CN110209446A

CN110209446A - The configuration method and device of slot position are combined in a kind of interactive system

Info

Publication number: CN110209446A
Application number: CN201910330314.0A
Authority: CN
Inventors: 张晴; 胡仁林; 刘畅; 张轶博
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2019-04-23
Filing date: 2019-04-23
Publication date: 2019-09-06
Anticipated expiration: 2039-04-23
Also published as: CN110209446B; WO2020216134A1

Abstract

The application provides the configuration method and device that slot position is combined in a kind of interactive system, is related to AI technical field, even if user's saying includes the expression of entity type exchange sequence in the combination slot position of user setting, the extraction of slot position still may be implemented.Concrete scheme includes: that Bot platform receives user in the first slot position of the first interface (i.e. for the interface of slot position to be arranged for the first intention in the first technical ability of Bot platform) configuration；First slot position be include according to the combination slot position of tactic N number of entity type set by user, N >=2, N are positive integer；N number of entity type is recombinated, M the second slot positions are obtained；M the second slot positions include the slot position that k entity type in N number of entity type arranges in any order, k ∈ { 1,2 ... ..., N }；According to one or more training corpus and M the second slot positions, one or more training corpus are trained, Bot platform is made to have the ability for extracting the M in user's saying the second slot position.

Description

The configuration method and device of slot position are combined in a kind of interactive system

Technical field

The invention relates to artificial intelligence (artificial intelligence, AI) technical fields, more particularly to The configuration method and device of slot position are combined in a kind of interactive system.

Background technique

AI is to realize the essential technology of artificial intelligence.Human-computer interaction (such as man machine language's interaction) is a kind of common production Product AI ability can be described as chat robots (ChatBot, abbreviation Bot).ChatBot can be divided into two kinds: open field (Open Domain) chat product and task orientation (Task Oriented) chat product.Wherein, task orientation chat product is with class It is the human-computer interaction product of guiding like the single task roles such as " ordering air ticket ", " ordering ", " inquiry weather ".

Wherein, user can be Bot platform configuration one or more technical ability in the Bot platform that product is chatted in task orientation (Skill), it is intended to (Intent) and slot position (Slot), for realizing human-computer interaction.For example, above-mentioned technical ability can for " ordering ", " paying one's fare " etc..Each technical ability may include one or more intentions, if technical ability " ordering " may include " buying hamburger " and " buying beverage " Etc. multiple intentions.Each intention may include one or more slot positions.Such as, it is intended that " buying hamburger " may include slot position " quantity ", with And slot position " hamburger@type " etc..

Referring to FIG. 1, by taking man machine language's interaction as an example, a kind of interactive voice process of task orientation can include: (1) Bot Microphone acquire voice data 101.(2) speech recognition module (automatic speech recognition, ASR) is to language Sound data 101 carry out speech recognition, identify corresponding text information 1.(3) semantic understanding (natural language Understanding, NLU) module understands the intention to be expressed of text information 1.Wherein, it is intended that can determine text information 1 instruction Bot platform executes the type of event.For example, it is assumed that the file information 1 be " I wants a wheat chicken with several spices leg fort ", NLU module by Text information 1 can obtain: be intended to=" buying hamburger ", i.e., text information 1 indicates that Bot is that user buys hamburger.(4) it is intended to be determined Afterwards, NLU module parsing text information 1 extracts core content, i.e. slot position.In conjunction with examples detailed above, NLU module is extractable out: slot position "@quantity "=" portion " and slot position " hamburger@type "=" wheat chicken with several spices leg fort ".(5) Bot platform is according to the slot position extracted It is responded to user.Wherein ,@is for identifying an entity type, without practical significance.

Currently, can not only be defined in Bot platform it is above-mentioned only include an entity type slot position, such as slot position "@quantity " With slot position " hamburger@type "；It can also support the design of combination slot position.It includes at least two entity types that combination slot position, which is a kind of, Slot position.For example, it includes '@quantity ' and ' hamburger@type ' two entity that above-mentioned two slot position can be combined into one by user The combination slot position " quantity hamburger type " of type.In this way, NLU module identifies that above-mentioned text information 1 is extractable out: slot position "@number Measure the hamburger@type "=" a wheat chicken with several spices leg fort ".

But the prior art only supports the extraction to the combination slot position for specifying built-up sequence in text information.If text The expression of entity type exchange sequence, Bot platform can not cannot then identify in combination slot position in information including user configuration.Example Such as.Assuming that defining combination slot position " quantity hamburger type " for intention " buying hamburger " in Bot platform.It is real in the combination slot position The sequence of body type are as follows: '@quantity ' preceding, ' hamburger@type ' is rear.So, Bot platform can only then identify and response text is believed " I wants a wheat chicken with several spices leg fort ", " giving me two parts of delicious and crisp cod-fish Hamburgs " etc., '@quantity ' are ceased in preceding, ' hamburger@type ' posterior table It reaches；And when the file information be " I wants two parts of delicious and crisp cod-fish Hamburg ", " give my three parts of wheat chicken with several spices leg fort " etc., '@quantity ' after, ' the@Chinese When fort type ' preceding expression, Bot cannot then extract corresponding slot position.

Summary of the invention

The application provides the configuration method and device that slot position is combined in a kind of interactive system, even if wrapping in user's saying The expression of entity type exchange sequence in the combination slot position of user setting is included, the extraction of slot position still may be implemented.

The application adopts the following technical scheme that

In a first aspect, the application provides the configuration method for combining slot position in a kind of interactive system, this method is applied to Bot platform.This method may include: that Bot platform can receive user for the first meaning in the first technical ability for Bot platform First slot position of the first interface configurations of figure setting slot position；First slot position is the combination slot position for including N number of entity type, N >= 2, N be positive integer；N number of entity type arranges in the first slot position according to sequence set by user；Then, Bot platform recombinates N A entity type obtains M the second slot positions；The M the second slot positions include k entity type in N number of entity type according to times Meaning sequentially arranges obtained slot position, k ∈ { 1,2 ... ..., N }；Finally, Bot platform is according to one or more training corpus and M Second slot position is trained one or more training corpus, and Bot platform is made to have M the second slots extracted in user's saying The ability of position.

It include one or more entities in multiple entity types due to recombinating obtained multiple second slot positions in the application The slot position that type arranges in any order, i.e. Bot platform can recombinate to obtain multiple entity type according to any suitable Slot position after sequence arrangement；Therefore, it is only wrapped in the sequence or user's saying of multiple entity type even if changing in user's saying An entity type is included, Bot platform can also extract corresponding slot position, reply user's saying.

With reference to first aspect, in a kind of possible design method, M the second slot positions include k in N number of entity type The slot position that entity type arranges in any order, k ∈ { 1,2 ... ..., N }.That is, k can be in [1, N] Any positive integer.

For example, Bot platform can select 1 (i.e. k) entity type, Bot platform from N number of entity type when k=1 1 entity type selected is as second slot position.Wherein, Bot platform can have the selection of N kind, it can obtain A second slot position.

When k=2, Bot platform can select 2 (i.e. k) entity types from N number of entity type, Bot platform selecting 2 entity types are combined into second slot position (i.e. combination slot position).Wherein, Bot platform can have N × (N-1) kind selection, i.e., It is availableA second slot position.

When k=3, Bot platform can select 3 (i.e. k) entity types from N number of entity type, Bot platform selecting 3 entity types are combined into second slot position (i.e. combination slot position).Wherein, Bot platform can have N × (N-1) × (N-2) kind Selection, it can obtainA second slot position.

When k=N-1, Bot platform can select N-1 (i.e. k) entity types, the choosing of Bot platform from N number of entity type The N-1 entity type selected is combined into second slot position (i.e. combination slot position).Wherein, Bot platform can have N × (N-1) × (N-2) × ... × 2 kinds of selections, it can obtain A Two slot positions.

When k=N, Bot platform can select N number of (i.e. k) entity type from N number of entity type, Bot platform selecting N number of entity type is combined into second slot position (i.e. combination slot position).Wherein, Bot platform can have N × (N-1) × (N-2) × ... × 2 × a kind of selection, it can obtain A Two slot positions.

By foregoing description it follows that

With reference to first aspect, in alternatively possible design method, the above method can also include: that Bot platform receives User is in the slot position title of the first interface configurations, and the slot position title is for identifying the first slot position.For example, above-mentioned first interface can be with Including " slot position title ", frame is set.Bot platform can receive what " slot position title " setting frame of the user in the first interface inputted "CompositeNE".I.e. above-mentioned slot position title may include " CompositeNE ".

With reference to first aspect, in alternatively possible design method, N number of entity type is recombinated in Bot platform, obtains M After a second slot position, Bot platform instructs one or more according to one or more training corpus and the M the second slot positions Before white silk corpus is trained, the present processes can also include: that slot position title is associated with by Bot platform with M the second slot positions Storage, and be the feature of corresponding M the second slot position label composite entity of slot position title, the feature of composite entity is used to indicate this M the second slot positions are the composite entity type after recombination, that is, the normalization composite entity (Composite Entity) after recombinating Type.For example, it is assumed that above-mentioned first slot position is " hamburger@quantity@type ".Bot platform recombinant C ompositeNE " the@quantity@Chinese Fort type ", available four the second slot positions: "@quantity ", " hamburger@type ", " hamburger@quantity@type " and " hamburger@class Type@quantity ".Bot platform can by slot position title (such as CompositeNE) and this four the second slot position associated storages, then for The feature of corresponding 4 the second slot positions label Composite Entity of CompositeNE (i.e. slot position title).

With reference to first aspect, in alternatively possible design method, above-mentioned Bot platform is according to one or more training language Material and M the second slot positions, are trained one or more of training corpus, comprising: Bot platform receive user be one or The true tag of each of multiple training corpus word addition；Bot platform is one or more training according to M the second slot positions Each of corpus word adds feature；Bot platform use deep learning algorithm, according to true tag and feature, to one or Multiple training corpus are trained；Alternatively, Bot platform is using deep learning algorithm conjugation condition with field (conditional Random field algorithm, CRF) algorithm instructs one or more training corpus according to true tag and feature Practice.

With reference to first aspect, in alternatively possible design method, above-mentioned deep learning algorithm includes shot and long term memory Network (long short-term memory, LSTM) algorithm.

With reference to first aspect, in alternatively possible design method, above-mentioned Bot platform can receive the one of user's input A or multiple training corpus relevant to above-mentioned M the second slot positions.The one or more training corpus is user for above-mentioned the One is intended to (such as buying hamburger) possible user's saying.For example, being directed to first intention " buying hamburger ", possible user's saying be can wrap It includes: " I will buy a wheat chicken with several spices leg fort ", " giving me three cod-fish Hamburgs ", " I wants four cod-fish Hamburgs " or " wheat chicken with several spices leg fort three It is a " etc..Then, Bot platform can receive the true tag that user is the word addition of each of the training corpus, and be each Word adds feature.Slot position title that the true tag of one word is used to indicate slot position corresponding to this word and this word are right Answer the position in slot position.The feature of one word is used to indicate the composite entity feature and this word of slot position corresponding to this word Position in corresponding slot position.Finally, Bot platform can use deep learning algorithm, it is training corpus addition according to user True tag, M slot position configuring in the feature that Bot platform is training corpus addition and Bot platform, to one or more A training corpus is trained.Alternatively, Bot platform can use deep learning algorithm combination CRF algorithm, it is the instruction according to user Practice the M slot position configured in the feature and Bot platform that true tag, the Bot platform that corpus adds are training corpus addition, One or more training corpus are trained.

