CN110069709A

CN110069709A - Intension recognizing method, device, computer-readable medium and electronic equipment

Info

Publication number: CN110069709A
Application number: CN201910284684.5A
Authority: CN
Inventors: 谭莲芝; 陈建荣
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-04-10
Filing date: 2019-04-10
Publication date: 2019-07-30
Anticipated expiration: 2039-04-10
Also published as: CN110069709B

Abstract

Embodiments herein provides a kind of intension recognizing method, device, computer-readable medium and electronic equipment.The intension recognizing method includes: the inquiry language for obtaining user's input；The term vector feature of the inquiry language is obtained according to the inquiry language, and obtains discrete type feature associated with the inquiry language；The multi-channel feature vector of the inquiry language is generated according to the term vector feature and the discrete type feature；The multi-channel feature vector is input to intention assessment model, obtains the intent information of the intention assessment model output.The technical solution of the embodiment of the present application can ensure the accuracy rate of intention assessment by multi-channel feature vector.

Description

Intension recognizing method, device, computer-readable medium and electronic equipment

Technical field

This application involves computer and fields of communication technology, in particular to a kind of intension recognizing method, device, meter Calculation machine readable medium and electronic equipment.

Background technique

Intention assessment, which refers to, carries out analysis and understanding to query (inquiry) the retrieval string of user's input, parses the meaning of user Figure is to facilitate the process for meeting user's search need.The accuracy of intention assessment is affected to query response as a result, right And the intention assessment scheme proposed in the related technology often has that intention assessment accuracy rate is lower.

Summary of the invention

Embodiments herein provides a kind of intension recognizing method, device, computer-readable medium and electronic equipment, into And the accuracy rate of intention assessment can be improved at least to a certain extent.

Other characteristics and advantages of the application will be apparent from by the following detailed description, or partially by the application Practice and acquistion.

According to the one aspect of the embodiment of the present application, a kind of intension recognizing method is provided, comprising: obtain user's input Inquire language；According to it is described inquiry language obtain it is described inquiry language term vector feature, and obtain it is associated with the inquiry language from Dissipate type feature；The multi-channel feature vector of the inquiry language is generated according to the term vector feature and the discrete type feature；It will The multi-channel feature vector is input to intention assessment model, obtains the intent information of the intention assessment model output.

According to the one aspect of the embodiment of the present application, a kind of intention assessment device is provided, comprising: first acquisition unit, For obtaining the inquiry language of user's input；Second acquisition unit, for according to it is described inquiry language obtain it is described inquiry language word to Measure feature, and obtain discrete type feature associated with the inquiry language；First generation unit, for special according to the term vector The discrete type feature of seeking peace generates the multi-channel feature vector of the inquiry language；Processing unit, for the multichannel is special Sign vector is input to intention assessment model, obtains the intent information of the intention assessment model output.

In some embodiments of the present application, be based on aforementioned schemes, the second acquisition unit be configured that acquisition with it is described Inquire the associated historical user feedback information of language；According to the historical user feedback information, count related to the inquiry language The distribution situation of each historical user feedback information of connection；It is raw according to the distribution situation of each historical user feedback information At the discrete type feature.

In some embodiments of the present application, aforementioned schemes are based on, the second acquisition unit is configured that described in identification and looks into That askes language is intended that precisely intention or fuzzy intention；Precisely it is intended to still fuzzy intention life according to being intended that for the inquiry language At the discrete type feature of the inquiry language.

In some embodiments of the present application, be based on aforementioned schemes, the second acquisition unit be configured to according to it is following because At least one factor in element identify it is described inquiry language be intended that precisely be intended to or fuzzy intention: the inquiry language with make a reservation for Intent information between text similarity, it is described inquiry language and scheduled intent information between semantic similarity, with it is described Inquire the distribution situation of the associated historical user feedback information of language.

In some embodiments of the present application, aforementioned schemes are based on, the second acquisition unit is configured that the inquiry Language carries out word segmentation processing and obtains at least one target word；Importance analysis is carried out at least one described target word, is obtained The importance score of each target word；The inquiry language is generated according to the importance score of each target word Discrete type feature.

In some embodiments of the present application, be based on aforementioned schemes, the second acquisition unit be configured to by it is following because At least one factor in element carries out importance analysis to each target word: the part of speech of each target word analyzes knot The interdependent syntactic analysis result between other target words, each target word in fruit, each target word and the inquiry language The tightness of language analyzes whether result, the reverse document-frequency of each target word, each target word are the entity words set, is each Whether a target word is off word.

In some embodiments of the present application, aforementioned schemes are based on, the second acquisition unit is configured that the inquiry Language is input to the term vector extraction model of pre-training, to obtain the term vector feature of the term vector extraction model output.

In some embodiments of the present application, aforementioned schemes are based on, the second acquisition unit is configured that be mentioned by feature Take the term vector feature that language is inquired described in network abstraction, wherein the feature extraction network passes through a variety of different size of convolution Core extracts the term vector feature of the inquiry language respectively.

In some embodiments of the present application, it is based on aforementioned schemes, the intention assessment device further include: third obtains Unit, for obtaining inquiry language sample；4th acquiring unit, for inquiring language sample according to the inquiry language sample acquisition Term vector feature and the associated discrete type feature of inquiry language sample；Second generation unit, for according to the inquiry The term vector feature of language sample and the associated discrete type feature of inquiry language sample generate training sample；Training unit is used The intention assessment model is trained in based on the training sample.

In some embodiments of the present application, aforementioned schemes are based on, the intention assessment model includes multitask engineering Practise model, the multitask machine learning model for exporting multistage intention, the multitask machine learning model include with respectively Grade is intended to corresponding full articulamentum and output layer, wherein output corresponding to the first intention in the multistage intent information Layer is connected to full articulamentum corresponding to second intention, and the level of the first intention is higher than the level of the second intention.

In some embodiments of the present application, aforementioned schemes are based on, the intention assessment model includes multitask engineering Practise model, the loss function of the multitask machine learning model are as follows:

Loss=∑ (c_i×loss_i)+λ

Wherein, Loss indicates the loss function of the multitask machine learning model；loss_iIndicate that i-stage is intended to correspond to Loss function；c_iFor hyper parameter, for indicating that i-stage is intended to the importance of corresponding loss function；λ indicates regularization term.

In some embodiments of the present application, aforementioned schemes are based on, the intention assessment model includes multitask engineering Practise model, the intention assessment device further include: response unit, for obtaining the intention letter of the intention assessment model output After breath, the inquiry language is responded by the intention of the lowest hierarchical level in the intent information, obtains response results；If the sound It answers result not to be able to satisfy response to require, then passes sequentially through upper one layer of intention and respond the inquiry language, until obtained response As a result meet the response requirement.

According to the one aspect of the embodiment of the present application, a kind of computer-readable medium is provided, computer is stored thereon with Program realizes such as above-mentioned intension recognizing method as described in the examples when the computer program is executed by processor.

