CN110059164A - Semantic expressiveness and realization for conversational system - Google Patents
Semantic expressiveness and realization for conversational system Download PDFInfo
- Publication number
- CN110059164A CN110059164A CN201910022976.1A CN201910022976A CN110059164A CN 110059164 A CN110059164 A CN 110059164A CN 201910022976 A CN201910022976 A CN 201910022976A CN 110059164 A CN110059164 A CN 110059164A
- Authority
- CN
- China
- Prior art keywords
- user
- semantic
- concept
- subgraph
- conversational system
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3325—Reformulation based on results of preceding query
- G06F16/3326—Reformulation based on results of preceding query using relevance feedback from the user, e.g. relevance feedback on documents, documents sets, document terms or passages
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/338—Presentation of query results
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
Abstract
The present invention relates to for conversational system semantic expressiveness and realization.The method, apparatus and computer program product of a kind of user interface of conversational system for rendering are described.Create context graph of the Uniform semantic expression of the conversation content between user and conversational system as concept and relationship.Based on current session activity, the subgraph component set of the semantic context figure is dynamically identified.The subgraph component set identified in user interface is as the graphic element collection for representing corresponding concept and relationship.
Description
Technical field
Present disclose relates generally to natural language processings.More specifically, this disclosure relates to be used for the nature of conversational system
Language Processing.
Background technique
The application that user encounters such as virtual protocol and chat robots etc becomes very universal, these are arrived using offer
Web content, application and the natural language interface in channel.In general, these applications use conversational system, these conversational systems use base
It interacts in the dialog prompt of natural language with end user to realize object-oriented task, such as online transaction.Although these
Using the huge potential value of offer, but they are restricted in the information of offer and in terms of helping type, because using to certainly
It the deficiency of right language understanding and is difficult to generate interface for each potential user's demand.Therefore, these systems usually will dialogue
Prompt be limited to user request direct and steady-error coefficient, without provide about why generation system response it is appropriate up and down
Text is explained.Unless by system designer it is contemplated that otherwise chat robots will usually lack in processing end user's feedback
The ability of specific project.
Existing conversational system usually not sufficiently inform end user, these conversational systems be used only based on natural language,
Result-oriented dialog prompt.Unless carrying out Comprehensive Designing, otherwise prompt may cause both system and end user and talk with
There is unexpected ambiguity and not found misunderstanding in period.
Further, since the sense of defeat experienced using system, end user loses the wish for further using conversational system,
This makes system have no chance to obtain the valuable user input that can be used for improving system.
Require further improvement computer assisted conversational system.
Summary of the invention
According to the disclosure, a kind of method, apparatus of the user interface of conversational system and computer journey for rendering are provided
Sequence product.Create context of the Uniform semantic expression of the conversation content between user and conversational system as concept and relationship
Figure.Based on current session activity, the subgraph component set of the semantic context figure is dynamically identified.Being identified in user interface
Subgraph component set as the graphic element collection for representing corresponding concept and relationship.
Foregoing has outlined certain more relevant characteristics of published subject.These characteristics should be construed as merely illustrative
's.By applying published subject in different ways or by modifying the present invention that will be described, can obtain many other
Advantageous result.
Detailed description of the invention
In order to which the present invention and its advantage is more fully understood, presently in connection with attached drawing with reference to being described below, these attached drawings are:
Fig. 1 shows the example that the distributed data processing environment of illustrative aspect of illustrative embodiments wherein may be implemented
Property block diagram;
Fig. 2 is the exemplary block diagram that the data processing system of illustrative aspect of illustrative embodiments wherein may be implemented;
Fig. 3 shows the framework of the frame for realizing the embodiment of the present invention;
Fig. 4 shows how system according to an embodiment of the invention will carry out during the dialog procedure with system
Sequence user spoken utterances are converted to more succinct semantic meaning figure;
Fig. 5 is the simplification Uniform semantic according to an embodiment of the invention using exemplary fields ontology (ontology)
The figure of figure;
Fig. 6 shows system according to an embodiment of the invention and highlights a part of context graph dynamically to draw
Play user feedback;
Fig. 7 show it is according to an embodiment of the invention how in dialog interface present surface semantic expressiveness (SSR);
Fig. 8 is the figure for showing several example user feed back inputs for one embodiment of the present of invention;
Fig. 9 is the figure for showing the one embodiment for making chat agency inquire correct problem;And
Figure 10 is the neural network frame wherein connected in an embodiment of the present invention using the modularization for dialogue management
The figure of frame.
Specific embodiment
High-level, the preferred embodiment of the present invention provides a kind of system, method and computer for conversational system
Program product, the conversational system provide the basic information that system is used to complete task to end user.By providing the information, boundary
Face can explain why conversational system is responded in a manner of it, and strategically attract end user, to use them
Feedback improve the availability of conversational system.In an embodiment of the present invention, based on the dialog prompt of natural language by frame
Enhancing, the frame are based on semantic context, domain knowledge and session activity, provide more information for end user's dynamic generation
Dialog prompt.
With regard to known to inventor, the present invention be systematically generated based on domain knowledge session activity semantic expressiveness and
The graphically present semantic expressiveness generated of user interface level is to cause the trial for the first time of user feedback.Of the invention
In embodiment, semantic expressiveness generated corresponds to corresponding session activity.Interface obtain user's input in implicit dialogue with
And rudimentary annotation is to be used for machine learning purpose.Because semantic expressiveness is dynamically generated, exports from multiple sources and from final
The orientation optimization of user, so the embodiment of the present invention represents important the changing of the integrated prior art work of related semantic content
Into.Will the problem in semantic intergration be represented in more contribution dynamic properties of the dialogue semantically indicated in the user interface.
In order to solve the limitation of the interaction based on natural language and improve the availability of conversational system, implementation of the invention
Example is provided Unified frame and is indicated with the grapheme for generating dialogue for object-oriented task.In addition, system dynamically identifies table
Subgraph in showing is to work as request, may be presented in the user interface with when necessary based on session activity and field logic.Specifically
Say that the embodiment of the present invention determines how system explains the input of user, and how system handles the information and system of rear end in ground
How the simplicity of explanation of field logic and query result is provided.
Compared with the traditional interface based on natural language, the embodiment of the present invention is utilized in the following manner based on figure
The ability to express of model: (1) content of text that standardizes indicates figure with generative semantics meaning;(2) entity and pass is can be explained into field
System is integrated with semantic matches technology with generative semantics context graph;(3) it is directed to identified conversational operation dynamically logos
The subgraph of adopted context graph;And the graphical representation of chosen content is presented in (4), such as is as graphic element collection, as dialogue
A part of the dialog prompt of system.
Realize that the process of (SSR) Lai Zengqiang dialog prompt is intended to effectively help system using underlying semantics or semantic surface
The transmitting of information and knowledge between end user, and end user is made to be capable of providing the feedbacks of various ranks.It is current right
The SSR of words has several embodiments of the invention movement on practical use, such as the website (a) or chat robots service
The experienced end user at interface;(b) the crowdsourcing worker in relation to dialogue annotation task;(c) turn in relation to the knowledge to system
The theme professional knowledge of shifting;And (d) based on the educational aid of domain knowledge.One importance of the frame is to cause finally
The feed back input of user.In an embodiment of the present invention, it feeds back for annotating purpose and by enhanced dialog prompt
Interaction characteristic receives.By simple post-processing, the feedback data of acquisition is applied to promote study, to pass through conversational system
To improve the future dialogue with user.
Fig. 1-2 is referred to referring now to the drawings and specifically, and the illustrative implementation that the disclosure wherein may be implemented is provided
The exemplary diagram of the data processing circumstance of example.It should be understood that Fig. 1-2 is merely exemplary, it is not intended that assert or suggest about
Any restrictions of the aspect of published subject or the environment of embodiment wherein may be implemented.It can be without departing from spirit of the invention
In the case where range, many modifications are carried out to shown environment.
Referring now to the drawings, Fig. 1 shows the example distributed number for the various aspects that illustrative embodiments wherein may be implemented
According to the graphical representation of processing environment.Distributed data processing system 100 may include that illustrative embodiments wherein may be implemented
The computer network of various aspects.Distributed data processing system 100 includes at least one network 102, and network 102 is for dividing
The medium of communication link is provided between the various equipment to link together in cloth data processing system 100 and computer.Network
102 may include connection, such as wired, wireless communication link or optical cable.