By taking training corpus is " I wants two pears of an apple " as an example, user can be the training in different combinations The true tag of each of corpus word addition, Bot platform can add feature for each of training corpus word.In reality In existing mode (1), Bot platform can be that each of training corpus word adds feature in a manner of " correct combination ".In reality In existing mode (2), Bot platform can be each of training corpus word in a manner of " correct combination "+" fine granularity combination " Add feature.In implementation (3), Bot platform can be in a manner of " exhaustion combination "+" fine granularity combination ", for training language Each of material word adds feature.In implementation (4), Bot platform can be in a manner of " exhaustion combination ", for training language Each of material word adds feature.

Wherein, in above-mentioned each implementation, user is each of training corpus word addition true tag, and Bot platform is that the specific method of each of training corpus word addition feature can be with reference to retouching in the embodiment of the present application in detail It states, it will not go into details here by the application.

With reference to first aspect, in alternatively possible design method, above-mentioned Bot platform recombinates N number of entity type, obtains M the second slot positions, comprising: Bot platform uses dynamic programming algorithm, recombinates N number of entity type, obtains M the second slot positions.

With reference to first aspect, in alternatively possible design method, above-mentioned Bot platform is according to one or more training language Material and M the second slot positions, are trained one or more training corpus, comprising: Bot platform uses single-point sorting algorithm and base In the dynamic programming algorithm of probability, according to one or more training corpus and M the second slot positions, to one or more training corpus It is trained.

With reference to first aspect, in alternatively possible design method, above-mentioned single-point sorting algorithm include at least support to Amount machine (support vector machine, SVM) model, maximum entropy model, Fast Text Classification algorithm (fasttext) mould Type, convolutional neural networks (convolution neural network, CNN) model, n-gram model (n-gram) model or Any one of Recognition with Recurrent Neural Network (recurrent neural network, RNN) model.

With reference to first aspect, in alternatively possible design method, above-mentioned single-point sorting algorithm is (bidirectional encoder representations from transformers, BERT) model, BERT model are Bi-directional language model.

With reference to first aspect, in alternatively possible design method, Bot platform is using single-point sorting algorithm and based on general The dynamic programming algorithm of rate carries out one or more training corpus according to one or more training corpus and M the second slot positions Training, comprising: for each training corpus in one or more training corpus, Bot platform is advised using dynamic based on probability Cost-effective method scans one training corpus is cut into one or more candidate entities in different location from right to left, cut Normalization composite entity and normalization composite entity number afterwards；Bot platform uses single-point sorting algorithm, according to M the second slot positions, Obtain the confidence level of the corresponding candidate entity in each position；Bot platform is according to the corresponding normalization composite entity in each position The confidence level of several and candidate entity, determines the cutting mode to a training corpus.

With reference to first aspect, in alternatively possible design method, N number of entity type is recombinated in Bot platform, obtains M After a second slot position, the present processes can also include: that Bot platform shows that second contact surface, the second contact surface include starting Training button, this starts that button is trained to be trained said one or multiple training corpus for triggering Bot platform；In response to User is to the clicking operation for starting trained button, and Bot platform is according to one or more training corpus and M the second slot positions, to one A or multiple training corpus are trained, and Bot platform is made to have the ability for extracting the M in user's saying the second slot position.I.e. originally In application, Bot platform can be in response to user to the clicking operation for starting trained button, and triggering Bot platform is trained.

Second aspect, the application provide a kind of Bot platform, which may include: processor, memory and display Device.Memory, display and processor coupling；For storing computer program code, computer program code includes memory Computer instruction, when processor computer instructions, Bot platform execute: display, for show the first interface, first Slot position is arranged in the first intention that interface is used in the first technical ability for Bot platform；Processor, it is aobvious in display for receiving user First slot position of the first interface configurations shown；First slot position is combination slot position, and the first slot position includes N number of entity type, N >=2, N For positive integer；N number of entity type arranges in the first slot position according to sequence set by user；Processor is also used to recombinate N number of reality Body type obtains M the second slot positions；M the second slot positions include k entity type in N number of entity type in any order Arrange obtained slot position, k ∈ { 1,2 ... ..., N }；Processor is also used to according to one or more training corpus and M the second slots Position, is trained one or more training corpus, and Bot platform is made to have the energy for extracting the M in user's saying the second slot position Power.

In conjunction with second aspect, in a kind of possible design method,

In conjunction with second aspect, in alternatively possible design method, above-mentioned processor is also used to receive user and is showing The slot position title for the first interface configurations that device is shown, slot position title is for identifying the first slot position.

In conjunction with second aspect, in alternatively possible design method, above-mentioned processor is also used to recombinating N number of entity Type, after obtaining M the second slot positions, according to one or more training corpus and M the second slot positions, to one or more training Before corpus is trained, by slot position title and M the second slot positions associated storage in memory, and it is corresponding for slot position title The feature of M the second slot position label composite entities, after the feature of composite entity is used to indicate the M the second slot positions as recombination Composite entity type.

In conjunction with second aspect, in alternatively possible design method, above-mentioned processor, for being instructed according to one or more Practice corpus and M the second slot positions, one or more training corpus are trained, comprising: processor is specifically used for receiving user For the true tag of each of one or more training corpus word addition；It is one or more instructions according to M the second slot positions Practice each of corpus word and adds feature；Processor, is also used to using deep learning algorithm, according to true tag and feature, One or more training corpus are trained；Alternatively, using deep learning algorithm conjugation condition with field CRF algorithm, according to true Real label and feature are trained one or more training corpus.

In conjunction with second aspect, in alternatively possible design method, above-mentioned deep learning algorithm includes LSTM algorithm.

In conjunction with second aspect, in alternatively possible design method, above-mentioned processor, for recombinating N number of entity type, Obtain M the second slot positions, comprising: processor is specifically used for using dynamic programming algorithm, recombinates N number of entity type, obtains M Second slot position.

In conjunction with second aspect, in alternatively possible design method, above-mentioned processor, for being instructed according to one or more Practice corpus and M the second slot positions, one or more training corpus are trained, comprising: processor is specifically used for using single-point Sorting algorithm and dynamic programming algorithm based on probability, according to one or more training corpus and M the second slot positions, to one or Multiple training corpus are trained.

In conjunction with second aspect, in alternatively possible design method, above-mentioned single-point sorting algorithm includes at least SVM mould Any one of type, maximum entropy model, fasttext model, CNN model, n-gram model or RNN model.

In conjunction with second aspect, in alternatively possible design method, above-mentioned single-point sorting algorithm is BERT model, BERT Model is bi-directional language model.

In conjunction with second aspect, in alternatively possible design method, above-mentioned processor, for using SVM algorithm and base In the dynamic programming algorithm of probability, according to one or more training corpus and M the second slot positions, to one or more training corpus It is trained, comprising: processor, specifically for for each training corpus in one or more training corpus, using being based on The dynamic programming algorithm of probability scans one training corpus is cut into one or more candidates in fact in different location from right to left Body normalizes composite entity and normalization composite entity number after being cut；Using single-point sorting algorithm, according to M second Slot position obtains the confidence level of the corresponding candidate entity in each position；According to the corresponding normalization composite entity number in each position With the confidence level of candidate entity, the cutting mode to a training corpus is determined.

In conjunction with second aspect, in alternatively possible design method, aforementioned display device is also used to recombinate N in processor A entity type after obtaining M the second slot positions, shows that second contact surface, second contact surface include starting to train button, starts to train Button is trained said one or multiple training corpus for triggering Bot platform.Processor is also used in response to user couple What display was shown starts to train the clicking operation of button, according to one or more training corpus and M the second slot positions, to one Or multiple training corpus are trained, and Bot platform is made to have the ability for extracting the M in user's saying the second slot position.

The third aspect, the application provide a kind of computer storage medium, which includes computer instruction, When the computer instruction is run on Bot platform, so that the Bot platform executes such as first aspect and its any possibility Design method described in the configuration method of slot position is combined in interactive system.

Fourth aspect, the application provide a kind of computer program product, when the computer program product on computers When operation, so that the computer executes the interactive system as described in first aspect and its any possible design method The configuration method of middle combination slot position.

It is to be appreciated that Bot platform described in the second aspect of above-mentioned offer, the storage of computer described in the third aspect is situated between Matter, the attainable beneficial effect of the institute of computer program product described in fourth aspect, can refer to first aspect and its it is any can Beneficial effect in the design method of energy, details are not described herein again.

Detailed description of the invention

Fig. 1 is a kind of interactive voice flow example of task orientation；

Fig. 2 is the hardware structural diagram of a kind of electronic equipment provided by the embodiments of the present application；

Fig. 3 is a kind of technical ability configuration interface schematic diagram of Bot platform provided by the embodiments of the present application；

Fig. 4 A is the technical ability configuration interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Fig. 4 B is the technical ability configuration interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Fig. 5 is a kind of intention configuration interface schematic diagram of Bot platform provided by the embodiments of the present application；

Fig. 6 is the intention configuration interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Fig. 7 is the configuration method flow chart that slot position is combined in a kind of interactive system provided by the embodiments of the present application；

Fig. 8 A is a kind of system frame for realizing the configuration of combination slot position in interactive system provided by the embodiments of the present application Figure；

Fig. 8 B is another system for realizing the configuration of combination slot position in interactive system provided by the embodiments of the present application Block diagram；

Fig. 9 is a kind of model training interface schematic diagram of Bot platform provided by the embodiments of the present application；

Figure 10 is the model training interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Figure 11 is the model training interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Figure 12 is the model training interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Figure 13 is the model training interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Figure 14 is the model training interface schematic diagram of another kind Bot platform provided by the embodiments of the present application；

Figure 15 is a kind of structure composition schematic diagram of Bot platform provided by the embodiments of the present application.

Specific embodiment

The embodiment of the present application provides the configuration method that slot position is combined in a kind of interactive system, and this method can be applied to Bot platform.Specifically it can be applied in the slot position configuration process of Bot platform.

Wherein, " the combination slot position " in the embodiment of the present application refers to the slot position including at least two entity types.For example, language Expect that the slot position extracted in " I wants two pears of an apple " includes " mono-@apple of@" and " two@pears of@"." mono-@apple of@" and Include in " two@pears of@" entity type '@quantity ' and ' hamburger@type '.Therefore, " mono-@apple of@" and " two@pears of@" It is combination slot position.

Illustratively, the Bot platform in the embodiment of the present application can be to be integrated with chat robots (Chat Bot) session The electronic equipment of formula application program.Alternatively, Bot platform can be for being capable of the Chat Bot conversational application that provides of login service device The electronic equipment of the webpage of procedure service.For example, above-mentioned Bot platform can be the personal computer for having above-mentioned function (personal computer, PC), laptop, portable computer (such as mobile phone), wearable electronic are (such as intelligence Wrist-watch), tablet computer, augmented reality (augmented reality, AR) virtual reality (virtual reality, VR) set The electronic equipments such as standby, vehicle-mounted computer, following embodiment do not do the concrete form of the Bot platform specifically limited.

Illustratively, in the embodiment of the present application by taking Bot platform is mobile phone as an example, the structural schematic diagram of Bot platform is shown.Such as Shown in Fig. 2, Bot platform 200 may include processor 210, external memory interface 220, internal storage 221, general serial Bus (universal serial bus, USB) interface 230, charge management module 240, power management module 241, battery 242, antenna 1, antenna 2, mobile communication module 250, wireless communication module 260, audio-frequency module 270, loudspeaker 270A, receiver 270B, microphone 270C, earphone interface 270D, sensor module 280, key 290, motor 291, indicator 292, camera 293, display screen 294 and Subscriber Identity Module (subscriber identification module, SIM) card interface 295 Deng.Wherein, sensor module 280 may include pressure sensor, gyro sensor, baroceptor, Magnetic Sensor, acceleration Spend sensor, range sensor, close to optical sensor, fingerprint sensor, temperature sensor, touch sensor, ambient light sensing Device, bone conduction sensor etc..

