cutting text data, obtains initial semantic understanding result；To initial semantic Understand that result is ranked up amendment；And the revised initial semantic understanding result of sequence of predetermined number is sequentially selected, as First semantic understanding result.
20. device according to claim 19, it is characterised in that the independent semantic understanding submodule is used for initial language Reason and good sense solution result is ranked up amendment, including：

Obtain the related data of initial semantic understanding result；

Degree of correlation feature is extracted according to current cutting text data, initial semantic understanding result and the related data；

According to the degree of correlation feature and the order models built in advance, amendment is ranked up to initial semantic understanding result.
21. device according to claim 20, it is characterised in that the degree of correlation feature includes at least one in following item ：

Initial semantic understanding result is with including crucial semanteme in current cutting text data same words number, initial semantic understanding result Slot number, with crucial semantic groove value temperature contained by the initial semantic understanding result of word identical in current cutting text data, current Cutting text data includes the quantity of the keyword related to semantic scene.
22. device according to claim 15, it is characterised in that the semantic understanding module includes：

Initial semantic understanding submodule, for carrying out initial semantic understanding to the cutting text data, obtain initial semantic reason Solve result；

Extracting sub-module, for the historical data according to the initial semantic understanding result and user mutual, extract the cutting The semantic feature of text data；

Secondary semantic understanding submodule, for according to the semantic feature and the semantic understanding sort result model built in advance, Secondary semantic understanding is carried out to the cutting text data, obtains secondary semantic understanding result；

Acquisition submodule, for obtaining the second semantic understanding result according to the secondary semantic understanding result.
23. device according to claim 22, it is characterised in that the acquisition submodule is specifically used for：

The secondary semantic understanding result of predetermined number is directly sequentially selected, as the second semantic understanding result；Or

Amendment is ranked up to the secondary semantic understanding result, sequentially selects the revised secondary semanteme of sequence of predetermined number Result is understood, as the second semantic understanding result.
24. device according to claim 23, it is characterised in that the acquisition submodule is used for the secondary semantic reason Solution result is ranked up amendment, including：

Obtain the related data of secondary semantic understanding result；

Degree of correlation feature is extracted according to current cutting text data, secondary semantic understanding result and related data；

Amendment is ranked up to secondary semantic understanding result according to degree of correlation feature and the order models pre-established.
25. device according to claim 22, it is characterised in that the historical data includes at least one in following item ：History semantic understanding result, the semantic understanding result of the feedback text data of last round of interaction.
26. device according to claim 25, it is characterised in that the semantic feature includes at least one in following item ：

Initial semantic understanding result feature, history semantic understanding result feature, semantic groove set contained by initial semantic understanding result With history semantic understanding result contained by the registration of semantic groove set, crucial semantic groove value contained by initial semantic understanding result with The matching degree of crucial semantic groove value contained by history semantic understanding result, semantic groove set contained by initial semantic understanding result with it is upper The registration of semantic groove set contained by the semantic understanding result of the feedback text data of one wheel.
27. device according to claim 24, it is characterised in that the degree of correlation feature includes at least one in following item ：

Secondary semantic understanding result is with including crucial semanteme in current cutting text data same words number, secondary semantic understanding result Slot number, with crucial semantic groove value temperature contained by the secondary semantic understanding result of word identical in current cutting text data, current Cutting text data includes the quantity of the keyword related to semantic scene, semantic groove set contained by initial semantic understanding result and The registration of semantic groove set contained by related data.