According to the one aspect of the embodiment of the present application, a kind of electronic equipment is provided, comprising: one or more processors； Storage device, for storing one or more programs, when one or more of programs are held by one or more of processors When row, so that one or more of processors realize such as above-mentioned intension recognizing method as described in the examples.

In the technical solution provided by some embodiments of the present application, by obtain inquiry language term vector feature and with Inquire the associated discrete type feature of language, according to term vector feature and discrete type feature generate the multi-channel feature of inquiry language to Amount, and the multi-channel feature vector is input to intention assessment model to obtain intent information, so that in the meaning of identification inquiry language Not only allow for the term vector feature of inquiry language when figure, and it is contemplated that with the associated discrete type feature of inquiry language, into And it can ensure the accuracy rate of intention assessment by multi-channel feature vector.

It should be understood that above general description and following detailed description be only it is exemplary and explanatory, not The application can be limited.

Detailed description of the invention

The drawings herein are incorporated into the specification and forms part of this specification, and shows the implementation for meeting the application Example, and together with specification it is used to explain the principle of the application.It should be evident that the accompanying drawings in the following description is only the application Some embodiments for those of ordinary skill in the art without creative efforts, can also basis These attached drawings obtain other attached drawings.In the accompanying drawings:

Fig. 1 is shown can be using the schematic diagram of the exemplary system architecture of the technical solution of the embodiment of the present application；

Fig. 2 shows the flow charts according to the intension recognizing method of one embodiment of the application；

Fig. 3 shows the process of the acquisition discrete type feature associated with inquiry language according to one embodiment of the application Figure；

Fig. 4 shows the process of the acquisition discrete type feature associated with inquiry language according to one embodiment of the application Figure；

Fig. 5 shows the process of the acquisition discrete type feature associated with inquiry language according to one embodiment of the application Figure；

Fig. 6 shows the training flow chart of the intention assessment model according to one embodiment of the application；

Fig. 7 shows the structural schematic diagram of the intention assessment model according to one embodiment of the application；

Fig. 8 shows the flow chart of the intention assessment process according to one embodiment of the application；

Fig. 9 shows the search result schematic diagram based on intention assessment in the related technology；

Figure 10 shows the intention assessment result schematic diagram according to one embodiment of the application；

Figure 11 shows the search result schematic diagram based on intention assessment of one embodiment according to the application；

The comparison between intention that the intention and technical scheme that the relevant technologies that Figure 12 show identify identify Effect diagram；

Figure 13 shows the block diagram of the intention assessment device according to one embodiment of the application；

Figure 14 shows the structural schematic diagram for being suitable for the computer system for the electronic equipment for being used to realize the embodiment of the present application.

Specific embodiment

Example embodiment is described more fully with reference to the drawings.However, example embodiment can be with a variety of shapes Formula is implemented, and is not understood as limited to example set forth herein；On the contrary, thesing embodiments are provided so that the application will more Fully and completely, and by the design of example embodiment comprehensively it is communicated to those skilled in the art.

In addition, described feature, structure or characteristic can be incorporated in one or more implementations in any suitable manner In example.In the following description, many details are provided to provide and fully understand to embodiments herein.However, It will be appreciated by persons skilled in the art that the technical solution of the application can be practiced without one or more in specific detail, Or it can be using other methods, constituent element, device, step etc..In other cases, it is not shown in detail or describes known side Method, device, realization or operation to avoid fuzzy the application various aspects.

Block diagram shown in the drawings is only functional entity, not necessarily must be corresponding with physically separate entity. I.e., it is possible to realize these functional entitys using software form, or realized in one or more hardware modules or integrated circuit These functional entitys, or these functional entitys are realized in heterogeneous networks and/or processor device and/or microcontroller device.

Flow chart shown in the drawings is merely illustrative, it is not necessary to including all content and operation/step, It is not required to execute by described sequence.For example, some operation/steps can also decompose, and some operation/steps can close And or part merge, therefore the sequence actually executed is possible to change according to the actual situation.

Fig. 1 is shown can be using the schematic diagram of the exemplary system architecture of the technical solution of the embodiment of the present application.

As shown in Figure 1, system architecture 100 may include terminal device (smart phone 101 as shown in fig. 1, tablet computer One of 102 and portable computer 103 are a variety of, naturally it is also possible to be desktop computers etc.), network 104 and service Device 105.Network 104 between terminal device and server 105 to provide the medium of communication link.Network 104 may include Various connection types, such as wired communications links, wireless communication link etc..

It should be understood that the number of terminal device, network and server in Fig. 1 is only schematical.According to realization need It wants, can have any number of terminal device, network and server.For example server 105 can be multiple server compositions Server cluster etc..

In one embodiment of the application, user such as can may be used by terminal device input inquiry language, the inquiry language To be the sentence for searching for some object (such as application program, person names).Server 105 is in the inquiry for getting user's input After language, the term vector feature of inquiry language can be obtained according to the inquiry language, and it is special to obtain discrete type associated with inquiry language Sign (the importance score of each word after search behavior feedback characteristic, the inquiry language of such as other users segment), and then basis The term vector feature and the discrete type feature generate the multi-channel feature vector of inquiry language, and the multi-channel feature vector is inputted To intention assessment model, the intent information of intention assessment model output is obtained.As it can be seen that inquiring language in identification in the embodiment of the present application Intention when not only allow for the term vector feature of inquiry language, and it is it is contemplated that special with the associated discrete type of inquiry language Sign, and then can ensure the accuracy rate of intention assessment by multi-channel feature vector.

In one embodiment of the application, server 105, can basis after the intent information for identifying inquiry language The intent information responds the inquiry language of user, and then response results are returned to terminal device, in order to which terminal device is fed back to User.

It should be noted that intension recognizing method provided by the embodiment of the present application is generally executed by server 105, accordingly Ground, it is intended that identification device is generally positioned in server 105.But in the other embodiments of the application, terminal device can also There is similar function with server, thereby executing intension recognizing method provided by the embodiment of the present application.

The realization details of the technical solution of the embodiment of the present application is described in detail below:

Fig. 2 shows the flow chart according to the intension recognizing method of one embodiment of the application, the intension recognizing methods It can be executed by server, which can be server shown in Fig. 1.Referring to shown in Fig. 2, the intension recognizing method Including at least step S210 to step S240, it is described in detail as follows:

In step S210, the inquiry language of user's input is obtained.

In one embodiment of the application, inquiry language can refer to the reference inquired, retrieve or searched for as information Any content, inquiry language can be content of text, may include one or more words, symbol or their combination.Than Such as, inquiry language can be the sentence for searching for some object (such as application program, person names).The inquiry language of user's input can be with Be it is by voice input, be also possible to by the input equipments such as keyboard, touch screen input.

In step S220, the term vector feature of the inquiry language is obtained according to the inquiry language, and is obtained and looked into described Ask the associated discrete type feature of language.