It is understood that the structure of the embodiment of the present application signal does not constitute the specific restriction to Bot platform 200.? In other embodiments, Bot platform 200 may include than illustrating more or fewer components, perhaps combine certain components or Split certain components or different component layouts.The component of diagram can be with hardware, and the combination of software or software and hardware is real It is existing.

Processor 210 may include one or more processing units, such as: processor 210 may include application processor (application processor, AP), modem processor, graphics processor graphics processing unit, GPU), image-signal processor (image signal processor, ISP), controller, memory, Video Codec, number Word signal processor (digital signal processor, DSP), baseband processor and/or neural network processor (neural-network processing unit, NPU) etc..Wherein, different processing units can be independent device, It can integrate in one or more processors.

Wherein, controller can be nerve center and the command centre of Bot platform 200.Controller can be grasped according to instruction Make code and clock signal, generates operating control signal, the control completing instruction fetch and executing instruction.

Memory can also be set in processor 210, for storing instruction and data.In some embodiments, processor Memory in 210 is cache memory.The memory can save the instruction that processor 210 is just used or is recycled Or data.If processor 210 needs to reuse the instruction or data, can be called directly from the memory.It avoids Repeated access, reduces the waiting time of processor 210, thus improves the efficiency of system.

In some embodiments, processor 210 may include one or more interfaces.Interface may include integrated circuit (inter-integrated circuit, I2C) interface, integrated circuit built-in audio (inter-integrated circuit Sound, I2S) interface, pulse code modulation (pulse code modulation, PCM) interface, universal asynchronous receiving-transmitting transmitter (universal asynchronous receiver/transmitter, UART) interface, mobile industry processor interface (mobile industry processor interface, MIPI), universal input export (general-purpose Input/output, GPIO) interface, Subscriber Identity Module (subscriber identity module, SIM) interface, and/or Universal serial bus (universal serial bus, USB) interface etc..

It is understood that the interface connection relationship of each intermodule of signal of the embodiment of the present invention, only schematically illustrates, The structure qualification to Bot platform 200 is not constituted.In other embodiments of the application, Bot platform 200 can also use upper State the combination of interface connection type or multiple interfaces connection type different in embodiment.

Charge management module 240 is used to receive charging input from charger.Wherein, charger can be wireless charger, It is also possible to wired charger.In the embodiment of some wired chargings, charge management module 240 can pass through USB interface 230 Receive the charging input of wired charger.In the embodiment of some wireless chargings, charge management module 240 can pass through Bot The Wireless charging coil of platform 200 receives wireless charging input.While charge management module 240 is that battery 242 charges, may be used also To be power electronic equipment by power management module 241.

Power management module 241 is for connecting battery 242, charge management module 240 and processor 210.Power management mould Block 241 receives the input of battery 242 and/or charge management module 240, is processor 210, internal storage 221, external storage Device, display screen 294, the power supply such as camera 293 and wireless communication module 260.

The wireless communication function of Bot platform 200 can pass through antenna 1, antenna 2, mobile communication module 250, wireless communication Module 260, modem processor and baseband processor etc. are realized.Electromagnetic wave is believed for transmitting and receiving for antenna 1 and antenna 2 Number.Each antenna in Bot platform 200 can be used for covering single or multiple communication bands.Different antennas can also be multiplexed, with Improve the utilization rate of antenna.Such as: antenna 1 can be multiplexed with to the diversity antenna of WLAN.In other embodiment In, antenna can be used in combination with tuning switch.

Mobile communication module 250, which can provide, applies wirelessly communicating on Bot platform 200 including 2G/3G/4G/5G etc. Solution.Mobile communication module 250 may include at least one filter, switch, power amplifier, low-noise amplifier (low noise amplifier, LNA) etc..Mobile communication module 250 can receive electromagnetic wave by antenna 1, and to received electricity Magnetic wave is filtered, and the processing such as amplification is sent to modem processor and is demodulated.Mobile communication module 250 can also be right The modulated modulated signal amplification of demodulation processor, switchs to electromagenetic wave radiation through antenna 1 and goes out.In some embodiments, it moves At least partly functional module of dynamic communication module 250 can be arranged in processor 210.In some embodiments, mobile logical At least partly functional module of letter module 250 can be arranged in the same device at least partly module of processor 210.

Modem processor may include modulator and demodulator.In some embodiments, modem processor can To be independent device.In further embodiments, modem processor can be independently of processor 210, with mobile communication Module 250 or other function module are arranged in the same device.

It includes WLAN (wireless that wireless communication module 260, which can be provided and be applied on Bot platform 200, Local area networks, WLAN) (such as Wireless Fidelity (wireless fidelity, Wi-Fi) network), bluetooth (bluetooth, BT), Global Navigation Satellite System (global navigation satellitesystem, GNSS), frequency modulation (frequency modulation, FM), the short distance wireless communication technology (near field communication, NFC) are red The solution of the wireless communications such as outer technology (infrared, IR).Wireless communication module 260 can be integrated into few communication One or more devices of processing module.Wireless communication module 260 receives electromagnetic wave via antenna 2, by electromagnetic wave signal frequency modulation And filtering processing, by treated, signal is sent to processor 210.Wireless communication module 260 can also connect from processor 210 Signal to be sent is received, frequency modulation is carried out to it, is amplified, is switched to electromagenetic wave radiation through antenna 2 and go out.

In some embodiments, the antenna 1 of Bot platform 200 and mobile communication module 250 couple, antenna 2 and wireless communication Module 260 couples, and allowing Bot platform 200, technology is communicated with network and other equipment by wireless communication.It is described wireless The communication technology may include global system for mobile communications (global system for mobile communications, GSM), general packet radio service (general packet radio service, GPRS), CDMA access (code Division multiple access, CDMA), wideband code division multiple access (wideband code division multiple Access, WCDMA), time division CDMA (time-division code division multiple access, TD- SCDMA), long term evolution (long term evolution, LTE), BT, GNSS, WLAN, NFC, FM and/or IR technology etc..Institute Stating GNSS may include GPS (global positioning system, GPS), global navigational satellite system It unites (global navigation satellite system, GLONASS), Beidou satellite navigation system (beidou Navigation satellite system, BDS), quasi- zenith satellite system (quasi-zenith satellite System, QZSS) and/or satellite-based augmentation system (satellite based augmentation systems, SBAS).

Bot platform 200 realizes display function by GPU, display screen 294 and application processor etc..GPU is at image The microprocessor of reason connects display screen 294 and application processor.GPU is calculated for executing mathematics and geometry, is used for figure wash with watercolours Dye.Processor 210 may include one or more GPU, execute program instructions to generate or change display information.

Display screen 294 is for showing image, video etc..Display screen 294 includes display panel.Display panel can use liquid Crystal display screen (liquid crystal display, LCD), Organic Light Emitting Diode (organic light-emitting Diode, OLED), active matrix organic light-emitting diode or active-matrix organic light emitting diode (active-matrix Organic light emitting diode's, AMOLED), Flexible light-emitting diodes (flex light-emitting Diode, FLED), Miniled, MicroLed, Micro-oLed, light emitting diode with quantum dots (quantum dot light Emitting diodes, QLED) etc..In some embodiments, Bot platform 200 may include 1 or N number of display screen 294, N For the positive integer greater than 1.

Bot platform 200 can be by ISP, camera 293, Video Codec, GPU, display screen 294 and at It manages device etc. and realizes shooting function.

ISP is used to handle the data of the feedback of camera 293.For example, opening shutter when taking pictures, light is passed by camera lens It is delivered on camera photosensitive element, optical signal is converted to electric signal, and camera photosensitive element passes to the electric signal at ISP Reason, is converted into macroscopic image.ISP can also be to the noise of image, brightness, colour of skin progress algorithm optimization.ISP can be with Exposure to photographed scene, the parameter optimizations such as colour temperature.In some embodiments, ISP can be set in camera 293.

Camera 293 is for capturing still image or video.Object generates optical imagery by camera lens and projects photosensitive member Part.Photosensitive element can be charge-coupled device (charge coupled device, CCD) or complementary metal oxide is partly led Body (complementary metal-oxide-semiconductor, CMOS) phototransistor.Photosensitive element turns optical signal It changes electric signal into, electric signal is passed into ISP later and is converted into data image signal.Data image signal is output to DSP by ISP Working process.Data image signal is converted into the RGB of standard, the picture signal of the formats such as YUV by DSP.In some embodiments, Bot platform 200 may include 1 or N number of camera 293, and N is the positive integer greater than 1.

NPU is neural network (Neural-Network, NN) computation processor, by using for reference biological neural network structure, Such as transfer mode between human brain neuron is used for reference, it, can also continuous self study to input information fast processing.Pass through NPU The application such as intelligent cognition of Bot platform 200 may be implemented, such as: image recognition, recognition of face, speech recognition, text understanding Deng.For example, NPU can run the AI model in the embodiment of the present application, above-mentioned image recognition is executed, recognition of face, speech recognition, The business such as text understanding.

External memory interface 220 can be used for connecting external memory card, such as MicroSD card, realize extension Bot platform 200 storage capacity.External memory card is communicated by external memory interface 220 with processor 210, realizes that data store function Energy.Such as by music, the files such as video are stored in external memory card.

Internal storage 221 can be used for storing computer executable program code, and the executable program code includes Instruction.Processor 210 is stored in the instruction of internal storage 221 by operation, thereby executing the various functions of Bot platform 200 Using and data processing.Internal storage 221 may include storing program area and storage data area.Wherein, storing program area Can storage program area, application program (such as sound-playing function, image player function etc.) needed at least one function etc.. Storage data area can store the data (such as audio data, phone directory etc.) etc. created in 200 use process of Bot platform.This Outside, internal storage 221 may include high-speed random access memory, can also include nonvolatile memory, for example, at least One disk memory, flush memory device, generic flash memory (universal flash storage, UFS) etc..For example, Memory (such as internal storage 221) can be used to save the model code of AI model.

Bot platform 200 can be by audio-frequency module 270, loudspeaker 270A, receiver 270B, microphone 270C, and earphone connects Mouth 270D and application processor etc. realize audio-frequency function.Such as music, recording etc..Audio-frequency module 270 is used for will be digital Audio-frequency information is converted into analog audio signal output, is also used for analog audio input being converted to digital audio and video signals.Audio mould Block 270 can be also used for audio-frequency signal coding and decoding.In some embodiments, audio-frequency module 270 can be set in processing In device 210, or the partial function module of audio-frequency module 270 is set in processor 210.

Key 290 includes power button, volume key etc..Key 290 can be mechanical key.It is also possible to touch-key. Bot platform 200 can receive key-press input, generate key signals related with the user setting of Bot platform 200 and function control Input.

Motor 291 can produce vibration prompt.Motor 291 can be used for calling vibration prompt, can be used for touching vibration Dynamic feedback.Indicator 292 can be indicator light, can serve to indicate that charged state, electric quantity change can be used for instruction and disappear Breath, missed call, notice etc..