In the example shown, server 104 and server 106 and networked storage units 108 are connected to network 102.This
Outside, client computer 110,112 and 114 is also connected to network 102.These client computer 110,112 and 114 for example can be intelligent electricity
Words, tablet computer, personal computer, network computer etc..In the example shown, server 104 is to client computer 110,112 and
114 provide data, such as guidance file, operation system image and application.In the example shown, client computer 110,112 and 114
It is the client computer of server 104.Distributed data processing system 100 may include additional server, client computer and other
Unshowned equipment.One or more server computers can be attached to the mainframe computer of network 102.Mainframe computer
Such as it can be the IBM System z mainframe of operation IBM z/OS operating system.Mass storage and work station (do not show
It may be coupled to mainframe out).Work station can be directly to through the personal computer of the mainframe of bus communication or
Person is the console terminal that mainframe is directly connected to via display port.
In the example shown, distributed data processing system 100 is internet, while network 102 indicates in global range
Using transmission control protocol/internet protocol (TCP/IP) protocol suite come the set of the network and gateway to communicate with one another.In internet
Core be the high-speed data communication lines between host node or host trunk, it includes thousands of business, government, religion
It educates and the computer system of other routing data and messages.Certainly, distributed data processing system 100 is also implemented as
Many different types of networks, such as Intranet, local area network (LAN), wide area network (WAN) etc..As described above, Fig. 1 is intended as
One example, rather than the framework limitation of the different embodiments to published subject, and therefore, specific member shown in Fig. 1
Part is not construed as the limitation about the environment that illustrative embodiments of the invention wherein may be implemented.
Referring now to Figure 2, showing the example data processing system that the various aspects of illustrative embodiments wherein may be implemented
Block diagram.Data processing system 200 is an example of computer, and the client computer 114 of example as shown in figure 1 realizes saying for the disclosure
The computer usable code of the process of bright property embodiment or instruction can be located therein.
Referring now to Figure 2, showing the block diagram that the data processing system of illustrative embodiments wherein may be implemented.Data processing
System 200 is an example of computer, and example server 104 as shown in figure 1 or client computer 110 realize illustrative embodiments
The computer usable program code of process or instruction can be located therein.In the illustrated examples, data processing system 200 is wrapped
Telecommunication optical fiber network 202 is included, in processor unit 204, memory 206, persistent storage 208, communication unit 210, defeated
Enter/export and provides communication between (I/O) unit (multiple) 212 and display 214.
Processor unit 204 is used to execute the instruction for the software that can be loaded into memory 206.Processor unit
204 can be the set comprising one or more processors or can be multiple processor cores, be specifically dependent upon specific implementation.
In addition it is possible to use one or more heterogeneous processor systems realize processor unit 204, wherein existing on a single chip
Primary processor and secondary processor.As another illustrated examples, processor unit 204 be can be comprising the more of same type
Symmetric multiprocessor (SMP) system of a processor.
Memory 206 and persistent storage 208 are to store the example of equipment.Storage equipment is can temporarily and/or forever
Any hardware of storage information long.In these examples, memory 206 for example can be random access memory or any other
Suitable volatibility or non-volatile memory device.Persistent storage 208 can take various forms, be specifically dependent upon specific
It realizes.For example, persistent storage 208 may include one or more components or equipment.For example, persistent storage 208 can
To be hard disk drive, flash memory, rewritable CD, rewritable tape or certain above-mentioned combination.Persistent storage 208 makes
Medium is also possible to mobile.For example, removable hard disk drive can be used for persistent storage 208.
In these examples, communication unit 210 provides the communication with other data processing systems or equipment.In these examples
In, communication unit 210 is network interface card.Communication unit 210 can be by using one of physics and wireless communication link or two
Person provides communication.
I/O unit 212 allow to input using the other equipment that may be coupled to data processing system 200 with it is defeated
Data out.For example, I/O unit 212 can provide connection by keyboard, mouse to carry out user's input.In addition, defeated
Enter/output unit 212 can send output to printer.In addition, I/O unit can be provided to for from user
Audio input microphone and connection to the loudspeaker for providing the audio output from computer.Display 214 mentions
For for showing the mechanism of information to user.
Persistent storage 208 can be located at for operating system and application or the instruction of program.These instructions can be with
It is loaded into memory 206 to be executed by processor unit 204.Computer implemented finger can be used in processor unit 204
It enables to execute the process of different embodiments, computer implemented instruction can be located in memory (such as memory 206).These
Instruction is referred to as program code, computer usable program code or computer readable program code, they can be by processor
Processor in unit 204 reads and executees.In different embodiments, program code can be embodied in different physics or tangible
On computer-readable medium (such as memory 206 or persistent storage 208).
Program code 216 is located on the computer-readable medium 218 that optionally moves in functional form, and can be with
Data processing system 200 is loaded into or is transferred to be executed by processor unit 204.In these examples, program code 216
Computer program product 220 is formed with computer-readable medium 218.In one example, computer-readable medium 218 can be adopted
Take tangible form, such as CD or disk, CD or disk be inserted into or be placed into be persistent storage 208 a part
In driver or other equipment, for transmission to storage equipment (such as the hard drive of a part for being persistent storage 208
Device).In tangible form, computer-readable medium 218 can also take the form of persistent storage, such as be connected to data
Hard disk drive, thumb actuator or the flash memory of processing system 200.The computer-readable medium 218 of tangible form is also claimed
For computer recordable storage medium.In some cases, computer recordable media 218 may be irremovable.
It is alternatively possible to pass through the communication link to communication unit 210 and/or the company by arriving I/O unit 212
It connects, program code 216 is transferred to data processing system 200 from computer-readable medium 218.In illustrated examples, communication
Link and/or connection can be physics or wireless.Computer-readable medium can also take the form of non-tangible media, such as
Communication link or wireless transmission comprising program code.It is not meant to for difference component shown in data processing system 200
Framework limitation is provided to the mode that different embodiments may be implemented.Different illustrative embodimentss can be in a data processing system
It realizes, in addition to, for component shown in data processing system 200, being somebody's turn to do for component described in data processing system 200 or substitution
Data processing system further includes other assemblies.Other components shown in Fig. 2 can the illustrated examples from shown in it is different.As
One example, the storage equipment in data processing system 200 is to can store any hardware device of data.Memory 206 is held
Long storage device 208 and computer-readable medium 218 are the examples of the storage equipment of tangible form.
In another example, bus system can be used to implement telecommunication optical fiber network 202, and may include one or
Multiple buses, such as system bus or input/output bus.Of course, it is possible to be realized using the framework of any suitable type total
Linear system system, the framework provide data transmission between the different components or equipment for being attached to bus system.In addition, communication unit can
To include one or more equipment for sending and receiving data, such as modem or network adapter.In addition, storage
Device for example can be memory 206 or the cache for example found in the interface and can reside in Communication ray fibre web
Storage control hub in network 202.
Computer program code for executing operation of the present invention can be with any combination of one or more programming languages
It writes, the programming language includes the programming language-such as Java of object-orientedTM、Smalltalk、C++、C#、
Objective-C etc., and conventional procedural programming languages-such as Python or C.Program code can fully with
It executes on the computer of family, partly execute on the user computer, being executed as an independent software package, partially in user's meter
Part executes on the remote computer or executes on a remote computer or server completely on calculation machine.It is being related to remotely counting
In the situation of calculation machine, remote computer can include local area network (LAN) or wide area network (WAN)-by the network-of any kind
It is connected to subscriber computer, or, it may be connected to outer computer (such as using ISP come by because of spy
Net connection).
One of ordinary skill in the art will be understood that the hardware in Fig. 1-2 can be different according to realization.In addition to figure
Hardware shown in hardware shown in 1-2 or alternate figures 1-2 can also use other internal hardware or peripheral devices, such as
Flash memory, equivalent nonvolatile memory or CD drive etc..In addition, in spirit and model without departing from published subject
In the case where enclosing, the process of illustrative embodiments can be applied to the multiprocessor number other than previously mentioned smp system
According to processing system.
Technology described herein can operate in conjunction in the standard client-server normal form shown in such as Fig. 1,
Wherein client computer and the addressable portal (it closes execution in the collection comprising one or more machines) based on Web in internet are logical
Letter.End user operation be able to access that the attachable equipment of portal and the internet interacted with portal (for example, desktop computer,
Notebook computer, the mobile device for supporting internet etc.).In general, each client computer or server machine are institutes in such as Fig. 2
The data processing system (including hardware and software) shown, and these entities pass through network (such as internet, Intranet, external connection
Net, dedicated network) or any other communication media or link communicate with one another.Data processing system generally includes one or more places
Manage device, operating system, one or more application and one or more utilities.