In one embodiment of the application, inquiry language can be input in the term vector extraction model of pre-training, into And get the term vector feature of term vector extraction model output.Wherein, the term vector extraction model of pre-training can be Fasttext (term vector and text classification tool of an open source) model, word2vec (for generating the model of term vector) mould Type, GloVe (Global Vectors for Word Representation) model etc..GloVe model is one based on complete The word characterization tool of office's word frequency statistics, a word can be expressed as the vector being made of real number by it, these vectors are caught Some feature of semanteme, such as similitude, analogy etc. between word are grasped.

In one embodiment of the application, the term vector feature of language can be inquired by feature extraction network abstraction, In, feature extraction network extracts the term vector feature of inquiry language by a variety of different size of convolution kernels respectively.Specifically, special Sign extracts network can extract the term vector feature of inquiry language simultaneously using different size of convolution kernel, available in this way to pass through The term vector feature that different size of convolution kernel is drawn into, for example can be divided using 1 × 1,2 × 2,3 × 3,4 × 4 convolution kernel The term vector feature of language Chou Qu not inquired.Feature extraction network can be convolutional neural networks.

In one embodiment of the application, as shown in figure 3, obtaining the mistake of discrete type feature associated with inquiry language Journey may include steps of S310 to step S330:

In step s310, historical user feedback information associated with inquiry language is obtained.

In one embodiment of the application, historical user feedback information associated with inquiry language is for indicating that history is used Feedback of the family to inquiry language response message.For example, inquiry language is the sentence of search for application, then related to inquiry language The historical user feedback information of connection can be the information for the application program that user is downloaded based on the inquiry language response message.More specifically Ground, it is assumed that inquiry language is " schoolgirl be suitble to play game ", after according to the inquiry language search user may download games 1, Game 2, game 3 etc., wherein the specific game of user's downloading is the feedback information for inquiring language " game that schoolgirl is suitble to play ".

In step s 320, it according to the historical user feedback information, counts associated with the inquiry language and each goes through The distribution situation of history field feedback.

In one embodiment of the application, the distribution of each historical user feedback information associated with inquiry language is counted Situation is distribution situation of the counting user for the feedback information of inquiry language response message.For example, inquiry language is search application The sentence of program, then the distribution situation of statistics each historical user feedback information associated with inquiry language can be statistics and use The distributed intelligence for the application program that family is downloaded based on the inquiry language response message.More specifically, it is assumed that inquiry language is that " schoolgirl is suitble to The game of object for appreciation ", after according to the inquiry language search user may download games 1, game 2, game 3 etc., then can unite Count the downloading accounting, the downloading accounting of game 2, the downloading accounting of game 3 etc. of game 1.

In step S330, according to the distribution situation of each historical user feedback information, it is special to generate the discrete type Sign.

In one embodiment of the application, spy can be generated according to the distribution situation of each historical user feedback information Vector is levied, using as discrete type feature.

In one embodiment of the application, as shown in figure 4, obtaining the mistake of discrete type feature associated with inquiry language Journey may include steps of S410 and step S420:

In step S410, identification inquiry language is intended that precisely intention or fuzzy intention.

It in one embodiment of the application, is precisely intended to indicate with specific object, fuzzy intention then indicates do not have Specific object.For example inquiry language is " schoolgirl be suitble to play game ", due to the game not downloaded clearly of inquiry language, because The intention of this inquiry language can be fuzzy intention；And inquire language " king's honor " due to be for specific game, Therefore the intention of the inquiry language can be accurate intention.

In one embodiment of the application, the meaning of inquiry language can be identified according at least one factor in following factor Figure is accurate is intended to or fuzzy intention: text similarity, inquiry language between inquiry language and scheduled intent information and predetermined Intent information between semantic similarity, with the distribution situation of the associated historical user feedback information of inquiry language.

In one embodiment of the application, if according to inquiry language and scheduled intent information between text similarity come Identification inquiry language is intended that precisely intention or fuzzy intention, then can be when text similarity is higher than given threshold, really The fixed inquiry language is intended that accurate intention.For example, inquiry language is " downloading king's honor ", due to the inquiry language and scheduled meaning Text similarity between figure information " king's honor " is higher, then can determine the inquiry language is intended that accurate intention.

In one embodiment of the application, if according to inquiry language and scheduled intent information between semantic similarity come Identification inquiry language is intended that precisely intention or fuzzy intention, then can be when the semantic similarity is higher than given threshold, really The fixed inquiry language is intended that accurate intention.For example, inquiry language is " downloading button button ", since the inquiry language and scheduled intention are believed The semantic similarity ceased between " QQ " is higher, then can determine the inquiry language is intended that accurate intention.

In one embodiment of the application, if according to the distribution feelings of historical user feedback information associated with inquiry language Condition inquires precisely intention or the fuzzy intention of being intended that of language to identify, then relatively high in accounting for for some historical user feedback information When, can determine the inquiry language is intended that accurate intention；If the accounting of each historical user feedback information is not much different, can Fuzzy intention is intended that with the determining inquiry language.For example, inquiry language is " king's honor ", if searching for it according to the inquiry language User downloads the ratio of " king's honor " app and is 11%, downloads it for the ratio of 80%, downloading " king's honor assistant " app afterwards The ratio of its application program is 9%, then since the ratio of user's download games " king's honor " app is higher, it may be considered that The inquiry language is the accurate intention for downloading " king's honor " app.

In one embodiment of the application, if according to inquiry language and scheduled intent information between text similarity, Inquire the distribution of the semantic similarity, historical user feedback information associated with inquiry language between language and scheduled intent information Two in situation are intended that precisely intention or fuzzy intention because of usually identification inquiry language, then can integrate the two factors Recognition result determine, if the recognition result than the two factors is all precisely to be intended to, it is determined that the intention of the inquiry language It is precisely to be intended to；If the recognition result of the two factors is all fuzzy intention, it is determined that the inquiry language is intended that fuzzy intention； If the recognition result of the two factors is not identical, can recognition result corresponding according to the two factors a possibility that size come Determine the intention of inquiry language.If than determining inquiry language according to the text similarity between inquiry language and scheduled intent information It is intended that accurate intention, and being intended that for inquiry language is determined according to the semantic similarity between inquiry language and scheduled intent information Fuzzy intention, if that the text similarity between the inquiry language and scheduled intent information is close to 1, then this can be determined Inquiry language is intended that accurate intention.

In one embodiment of the application, if according to inquiry language and scheduled intent information between text similarity, Inquire the distribution of the semantic similarity, historical user feedback information associated with inquiry language between language and scheduled intent information It all is intended that precisely intention or fuzzy intention because of usually identification inquiry language in situation, then can integrate whole factors Recognition result determines, if being precisely to be intended to than having the recognition result of more than half factor all, it is determined that the meaning of the inquiry language Figure is precisely to be intended to；If the recognition result of more than half factor is all fuzzy intention, it is determined that the inquiry language is intended that mould Paste is intended to.