SIM card interface 295 is for connecting SIM card.SIM card can be by being inserted into SIM card interface 295, or from SIM card interface 295 extract, and realization is contacting and separating with Bot platform 200.Bot platform 200 can support 1 or N number of SIM card interface, N are Positive integer greater than 1.SIM card interface 295 can support Nano SIM card, Micro SIM card, SIM card etc..The same SIM card Interface 295 can be inserted into multiple cards simultaneously.The type of multiple cards may be the same or different.SIM card interface 295 Different types of SIM card can be compatible with.SIM card interface 295 can also be with compatible external storage card.Bot platform 200 passes through SIM card And network interaction, realize the functions such as call and data communication.In some embodiments, Bot platform 200 uses eSIM, it may be assumed that embedding Enter formula SIM card.ESIM card can cannot separate in Bot platform 200 with Bot platform 200.

Wherein, the Bot platform in the embodiment of the present application is the Bot platform of task orientation chat product.User can be in the Bot In platform, it is Bot platform configuration one or more technical ability (Skill), is intended to (Intent) and slot position (Slot), for realizing people Machine interaction.The process that the embodiment of the present application creates technical ability, intention and slot position here in conjunction with attached drawing to user in Bot platform carries out Explanation.

Referring to FIG. 3, its technical ability configuration interface 301 for showing a kind of Bot platform provided by the embodiments of the present application.Such as Fig. 3 It is shown, technical ability configuration interface 301 may include: " basic configuration " item, " skill designations " setting frame, " technical ability classification " be arranged frame, " icon " setting option, " technical ability request word " setting frame, " technical ability welcome words " setting frame and " technical ability description " setting frame etc..

Wherein, " skill designations " setting frame is used to input the skill designations that user wants creation.For example, being somebody's turn to do " skill designations " Setting frame can input the skill designations such as " ordering " or " paying one's fare "." technical ability classification " setting frame is for being arranged above-mentioned " skill designations " The type that the skill designations inputted in frame correspond to technical ability is set.For example, above-mentioned skill designations " ordering " or " paying one's fare " etc. are corresponding Technical ability belongs to tool assistant shown in Fig. 3." icon " setting option is used to that the technical ability name inputted in frame to be arranged for " skill designations " Claim corresponding technical ability that icon is set." technical ability request word " setting frame starts asking for above-mentioned technical ability for controlling Bot platform for inputting Seek word.For example, as shown in figure 3, " technical ability request word " setting frame can input " Xiao Ming ".Language of the Bot land identification to user Sound data or when text data " Xiao Ming ", can start above-mentioned " ordering " technical ability.Frame is arranged for defeated in " technical ability welcome words " Enter the welcome words of above-mentioned technical ability, which is used for when system can not identify user's saying, the content is broadcasted automatically, to user Energy conservation is introduced, and guide user how to express to be identified." technical ability description " setting frame is used to input the introduction to above-mentioned technical ability, The problem of can solve such as above-mentioned energy conservation can be the service etc. that user provides.

Bot platform response is in user to the clicking operation of " preservation " button in " basic configuration " item shown in Fig. 3, Ke Yixian Be intended to configuration interface 401 shown in diagram 4A, guidance user be the technical ability that configures in above-mentioned technical ability configuration interface 301 of user (such as " ordering ") the one or more intentions of configuration.

As shown in Figure 4 A, it is intended that configuration interface 401 include: " be intended to title " setting frame, " being intended to Chinese name " setting frame, " more wheel dialogues " setting frame and " user's saying " setting frame etc..

Wherein, " be intended to title " setting frame be used to input user want creation intention title (such as English name, or Title including letter and/or data)." buyHum " is inputted for example, " being intended to title " shown in Fig. 4 A and being arranged in frame." it is intended to Chinese name " setting frame is used to input the Chinese for the intention that user wants creation.For example, " being intended to Chinese name " shown in Fig. 4 A It is arranged in frame and inputs " buying hamburger ".The side for more wheels dialogue that " more wheel dialogues " setting frame is used to be arranged between Bot platform and user Formula." user's saying " setting frame is for inputting user for above-mentioned intention " buying hamburger " possible saying (i.e. user's saying).Example Such as, user can " user's saying " be arranged frame input " I will buy a wheat chicken with several spices leg fort ", " giving me three cod-fish Hamburgs " or Users' sayings such as " three, wheat chicken with several spices leg forts ".Wherein, the user's saying inputted in " user's saying " setting frame is referred to as instructing Practice corpus.The training corpus is used for one or more slot positions (including combination slot position) that user is the intention " buying hamburger " setting It is trained, such as carries out artificial intelligence (Artificial Intelligence, AI) model training.

Wherein, as shown in Figure 4 A, it is intended that configuration interface 401 can also include " being intended to save " item 403 and " configuration selection Column " 402.It is inputted in response to user in " being intended to title " setting frame and is intended to title, Bot platform can be in " being intended to save " item 403 Show the intention title of user's input.For example, " buyHum " is inputted in " being intended to title " setting frame in response to user, such as Fig. 4 A institute Show, Bot platform shows " buyHum " in " being intended to save " item 403." preservation " in " being intended to save " item 403 is pressed in response to user The clicking operation of button, Bot platform can save user and be intended to the intention and relevant parameter that configuration interface 401 is arranged.

" configuration selectionbar " 402 may include " configurations " options, " intention " options and " type of slots " selection Multiple options such as item and " training " options.After user is intended to the creation intention of configuration interface 401 shown in Fig. 4 A, Ke Yidian Hit " type of slots " options in " configuration selectionbar " 402.In response to user to " the slot position class in " configuration selectionbar " 402 The clicking operation of type " options, Bot platform can show slot position configuration interface 404 shown in Fig. 4 B.Slot position configuration interface 404 In include " type of slots " setting option 405.It include " system type of slots " option in " type of slots " setting option 405 and " newly-built Type of slots " option.In response to user to the clicking operation of " system type of slots " option, Bot platform can determine that user thinks The type of slots to be configured is system type of slots, i.e. preconfigured type of slots in Bot platform.In response to user to " new Build type of slots " clicking operation of option, the type of slots that Bot platform can determine that user wants configuration is newly-built slot position class The customized type of slots of type, i.e. user.As shown in Figure 4 B, " newly-built type of slots " choosing is selected with user in the embodiment of the present application For.In response to user to the clicking operation of " newly-built type of slots " option, Bot platform can show slot position shown in fig. 5 Configuration interface 501, guidance user are intention (such as " buying hamburger ") configuration one that user configures in above-mentioned intention configuration interface 401 Or multiple slot positions.

As shown in figure 5, slot position configuration interface 501 may include: that frame 503 is arranged in " slot position title ", " slot position Chinese name " is set Set frame 504 and " entity type " setting frame 505 etc..

Wherein, " slot position title " setting frame is used to input the name that user is intended for the slot position of above-mentioned intention " buying hamburger " configuration Claim (such as English name, or the title including letter and/or data).For example, as shown in figure 5, in " slot position title " setting frame It inputs " CompositeNE "." slot position Chinese name " setting frame is used to input the Chinese for the slot position that user wants configuration.Example Such as, as shown in figure 5, inputting Chinese " combination slot position 1 " in " slot position Chinese name " setting frame.

It is appreciated that slot position (including combination slot position) can be made of one or more entity types.Above-mentioned " entity class Type " setting frame 505 is used to configure the entity type in above-mentioned slot position " combination slot position 1 ".In the embodiment of the present application, user can be The entity type configured in " entity type " setting frame 505 may include: that system entity type (is pre-configured in Bot platform In entity type) and newly-built entity type (i.e. user's custom entities type).

As shown in figure 5, including " system entity type " button 506, " newly-built entity class in " entity type " setting frame 505 Type " button 507, " example types " input frame 508 and " newly-increased entity type " button 509." example types " input frame 508 is used for Input user is intended for one or more entity types of above-mentioned slot position " combination slot position 1 " configuration.In response to user to " system is real The clicking operation of body type " button 506, Bot platform can determine that user wants the reality configured in " example types " input frame 508 Body type is system entity type.In response to user to the clicking operation of " newly-built entity type " button 507, Bot platform can be with Determine that the entity type that user wants to configure in " example types " input frame 508 is newly-built entity type.As shown in figure 5, this Shen It please be in embodiment by taking user selectes " newly-built entity type " button 507 as an example." newly-increased entity type " button 509 is for triggering Bot platform shows " example types " input frame 508 for configuring new entity type.

As shown in figure 5, slot position configuration interface 501 can also include " slot position preservation " item 502.In response to user in " slot position Title " is arranged frame and inputs slot position title, and Bot platform can show the slot position title that user inputs in " slot position preservation " item 502.Example Such as, " CompositeNE " is inputted in " slot position title " setting frame in response to user, as shown in fig. 6, Bot platform is in " slot position guarantor Deposit " display of item 502 " CompositeNE ".

As shown in fig. 6, user has input slot position Chinese name " combination slot position 1 ", user in " slot position Chinese name " setting frame 504 Two example types '@quantity ' and ' hamburger@type ' are configured in " example types " input frame 508.In general, if user " preservation " button in " slot position preservation " item 502 shown in fig. 6 is clicked, the click of " preservation " button is grasped in response to user Make, combination slot position 1 " quantity hamburger type " can be generated in Bot platform.The sequence of entity type in the combination slot position are as follows: ' number Amount ' preceding, ' hamburger@type ' is rear.In this way, Bot platform can only then identify and response text information " I wants a wheat chicken with several spices leg Fort ", " giving me two parts of delicious and crisp cod-fish Hamburgs " etc., '@quantity ' are in preceding, ' hamburger@type ' posterior expression.And when user's saying is " I wants two parts of delicious and crisp cod-fish Hamburg ", " giving me three parts of wheat chicken with several spices leg fort " etc., '@quantity ' is in rear, ' hamburger@type ' preceding expression When, Bot platform cannot then extract corresponding slot position.

If realizing that user's saying be " I wants two parts of delicious and crisp cod-fish Hamburg ", " give my three parts of wheat chicken with several spices leg fort " etc., '@ Quantity ' rear, the slot position of ' hamburger type ' preceding expression is extracted, then user then need to be reconfigured a combination slot position " Hamburger type@quantity ".User's saying that i.e. user is directed to same meaning needs to configure the combination slot position of multiple specified sequences, uses Family is cumbersome.

To solve the above-mentioned problems, the embodiment of the present application provides the configuration side that slot position is combined in a kind of interactive system Method, this method can be applied to robot Bot platform.In this method, Bot platform can receive the combination slot position of user configuration, The combination slot position may include multiple entity types.Although this multiple entity type is in the combination slot position according to set by user Sequence arranges；But in the embodiment of the present application, Bot platform can recombinate multiple entity type, obtain multiple combination slot positions. Multiple combination slot position includes that one or more entity types in above-mentioned multiple entity types arrange obtain in any order Slot position.Finally, Bot platform can according to the one or more training corpus and the obtained multiple slot positions of recombination that user inputs into Row model training.

It include one or more entity types in above-mentioned multiple entity types due to recombinating obtained multiple combination slot positions The slot position arranged in any order, i.e. Bot platform can recombinate to obtain multiple entity type arranges in any order Slot position after column；Therefore, even if changing the sequence of multiple entity type in user's saying, Bot platform can also be extracted pair The slot position answered.

Also, in the embodiment of the present application, it is only necessary to which it includes that user configuration combines in slot position for which entity type, is not required to Want the sequence of entity type in user's given combination slot position.

In order to make it easy to understand, below in conjunction with attached drawing to combination slot position in interactive system provided by the embodiments of the present application Configuration method describes in detail.