Session between system and user, the embodiment of the present invention generate the unification of conversation content using frame
Model, and related content is dynamically selected from model to present in the user interface.Specifically, embodiment determines system
How semantically to explain user spoken utterances, request is handled in backend application and requests user feedback.In different implementations
In example, according to user preference and if necessary, user feedback is requested when it is possible.Because user interface preferably notifies finally
User provides the feedback of various ranks, it is possible that having the more data annotated by end user, collects these numbers at any time
Conversational system is improved accordingly.
Fig. 3 shows the framework of the frame for realizing the embodiment of the present invention.In general, for realizing frame of the invention
Frame should provide the semantic interpretation of four core function (1) user spoken utterances;(2) semantic intergration;(3) dynamic content selects;And
(4) semantic surface is realized.The first two component corresponds to the pass the semantic parsing result collection for entering text into (such as user spoken utterances)
At to the field concept in conversational system and the drawing of seeds for being embodied in domain body, the unified model of semantic content is constructed jointly.
After the Uniform semantic of creation conversation content indicates as the context graph of concept and relationship, system is dynamically identified due to working as
Preceding session activity and will to end user present semantic context figure subgraph.Then, system makes in user interface level
The content of subgraph is usually presented with corresponding pattern primitive.By what is discussed in greater detail below, in the preferred embodiment of the present invention
" semantic surface realization " (SSR) of middle presentation and the dialogue of user.Then processing, storage and use are anti-from the received user in interface
Feedback is to improve conversational system.
Turning now to Fig. 3, user interacts with user interface 301.User interface capture annotated feedback 302 and user with
Both normal dialogs (for example, user spoken utterances 303) between system.In a preferred embodiment, the session proxy of capture systems is gone back
Response, that is, the components of system as directed talked with.Natural language utterances are assessed by one group of semantic analyzer 305, semantic analyzer 305 generates
It can be by the official interpretation for the natural language utterances meaning that the rest part of conversational system uses.Because user is based in user interface
Graphical configuration annotated feedback is provided, so annotated feedback can be directly inputted to atlas and grow up to be a useful person 307, atlas at
Device 307 is responsible for creating for newest " round " in user/system dialog from the input from multiple system components and information source
Context graph.
Semantic interpretation 308 from semantic analyzer 305, which is supplied to meaning, indicates processor 309, meaning expression processing
Device 309 will explain the semantic expressiveness for being converted into being suitable for incorporation into context graph.Context resolution device assembly 311 to atlas at
Device 307 provides the input of related previous user input (for example, previous user language), so that can be according to user/system pair
The current context of words carrys out structure figures.By what is discussed in greater detail below, current utterance is assessed by reference to previous language
Context, illustrate certain natural language meanings.The grow up to be a useful person semantic meaning of 307 generation sentences of atlas indicates Figure 31 2 (MR figure), with
It is integrated for the newest round in talking with and with Uniform semantic Figure 32 1.The each user spoken utterances captured in dialogue are successively
It is converted into the sentence meaning figure of own.Although word " sentence " means, those skilled in the art related to sentence
It will be recognized and not all user spoken utterances all will be to comply fully with the sentence of syntax rule, and user spoken utterances may include being more than
One sentence.Its use in the description is intended to express according to language (that is, for most of (if not complete in dialogue
Portion) user spoken utterances) create one or more meaning figures.
In a preferred embodiment, statement semantics meaning expression is converted into corresponding sentence " concept " figure.Provide user's words
Sentence in language, MR figure be using the semantic analysis of the sentence of semantic label (more than concept), and concept map be based on field it is general
It reads.
Context resolver 311 can also access " rear end " using 319, and for backend application 319, conversational system is
" front end ".Backend application includes several databases 313,315,317, they include the specific information in field.It is of the invention not
With in embodiment, there will be only certain databases.When user, which currently participates in backend application, is designed to completing for task, field
Specific information is used to eliminate the ambiguity of user spoken utterances.Context resolver 311 can produce the inquiry to these databases, with
Obtain the specific information in field for being used for semantic basis." semantic basis " refers to that (such as field is general from content of text to relevant knowledge
Thought/relationship) mapping.
Semantic meaning indicates that Figure 31 2 is bonded in Uniform semantic Figure 32 1, and Uniform semantic Figure 32 1 is the upper of conversation content
Lower texts and pictures.In reference implementation example, merge Figure 31 2 and 321, as it is following at entitled " semantic intergration for conversation content "
Described in part.Relevant information (given user's meaning is integrated by the integrating process of any or several known types
Figure), these integrating process for example including across sentence, back gear time, across interlocutor and across knowledge base.By semantic meaning obtained
Figure, relevant semantic content is identified based on field database, so that inquiry or order can be formed to complete task.Two
A different stage executes semantic matches, and one in element rank;One in structure rank.For graphic element, system-computed neck
The Semantic Similarity between nodename in domain concept and MR figure.If similarity score, which is higher than specific threshold, (passes through practice
Determine), then graphical nodes are mapped with field concept.For graphic structure, system be based on equivalence, partly overlap, superset with
And subset considers semantic dependency.
Graphical configuration 322 (for example, subgraph) is presented to be provided as dialog prompt as a part of user interface 301
323 for user comment.Will be as discussed below, the graphical configuration presented needs not be continuous general from Uniform semantic figure
Thought and relationship, but selected relationship and concept that most probable causes user feedback can be predicted to be.It user knowledge 322 and looks into
Result 324 is ask for contributing to Uniform semantic Figure 32 1.User knowledge is for improving existing conversational system and user experience
Important sources.For example, if end user frequently refers to certain contents, but it is in domain knowledge base.Identify this use
Family knowledge and to be added to domain knowledge base very useful.
Realize the interface (SSR) as the system with user as described above, semantic surface is presented in the preferred embodiment of the present invention
A part of dialogue.The portion identification system of user interface intermediate semantic expressiveness currently in use, to instruct its object-oriented
Dialogue a part.That is, system actually tells user why it is being presented specific selection to user.As user circle
Semantic interpretation can be performed for finally using in the basis that the graphical representation for the more structuring that a part in face is presented allows user to request
Family is visible and is appreciated that enough.In this way, user can check how system handles and explain contextual information.It is also
User is allowed to provide feedback to dialog manager via chat interface, such as whether the hypothesis that system is made is well to assume.?
The semantic expressiveness presented at user interface is figure, and therefore more more intuitive than interminable dialogue explanation, therefore allows final use
The feedback at family is also figure.That is, user can interact with graphical interfaces.By showing correspondence by means of intuitive graphics feature
In the semantic interpretation of newest dialogue state, the interface SSR be should be readily appreciated that (especially for experienced user), while visually
It is intuitive.
In order to develop the intermediate representation encoded to various semantemes, the embodiment of the present invention includes for by semantic interpretation
Convert and be integrated into the frame in unified model.In the ideal case, for create unified model method should can across application and
Field is extensive, and is semantically expressing to capture the meaning of various inquiries.Other required attributes include supporting that definition is clear
Criterion calculation technology convenience of calculation, with main rear end storage (such as relational database and chart database) compatibility,
And interoperability and reusability for different application.
In an embodiment of the present invention, using the intermediate semantic expressiveness for generating dialogue based on the method for figure.One challenge
It is to handle the context semanteme based on heterogeneous resource and be integrated into unified model.For object-oriented chat, on
It is hereafter semantic to be determined by both " unofficial " requestor (i.e. end user) and " formal " respondent (i.e. conversational system).More specifically
Say that the semanteme that user is intended to can be embedded into user spoken utterances, and user spoken utterances may include following information, such as specific objective in ground
Or the letter for being intended to (for example, " finding course "), supporting information (for example, " course with 3 credits ") and customer-centric
Breath (for example, " preferred theory course ").The neck of the fact that also using corresponding at conversational system rear end or ontology knowledge base (KB)
The interpretable semanteme in domain.The information is generally stored inside in relationship and/or chart database, and uses in a preferred embodiment of the invention
Information is provided in explaining user spoken utterances and being additionally in response to user query.Another challenge is selected from intermediate semantic expressiveness
Subgraph component needs enough intuitively to present at user interface to end user.In a preferred embodiment, by combining tuple
The characteristic of calculus of relation (TRC) and Domain relation calculation both (DRC), mark represent the simplicity of selected subgraph component and intuitive
Visual configuration collection.The variable that the clearly expression of TRC and DRC uses one group of succinct conjunction and specifies, and conjunction and variable
The node and Bian Lai that can be indicated by figure are shown.Important core subtask be generate user spoken utterances semantic expressiveness, will be upper
Lower texts and pictures integrate and are dynamically selected sub-picture content with complete semantic expressiveness (having the interpretable semanteme of user's intention and field)
It is realized with preparing to carry out the surface of subgraph component in interface.