With continued reference to shown in Fig. 4, in the step s 420, precisely it is intended to or obscures meaning according to being intended that for the inquiry language Figure generates the discrete type feature of the inquiry language.

In one embodiment of the application, can using inquire language be intended that precisely be intended to or fuzzy intention as one A feature vector, to obtain discrete type feature.

In one embodiment of the application, as shown in figure 5, obtaining the mistake of discrete type feature associated with inquiry language Journey may include steps of S510 to step S530:

In step S510, word segmentation processing is carried out to inquiry language and obtains at least one target word.

It, can the segmenting method based on string matching, the participle side based on understanding in one embodiment of the application Method carries out word segmentation processing to inquiry language based on the segmenting method of statistics.

In step S520, importance analysis is carried out at least one described target word, obtains each target word The importance score of language.

It, can be with since the significance level of different terms in a sentence is different in one embodiment of the application By carrying out importance analysis to the word obtained after inquiry language participle, with each word of determination to the importance for understanding inquiry language Score.

It, can be by least one factor in following factor to each target word in one embodiment of the application Carry out importance analysis: part of speech analysis result, each target word and the other mesh inquired in language of each target word Interdependent syntactic analysis result, the tightness of each target word analysis result between mark word, each target word are inversely literary Part frequency, each target word whether be setting entity word, each target word whether be off word.

In one embodiment of the application, above-mentioned each factor can be generated into feature vector as feature, so Machine learning model is trained afterwards, and then can determine the important of each word by the machine learning model after training Property score.

In one embodiment of the application, above-mentioned each factor can also be quantified, and each factor is set Weight, and then determine based on the quantized result of the weight of each factor and each factor the importance score of each word.

In step S530, the discrete type for generating the inquiry language according to the importance score of each target word is special Sign.

In one embodiment of the application, feature vector can be generated according to the importance score of each target word, Using as discrete type feature.

With continued reference to shown in Fig. 2, in step S230, institute is generated according to the term vector feature and the discrete type feature State the multi-channel feature vector of inquiry language.

, can be using each term vector feature as the feature in a channel in one embodiment of the application, and it will be every A discrete type feature generates multiple channel characteristics vectors as the feature in a channel.

In step S240, the multi-channel feature vector is input to intention assessment model, obtains the intention assessment The intent information of model output.

In one embodiment of the application, it is intended that identification model be in advance it is trained, training process can be as Shown in Fig. 6, include the following steps:

Step S610 obtains inquiry language sample.

Step S620, according to the term vector feature of inquiry language sample described in the inquiry language sample acquisition and the inquiry language The associated discrete type feature of sample.

In one embodiment of the application, term vector feature and the inquiry language sample for obtaining inquiry language sample are associated The process of the process of discrete type feature discrete type feature associated with the term vector feature of above-mentioned acquisition inquiry language and inquiry language It is identical, it repeats no more.

Step S630, according to the term vector feature and the associated discrete type of the inquiry language sample of the inquiry language sample Feature generates training sample.

In one embodiment of the application, each term vector feature of language sample can will be inquired as channel Feature, and each discrete type feature for inquiring language sample is generated into multiple channel characteristics vectors as the feature in a channel, Then the intent information for inquiring language sample is generated into training sample as label.

Step S640 is trained the intention assessment model based on the training sample.

In one embodiment of the application, it is intended that identification model can be multitask machine learning model, the multitask For machine learning model for exporting multistage intention, which includes full connection corresponding with intentions at different levels Layer and output layer, wherein output layer corresponding to the first intention in the multistage intent information is connected to corresponding to second intention Full articulamentum, and the level of the first intention be higher than the second intention level.The technical solution of the embodiment by make compared with Output layer corresponding to the intention of high-level is connected to full articulamentum corresponding to the intention of lower-level, so that higher levels The supervision message of intention labels can assist the fitting of the intention labels of lower-level to learn well, and then intention can be improved The accuracy rate of identification.For example, if inquiry language is the sentence " downloading king's honor " of search for application, then be intended to can be with for level-one It is " game ", second level intention can be " game-sports class ", three-level intention can be specific application name.

In one embodiment of the application, it is intended that identification model includes multitask machine learning model, the multiplexer The loss function of device learning model are as follows:

Loss=∑ (c_i×loss_i)+λ

Wherein, Loss indicates the loss function of multitask machine learning model；loss_iIndicate that i-stage is intended to corresponding damage Lose function；c_iFor hyper parameter, for indicating that i-stage is intended to the importance of corresponding loss function；λ indicates regularization term.

The loss function constructed in the embodiment is multi-tag Classification Loss function, and this mode can preferably capture use More intentions at family are distributed, and improve the accuracy of intention assessment.

In one embodiment of the application, in the case where being intended to identification model includes multitask machine learning model, After obtaining the intent information of intention assessment model output, it can be responded and be looked by the intention of the lowest hierarchical level in intent information Language is ask, response results are obtained；If response results are not able to satisfy response and require, upper one layer of intention response inquiry is passed sequentially through Language, until obtained response results meet response and require.

In one embodiment of the application, most fine intention is intended that due to lowest hierarchical level, it can be preferential The inquiry language of user's input is responded using the intention of lowest hierarchical level.Wherein, response is required can be and be recalled according to inquiry language Response results quantity, user are to satisfaction of response results recalled according to inquiry language etc..For example, if according to inquiry language output Level-one is intended that " game ", second level is intended that " game-sports class ", three-level is intended that specific application name, then such as Fruit is less according to the number of applications that three-level intention is recalled or is not able to satisfy the requirement of user, then can be intended to by second level " game-sports class " continues to recall, if the number of applications recalled is still less or is not able to satisfy wanting for user It asks, then passes through and be intended to again " game " and recalled.

The technical solution of the above embodiments of the present application to not only allow for inquiry language in the intention of identification inquiry language Term vector feature, and it is contemplated that with the associated discrete type feature of inquiry language, so can by multi-channel feature to Measure the accuracy rate to ensure intention assessment.

Below by taking user inputs the inquiry language (query) of search for application as an example, to the technical side of the embodiment of the present application Case is described in detail:

In one embodiment of the application, it is intended that the process of identification is mainly intended to polytypic process, uses obtaining After the query of family input, by being handled query the intention labels identified.Specifically, the embodiment of the present application Technical solution mainly include multi-channel feature fusion, model training and the several parts of model prediction, carry out individually below detailed Explanation.

1, multi-channel feature merges:

In one embodiment of the application, Embedding (insertion) vector spy has mainly been merged in multi-channel feature fusion Sign, user side search behavior feedback characteristic and the side query correlated characteristic.

In one embodiment of the application, Embedding vector characteristics can obtain in the following manner:

(1) using the term vector of the fasttext model extraction of pre-training, the CBOW (continuous of fasttext model Bag-of-words, continuous bag of words) model be a kind of supervised learning model, target be by have supervision pre-training obtain more Reasonable query vector characterization is so that the accuracy rate of intention labels classification is higher.