The embodiment of the present application provides the configuration method that slot position is combined in a kind of interactive system, as shown in fig. 7, this method May include S701-S703:

S701, Bot platform receive user in the first slot position of the first interface configurations.First slot position is combination slot position, should First slot position includes N number of entity type, and N >=2, N are positive integer.

Wherein, which arranges in the first slot position according to sequence set by user.For example, the first slot position can Think said combination slot position 1 " hamburger@quantity@type ".It include two example types '@quantity ' and ' hamburger@in first slot position Type ', i.e. N=2.The sequence of entity type is set by the user in first slot position, specific order are as follows: '@quantity ' is in preceding, ' the@Chinese Fort type ' rear.

Slot position is arranged in the first intention that above-mentioned first interface is used in the first technical ability for Bot platform.For example, the first interface It can be Fig. 5 or slot position configuration interface shown in fig. 6.Correspondingly, the first technical ability can be above-mentioned technical ability " ordering ", first intention Can be above-mentioned intention " buying hamburger ".

S702, Bot platform recombinate N number of entity type, obtain M the second slot positions.The M the second slot positions include N number of entity The slot position that k entity type in type arranges in any order, k ∈ { 1,2 ... ..., N }.

For example, after Bot platform receives combination slot position 1 " quantity hamburger type ", this can be recombinated in conjunction with examples detailed above Two example types '@quantity ' and ' hamburger@type ', obtain four slot positions: "@quantity ", " hamburger@type ", " hamburger@quantity@ Type " and " hamburger@type@quantity ".In this way, even if user's saying is " I wants two parts of delicious and crisp cod-fish Hamburg ", " gives me wheat chicken with several spices leg Three parts of fort " etc., '@quantity ' are rear, and ' when type ' the preceding expression of the hamburger@, Bot platform can also extract corresponding slot position.

It should be noted that the@in the embodiment of the present application is used to identify an entity type, without physical meaning.This Shen ' ' and " " difference entity type and slot position please be used in embodiment.For example, '@quantity ' presentation-entity type, and "@quantity " indicates Slot position.

In the embodiment of the present application, M the second slot positions include k entity type in N number of entity type in any order Arrange obtained slot position, k ∈ { 1,2 ... ..., N }.That is, k can be any positive integer in [1, N].

When k=N-1, Bot platform can select N-1 (i.e. k) entity types, the choosing of Bot platform from N number of entity type The N-1 entity type selected is combined into second slot position (i.e. combination slot position).Wherein, Bot platform can have N × (N-1) × (N-2) × ... × 2 kinds of selections, it can obtain A second Slot position.

By foregoing description it follows that

For example, with N=3, i.e., include in above-mentioned first slot position 3 entity types (such as entity type ' A ', ' B ' and '@C ') for.Bot platform recombinates this 3 entity types, a second slot position of available following 15 (i.e. M=15).Specifically, k When=1, available 3 the second slot positions of Bot platform: "@A ", "@B " and "@C ".When k=2, available 6, Bot platform Two slot positions: "@A@B " and "@A@C "；"@B@A " and "@B@C "；And "@C@A " and "@C@B ".When k=3, Bot platform can be obtained To 6 the second slot positions: "@A@B@C " and "@A@C@B "；"@B@A@C " and "@B@C@A "；And "@C@A@B " and "@C@B@A ".Its In, M=3+6+6=15.

It should be noted that user can not also set the sequence of entity type in the first slot position.In this case, it uses It includes in the first slot position that family can be only arranged for which entity type, suitable in the first slot position without setting these entity types Sequence.That is, in the embodiment of the present application, no matter whether user sets the sequence of entity type in the first slot position, and Bot platform is all N number of entity type can be recombinated, M the second slot positions are obtained.

In some embodiments, Bot platform can use dynamic programming algorithm, recombinate above-mentioned N number of entity type, obtain State M the second slot positions.

S703, Bot platform are according to one or more training corpus and M the second slot positions, to one or more training corpus It is trained, Bot platform is made to have the ability for extracting the M in user's saying the second slot position.

Wherein, Bot platform can show second contact surface.The second contact surface include start train button, this start training by Button is trained for triggering Bot platform.Start the clicking operation of trained button to this in response to user, Bot platform can root According to the one or more training corpus and M the second slot positions of user's input, one or more training corpus are trained, are made Bot platform has the ability for extracting the M in user's saying the second slot position.

For example, in response to user's clicking operation shown in fig. 6 to " training " options in " configuration selectionbar " 601, Bot platform can show second contact surface 901 shown in Fig. 9, and guidance user triggers Bot platform and is trained.Second contact surface 901 wraps It includes and starts to train button 902.Start the clicking operation of trained button 902 to this in response to user, Bot platform can execute S703 It is trained.

In some embodiments, it is the configuration of above-mentioned first slot position that Bot platform, which can also receive user at above-mentioned first interface, Slot position title.The slot position title is for identifying above-mentioned first slot position.For example, to can receive user shown in Fig. 6 for Bot platform In the first interface " CompositeNE " and user that " slot position title " setting frame in first interface inputs shown in Fig. 6 " slot position Chinese name " setting frame " combination slot position 1 " that inputs.I.e. above-mentioned slot position title may include " CompositeNE " and " combination slot position 1 ".

In some embodiments, Bot platform can also be by above-mentioned slot position title and M the second slot position associated storages, and is The feature of corresponding M the second slot position label composite entity (Composite Entity) of slot position title.The composite entity It is normalized composite entity (Composite that the feature of (Composite Entity), which is used to indicate the M the second slot positions, Entity) type.

In conjunction with examples detailed above, Bot platform recombinant C ompositeNE " hamburger@quantity@type ", available four second Slot position: "@quantity ", " hamburger@type ", " hamburger@quantity@type " and " hamburger@type@quantity ".Bot platform can be by slot position Title (such as combination slot position 1) and this four the second slot position associated storages, are then combination slot position 1 (i.e. slot position title) corresponding 4 The feature of a second slot position label Composite Entity.

In some embodiments, Bot platform can receive the one or more and above-mentioned M the second slot position phases of user's input The training corpus of pass.For example, Bot platform can receive user be intended to shown in Fig. 4 A in configuration interface 401 " user says One or more training corpus that frame inputs are arranged in method ".The one or more training corpus is user for above-mentioned first intention (such as buying hamburger) possible user's saying.For example, being directed to first intention " buying hamburger ", possible user's saying may include: " I Buy a wheat chicken with several spices leg fort ", " giving me three cod-fish Hamburgs ", " I wants four cod-fish Hamburgs " or " three, wheat chicken with several spices leg fort " etc.. Then, Bot platform can receive the true tag that user is the word addition of each of the training corpus, and according to M each second Slot position is that each word in the training corpus adds feature.The true tag of one word is used to indicate slot position corresponding to this word Position in corresponding slot position of slot position title and this word.The feature of one word is used to indicate slot position corresponding to this word Position in corresponding slot position of composite entity feature and this word.Wherein, Bot platform can be according to above-mentioned M the second slots It such as " quantity ", " fruit type ", " quantity fruit type " and " fruit type quantity " determines and training corpus position Each of the matched feature of word, then be each word add feature.Finally, Bot platform can be the instruction according to user Practice the M slot position configured in the feature and Bot platform that true tag, the Bot platform that corpus adds are training corpus addition, One or more training corpus are trained.In one implementation, Bot platform can use deep learning algorithm, root According to the true tag of above-mentioned addition and the feature of addition, one or more training corpus are trained.

In another implementation, Bot platform can use deep learning algorithm combination CRF algorithm, be added according to above-mentioned The feature of the true tag and addition that add is trained one or more training corpus.Specifically, Bot platform can compare Bot platform is that the feature of each word addition and user are each word in training corpus (such as " I wants two pears of an apple ") The true tag of addition, study to the ability that feature corresponding with the true tag that user adds is added for the training corpus.Such as This can be extracted correctly after Bot platform receives corresponding user's saying (such as " I wants two pears of an apple ") again Slot position.

Optionally, said one or multiple training corpus can be input by user.Alternatively, said one or multiple instructions Practice training corpus and Bot platform that corpus may include: user's input and extensive obtain is carried out to the training corpus that user inputs One or more training corpus.For example, it is assumed that Bot platform receives the training corpus that user inputs, " it is peppery that I will buy a wheat Chicken leg fort ", Bot platform can be extensive to training corpus progress, obtain one or more similar training corpus.For example, Bot Platform can be extensive to training corpus " I will buy a wheat chicken with several spices leg fort " progress, and obtaining training corpus, " I will buy three delicious and crisps Cod-fish Hamburg " and " I will buy five wheat chicken with several spices leg forts " etc..

Illustratively, by taking above-mentioned first slot position is " quantity fruit type " as an example.So, above-mentioned M the second slot positions can To include: " quantity ", " fruit type ", " quantity fruit type " and " fruit type quantity "."@quantity ", "@fruit Type ", " quantity fruit type " and " fruit type quantity " can be normalized to " Composite Entity " entity class Type, i.e. the composite entity feature of this four the second slot positions are Composite Entity.

By taking training corpus is " I wants two pears of an apple " as an example, user can be each of training corpus word True tag is added, Bot platform can add feature for each of training corpus word.In the embodiment of the present application, Yong Huwei Each of training corpus word adds true tag, when Bot platform is that each of training corpus word adds feature, " O " Indicating corresponding word not is the corresponding word of the second slot position, and " B " indicates that corresponding word is the first character of the second slot position, and " I " is indicated Corresponding word is other words (such as second word or third word) in the second slot position in addition to first character.

Specifically, if Bot platform extracts correct slot position from " I wants two pears of an apple ", the slot extracted Position is " apple " and " two pears ", i.e., " quantity fruit type ".Based on this, user can according to combination slot position " Quantity fruit type " is that " I wants two pears of an apple " adds in Figure 10, Figure 11, Figure 12 or Figure 13 shown in either figure True tag.For example, as shown in either figure in Figure 10, Figure 11, Figure 12 or Figure 13, user can for " I " word, " wanting " word and " buying " word adds true tag " O "；True tag " B-composite Slot " is added for " one " word；It is true for the addition of " a " word Label " I-composite Slot "；True tag " I-composite Slot " is added for " apple " word；It is true for the addition of " fruit " word Real label " I-composite Slot "；True tag " B-composite Slot " is added for " two " word；For the addition of " a " word True tag " I-composite Slot "；True tag " I-composite Slot " is added for " pears " word.

Wherein, it may include multiple same slot positions (can be combination slot position) in some training corpus or user's saying. It in this case, include List slot position in the training corpus or user's saying.For example, corpus " I wants two pears of an apple " The slot position of middle extraction includes " mono-@apple of@" and " two@pears of@"." mono-@apple of@" and " two@pears of@" corresponds to combination slot Position " quantity fruit type ".It include two slot position "@quantity@fruits i.e. in the corpus " I wants two pears of an apple " Type " then includes List slot position in the corpus " I wants two pears of an apple ".

Illustratively, Bot platform can add feature in different combinations for each of training corpus word.

In implementation (1), Bot platform can be each of training corpus word in a manner of " correct combination " Add feature.

Specifically, if Bot platform extracts correct slot position from " I wants two pears of an apple ", the slot extracted Position is " apple " and " two pears ", i.e., " quantity fruit type ".Based on this, Bot platform can be according to combination slot Position " quantity fruit type " is that " I wants two pears of an apple " adds feature a shown in Fig. 10.For example, as shown in Figure 10, Bot platform can add feature " B-composite Entity " for " one " word；Feature " I-composite is added for " a " word Entity"；Feature " I-composite Entity " is added for " apple " word；Feature " I-composite is added for " fruit " word Entity"；Feature " B-composite Entity " is added for " two " word；Feature " I-composite is added for " a " word Entity"；Feature " I-composite Entity " is added for " pears " word.