The explanation of user spoken utterances
User spoken utterances include important contextual information, which usually determines the dialogue between system and user
Process.For purposes of illustration, " user spoken utterances " include by the utterance of speech recognition system explanation and to conversational system
Written responses and inquiry.One core missions is that user spoken utterances are converted to more standard, formal and specification expression or semantic table
Show, this is closely related with semantic analysis task.In an embodiment of the present invention, explain that user talks about based on semantic analysis result
Language.From explanation results, conversational system generates the concept map represented for completing the related content of task.In implementation of the invention
Various types of semantic analysis mechanism are used in example.
Specifically, in a preferred embodiment, being abstracted meaning using the meaning representation language (MRL)-being recently introduced indicates
(AMR).AMR is a kind of analysis mechanisms for multilayer semantic interpretation, abstract representation and unified simple data structure.AMR pairs
Whole sentence semantics are formalized, and are specifically designed as normalized language and are indicated its meaning.It is equipped with extensive
English sentence semanteme field General Comments library.AMR expresses the meaning of sentence in figure, wherein node indicate concept (for example,
Event, entity, attribute), side indicates relationship (for example, part, agency, position).The semantic relation encoded in AMR figure can be by
It is construed to the combination of logical proposition or triple.AMR figure is that have root, oriented, acyclic, side label, leaf label figure, is designed
It annotates and reads and be easy to computer program to be easy to the mankind and calculate.It, which is explained, asserts that (" whom who has done assorted
"), mark concept, value and name entity.
Therefore, because these advantages, the semanteme that the preferred embodiment of the present invention expresses user spoken utterances using AMR figure contains
Justice.Preferably, AMR figure is adjusted by being stored in the domain knowledge at conversational system rear end.Analytic process includes looking into text
The label (token) for asking (i.e. user spoken utterances) is mapped between various ontology elements, such as concept, attribute and corresponding concepts
Relationship.Closely related with query construction by several semantic aspects of AMR annotation, the query construction is for example including entity/value, ratio
Compared with, polymerization, quantifier, conjunction, to potentially form the inquiry with labyrinth and implicit dialogue stream.
One of AMR annotation, which is noteworthy characterized by it, makes the element (such as word order and form syntactic marker) of surface syntactic structure
Abstract.Therefore, AMR figure can be converted into the concept map encoded to main semantic content.Nearest discussion
System dialog from AMR to first order logic.Natural language, which is converted to, formally indicates critically important, so that conversational system can make
With formally indicating to make inferences.First order logic computationally facilitates to be made inferences automatically.Therefore, AMR is very suitable to the purpose.
In order to which user explains that the semantic meaning of request is explained in request or system, it is sometimes desirable to across the semanteme of several sentences
Analysis.In the case where needing the operation, in a preferred embodiment, system runs semantic analyzer sequentially for sentence first
And obtain orderly semantic atlas.Depending on dialogue, can there are semantic overlapping or connection (discourse analysis/rhetoric hand among the drawings
Method).Based on the association between sentence, the update of the attribute of the same or like concept between user spoken utterances is executed.In the present invention
Preferred embodiment in, integrate the figure in dialog procedure.Several operations based on figure help for individual sentence figure to be integrated into
In one figure:
Merging-combination has the node of identical semantic meaning, such as entity, the semantic frame of the node, mark quoted jointly
Frame, Wiki.
Folding-processing name entity syntax rule, that is, hide no longer activity or in semantically related node.
Expansion-addition analyzer is not with language form presentation or the implicit nodes identified and side.
If relationship concatenate-is not detected, this is connected to generate the sequence of two figures with virtual ROOT node
Two figures.
Relationship between reconstruct-concept transfer, including separation and attachment edge.
Alignment-urtext index and concept node/elasticity search are with fast search subgraph.
If i is concept node, and i- > r- > j is the side with relationship r from i to j.In this case, by into
Enter while/it is outgoing while E (in) and the node collection of E (out) (path between i and j) connection will be (i ... j).It therefore, if can be with
Folding or the path between merge node i and j, then path becomes ij.
How Fig. 4 shows system by multiple user spoken utterances (sequence user carried out i.e. during the dialog procedure with system
Language) be converted to more succinct semantic meaning figure (Uniform semantic Figure 32 1 in Fig. 3).Semantic meaning Figure 40 0 includes multiple sections
Point, these nodes are themes relevant to college course selection website.As conventional, node by the circle that is connected by line or
Ellipse representation, these lines indicate relationship or the side of figure.According to the several differences carried out during the dialog procedure with conversational system
User spoken utterances edit original graph 400.The embodiment of the present invention is it will be recognized that in the presence of for being integrated into integration for original graph 400
Chance in Figure 40 1.For example, a part as subgraph union operation 403, can integrate two " course " nodes, thus
Single " course " node 405 is generated in integration map 401.It, can be in addition, as shown, a part as subgraph folding operation
Single " the credit: 3 " node 409 " credit " and " 3 " node 407 from original graph 400 being merged into integration map 401.This
Outside, subgraph expansion operation 411 can merge " algorithm " and " theory " node from original graph 400.The subgraph of generation is by integrating
" theory " node 413 in Figure 40 1 is used as title.
In a preferred embodiment of the invention, before it will integrate sentence figure and be merged into Uniform semantic figure, by sentence figure pressure
It shortens into and integrates sentence figure.
The semantic intergration of conversation content
The purpose of semantic intergration is the relevant information collected from various sources for completing task.
It is integrated into backend application specifically, the process is usually directed to the explanation that system is requested user and inquired
One or more databases.In a preferred embodiment, which further includes compiling autocommand and intermediate or final query result
It collects into unified format.These embodiments use the uniform context figure based on the semantic meaning figure from user spoken utterances.It is this logical
It can be constructed on different conversational systems with method.
Prior knowledge (such as core realm ontology, conversation tasks or main users are intended to) can be used to collect related letter
Breath.It can be given to export according to a plurality of types of integrated technologies (such as across sentence, back gear time, across interlocutor and across knowledge base)
User is intended to.Method as described above or its variant are based preferably on to obtain semantic meaning figure.
Next, system identifies related semantic content based on the information in field database, so that use can be formed
In the inquiry or order of the task of completion.In a preferred embodiment, mark is to complete in two different stages as semantic matches
, one in element rank;One in structure rank.For pel element, system-computed field node and sentence MR figure (or integration
Sentence figure) in nodename between Semantic Similarity.If similarity score is higher than specific threshold (determining by practice),
Then the node of graph in MR figure is mapped with the specific area concept in domain knowledge figure.For graph structure, in embodiment, system
Based on equivalence, partly overlap, superset and subset consider semantic dependency.If similarity score is higher than specific threshold,
Subgraph is mapped with field proposition, and field proposition generally corresponds to query graph.
The similarity score equation used in an embodiment of the present invention is given below:
Score (i, j)=a*equal (i, i) b*overlap (i, j)
+ c*superset (i, j)+d*subset (i, j)
In an embodiment of the present invention, using the inquiry based on standard drawing.Preferably, inquiry independently of be coupled to dialogue system
The type of the back-end system of system.By keeping inquiry independent, this facilitates the unnecessary details for inhibiting other modules, and in number
According to the robustness for improving frame when the patterns of change of library.Different from the level query language of such as SQL etc, which is designed to
There is no the simplification of the modeling process of specific syntax but more intuitively indicates.
The integrating process for generating context graph for the embodiment of the present invention is described in table 1.Assuming that K is core
Field concept, S are field proposition (triples).
Table 1: the algorithm integrated for grapheme
Table 1 describes given sentence sequence S and sky or existing Uniform semantic figure G, and the sentence in S is integrated with G.It is first
First, the direct overlapping nodes between system banner gi and G (b) update accordingly G;Then semantically by gi and domain knowledge K
Match, and updates accordingly Uniform semantic figure G.