In one embodiment of the application, the parameter that fasttext model uses can be such that minimum word frequency can be for 5；Term vector dimension can be 128；Iteration cycle number can be 150；Training window size can be 5；N-gram (n-gram) It can be set to 4；And loss function can be constructed using negative sampling (negative sampling) algorithm, to ensure To pre-training term vector can preferably indicate the term (indicate word, term) of query, accelerate the convergence of training network.

(2) term vector of random initializtion, and as the training of convolutional neural networks carries out fine-tune (adjustment), lead to It crosses while can be captured using the convolution kernel of different size (such as can be using 1 × 1,2 × 2,3 × 3,4 × 4) more in query The local message of a difference n-gram, while being different from general shallow-layer neural network, the training of convolutional neural networks can be into One step takes out the expression of word-> word-> sentence higher order vector.

In addition, can also be obtained using bag of words such as word2vec, GloVe in one embodiment of the application The term vector of query.

In one embodiment of the application, user side search behavior feedback characteristic mainly includes the corresponding downloading of query Distribution situation.Specifically, each query has different downloading distributions for all app_id (application name), and can Downloading accounting will be normalized to the download of each app, therefore query-app downloading distribution multi-hot can be constructed (more heat) feature.For example, the downloading accounting of query " king's honor " is distributed are as follows: " king's honor: 0.8496 ", " king's honor helps Hand: 0.1297 ", " sword of heroic spirit: 0.0025 ".

In one embodiment of the application, the side query correlated characteristic may include: query precisely and ambiguous identification, The features such as the term different degree after query participle.

In one embodiment of the application, query is precisely and ambiguous identification is for indicating that query's is intended that accurate meaning Figure or fuzzy intention.Wherein, accurate query is to refer to that user has the intention for clearly downloading some app, obscures query table Show that user does not download the demand of specific app clearly, may just hope downloading certain one kind app, it is not true for specifically downloading which app Fixed.When determining query is accurate query or fuzzy query, can comprehensively consider user download entropy (i.e. with query phase The app of pass downloads distribution), the text similarity of query and app title, the semantic similarity etc. of query and app title it is multiple The factor carries out decision judgement.If it is determined that query is accurate query, then model should be intended to more to accurate query is corresponding Tendency；If it is determined that query is fuzzy query, then the corresponding more intention distributions of fuzzy query can be increasingly focused on.This In construct and be characterized in: if query is accurate query, characteristic value can be 1, otherwise can be set to 0.

In one embodiment of the application, term different degree is characterized in through part of speech analysis, interdependent sentence after query participle Whether method analysis the analysis of term tightness, IDF (Inverse Document Frequency, reverse document-frequency), is app Title word, multiple dimension construction features training term hierarchy models such as whether be off word come the term after being segmented to query into Row importance analysis provides the corresponding weight score value of each term.For example the query of user's input is " to download and be suitble to female The raw game played ", then the term importance analysis result finally obtained may is that " downloading: 0.126；It is suitble to: 0.135；Female It is raw: 0.452；It plays: 0.03；: 0.01；Game: 0.504 ".

2, model training

In one embodiment of the application, it is intended that identification model can using Multi-TextCNN (multichannel TextCNN) model, wherein TextCNN model is the convolutional neural networks model for carrying out text classification.As shown in fig. 7, In one embodiment of the application, Multi-TextCNN model specifically includes that multichannel input layer, convolutional layer, maximum pond Layer, full articulamentum, multitask softmax (normalization exponential function) probability output layer.Wherein, multichannel input layer includes (such as user side search behavior is fed back for the static channel feature (such as Embedding vector characteristics) of query and non-static channel characteristics Feature and the side query correlated characteristic).

In one embodiment of the application, there is over-fitting when Multi-TextCNN model training to be further reduced Probability, batch normalize (batch normalizes) layer is also introduced in model, and can also introduce in full articulamentum Dropout (random inactivation) mechanism.

In one embodiment of the application, as shown in Figure 7, it is assumed that the output of Multi-TextCNN model is three-level meaning Figure is (as level-one is intended to " game ", second level is intended to that " game _ chess and card " center ", three-level are intended to " game _ chess and card center _ bucket It is main "), then the probability output vector layer grade that level-one in Multi-TextCNN model is intended to can be connected to second level and three-level The full articulamentum being intended to, and the probability output vector layer grade that second level is intended to is connected to the full articulamentum that three-level is intended to, and then make The training that the supervision message that level-one is intended to can feed back to second level, three-level is intended to is obtained, the supervision message that second level is intended to can be fed back To the training that three-level is intended to, to further promote the accuracy rate of Intention Anticipation.

In one embodiment of the application, due to being related to the classification problem of level intention in the embodiment of the present application, because This can construct the loss function of MultiLabel&MultiTask (multi-tag & multitask), for example the formula of loss function can With as follows:

Loss=α × loss₁+β×loss₂+γ×loss₃+λ||Θ||₂

Wherein, α, β, γ indicate model hyper parameter, are respectively used to measure the phase of level-one label, second level label and three-level label To importance, such as α=0.5, β=0.3, γ=0.2；λ||Θ||₂Indicate regularization term.

In one embodiment of the application, loss₁、loss₂And loss₃It can be two classification cross entropy loss functions, than Such as it can beWherein n indicates number of training, and x indicates one Bar sample, y indicate that the true tag of the sample, σ () function representation sigmoid activation primitive, logits indicate network output layer Feedforward output valve: logits=Wx+b, W be output layer weight parameter, b be output layer offset parameter.

3, model prediction

In one embodiment of the application, after being completed to model training, if receiving the query of user's input, The multi-channel feature of query can be then extracted, which has equally merged Embedding vector characteristics, user side is searched Then the multi-channel feature of query is input in the model after training by Suo Hangwei feedback characteristic and the side query correlated characteristic, Obtain the intent information of model output.In one embodiment of the application, as shown in figure 8, intention assessment process may include Following steps:

Pretreatment: mainly to the query that user inputs carries out capital and small letter conversion, full-shape turns half-angle, either traditional and simplified characters conversion, spy The normalizeds such as different filtered symbol.

Slot position parsing: mainly by default rule template to query carry out intents process, target be from Corresponding intention slot position and slot position value are extracted in query.For example the query of user's input is " to download and schoolgirl is suitble to play Game " parses available " [D:user]: schoolgirl " and " [D:type]: game " two intention slot positions by slot position.

Query error correction: error detection and correction mainly is carried out to the wrong query that user inputs.

Query participle: the other dicing process of word-level mainly is carried out to the query sentence sequence that user inputs.? After query word segmentation processing, query correction process can also be carried out again.

Query extension: the processing that semantic relevant query recommends mainly is carried out to the query that user inputs, such as: " king Person's honor " expands the relevant query such as " king's honor assistant ", " moba game ", so that user be helped to carry out interest exploration And expand call back number.

Term importance analysis: importance measurement mainly is carried out to each term after query participle, in query not It is different with significant contribution degree of the term when carrying out text and recalling, it should be distinguish.