In implementation (2), Bot platform can be training corpus in a manner of " correct combination "+" fine granularity combination " Each of word add feature.

If Bot platform extracts correct slot position from " I wants two pears of an apple ", the slot position extracted is "@ One apple " and " two pears ", i.e. " quantity fruit type ".Based on this, Bot platform can be according to combination slot position " number Amount fruit type " is each word addition feature in " I wants two pears of an apple ".I.e. Bot platform can be with " correct group Conjunction " mode is that each word in " I wants two pears of an apple " adds feature.But the entity type in " mono-@apple of@" ' one ' and ' apple ' can also be extracted as a slot position, such as " one " and " apple ", in " two pears " Entity type ' two ' and ' pears ' can also be extracted as a slot position, such as " two " and " pears ".Namely It says, said combination slot position " mono-@apple of@" and " two@pears of@" can carry out fine-grained division.For example, as shown in figure 11, Bot platform can add feature a and feature b for " I wants two pears of an apple ".In conjunction with Figure 10, as shown in figure 11, Bot platform Feature " #B-composite Entity " can also be added for " one " word；Feature " #I-composite is added for " a " word Entity"；Feature " #B-composite Entity " is added for " apple " word；Feature " #I-composite is added for " fruit " word Entity"；Feature " #B-composite Entity " is added for " two " word；Feature " #I-composite is added for " a " word Entity"；Feature " #B-composite Entity " is added for " pears " word.

In implementation (3), Bot platform can be training corpus in a manner of " exhaustion combination "+" fine granularity combination " Each of word add feature.

Wherein, " the@apple@two " in training corpus " I wants two pears of an apple " is in above-mentioned M the second slot positions " fruit type quantity ".Therefore, Bot platform can be in conjunction with above-mentioned " correct combination " scheme, for " I wants apple two The corresponding feature of pears " addition combination slot position " apple two ".

As shown in figure 12, Bot platform can add feature a, feature b and feature c for " I wants two pears of an apple ".Knot Figure 11 is closed, as shown in figure 12, Bot platform can also add feature " #B-composite Entity " for " apple " word；For " fruit " word It adds feature " #I-composite Entity "；Feature " #B-composite Entity " is added for " two " word；For " a " word It adds feature " #I-composite Entity "；Feature " #B-composite Entity " is added for " pears " word.

In implementation (4), Bot platform can be each of training corpus word in a manner of " exhaustion combination " Add feature.For example, as shown in figure 13, Bot platform can add feature a and feature c for " I wants two pears of an apple ".Its In, the detailed content of feature a and feature c can be with reference to the descriptions in above-described embodiment, and it will not go into details here for the embodiment of the present application.

It should be noted that can make since Bot platform uses the mode of " exhaustion combination " to add feature for training corpus Bot platform study all combined situations of each word into training corpus；" fine granularity combination " will increase each word not in Figure 12 The uncertainty of necessity combination influences the accuracy for extracting result.And corresponding composition slot position " the@apple of feature c shown in Figure 13 Fruit@two " is the assemblage characteristic of ambiguity, and the learning effect of Bot platform can be enhanced, and promotes Bot platform and extracts the accurate of result Property.Therefore, Bot platform is that the word addition of each of training corpus is special by the way of " exhaustion combination " in implementation (4) Sign, the ambiguity enumerator after can reducing each word combination, can be improved the accuracy rate of slot position extraction.

In some scenes, some words in user's saying may have ambiguity.Specifically, some in user's saying Word can may belong to a variety of entity types.For example, " apple " in user's saying " I will buy apple " may be product name, It may be fruit type.Some entities i.e. in user's saying the case where there may be ambiguities itself.It is extracted to promote slot position Accuracy rate, when being trained, user can also add the training corpus including ambiguity entity in Bot platform, and to instruction The ambiguity entity practiced in corpus adds entity tag, and Bot platform is enable to learn the ability to identification ambiguity entity.

For example, training corpus can be " I will buy apple, I will buy an apple ".Bot platform not only can be for " I wants Buy apple, I will buy an apple " in each word addition feature, entity tag can also be added.Wherein, entity tag is used for Indicate that corresponding word is the word in product name.For example, as shown in figure 14, user can add entity tag " B- for " apple " word Product Name " adds entity tag " I-product Name " for " fruit " word.

It is appreciated that increasing the study to the training corpus for including entity ambiguity when carrying out model training, helping to mention When liter Bot platform extracts the slot position in the user's saying for including entity ambiguity, the accuracy of slot position extraction.

In another implementation, Bot platform can use single-point sorting algorithm (i.e. single-point disaggregated model) and be based on The dynamic programming algorithm of probability, according to one or more training corpus of user's input and M the second slot positions, to one or more Training corpus is trained, and Bot platform is made to have the ability for extracting the M in user's saying the second slot position.

Illustratively, Bot platform can use dynamic programming algorithm based on probability, be scanned from right to left in different location One training corpus is cut into one or more candidate entities, composite entity number is normalized after being cut.Then, Bot Platform uses single-point sorting algorithm, according to M the second slot positions, obtains the confidence level of the corresponding candidate entity in each position.Finally, Bot platform is determined according to the confidence level of each position corresponding normalization composite entity number and candidate entity to the training language The cutting mode of material.

In some embodiments, Bot platform can use the dynamic programming algorithm of probability, minimize to training corpus Cutting obtains the number of multiple composite entities and composite entity.In the embodiment of the present application, Bot platform carries out training corpus The purpose for minimizing switching is maximumlly to be combined the entity in training corpus.For example, with " two pears of an apple " For.It include four candidate entities: (one), (apple), (two) and (pears) in " two pears of an apple ".Bot platform can To cut to " two pears of an apple ", multiple composite entities are obtained, such as " (one) (apple) " " (two) (pears) ", or Person " (one) ", " (apple) " and " (two) (pears) " etc..Wherein, " (one) (apple) " " (two) (pears) " corresponding cutting In mode, " two pears of an apple " cutting is obtained into two composite entities, such as " (one) (apple) " and " (two) (pears) ", The number for normalizing composite entity is 2.It, will be " one in " (one) " " (apple) " " (two) (pears) " corresponding cutting mode Two pears of apple " cutting obtains a composite entity, and such as " (two) (pears) ", the number for normalizing composite entity is 1.Wherein, The number of the normalization composite entity of " (one) (apple) " " (two) (pears) " is greater than " (one) " " (apple) " " (two) The number of the normalization composite entity of (pears) "；Therefore, " (one) (apple) " " (two) (pears) " corresponding cutting mode is better than " (one) " " (apple) " " (two) (pears) " corresponding cutting mode.

Dynamic programming process in the embodiment of the present application: the composite entity number obtained after cutting is minimized.

If (Max comparison phase):

F (i+1) '=Max p (i-1) * F (i-1)+1, p (i-1) * F (i-1),

...

p(i-k)*F(i-k)+1,p(i-k)*F(i-k),

K=1,2,3 ... Ki }

If (F (i+1) assignment phase):

F (i+1)=F (i+1) '/p removes probability value when Max compares

Wherein, Ki is " the most major term length+1 " in the list of entities extended to the left of i-th of position.For example, for For " pears " word in " two pears of an apple ", when being designated as 0 under i starting, " 6 " are designated as under " pears " are corresponding.Bot platform calculates It when (6) F, needs from this position i=6, finds candidate entity respectively to the left, candidate entity herein is " pears " and " two Pears ".For " pears " and " two pears ", the two candidate entities, are also predefined whether candidate's entity corresponds to target slot position, because One entity type may correspond to multiple slot positions.For example sys.location may not only correspond to departure place but also corresponding destination.One As for, " normalization " composite entity type here equally will appear the situation, therefore need in dynamic programming process to time The certainty of entity itself is selected to account for.In the embodiment of the present application, candidate entity can be calculated using single-point sorting algorithm Confidence level p (i-1), the measure object of i-1 is F (i-1) corresponding candidate entity.

Wherein, F (i) is the normalization group after cutting, being formed by candidate entity " in conjunction with " until i-th of position Close the number of entity.It wherein, include at least two candidate entities in the normalization composite entity that " in conjunction with " is formed.For example, " one Two pears of apple ", according to F (i) meaning in the first Max, " (one) " " (apple) (two) " " (pears) " pass through candidate real The normalization composite entity number that body " in conjunction with " is formed is " 1 ", and the normalization composite entity that " in conjunction with " is formed is " (apple) (two It is a) "；" (one) (apple) " " (two) (pears) " are by the normalization composite entity number that candidate entity " in conjunction with " is formed " 2 ", normalization composite entity are " (one) (apple) " and " (two) (pears) ".

Wherein, the candidate entity of include in the normalization composite entity that " in conjunction with " is formed at least two can correspond to different Entity type.For example, by taking the normalization composite entity " (one) (apple) " that " in conjunction with " is formed as an example.The normalizing that " in conjunction with " is formed Change in composite entity " (one) (apple) " includes candidate entity (one) and candidate entity (apple).Wherein, candidate entity (one It is a) correspondent entity type ' quantity ', candidate entity (apple) correspondent entity type ' fruit type '.

For F (i) meaning in the first Max, when Bot platform can be according to the value maximum of F (i), corresponding cutting Mode cuts training corpus.Such as, " two pears of an apple " are cut according to " (one) (apple) " " (two) (pears) " Obtained normalization composite entity number is maximum (for " 2 "), and therefore, Bot platform can be according to " (one) (apple) " " (two) The corresponding cutting mode cutting " two pears of an apple " of (pears) ", obtains the normalization composite entity " (one) of " in conjunction with " formation (apple) " and " (two) (pears) ".It should be noted that in above-mentioned formula in Max mode as an example.

If the Max in formula is changed to Min, the number that F (i) is all composite entities after cutting can be defined. It wherein, may include one or more candidate entities in the composite entity.For example, " two pears of an apple ", according to second F (i) meaning in Min, " (one) " " (apple) (two) " " (pears) " pass through the composite entity that candidate entity " in conjunction with " is formed Number is " 3 ", and composite entity is " (one) ", " (apple) (two) " and " (pears) "；" (one) (apple) " " (two) It is " 2 " that (pears) ", which pass through the composite entity number that candidate entity " in conjunction with " is formed, and composite entity is " (one) (apple) " and " (two It is a) (pears) ".

For F (i) meaning in second of Min, when Bot platform can be according to the value minimum of F (i), corresponding cutting Mode cuts training corpus.Such as, " two pears of an apple " are cut according to " (one) (apple) " " (two) (pears) " Obtained normalization composite entity number is minimum (for " 2 "), and therefore, Bot platform can be according to " (one) (apple) " " (two) (pears) " corresponding cutting mode cutting " two pears of an apple ", obtain normalization composite entity " (one) (apple) " and " (two) (pears) ".

In dynamic programming process, for i-th of position, to the left for each candidate entity, there are two types of situations: situation (1) entity is cut, formula shows as+1；Situation (2) does not cut the entity, and formula shows as+0.In varied situations, The selection of next position when recursion is different to the left.For example, next position when recursion is candidate to the left in situation (1) One, the starting subscript left side position of entity.For example, in " two pears of an apple ", the 6th position (i.e. " pears ") recursion to the left When next position be candidate entity " pears " one, starting subscript left side position " a ", formula shows as+1.And situation (2) In, next position when recursion is one, target left side position under candidate entity end to the left.For example, " one apple two In pears ", the 3rd position (i.e. " fruit ") to the left recursion when next position be candidate entity " apple " the starting subscript left side one A position " a ", formula shows as+0.