Content selection
Content selection is intended to dynamically identify semantic expressiveness or the subgraph of context graph to present in the user interface.More
Say to body, system prediction when in interface to end user show what information with facilitate assisted by conversational system it is same
Shi Shixian target.Second target is to present to be predicted to be the information that most probable collects user feedback, such as based on using the past
The study of family session and predict.In principle, semantic expressiveness corresponds mainly to current session movement.For example, if current session is dynamic
Work is user to system offer information, then selectes subgraph and correspond to how system is based on field concept and logic to explain newest use
Family language.If current session movement is that system is made to provide the simplicity of explanation of query result, corresponding subgraph will be database
The expression of inquiry.Optionally, if initial data library inquiry do not generate effectively as a result, if system raw data base can be presented look into
The variant of inquiry.
But the subgraph for clearly corresponding to dialogue movement may be unavailable.In these cases, in preferred reality of the invention
Apply in example, use merotype is ranked up candidate subgraph, this be based on two main aspects: (1) give user be intended to, accordingly
How subgraph and user are intended to semantically related;(2) corresponding subgraph is given, user will have more a possibility that providing feedback
Greatly.Candidate son is obtained by the jump of the predetermined quantity for the concept node being intended to far from the user represented in Uniform semantic expression
Figure.If not providing user's intention, default content is the semantic meaning figure based on newest user spoken utterances.In preferred embodiment
In, system award is soundd out for the node and graph structure of conversation content design.Merotype is obtained to be provided by following equation:
For the node i of subgraph, there are gain (being expressed as q (i)), on condition that
Node I had previously occurred;
Node I is that domain is interpretable;
Node I can semantically edited/annotation by end user;And
Node I semantically with prior art conceptual dependency.
For the side (I, j) of semantic context figure, if there is information gain (being expressed as p (I, j)), then on condition that:
Side (I, j) does not occur previously;
Field can be used to explain in side (I, j);
Side (I, j) can be edited/annotation by end user;
Side (I, j) and preceding node or side have semantic dependency;
(I, j) is used to form inquiry on side;And
The preceding value of side (I, j) instruction concept.
System selects candidate subgraph relevant to current session activity first.If it is not then based on score (V',
E') subgraph is ranked up, and the highest subgraph of selected and sorted.Alternative embodiment of the invention is in different scoring equatioies
Quantify subgraph using similar set of factors, these factors include at least one of the following: concept level features, relationship level features
Or level features are discussed, and be then based on quantization factor and be ranked up to sub-collective drawing.
Graphical representation at user interface
The preferred embodiment of the present invention using can apparent structure collection with the intuitive of underlying semantics for rendering and interactive interface
It explains.The interface is used to collect the feedback of user, as described above, the interface is the semantic surface realization for conversational system
(SSR) interface.The vision that the task is related to ontology knowledge is presented, the dynamic of dialogue state updates (if providing knowing in interface
The time of knowledge and space distribution).
In order to which semantic expressiveness is presented to end user, conceptual simplicity and the maximum information in relation to current action are emphasized: (1)
End user should be appreciated that presentation;(2) session activity should have good covering;(3) design should indicate that dialogue state changes
Clearly identification;(4) interface should facilitate user to input.It in an embodiment of the present invention, can according to the type of user or specially
Industry know-how selects the type and quantity of graphic element.For example, can be to training the subject matter expert of conversational system to be in
Subgraph now more more dense than novice users (several subgraph components only can be presented to novice users).In efficiency, (more multielement makes
More feedbacks must be easier to provide with) with deposited between user friendly (more multielement makes us puzzled, especially for novice users)
Weighing.
In fact, session activity may be extremely complex, and whole figure or even only related subgraph will all reflect
This complexity.The option used in certain embodiments is characterization session activity, and they are set with user interface
Meter person prediction matches corresponding graphics feature most important to user.Presentation in interface is covered by end user and system two
The context that person generates.The purpose for indicating certain inputs of user again to user is how the system that informs the user understands them
Input so that they can agree or disagree with analysis result.Sometimes, notify task schedule also very to end user
It is important, so that user understands that task status and proposition can be used for completing the alternative of task.In addition to tradition talks with boundary
Except face, also addition visual configuration collection is to support the presentation of semantic information and request various forms of feedbacks, such as institute in Fig. 7
Show.
Another option used in other embodiments is the maximum language which subgraph element to have current action based on
Adopted expressivity, it is contemplated that calculate and plan the presentation in graphic user interface.This is a part of semantic integrity criterion, that is, which
Subgraph element set shows " the best overall condition " of the current state of the dialogue between user and conversational system.
In order to further increase performance, system optimization can be used for presenting the viewing area of subgraph under time and space constraint
Domain.That is, system constrains to optimize content based on room and time, wherein p if all related subgraphs (multiple) cannot be presented
(i) space occupied as node i, the space that q (i, j) is used as side (i, j) to occupy, S is as total available space.In general,
Real-time visual is time-sensitive, because user, which will say new language or desirable system, will reply previous language.Therefore, when
When user quickly provides new language, from the perspective of user satisfaction, complex figure expression will likely be unacceptable.Separately
On the one hand, when system is in test and subject matter expert is with interface alternation with the hypothesis of correcting system, the speed of dialogue
Degree may be slower, it is thus possible to potentially provide more information.Therefore, in an embodiment of the present invention, using nearest dialogue
Speed determines the time-constrain of presentation.
Although further constraint be present for a user must it is understood that therefore preferentially show most important node with
Side can also then show not too important node and side but if not too important node and side are to present to increase meaning.In table
Which provided in 2 for calculating the equation that show element in interface.
xi, yI, j∈ { 0,1 }
xi≥yI, j;xj≥yI, j
∑xi≤N
∑xi·f(i)+∑yI, jG (i, j)≤S
Table 2: the dynamic integrity of contextual information and content selection
Fig. 5 is the figure of the simplification Uniform semantic figure according to an embodiment of the invention for exemplary fields ontology.Show
Example ontology is related to college course selection website.The figure has multiple node 501-545, they indicate the theme in website, such as
The course node 509 and student's node 519 linked together by multiple sides (unnumbered).While holding the pass between respective nodes
The value of system.In an embodiment of the present invention, Uniform semantic figure can have more nodes and side.Conversational system operation it
Before, Uniform semantic figure is a domain body, by developer's manual construction of conversational system or from the data of back-end system
Library is derived automatically from.By traversing knowledge graph, it is related to other nodes which respective nodes is system can find out.In the operation of system
Period, with the meaning expression figure for sending each user spoken utterances of being grown up to be a useful person by atlas, each section of meaning expression figure is included into unification
In grapheme.
Provide two sentence examples: the first user spoken utterances " I needs to register other 3 credits " and second user language " I
Preferred theory course ".Using the sentence in the first user spoken utterances, MR figure is generated, and it is matched with field concept.So
Afterwards, can scheme to generate Uniform semantic figure from MR.After receiving second user language, another MR figure is generated, by itself and field
Concept matches, and new concept " theory " is correspondingly integrated into existing Uniform semantic figure.
Fig. 6 shows system and highlights a part of context graph dynamically to cause the feedback of user.In the figure, exist
User 601 is shown with 600,602 period of two states of the dialogue of conversational system.A part (i.e. subgraph) of context graph includes
Node 605,607 and 609, and be illustrated in the part SSR at interface of conversational system.Under state 600, user has started
The dialogue in relation to course selection is carried out with system.Based on 611 " showing computer science (CS) course to me " of inquiry, system is prominent
Show the course node 605 with value 30.User has executed annotation 613, and method is to highlight credit: 3 nodes 607, or
The side between course node 605 and credit node 607 is drawn to indicate that it should be a part of inquiry.Then in dialogue,
Arrival state 602, in state 602, user has inquired theoretical course, therefore highlights theoretical node 609 now.In state
Under 602, credit: 3 nodes 607 are no longer highlighted, this either cancels selection credit node because of user in interface,
Either user is via dialogue instruction to the unconcerned of credit hour number.