Intention assessment: be mainly based upon term importance analysis result and query other features (such as Embedding to Measure feature, user side search behavior feedback characteristic and the side query correlated characteristic, identify query's by intention assessment model It is intended to.

Yellow anti-identification can also be carried out after intention assessment, that is, identifies that relate to Huang etc. is intended in violation of rules and regulations, and can also be into Row manual intervention adjustment.

It, can be based on the meaning identified after identifying the intent information of query in one embodiment of the application Scheme the query of information response user.

In one embodiment of the application, as shown in figure 9, when user inputs query " uncle in app search column 901 Do me a favour " after, intention is identified according to the scheme of the relevant technologies and responds result and user's input that query is obtained The correlation of query is poor.And by the technical solution of the embodiment of the present application to intention assessment result such as Figure 10 of the query It is shown, wherein the accurate fuzzy Judgment result of the query is accurate query；The importance of " uncle " that fine granularity segments It is scored at 0.5556, the importance of " doing me a favour " is scored at 0.4444；The result of coarseness participle is " uncle does me a favour "；Identification The probability for being intended to " social activity _ love and marriage _ blind date/appointment " to the query is 0.4256, it is intended that for " social _ chatting with friends _ is chatted The probability of its tool " is 0.1863.And then the associated app result that the intention by identifying is recalled is as shown in figure 11, it is seen that Accurate intention can recognize that using the technical solution of the embodiment of the present application, and then can recall and more meet user's intention APP, be conducive to promoted user search experience.

In another embodiment of the application, as shown in figure 12, for query: Shanghai Volkswagen automobile data recorder client End, it is 0.6094 that identifies in the related technology, which is intended to the probability of " video _ Online Video _ comprehensive video ", it is intended that is " raw The probability of work _ buying car _ information/strategy " is 0.1992；And it is intended to using what the technical solution of the embodiment of the present application identified The probability of " trip _ trip service _ driving recording " is 0.9540.Does for query: how rear Yi, a legendary monarch of Youqiong State in the xia Dynasty go out to fill? game strategy just asks it, It is 0.1699 that identifies in the related technology, which is intended to the probability of " game _ movement venture _ cool run ",；And use the embodiment of the present application The probability for being intended to " game _ peripheral game _ game community " that identifies of technical solution be 0.5001, it is intended that be " game _ trip The probability of play periphery _ game strategy " is 0.1289.For query: blurred picture can be allowed to become clearly software, the relevant technologies In the probability for being intended to " photography _ picture sharing _ U.S. figure " that identifies be 0.5645；And use the technical side of the embodiment of the present application The probability that case identified be intended to " photography _ editor's beautification _ editor's beautification " is 0.8348.For query: China Unicom stock Code is how many, and it is 0.2671 that identifies in the related technology, which is intended to the probability of " communication _ business hall _ connection business hall ", meaning Figure is that the probability of " shopping _ shopping payment _ payment " is 0.2599；And the meaning for using the technical solution of the embodiment of the present application to identify Figure is that the probability of " financing _ investment _ stock " is 0.2452, it is intended that is for the probability of " communication _ business hall _ integrated service " 0.2377.As it can be seen that more accurate using the intention that the technical solution of the embodiment of the present application identifies.

To sum up, the technical solution of the embodiment of the present application has the advantages that

1, end-to-end study:

More fusion features that the technical solution of the embodiment of the present application proposes are not needed by too complicated Feature Engineering.

2, multichannel inputs:

The technical solution of the embodiment of the present application in addition to using random initializtion and carry out fine-tune term vector feature it Outside, it is also contemplated that other channel characteristics: (1) term vector being drawn into using the fasttext of pre-training, model is by from query It can learn in input and its corresponding intention labels supervision message to more reasonable term vector to characterize.(2) using some discrete Feature is as input, and (the corresponding app of such as query downloads distributed intelligence to user's search behavior feedback characteristic including user side Deng) and the features such as the accurate fuzzy Judgment mark of the side query, term importance, keyword.The technology of the embodiment of the present application Scheme is trained using the feature of multiple channels fusion.

3, Multi-label multi-tag is classified:

Since the intention that user inputs the behind query may be diversified, such as: " the cologne marshal " may be corresponding " celestial A variety of intentions such as chivalrous class game ", " novel class ", " video class ", therefore use the more classification methods pair of general Multi-classes The prediction of intention be it is exclusive, cannot capture well user more intentions distribution.Therefore the technical side of the embodiment of the present application Case puts forward the multi-tag Classification Loss function based on multi-label.

4, multi-task learning:

Since the sample size that different three-level intention labels are covered has a certain difference, such as three-level intention labels Sample size it is less, it is thus possible to will lead to some three-levels and be intended to study and be not enough and cause forecasting inaccuracy true.In this Shen By introducing multi-task learning frame in embodiment please, and the supervisory signals for learning I and II intention labels are very It assists the fitting of three-level intention labels to learn well, therefore the accuracy rate of intention assessment can be promoted.

It should be noted that scene of the technical solution of the embodiment of the present application in addition to being suitable for search for application, also suitable In the scene being intended to for other any required identification users, such as the application such as Webpage search, human-machine intelligence's dialogue, voice control Scene.

The Installation practice of the application introduced below can be used for executing the intention assessment side in the above embodiments of the present application Method.For undisclosed details in the application Installation practice, the embodiment of the above-mentioned intension recognizing method of the application is please referred to.

Figure 13 shows the block diagram of the intention assessment device according to one embodiment of the application.

Referring to Fig.1 shown in 3, according to the intention assessment device 1300 of one embodiment of the application, comprising: first obtains list Member 1302, second acquisition unit 1304, the first generation unit 1306 and processing unit 1308.

Wherein, first acquisition unit 1302 is used to obtain the inquiry language of user's input；Second acquisition unit 1304 is used for root The term vector feature of the inquiry language is obtained according to the inquiry language, and obtains discrete type feature associated with the inquiry language； First generation unit 1306 is used to generate the multichannel of the inquiry language according to the term vector feature and the discrete type feature Feature vector；Processing unit 1308 is used to the multi-channel feature vector being input to intention assessment model, obtains the intention The intent information of identification model output.

In some embodiments of the present application, be based on aforementioned schemes, second acquisition unit 1304 be configured that acquisition with it is described Inquire the associated historical user feedback information of language；According to the historical user feedback information, count related to the inquiry language The distribution situation of each historical user feedback information of connection；It is raw according to the distribution situation of each historical user feedback information At the discrete type feature.

In some embodiments of the present application, aforementioned schemes are based on, second acquisition unit 1304 is configured that described in identification and looks into That askes language is intended that precisely intention or fuzzy intention；Precisely it is intended to still fuzzy intention life according to being intended that for the inquiry language At the discrete type feature of the inquiry language.