F (i+1)=F (i+1) '/p purpose is that the p in counting meaning F (i+1)=F (i+1) '/p for restore F is Max { } When comparing, maximum value in { }, in this way in next round iterative process, F (i+1) can keep number meaning, the influence of probability Only work when Max { } compares.Wherein, (0)=0 F.

In the above probability Dynamic Programming frame, it is the uncertainty for measuring candidate entity that Probability p (i), which introduces mesh,.For i-th A position, prediction be candidate entity confidence level.It is not limited to the confidence level using the SVM model prediction entity, can be and appoint Meaning model, such as the single-points disaggregated model such as maximum entropy, fasttext, CNN, or can be the language models such as n-gram, RNN. Wherein, it is using the advantage of single-point disaggregated model, in Practical Project realization, predetermined speed is fast and only needs a small amount of sample Complete effectively training.

But single-point model infirmities are not accounting for relationship between sequences.To solve this problem, in combination with single-point model The few advantage of required training sample, the embodiment of the present application give using dynamic programming method based on probability predetermined speed fastly. Using dynamic programming method based on probability, the limitation (maximum journey that the prediction result of single-point model can be integrated in decoding Degree combines), realize the effect of sequence labelling model.In the frame, above-mentioned single-point disaggregated model can be extended to fusion Sentence-level Model of sequence contextual feature, such as the single-point disaggregated model based on BERT etc..Although it is special to merge Sentence-level sequence context The model (such as single-point disaggregated model based on BERT) of sign is also single-point prediction to be carried out to candidate entity, but the model is a kind of Bi-directional language model, while the confidence level of predicting candidate entity, it is contemplated that Sentence-level contextual information.

For example, for the embodiment of the present application calculates F (6) in the calculating " two pears of an apple " here, to above-mentioned calculating Process is illustrated:

Step 1: F (6)=Max { according to scanning to the left, generating candidate entity: " pears ", " two pears " } -> determine F (i-k) In k, F (6)=Max { F (a)+1, F (fruit)+1, F (a) }, wherein F is (a) not cut situation to candidate's entity. After " a " determination, due to including a word " pears ", k=1 after " a "；After " fruit " determines, due to including three after " fruit " A word " two pears ", therefore k=3.

Step 2: being substituted into according to determining k, F (6)=Max { F (a)+1, F (fruit)+1, F (a) } -> F (6)=Max { F (5)+1, F (3)+1, F (5) } -> calculating corresponding Probability p (" pears "), p (" two pears ") -> p (5), p (3), the p (i-k) refer to pair Candidate physical operation obtains i-th-k location of recursion to the left, the probability of corresponding candidate's entity.

Step 3: combining the probability value after calculating, Max, F (6)=Max { F (5)+1, F (3)+1, F (5) } are calculated.

In this way, iterative calculation, may finally obtain F (6) by F (5), F (4), F (3), F (2), F (1), according to Dynamic Programming It takes Max to operate corresponding operation, can reversely whether be cut and cutting position (i.e. cutting mode).

As shown in Figure 8 A, the method for the embodiment of the present application may include: " slot position creation and configuration " process and " model is instructed Practice " process.

Step 1 shown in the corresponding S701 and Fig. 8 B shown in Fig. 7 of " slot position creation with configure " process shown in Fig. 8 A: User interface configuration.

Wherein, as shown in Figure 8 B, in step 1, Bot platform can carry out user interface configuration (i.e. execution S701).For example, User can carry out user interface configuration in Fig. 5 or slot position configuration interface shown in fig. 6 (i.e. the first interface).Specifically, step 1 It may include: that user selects " combination slot position ", configures the entity type in " combination slot position ".For example, user can be shown in Fig. 5 " entity type " setting frame 505 configure multiple entity types, indicate user's selection " combination slot position ".User can be in Fig. 5 institute " entity type " setting frame 505 shown configures the entity type in combination slot position " combination slot position 1 ", ' quantity ' as shown in FIG. 6 ' hamburger@type '.

" model training " process shown in Fig. 8 A corresponds to step 2 shown in S702-S703 and Fig. 8 B shown in Fig. 7: mould Type training.

In step 2 shown in " model training " process or Fig. 8 B shown in Fig. 8 A, Bot platform can recombinate above-mentioned N number of reality Body type obtains above-mentioned M the second slot positions.Also, Bot platform can be using deep learning algorithm (for example, Fig. 8 A or Fig. 8 B institute The LSTM algorithm shown) CER algorithm is combined, or using SVM algorithm and Dynamic Programming based on probability shown in Fig. 8 A or Fig. 8 B Algorithm is trained one or more training corpus.

As shown in Figure 8 A, the method for the embodiment of the present application can also include: " model prediction " process." mould shown in Fig. 8 A Step 3 shown in type prediction " process corresponding diagram 8B: model prediction.

Wherein, after " model training " process, Bot platform can receive user's saying (such as voice of user's input Information or text information), and understand user's saying intention to be expressed；After intention is determined, Bot platform can be to the use Family saying carries out slot position extraction.It wherein, can be by the voice if user's saying that Bot platform receives is voice messaging Information is converted to text information, then carries out being intended to understanding and slot position extraction to text information.

Illustratively, after Bot platform completes AI model training, it can receive user's " user's saying " shown in Fig. 4 A and set Set the user's saying inputted in frame.Then, Bot platform can be arranged user's saying that frame inputs to " user's saying " and be intended to Understand and slot position is extracted.Determining for the slot position extracted is operated finally, Bot platform can receive user, i.e. Bot platform can be with The slot position that confirmation extracts is interacted with user.As shown in Figure 8 B, step 3 may include " to user's saying carry out slot position extraction " and " user's interaction confirmation ".

Wherein, after " model prediction " process, above-mentioned Bot platform can be used formally.Wherein, Bot platform is being just The method that formula uses is similar with the application method of Bot platform in " model prediction " process, and it will not go into details here for the embodiment of the present application. It should be noted that the user in above-mentioned " slot position creation and configuration " process, " model training " process and " model prediction " process It is properly termed as the first user；User in the above-mentioned formal use process of Bot platform is properly termed as second user.

In some embodiments, above-mentioned first user can be the developer or tester of Bot platform.Exploit person Member or tester can according to the demand of the owner of Bot platform, be Bot platform configuration " technical ability ", " intention " and " slot position ", And carry out model training.Wherein, the owner of above-mentioned Bot platform can be dining room, convenience store or the staff in market etc..

In further embodiments, above-mentioned first user can be above-mentioned Bot platform owner (such as dining room, convenience store or The staff etc. in person market).The staff in dining room, convenience store or market can be according to demand Bot platform configuration " technical ability ", " intention " and " slot position ", and carry out model training.

In some embodiments, above-mentioned second user can be the owner of Bot platform, such as dining room, convenience store or market Staff etc..In this embodiment, from the owner of Bot platform (such as dining room, convenience store or market staff) to Bot Platform inputs user's saying (such as voice messaging or text information).Bot platform can receive user's saying, and understand that user says Method intention to be expressed；After intention is determined, Bot platform can carry out slot position extraction to user's saying；Finally, Bot is flat Platform can make feedback according to the slot position extracted.

In further embodiments, above-mentioned second user can be the consumer in above-mentioned dining room, convenience store or market.? In the embodiment, user's saying (such as voice messaging or text information) is inputted from consumer to Bot platform.Bot platform can be with User's saying is received, and understands user's saying intention to be expressed；After intention is determined, Bot platform can be said the user Method carries out slot position extraction；Finally, Bot platform can make feedback according to the slot position extracted, realize that the self-service of consumer disappears Take.

In interactive system provided by the embodiments of the present application combine slot position configuration method in, due to recombination obtain it is more A second slot position includes the slot position that one or more entity types in multiple entity types arrange in any order, i.e., Bot platform can recombinate to obtain the slot position after multiple entity type arranges in any order；Therefore, even if in user's saying Changing only includes an entity type in the sequence or user's saying of multiple entity type, and Bot platform can also extract Corresponding slot position replys user's saying.

Also, for the complex information in user's saying, such as " two pears of an apple ", Bot platform still can be correct Extract combination slot position " apple " and " two pears ".By the method for the embodiment of the present application, Bot platform can be improved to multiple The understandability of miscellaneous user information can promote the usage experience that user uses Bot platform.

Illustratively, the embodiment of the present application states " technical ability " herein above as shopping, and " intention " is " buying fruit ", Yong Hu The combination slot position of Bot platform configuration is " quantity fruit type ", and training corpus is for " I will buy two pears of an apple ". Above-mentioned " slot position creation and configuration " process, " model training " process and " model prediction " process are illustrated.

Firstly, group can be recombinated after Bot platform receives the combination slot position " quantity fruit type " of the first user configuration Two entity types ' quantity ' in slot position " quantity fruit type " and ' fruit type ' are closed, four slot positions are obtained: " number Amount ", " fruit type ", " quantity fruit type " and " fruit type quantity ".

Then, Bot platform can be according to training corpus " I will buy two pears of an apple " and slot position "@quantity ", "@ Fruit type ", " quantity fruit type " and " fruit type quantity " are trained one or more training corpus, make Bot platform has the ability for extracting the slot position in user's saying " I will buy two pears of an apple ".

Illustratively, with Bot platform use above-mentioned deep learning algorithm to one or more training corpus be trained for Example.For example, as shown in Figure 10, Bot platform can receive the true tag that user is the word addition of each of the training corpus, It is that each of training corpus word adds feature a and by the way of " the correct combination " in above-mentioned implementation (1).Such as This, Bot platform can learn to the true mark for training corpus " I will buy two pears of an apple " addition and user's addition Sign the ability of corresponding feature.For example, as shown in Figure 10, Bot platform can add feature " B-composite for " one " word Entity"；Feature " I-composite Entity " is added for " a " word；Feature " I-composite is added for " apple " word Entity"；Feature " I-composite Entity " is added for " fruit " word；Feature " B-composite is added for " two " word Entity"；Feature " I-composite Entity " is added for " a " word；Feature " I-composite is added for " pears " word Entity".It is "@mono- that i.e. Bot platform, which may learn correct slot position in training corpus " I will buy two pears of an apple ", A@apple " and " two@pears of@".In this way, after Bot platform receives user's saying " I wants two pears of an apple ", it can Extract slot position " mono-@apple of@" and " two@pears of@".I.e. Bot platform can " I wants apple two from user's saying Pears " extract correct slot position.

It is understood that above-mentioned Bot platform is in order to realize the above functions, it comprises executing, each function is hard accordingly Part structure and/or software module.Those skilled in the art should be readily appreciated that, retouch in conjunction with the embodiments described herein Each exemplary unit and algorithm steps stated, the embodiment of the present application can be with the combining forms of hardware or hardware and computer software To realize.Some functions is executed in a manner of hardware or computer software driving hardware actually, depending on technical solution Specific application and design constraint.Professional technician can to each specific application come using distinct methods to realize The function of description, but this realization is it is not considered that exceed the range of the embodiment of the present application.

The embodiment of the present application can carry out the division of functional module according to above method example to above-mentioned Bot platform, for example, The each functional module of each function division can be corresponded to, two or more functions can also be integrated in a processing mould In block.Above-mentioned integrated module both can take the form of hardware realization, can also be realized in the form of software function module. It should be noted that be schematical, only a kind of logical function partition to the division of module in the embodiment of the present application, it is practical There may be another division manner when realization.