Fig. 7 illustrates how one embodiment that surface semantic expressiveness (SSR) is presented in dialog interface.In the figure, boundary
The first part 701 in face is exclusively used in the dialogue between user and conversational system.In the second part at interface, SSR shows that 703 show
A part of the related subgraph changed out with dialogue dynamic.At round (1), SSR shows that 703' corresponds to how system is analyzed
User inputs " there are also other 3 credit corusers for I " and shows two entities, i.e. course node 709 and credit node 711, and
Show entity/value to (credit: 3) in credit node.At each specific round, such as round (2) " finds 9 subjects, example
Such as AA, BB, CC " and round (3) " uh? how is those of Preston professor course? ", figure is dynamically presented in the interface SSR
Element 713,715,717, they are presented new nodes and remove old node (not shown) how to explain pair to user display system
It talks about and executes system response.For example, system can be in by the graphic element in the subgraph component for representing Uniform semantic figure
How existing system forms the inquiry to rear end.
In an embodiment of the present invention, independent graphic element is emphasized to indicate the node with certain semantic meaning.It uses
Color indicates which node is contributed, contributed by original bulk or by the inquiry institute tribute to back-end data base by dialogue
It offers.User can be handed over by selecting or cancelling selection respective element with the graphic element of the node or side that represent shown subgraph
Mutually.For example, user can draw the new line of representative edge to indicate that given node should include in the search.In no sufficient space
In the case where graphic element to show all interdependent nodes, line, such as dotted line can be presented in different ways, thus instruction two
A node does not contact directly in grapheme.User can choose the line to change the interface SSR to be in represent subgraph component
Element.Those skilled in the art will recognize that there are many alternate ways to highlight and select graphical interfaces not
Same element.
The extra elements at the interface used in certain embodiments include related User Status 719 or system mode 721
Simplicity of explanation.
In selectable alternate views 703 " in, user can choose the previous user language from dialogue come the system of checking
The whole contextual information used.In the figure, contextual information is arranged according to timeline 725, so that older context
Information is located at left side.In addition, in an embodiment of the present invention, one or more indicators 722,723,724 may be used to indicate
After latest content information is added as search criteria (name Preston), as a result how to become 0.
The surface semantic expressiveness of conversational system
In a preferred embodiment of the invention, surface semantic expressiveness (SSR) is used as the user interface of conversational system.It is generated
The visual representation of basic significant semantic information, times that session proxy completes user using the semantic information and system is being talked
Business.By the way that chat interface and SSR to be integrated, by disclosing how prediction task relevant information, and enable a system to pass through
Talk with user feedback of the direct request in relation to prediction (or hypothesis), end user further participates in than individually dialogue.It is this
Targetedly feedback is very valuable for training statistics conversational system.In a preferred embodiment, object-oriented dialogue conduct
The basis at the interface SSR, these dialogues are related to exchanging factural information between user and system and based on domain knowledge base.
Feedback based on SSR
The interface SSR uses graphic element collection abundant in which can be convenient, and information is presented to end user in these graphic elements
And it collects and feeds back from end user.Fig. 8 shows range and is clipped to the several example user feed back inputs for discussing rank from lexical level.
If providing new user spoken utterances 801 " I only needs other 3 credit theory course ", several system outputs are then triggered.?
In the example, system output includes position (slot) credit 803 and 3 as from location tags device (slot labeller)
Output.Another output is the search inquiry 805 in relation to " course ", indicates how to form back-end data library inquiry.Another
Output is that the invisible semantic relation 807 between " course " and " theory " is confirmed or specified by user.
Even subtask 809 in the past effectively can be presented to end user.Present in the past subtask purpose be using
Family is by the context relation of current utterance (" I only needs other 3 credit theory course ") and previous subtask, i.e. current utterance
It is the continuity of previous tasks.It (is usually taken and is filled out using static state intention with the art methods of the annotation of dialogue data for rendering
The simple more wheel positions filled) it compares, the dynamical feedback based on SSR used in an embodiment of the present invention is more general and flexible,
It is enable to annotate more complicated dialog strategy.
Fig. 9 shows following user/system dialog behind system logic in one embodiment of the invention, wherein U
Indicate that user, S indicate system, and the sequence of digital indication dialogue:
U1:, I is look for the CS course to register in next term.(P1)
Do S1: you want what kind of course, theoretical or programming? (P1)
Does U2: theoretical course provide how many credit? (P2)
S2 (a): sorry, I less understands.Are you which type of preferred course? (P1)
S2 (b): the theoretical course of all three of next term all provides three credits.(P2)
U3: how is programming course? (P2)
S3: two programming courses provide two credits, two offers, three credits.(P2)
U4: good, I am preferred to program course.(P1)
Dialog strategy is for determining how system generates next natural language response.In response to user query U1, chat
Agency inquires correct problem S1.But it is not to provide answer desired by system, but user's inquiry facilitates answer system
The problem of inquiring S1 U2.When sharing position (that is, being number and term number), the intention of U2 semantically with the intention phase of U1
It closes, but because specified class number is not necessarily identical as previous user intention, the intention of U2 is different.If assuming that quiet
State user optimizes dialog strategy in the case where being intended to, then conversational system still may be answered to attempt to fill by using S2 (a)
Position classification.But better dialog strategy will can respond new associated user by providing such as the information in S2 (b)
It is intended to, because it can trace back to the intention of the history in previous language.
By the inspiration for starting concept in human mind, system is used for being modeled to more complicated dialog strategy
" context starts (contextual priming) " process.Each starting corresponds to previous user and is intended to, these users are intended to
Be currently up or user be intended to share it is identical constraint or position/value collection.End user can provide related current utterance with
The relevant feedback of previously started or new (nearest) starting in dialogue.It, can be according to right by using context activation procedure
New user in words is intended to generate dialog strategy with Historical remarks, without the newest language for being limited to be used only in dialogue.
Task
The chat carried out for object-oriented task with end user interacts largely true by dialog strategy
Fixed, which is predetermined meter or pre-training in given field.Dialog strategy is set to adapt to practical application there are many challenges,
Especially in a case where: 1) basic field or task are frequently unfolded or 2) are difficult to a priori construct complicated dialogue management
Device.In addition, offline manual annotation is expensive and noisy.Feedback scheme based on SSR is effectively encouraged end user and is provided
Various user feedback mechanisms are to improve pre-designed or training dialog strategy.In an embodiment of the present invention, dialogue management is counted
For the user feedback based on SSR to be integrated in dialog strategy.
Dialogue management corresponds to two subtasks: dialogue state tracking and dialog strategy study.When being communicated with user, system
The distribution in the conversational system usually possible dialogue state of maintenance during referred to as dialogue state tracks is counted, which is used for
It is docked with domain knowledge base.It is also dialog strategy study preparation component.Strategy can indicate by the transition probabilities between state,
Wherein state is the expression of dialogue.In an embodiment of the present invention, state includes newest user/system dialog movement, such as is asked
It asks, information is shown, etiquette and corresponding position/value information.Dialog strategy directly determines how system generates next sound
It answers.
The model of proposition
In an embodiment of the present invention, using method neural network based, so that model framework can be constructed in sequence
On column flag data collection, without hand-made data.By combining multiple inputs (including user spoken utterances, associated
Domain location/value of dialogue movement, the starting of each context), model prediction is in the dialogue for semantically determining optimizer system response
Activity.
The frame of the modularization connection for dialogue management is shown in FIG. 10.State tracker 1001 is neural network,
It obtains user spoken utterances 1003Ut=(w1, w2 ..., wi) and generates an output of model, which is location tags device
1005 location tags sequence.Another output of model is the intention labels collection generated by intention marker 1007 and by starting
The starting label that tag 1009 generates.Policy learning device 1013 is another neural network, has and acts for dialogue
1015 and inquiring position 1017 output layer.Output dialogue movement 1019 and inquiring position 1021 are passed to from policy learning device
State tracker 1001.
Semantic coding
Language coding-sequence mark framework is embedded in capture similitude using word, but when processing is previously invisible or rare
It is affected when the word seen.The embodiment of the present invention uses average value bag for word insertion and Recognition with Recurrent Neural Network (RNN)
(bag-of-means).If being given at the language Ut=(w1, w2...wi) of time t, then corresponding to vector indicates to divide in RNN
The shot and long term not being encoded to backward in time t remembers (LSTM) hidden state.
Dialogue coding-each position/value is designated as<s=(m, d, g), and v>, wherein s is the position that type is m ∈ M;D is
The directionality and d ∈ { user → agency, agency → user } of information;The type of g expression variation, such as+,-,V is from g
Latest result value.When v is entity name (such as " condo " (for attribute type) or " New York (knob based on character string
About) " (it is used for place)) when, then the insertion of v is calculated as the insertion of character string text.Embodiment operating specificationization label indicates to come
The insertion of replacement values.For example, operating specificationization indicates " the time of meeting " replacement " 5pm ".Position s is encoded as location type dictionary
Index in Dm, and mutually concatenated with the index in change type dictionary Dg and a thermal potential in relation to directionality.Each round
A context starting Pi is generally corresponded to, semantically by the constraint of one group of s.Therefore, context starting is encoded as having
Have newest v relevant s concatenation.System also safeguards the look-up table of the contextual history of the s for each P for specific
Ground forms inquiry.