In some embodiments of the present application, be based on aforementioned schemes, second acquisition unit 1304 be configured to according to it is following because At least one factor in element identify it is described inquiry language be intended that precisely be intended to or fuzzy intention: the inquiry language with make a reservation for Intent information between text similarity, it is described inquiry language and scheduled intent information between semantic similarity, with it is described Inquire the distribution situation of the associated historical user feedback information of language.

In some embodiments of the present application, aforementioned schemes are based on, second acquisition unit 1304 is configured that the inquiry Language carries out word segmentation processing and obtains at least one target word；Importance analysis is carried out at least one described target word, is obtained The importance score of each target word；The inquiry language is generated according to the importance score of each target word Discrete type feature.

In some embodiments of the present application, be based on aforementioned schemes, second acquisition unit 1304 be configured to by it is following because At least one factor in element carries out importance analysis to each target word: the part of speech of each target word analyzes knot The interdependent syntactic analysis result between other target words, each target word in fruit, each target word and the inquiry language The tightness of language analyzes whether result, the reverse document-frequency of each target word, each target word are the entity words set, is each Whether a target word is off word.

In some embodiments of the present application, aforementioned schemes are based on, second acquisition unit 1304 is configured that the inquiry Language is input to the term vector extraction model of pre-training, to obtain the term vector feature of the term vector extraction model output.

In some embodiments of the present application, aforementioned schemes are based on, second acquisition unit 1304 is configured that be mentioned by feature Take the term vector feature that language is inquired described in network abstraction, wherein the feature extraction network passes through a variety of different size of convolution Core extracts the term vector feature of the inquiry language respectively.

In some embodiments of the present application, it is based on aforementioned schemes, the intention assessment device 1300 further include: third Acquiring unit, for obtaining inquiry language sample；4th acquiring unit, for inquiring language according to the inquiry language sample acquisition The term vector feature of sample and the associated discrete type feature of inquiry language sample；Second generation unit, for according to The term vector feature and the associated discrete type feature of inquiry language sample of language sample are inquired, training sample is generated；Training is single Member, for being trained based on the training sample to the intention assessment model.

Loss=∑ (c_i×loss_i)+λ

In some embodiments of the present application, aforementioned schemes are based on, the intention assessment model includes multitask engineering Practise model, the intention assessment device 1300 further include: response unit, for obtaining the meaning of the intention assessment model output After figure information, the inquiry language is responded by the intention of the lowest hierarchical level in the intent information, obtains response results；If institute It states response results and is not able to satisfy response requirement, then pass sequentially through upper one layer of intention and respond the inquiry language, until obtain Response results meet the response requirement.

It should be noted that the computer system 1400 of the electronic equipment shown in Figure 14 is only an example, it should not be to this The function and use scope for applying for embodiment bring any restrictions.

As shown in figure 14, computer system 1400 include central processing unit (Central Processing Unit, CPU) 1401, it can be according to the program being stored in read-only memory (Read-Only Memory, ROM) 1402 or from depositing It stores up the program that part 1408 is loaded into random access storage device (Random Access Memory, RAM) 1403 and executes each Kind movement appropriate and processing, such as execute method described in above-described embodiment.In RAM 1403, it is also stored with system behaviour Various programs and data needed for making.CPU 1401, ROM 1402 and RAM 1403 are connected with each other by bus 1404.It is defeated Enter/export (Input/Output, I/O) interface 1405 and is also connected to bus 1404.

I/O interface 1405 is connected to lower component: the importation 1406 including keyboard, mouse etc.；Including such as cathode Ray tube (Cathode Ray Tube, CRT), liquid crystal display (Liquid Crystal Display, LCD) etc. and loudspeaking The output par, c 1407 of device etc.；Storage section 1408 including hard disk etc.；And including such as LAN (Local Area Network, local area network) card, modem etc. network interface card communications portion 1409.Communications portion 1409 is via such as The network of internet executes communication process.Driver 1410 is also connected to I/O interface 1405 as needed.Detachable media 1411, such as disk, CD, magneto-optic disk, semiconductor memory etc., are mounted on as needed on driver 1410, in order to It is mounted into storage section 1408 as needed from the computer program read thereon.

Particularly, according to an embodiment of the present application, it may be implemented as computer above with reference to the process of flow chart description Software program.For example, embodiments herein includes a kind of computer program product comprising be carried on computer-readable medium On computer program, which includes the program code for method shown in execution flow chart.In such reality It applies in example, which can be downloaded and installed from network by communications portion 1409, and/or from detachable media 1411 are mounted.When the computer program is executed by central processing unit (CPU) 1401, executes in the system of the application and limit Various functions.

It should be noted that computer-readable medium shown in the embodiment of the present application can be computer-readable signal media Or computer readable storage medium either the two any combination.Computer readable storage medium for example can be with System, device or the device of --- but being not limited to --- electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor, or it is any more than Combination.The more specific example of computer readable storage medium can include but is not limited to: have one or more conducting wires Electrical connection, portable computer diskette, hard disk, random access storage device (RAM), read-only memory (ROM), erasable type are programmable Read-only memory (Erasable Programmable Read Only Memory, EPROM), flash memory, optical fiber, Portable, compact Disk read-only memory (Compact Disc Read-Only Memory, CD-ROM), light storage device, magnetic memory device or The above-mentioned any appropriate combination of person.In this application, computer readable storage medium can be it is any include or storage program Tangible medium, which can be commanded execution system, device or device use or in connection.And in this Shen Please in, computer-readable signal media may include in a base band or as carrier wave a part propagate data-signal, In carry computer-readable program code.The data-signal of this propagation can take various forms, including but not limited to Electromagnetic signal, optical signal or above-mentioned any appropriate combination.Computer-readable signal media can also be computer-readable Any computer-readable medium other than storage medium, the computer-readable medium can send, propagate or transmit for by Instruction execution system, device or device use or program in connection.The journey for including on computer-readable medium Sequence code can transmit with any suitable medium, including but not limited to: wireless, wired etc. or above-mentioned is any appropriate Combination.

Flow chart and block diagram in attached drawing are illustrated according to the system of the various embodiments of the application, method and computer journey The architecture, function and operation in the cards of sequence product.Wherein, each box in flowchart or block diagram can represent one A part of a part of a module, program segment or code, above-mentioned module, program segment or code is used for comprising one or more The executable instruction of logic function as defined in realizing.It should also be noted that in some implementations as replacements, being marked in box Function can also occur in a different order than that indicated in the drawings.For example, two boxes succeedingly indicated actually may be used To be basically executed in parallel, they can also be executed in the opposite order sometimes, and this depends on the function involved.It is also noted that , the combination of each box in block diagram or flow chart and the box in block diagram or flow chart can be as defined in executing The dedicated hardware based systems of functions or operations is realized, or can be come using a combination of dedicated hardware and computer instructions It realizes.

Being described in unit involved in the embodiment of the present application can be realized by way of software, can also be by hard The mode of part realizes that described unit also can be set in the processor.Wherein, the title of these units is in certain situation Under do not constitute restriction to the unit itself.