The embodiment of the present application also provides a kind of Bot platforms.As shown in figure 15, which may include: slot position Configuration module 1501, model training module 1502 and model prediction module 1503.Wherein, slot position configuration module 1501 is for executing Above-mentioned " slot position creation and configuration " process, model training module 1502 is for executing above-mentioned " model training " process, model prediction Module 1503 is for executing above-mentioned " model prediction " process.

It is appreciated that the partial function and model training module 1502 and model prediction mould of slot position configuration module 1501 The function of block 1503 can integrate to be realized in processor 210 shown in Fig. 2.It is above-mentioned for showing in slot position configuration module 1501 It is realized in the display 294 that the function at the first interface can be shown in Fig. 2.

The embodiment of the present application also provides a kind of computer storage medium, which includes computer instruction, When the computer instruction is run on above-mentioned Bot platform, so that Bot platform executes Bot in the description of above-described embodiment and puts down Each function that platform executes.

The embodiment of the present application also provides a kind of computer program product, when the computer program product is run on computers When, so that computer executes each function that Bot platform executes in the description of above-described embodiment.

Through the above description of the embodiments, it is apparent to those skilled in the art that, for description It is convenienct and succinct, only the example of the division of the above functional modules, in practical application, can according to need and will be upper It states function distribution to be completed by different functional modules, i.e., the internal structure of device is divided into different functional modules, to complete All or part of function described above.The specific work process of the system, apparatus, and unit of foregoing description, before can referring to The corresponding process in embodiment of the method is stated, details are not described herein.

In several embodiments provided by the present embodiment, it should be understood that disclosed system, device and method can To realize by another way.For example, the apparatus embodiments described above are merely exemplary, for example, the module Or the division of unit, only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple lists Member or component can be combined or can be integrated into another system, or some features can be ignored or not executed.Another point, Shown or discussed mutual coupling, direct-coupling or communication connection can be through some interfaces, device or unit Indirect coupling or communication connection, can be electrical property, mechanical or other forms.

The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme 's.

In addition, each functional unit in each embodiment of the present embodiment can integrate in one processing unit, it can also To be that each unit physically exists alone, can also be integrated in one unit with two or more units.It is above-mentioned integrated Unit both can take the form of hardware realization, can also realize in the form of software functional units.

If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product When, it can store in a computer readable storage medium.Based on this understanding, the technical solution essence of the present embodiment On all or part of the part that contributes to existing technology or the technical solution can be with the shape of software product in other words Formula embodies, which is stored in a storage medium, including some instructions are used so that a calculating Machine equipment (can be personal computer, server or the network equipment etc.) or processor execute each embodiment the method All or part of the steps.And storage medium above-mentioned includes: flash memory, mobile hard disk, read-only memory, arbitrary access The various media that can store program code such as memory, magnetic or disk.

The above, the only specific embodiment of the present embodiment, but the protection scope of the present embodiment is not limited thereto, Any change or replacement in the technical scope that the present embodiment discloses, should all cover within the protection scope of the present embodiment. Therefore, the protection scope of the present embodiment should be based on the protection scope of the described claims.

Claims

1. combining the configuration method of slot position in a kind of interactive system, which is characterized in that be applied to robot Bot platform, institute The method of stating includes:

The Bot platform receives user in the first slot position of the first interface configurations；First slot position is combination slot position, described the One slot position includes N number of entity type, and N >=2, N are positive integer；N number of entity type is in first slot position according to user The sequence of setting arranges；Slot position is arranged in the first intention that first interface is used in the first technical ability for the Bot platform；

The Bot platform recombinates N number of entity type, obtains M the second slot positions；The M the second slot positions include described N number of The slot position that k entity type in entity type arranges in any order, k ∈ { 1,2 ... ..., N }；

The Bot platform is according to one or more training corpus and the M the second slot positions, to one or more of trained languages Material is trained, and the Bot platform is made to have the ability for extracting the M in user's saying the second slot position.

2. the method according to claim 1, wherein

3. method according to claim 1 or 2, which is characterized in that the method also includes:

The Bot platform receives user in the slot position title of first interface configurations, and the slot position title is described for identifying First slot position.

4. according to the method described in claim 3, it is characterized in that, being obtained in Bot platform recombination N number of entity type To after M the second slot positions, the Bot platform is according to one or more training corpus and the M the second slot positions, to described one Before a or multiple training corpus are trained, the method also includes:

The slot position title and the M the second slot position associated storages are that the slot position title is corresponding by the Bot platform The feature of the M the second slot position label composite entities, the feature of the composite entity are used to indicate the M the second slot positions and are Composite entity type after recombination.

5. according to the method described in claim 4, it is characterized in that, the Bot platform according to one or more training corpus and The M the second slot positions, are trained one or more of training corpus, comprising:

The Bot platform receives the true tag that user is the word addition of each of one or more of training corpus；

The Bot platform is that each of one or more of training corpus word adds spy according to the M the second slot positions Sign；

The Bot platform uses deep learning algorithm, according to the true tag and the feature, to one or more of instructions Practice corpus to be trained；Alternatively, the Bot platform uses deep learning algorithm conjugation condition with field CRF algorithm, according to described true Real label and the feature, are trained one or more of training corpus.

6. according to the method described in claim 5, it is characterized in that, the deep learning algorithm includes shot and long term memory network LSTM algorithm.

7. method according to any one of claim 1-3, which is characterized in that the Bot platform is according to one or more Training corpus and the M the second slot positions, are trained one or more of training corpus, comprising:

The Bot platform uses single-point sorting algorithm and dynamic programming algorithm based on probability, according to one or more of instructions Practice corpus and the M the second slot positions, one or more of training corpus are trained.

8. the method according to the description of claim 7 is characterized in that the single-point sorting algorithm includes at least support vector machines SVM model, maximum entropy model, Fast Text Classification algorithm model, convolutional neural networks CNN model, n-gram model or circulation Any one of neural network RNN model.

9. the method according to the description of claim 7 is characterized in that the single-point sorting algorithm is BERT model, the BERT Model is bi-directional language model.

10. the method according to any one of claim 7-9, which is characterized in that the Bot platform is calculated using single-point classification Method and dynamic programming algorithm based on probability, according to one or more of training corpus and the M the second slot positions, to described One or more training corpus are trained, comprising:

For each training corpus in one or more of training corpus,

The Bot platform uses the dynamic programming algorithm based on probability, scans instruct one in different location from right to left Practice corpus and be cut into one or more candidate entities, composite entity and normalization composite entity number are normalized after being cut；

The Bot platform uses the single-point sorting algorithm, according to the M the second slot positions, obtains the corresponding time in each position Select the confidence level of entity；

Confidence level of the Bot platform according to each position corresponding normalization composite entity number and candidate entity, determination pair The cutting mode of one training corpus.

11. method according to claim 1 to 10, which is characterized in that recombinated in the Bot platform described N number of Entity type, after obtaining M the second slot positions, the method also includes:

The Bot platform shows second contact surface, and the second contact surface includes starting to train button, described to start that button is trained to be used for The Bot platform is triggered to be trained one or more of training corpus；

In response to user to the clicking operation for starting to train button, the Bot platform is according to one or more of training Corpus and the M the second slot positions, are trained one or more of training corpus, the Bot platform are made to have extraction The ability of the M the second slot positions in user's saying.

12. a kind of robot Bot platform, which is characterized in that the Bot platform includes: processor, memory and display；Institute State memory, the display and processor coupling；The memory is for storing computer program code, the calculating Machine program code includes computer instruction, and when the processor executes the computer instruction, the Bot platform is executed:

The display, for showing that the first interface, first interface are used for the in the first technical ability for the Bot platform One is intended to setting slot position；

The processor, for receiving the first slot position of first interface configurations that user shows in the display；It is described First slot position is combination slot position, and first slot position includes N number of entity type, and N >=2, N are positive integer；N number of entity type It is arranged in first slot position according to sequence set by user；

The processor is also used to recombinate N number of entity type, obtains M the second slot positions；The M the second slot positions include The slot position that k entity type in N number of entity type arranges in any order, k ∈ { 1,2 ... ..., N }；

The processor is also used to according to one or more training corpus and the M the second slot positions, to one or more of Training corpus is trained, and the Bot platform is made to have the ability for extracting the M in user's saying the second slot position.

13. Bot platform according to claim 12, which is characterized in that

14. Bot platform according to claim 12 or 13, which is characterized in that the processor is also used to receive user and exists The slot position title for first interface configurations that the display is shown, the slot position title is for identifying first slot position.

15. Bot platform according to claim 14, which is characterized in that the processor is also used to described N number of in recombination Entity type, after obtaining the M the second slot positions, according to one or more of training corpus and the M the second slot positions, Before being trained to one or more of training corpus, the slot position title is deposited with the M the second slot positions described Associated storage in reservoir, and the feature of composite entity is marked for the corresponding M the second slot positions of the slot position title, described group It is the composite entity type after recombination that the feature for closing entity, which is used to indicate the M the second slot positions,.

16. Bot platform according to claim 15, which is characterized in that the processor, for being instructed according to one or more Practice corpus and the M the second slot positions, one or more of training corpus be trained, comprising:

The processor receives the true tag that user is the word addition of each of one or more of training corpus；Root It is that each of one or more of training corpus word adds feature according to the M the second slot positions；

The processor, is also used to using deep learning algorithm, according to the true tag and the feature, to one or Multiple training corpus are trained；Alternatively, using deep learning algorithm conjugation condition with field CRF algorithm, according to the true mark Label and the feature, are trained one or more of training corpus.

17. Bot platform according to claim 16, which is characterized in that the deep learning algorithm includes shot and long term memory Network LSTM algorithm.

18. Bot platform described in any one of 2-15 according to claim 1, which is characterized in that the processor is used for basis One or more of training corpus and the M the second slot positions, are trained one or more of training corpus, wrap It includes:

The processor is specifically used for using single-point sorting algorithm and dynamic programming algorithm based on probability, according to one Or multiple training corpus and the M the second slot positions, one or more of training corpus are trained.

19. Bot platform according to claim 18, which is characterized in that the single-point sorting algorithm include at least support to Amount machine SVM model, maximum entropy model, Fast Text Classification algorithm model, convolutional neural networks CNN model, n-gram model or Any one of Recognition with Recurrent Neural Network RNN model.

20. Bot platform according to claim 18, which is characterized in that the single-point sorting algorithm is BERT model, described BERT model is bi-directional language model.

21. Bot platform described in any one of 8-20 according to claim 1, which is characterized in that the processor, for being used for Using single-point sorting algorithm and dynamic programming algorithm based on probability, according to one or more of training corpus and the M Second slot position is trained one or more of training corpus, comprising:

The processor, specifically for each training corpus being directed in one or more of training corpus, using the base In the dynamic programming algorithm of probability, scans one training corpus is cut into one or more candidates in different location from right to left Entity normalizes composite entity and normalization composite entity number after being cut；Using the single-point sorting algorithm, according to institute M the second slot positions are stated, the confidence level of the corresponding candidate entity in each position is obtained；According to the corresponding normalization combination in each position The confidence level of entity number and candidate entity, determines the cutting mode to one training corpus.

22. Bot platform described in any one of 2-21 according to claim 1, which is characterized in that the display is also used to The processor recombinates N number of entity type, after obtaining the M the second slot positions, shows second contact surface, second boundary Face includes starting to train button, described to start to train button for triggering the Bot platform to one or more of trained languages Material is trained；

The processor is also used to the clicking operation for starting to train button shown in response to user to the display, According to one or more of training corpus and the M the second slot positions, one or more of training corpus are trained, The Bot platform is set to have the ability for extracting the M in user's saying the second slot position.