State tracker
Status tracking task is embodied as multitask Sequence Learning problem by the embodiment of the present invention.Make in an alternative embodiment
With the various methods for sequence mark task.Neural model updates the probability distribution p (sm) of the candidate value of location type, such as
It is one of new context starting or previous context starting.For the round t of each user, bidirectional valve controlled cycling element is used
(GRU) codings of user spoken utterances is calculated, concatenation ht=GRU (xt, ht-1) as the hidden state that forward and backward calculates.
The hiding activation of each s is calculated using another two-way GRU.
The supervised learning of dialog strategy
For use state tracking tags as input characteristics, the target of dialog strategy is to minimize label and shared network parameter
Associated losses function between the prediction p of θ (theta):
L (θ)=∑ H (yd,pd)
d∈{a,u,Ds}
Wherein a is dialogue movement, and u is intended to the classification distribution of (it is expected that entity), and Ds is the binary value of position.
ht=tanh (Wxt+U ht-1)
rt=σ (Wr xt+Ur ht-1)
h~ t=tanh (W xt+rtΘ(U ht-1))
zt=σ (Wz xt+U zht-1)
ht=(1-zt)Θht-1+ztΘ h~t
Wherein xtIt is the input in time t, htIt is the hidden state in time t, W and U are input and previous hidden state
Transformation matrix.Variable r and z are reset gate respectively and are updated.
The intensified learning of dialog strategy
In an embodiment of the present invention using intensified learning (RL) with the best of the conversational system for learning oriented mission
Dialog strategy.In order to combine the online feedback in relation to dialog strategy, using the method based on RL come optimisation strategy network.Target is
Maximize the reward J (θ) of dialogue
T=T
J (θ)=E [∑ γtR(st,at)|θ]
T=0
Wherein γt∈ [0,1) it is discount factor, R (at,st) it is when the prize for being directed to the movement a of state s in the activity of time t when
It encourages.
Depth Q network (DQN) parameterizes Q value function Q (a, s, P using deep neural network;θ).Network take when
Between t observation ot.Cycling element updates its hidden state based on both history and the insertion of current round.Then, model exports
The Q value of everything.Specifically, one comes from end user, and one from neck for two possible observations using reward
Domain knowledge base.The user feedback o observed via SSRUBased on (1) round rank success, that is, if current system response for
Completion task is useful;(2) Status Level success, that is, if dialogue state is correctly marked.The query result o observedQBy
It is determined for the restrained inquiry q of most likely location/value.Therefore, o is observedtIt can be by at、ot UAnd ot QDefinition.Use LSTM
Polymerization run bt=LSTM (ot,bt-1) on contextual information.
It applies a major issue of the method based on RL to be the large space due to probable value in practice and causes to restrain
Slowly.In the present invention, system can be substantially reduced empty for the search of movement based on the user feedback in relation to dialogue state
Between size.The model shields movement using user feedback, as confirmation (for example, user indicates "Yes" or "No") and specifies
(for example, user needs to specify the value).Model exports the Q value of all dialogue movements.
Although having described preferred operations environment and use-case, technology herein can be used for needing deployment services it is any its
In its operating environment.
As already described, above-mentioned function may be implemented as independent solution, such as be held by one or more hardware processors
The capable software-based functions of one or more, or may be used as managed service and (including be used as via SOAP/XML or REST
The Web service of formula interface).Being merely to illustrate property of specific hardware and software realization details purpose described herein and do not mean that limit
Make the range of described theme.
In more general terms, the calculating equipment in the context of published subject is the data processing for including hardware and software
System, and these entities are situated between by network (such as internet, Intranet, extranet, dedicated network) or any other communication
Matter or link communicate with one another.Application in data processing system provides primary branch for Web and other known service and agreement
It holds, including but not limited to the support to HTTP, FTP, SMTP, SOAP, XML, WSDL, UDDI and WSFL etc..About SOAP,
The information of WSDL, UDDI and WSFL can be obtained from World Wide Web Consortium (W3C), which is responsible for developing and safeguarding these standards;It closes
It can be obtained from internet engineering task group (IETF) in the further information of HTTP, FTP, SMTP and XML.
Other than environment based on cloud, technology described herein can be realized or be tied in various server-side architectures
The realization of these frameworks is closed, these frameworks include simple n-layer framework, web portal, association system etc..
In more general terms, theme described herein can take complete hardware embodiment, complete software embodiment or packet
The form of the embodiment of both elements containing hardware and software.In a preferred embodiment, functions of modules is implemented in software, and software includes
But be not limited to firmware, resident software, microcode etc..In addition, interface and function can take the form of computer program product, it should
Computer program product can be used from the computer for providing program code or computer-readable medium accesses so as to by computer
Or any instruction execution system use or in connection.For purposes of this description, computer is available or computer can
Reading medium can be any device that can include or store program, which can be commanded execution system, device or set
Standby use or in connection.Medium can be electricity, magnetic, optical, electromagnetic, infrared ray or semiconductor system (or device or
Equipment).The example of computer-readable medium includes semiconductor or solid-state memory, tape, movable computer disk, arbitrary access
Memory (RAM), read-only memory (ROM), hard disc and CD.The present exemplary of CD includes compact disk-read only memory
(CD-ROM), compact disk-read/write (CD-R/W) and DVD.Computer-readable medium is tangible non-transient article.
Computer program product can be following product: it is with program instruction (or program code) to realize one or more
A function.These instructions or code after remote data processing system downloading, can be stored in data by network
In computer readable storage medium in processing system.Alternatively, these instructions or code can store in server data processing
In computer readable storage medium in system, and it is suitable for downloading to Remote data processing in remote system by network
Computer readable storage medium in use.
In the exemplary embodiment, these technologies are realized in dedicated computing platform, preferably by one or more
It is realized in the software that reason device executes.With in the associated one or more data repositories of one or more processors or memory
Safeguard software, and software may be implemented as one or more computer programs.Generally speaking, this specialized hardware and software
Including above-mentioned function.
In a preferred embodiment, function is implemented as the attachment of existing cloud computing deployment rwan management solution RWAN provided herein
Or extension.
Although described above is the particular orders of the operation executed by certain embodiments of the present invention, it is to be understood that this
Kind sequence is exemplary, because alternative embodiment can combine certain operations perform the operation in a different order, is overlapped certain behaviour
Make etc..The embodiment, which may include specific feature, structure or feature, to be indicated to the reference of given embodiment in specification,
It but is not each embodiment centainly including the specific feature, structure or feature.
Finally, although having described the given component of system respectively, ordinarily skilled artisan will understand that, it can refer to given
It enables, combine or share certain functions in agenda, code section etc..
After describing the present invention, presently claimed content is as follows.
Claims (22)
1. a kind of method of the user interface of conversational system for rendering, comprising:
Create context graph of the Uniform semantic expression of the conversation content between user and conversational system as concept and relationship;
Based on current session activity, the subgraph component for the semantic context figure that dynamically mark will be presented to the user
Collection;And
Identified subgraph component set is presented in the user interface as the graphic element collection for representing corresponding concept and relationship.
2. according to the method described in claim 1, wherein, the subgraph component set is to be based on which concept and relationship currently
The system response in the current session activity is used to form by the conversational system to identify.
3. according to the method described in claim 2, further comprising:
Dynamically identify multiple subgraph components relevant to the current session activity of the semantic context figure;
Mark prevents that all constraint set of multiple subgraph components as corresponding multiple graphic elements are presented to the user;And
Based on semantic integrity criterion, optimize the subgraph component set identified in the multiple subgraph component to help finally to use
Family understands the graphic element collection in the user interface.
4. according to the method described in claim 3, wherein, the constraint is identified for presenting in the user interface
The time of subgraph component set and space constraint collection, the method further includes:
If it is corresponding to identify a possibility that providing feedback with the user in the graphic element for representing corresponding subgraph component
Subgraph component;
Compared with for total available space of the graphic element component set in the user interface, corresponding figure is presented in estimation
Space needed for shape element;And
Time needed for corresponding graphic element is presented in estimation in the user interface.