As on the other hand, present invention also provides a kind of computer-readable medium, which be can be Included in electronic equipment described in above-described embodiment；It is also possible to individualism, and without in the supplying electronic equipment. Above-mentioned computer-readable medium carries one or more program, when the electronics is set by one for said one or multiple programs When standby execution, so that the electronic equipment realizes method described in above-described embodiment.

It should be noted that although being referred to several modules or list for acting the equipment executed in the above detailed description Member, but this division is not enforceable.In fact, according to presently filed embodiment, it is above-described two or more Module or the feature and function of unit can embody in a module or unit.Conversely, an above-described mould The feature and function of block or unit can be to be embodied by multiple modules or unit with further division.

Through the above description of the embodiments, those skilled in the art is it can be readily appreciated that example described herein is implemented Mode can also be realized by software realization in such a way that software is in conjunction with necessary hardware.Therefore, according to the application The technical solution of embodiment can be embodied in the form of software products, which can store non-volatile at one Property storage medium (can be CD-ROM, USB flash disk, mobile hard disk etc.) in or network on, including some instructions are so that a calculating Equipment (can be personal computer, server, touch control terminal or network equipment etc.) is executed according to the application embodiment Method.

Those skilled in the art will readily occur to the application after considering specification and practicing embodiment disclosed herein Other embodiments.This application is intended to cover any variations, uses, or adaptations of the application, these modifications are used Way or adaptive change follow the application general principle and including the application it is undocumented in the art known in Common sense or conventional techniques.

It should be understood that the application is not limited to the precise structure that has been described above and shown in the drawings, and And various modifications and changes may be made without departing from the scope thereof.Scope of the present application is only limited by the accompanying claims.

Claims

1. a kind of intension recognizing method characterized by comprising

Obtain the inquiry language of user's input；

The term vector feature of the inquiry language is obtained according to the inquiry language, and obtains discrete type associated with the inquiry language Feature；

The multi-channel feature vector of the inquiry language is generated according to the term vector feature and the discrete type feature；

The multi-channel feature vector is input to intention assessment model, obtains the intention letter of the intention assessment model output Breath.

2. intension recognizing method according to claim 1, which is characterized in that obtain associated discrete with the inquiry language Type feature, comprising:

Obtain historical user feedback information associated with the inquiry language；

According to the historical user feedback information, point of each historical user feedback information associated with the inquiry language is counted Cloth situation；

According to the distribution situation of each historical user feedback information, the discrete type feature is generated.

3. intension recognizing method according to claim 1, which is characterized in that obtain associated discrete with the inquiry language Type feature, comprising:

Identify the inquiry language is intended that precisely intention or fuzzy intention；

According to the discrete type feature for being intended that precisely intention or fuzzy intention and generating the inquiry language of the inquiry language.

4. intension recognizing method according to claim 3, which is characterized in that according at least one factor in following factor Identify the inquiry language is intended that precisely intention or fuzzy intention:

It is described to inquire between the text similarity between language and scheduled intent information, the inquiry language and scheduled intent information Semantic similarity, with the distribution situation of the associated historical user feedback information of the inquiry language.

5. intension recognizing method according to claim 1, which is characterized in that obtain associated discrete with the inquiry language Type feature, comprising:

Word segmentation processing is carried out to the inquiry language and obtains at least one target word；

Importance analysis is carried out at least one described target word, obtains the importance score of each target word；

The discrete type feature of the inquiry language is generated according to the importance score of each target word.

6. intension recognizing method according to claim 5, which is characterized in that pass through at least one factor in following factor Importance analysis is carried out to each target word:

The part of speech of each target word is analyzed between other target words in result, each target word and the inquiry language Interdependent syntactic analysis result, the tightness of each target word analyze result, the reverse document-frequency of each target word, each mesh Mark word whether be setting entity word, each target word whether be off word.

7. intension recognizing method according to claim 1, which is characterized in that obtain the inquiry language according to the inquiry language Term vector feature, comprising:

The inquiry language is input to the term vector extraction model of pre-training, to obtain the word of the term vector extraction model output Vector characteristics.

8. intension recognizing method according to claim 1, which is characterized in that obtain the inquiry language according to the inquiry language Term vector feature, comprising:

By the term vector feature for inquiring language described in feature extraction network abstraction, wherein the feature extraction network passes through a variety of Different size of convolution kernel extracts the term vector feature of the inquiry language respectively.

9. intension recognizing method according to claim 1, which is characterized in that further include:

Obtain inquiry language sample；

It is associated according to the term vector feature of inquiry language sample described in the inquiry language sample acquisition and the inquiry language sample Discrete type feature；

According to the term vector feature and the associated discrete type feature of inquiry language sample of the inquiry language sample, training is generated Sample；

The intention assessment model is trained based on the training sample.

10. intension recognizing method according to claim 9, which is characterized in that the intention assessment model includes multitask Machine learning model, the multitask machine learning model is for exporting multistage intention, the multitask machine learning model packet Include full articulamentum corresponding with intentions at different levels and output layer, wherein corresponding to the first intention in the multistage intent information Output layer be connected to full articulamentum corresponding to second intention, the level of the first intention is higher than the layer of the second intention Grade.

11. intension recognizing method according to claim 9, which is characterized in that the intention assessment model includes multitask Machine learning model, the loss function of the multitask machine learning model are as follows:

Loss=∑ (c_i×loss_i)+λ

Wherein, Loss indicates the loss function of the multitask machine learning model；loss_iIndicate that i-stage is intended to corresponding loss Function；c_iFor hyper parameter, for indicating that i-stage is intended to the importance of corresponding loss function；λ indicates regularization term.

12. intension recognizing method according to any one of claim 1 to 11, which is characterized in that the intention assessment mould Type includes multitask machine learning model, the intension recognizing method further include:

After the intent information for obtaining the intention assessment model output, pass through the meaning of the lowest hierarchical level in the intent information Figure responds the inquiry language, obtains response results；

If the response results are not able to satisfy response and require, passes sequentially through upper one layer of intention and respond the inquiry language, directly Meet the response requirement to obtained response results.

13. a kind of intention assessment device characterized by comprising

First acquisition unit, for obtaining the inquiry language of user's input；

Second acquisition unit for obtaining the term vector feature of the inquiry language according to the inquiry language, and is obtained and is looked into described Ask the associated discrete type feature of language；

Generation unit, for generating the multi-channel feature of the inquiry language according to the term vector feature and the discrete type feature Vector；

Processing unit obtains the intention assessment model for the multi-channel feature vector to be input to intention assessment model The intent information of output.

14. a kind of computer-readable medium, is stored thereon with computer program, which is characterized in that the computer program is located Manage the intension recognizing method realized as described in any one of claims 1 to 12 when device executes.

15. a kind of electronic equipment characterized by comprising

One or more processors；

Storage device, for storing one or more programs, when one or more of programs are by one or more of processing When device executes, so that one or more of processors realize the intention assessment side as described in any one of claims 1 to 12 Method.