5. according to the method described in claim 1, wherein, the graphic element collection for representing corresponding concept and relationship is user
It can annotate, so that providing the user feedback to the conversational system from interacting with the user of the graphic element collection.
6. according to the method described in claim 1, further comprising: in response to subgraph group corresponding with the current session activity
Part is unavailable, is scored according to following item candidate subgraph component set: corresponding candidate's subgraph component and active user are intended to
In degree of correlation semantically;And if showing corresponding subgraph component, the user can by provide feedback
It can property;
Wherein, the candidate subgraph component passes through far from the general of the active user intention represented in the Uniform semantic expression
The jump of the predetermined quantity of node is read to obtain.
7. according to the method described in claim 1, wherein, the graphic element collection for representing corresponding subgraph component receives user
Input, the method further includes:
Corresponding graphic element is highlighted in the user interface;
The first user received for corresponding graphic element inputs;
First user input is supplied to the generation meaning table corresponding with the first user input of the conversational system
The atlas generator component of diagram;And
Being changed the Uniform semantic based on the meaning expression figure is indicated.
8. according to the method described in claim 1, wherein, the subgraph component set identified is according to the professional knowledge water of the user
It puts down to identify.
9. a kind of device, comprising:
Processor;
Computer storage is saved and is executed by the processor to search for the computer journey for the user interface that conversational system is presented
Sequence instruction, the computer program instructions include:
Program code is used to execute step according to any one of claim 1 to 8.
10. a kind of computer program for using in a data processing system in non-transient computer-readable medium produces
Product, the computer program product, which is saved, to be executed by the data processing system the calculating of the user interface of conversational system is presented
Machine program instruction, the computer program instructions include:
Program code is used to execute step according to any one of claim 1 to 8.
11. a kind of system of the user interface of conversational system for rendering, the system comprises for executing according to claim 1
To the device of step described in any one of 8.
12. a kind of method of the user interface of conversational system for rendering, comprising:
For each user spoken utterances that the user spoken utterances generated in the dialogue with the conversational system are concentrated, semantic meaning is determined
It indicates;
Semantic meaning expression is converted into corresponding sentence concept map;
First sentence concept map is integrated into uniform context figure;And
When with the dialogue of conversational system progress, the uniform context is updated based on new sentence concept map
Figure.
13. according to the method for claim 12, further comprise: based on be stored in the database of the conversational system
The semantic matches of domain knowledge update the uniform context figure.
14. according to the method for claim 12, further comprising: dynamically identifying the content in the uniform context figure
To present in the user interface.
15. according to the method for claim 14, further comprising: in the dialogue of mark and the conversational system most
New session activity is in semantically related concept and relationship.
16. according to the method for claim 13, further comprising:
Identify in the first sentence concept map in the uniform context figure concept and relationship semantically related general
Thought and relationship;And
The inquiry of the database to the conversational system is constructed according to the concept and relationship that are identified.
17. according to the method for claim 14, further comprising:
It is inputted based on newest user, identifies the change collection to concept, concept value and concept status;
Identify the associated component of the concept, the concept value and the concept status in the uniform context figure;With
And
It is based on the inquiry from the database to the conversational system as a result, mark to the correlation in the uniform context figure
The change of component.
18. according to the method for claim 17, further comprising:
Quantization with the associated set of factors of corresponding subgraph, the set of factors includes concept level features, relationship level features or opinion
State at least one of level features;And
Based on the factor quantified, the sub-collective drawing in the uniform context figure is ranked up.
19. according to the method for claim 14, further comprising: being intended to according to the new user in the dialogue and history is talked about
Language rather than the newest language in the dialogue generate dialog strategy.
20. a kind of device, comprising:
Processor;
Computer storage is saved and is referred to by processor execution with the computer program that the user interface of conversational system is presented
It enables, the computer program instructions include:
Program code is used to execute step described in any one of 2 to 19 according to claim 1.
21. a kind of computer program for using in a data processing system in non-transient computer-readable medium produces
Product, the computer program product, which is saved, to be executed by the data processing system the calculating of the user interface of conversational system is presented
Machine program instruction, the computer program instructions include:
Program code is used to execute step described in any one of 2 to 19 according to claim 1.
22. a kind of system of the user interface of conversational system for rendering, the system comprises for executing according to claim
The device of step described in any one of 12 to 19.
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US15/868987 | 2018-01-11 | ||
US15/868,987 US10845937B2 (en) | 2018-01-11 | 2018-01-11 | Semantic representation and realization for conversational systems |
US15/869000 | 2018-01-11 | ||
US15/869,000 US20190213284A1 (en) | 2018-01-11 | 2018-01-11 | Semantic representation and realization for conversational systems |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110059164A true CN110059164A (en) | 2019-07-26 |
CN110059164B CN110059164B (en) | 2023-06-06 |
Family
ID=67315866
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910022976.1A Active CN110059164B (en) | 2018-01-11 | 2019-01-10 | Method and system for presenting a user interface of a dialog system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110059164B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112507065A (en) * | 2020-11-18 | 2021-03-16 | 电子科技大学 | Code searching method based on annotation semantic information |
WO2021137941A1 (en) * | 2019-12-31 | 2021-07-08 | Microsoft Technology Licensing, Llc | Plan-guided response providing |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080235023A1 (en) * | 2002-06-03 | 2008-09-25 | Kennewick Robert A | Systems and methods for responding to natural language speech utterance |
US20090027392A1 (en) * | 2007-06-06 | 2009-01-29 | Apurva Rameshchandra Jadhav | Connection sub-graphs in entity relationship graphs |
CN101566988A (en) * | 2008-04-24 | 2009-10-28 | 华为技术有限公司 | Method, system and device for searching fuzzy semantics |
US20140379755A1 (en) * | 2013-03-21 | 2014-12-25 | Infosys Limited | Method and system for translating user keywords into semantic queries based on a domain vocabulary |
WO2015175313A1 (en) * | 2014-05-12 | 2015-11-19 | Google Inc. | Disambiguation of queries implicit to multiple entities |
US20160283463A1 (en) * | 2015-03-26 | 2016-09-29 | Tata Consultancy Services Limited | Context based conversation system |
-
2019
- 2019-01-10 CN CN201910022976.1A patent/CN110059164B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080235023A1 (en) * | 2002-06-03 | 2008-09-25 | Kennewick Robert A | Systems and methods for responding to natural language speech utterance |
US20090027392A1 (en) * | 2007-06-06 | 2009-01-29 | Apurva Rameshchandra Jadhav | Connection sub-graphs in entity relationship graphs |
CN101566988A (en) * | 2008-04-24 | 2009-10-28 | 华为技术有限公司 | Method, system and device for searching fuzzy semantics |
US20140379755A1 (en) * | 2013-03-21 | 2014-12-25 | Infosys Limited | Method and system for translating user keywords into semantic queries based on a domain vocabulary |
WO2015175313A1 (en) * | 2014-05-12 | 2015-11-19 | Google Inc. | Disambiguation of queries implicit to multiple entities |
US20160283463A1 (en) * | 2015-03-26 | 2016-09-29 | Tata Consultancy Services Limited | Context based conversation system |
Non-Patent Citations (2)
Title |
---|
ZE-QI LIN等: "Intelligent Development Environment and Software Knowledge Graph", 《JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY》 * |
黄鹏程等: "面向自然语言查询的知识搜索关键技术研究", 《优秀硕士论文全文数据库》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021137941A1 (en) * | 2019-12-31 | 2021-07-08 | Microsoft Technology Licensing, Llc | Plan-guided response providing |
CN112507065A (en) * | 2020-11-18 | 2021-03-16 | 电子科技大学 | Code searching method based on annotation semantic information |
CN112507065B (en) * | 2020-11-18 | 2022-07-12 | 电子科技大学 | Code searching method based on annotation semantic information |
Also Published As
Publication number | Publication date |
---|---|
CN110059164B (en) | 2023-06-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11797609B2 (en) | Semantic representation and realization for conversational systems | |
US20220004702A1 (en) | Robust expandable dialogue system | |
US11688022B2 (en) | Semantic representations using structural ontology for assistant systems | |
US10845937B2 (en) | Semantic representation and realization for conversational systems | |
US11210836B2 (en) | Applying artificial intelligence to generate motion information | |
US20230394102A1 (en) | Automatic navigation of interactive web documents | |
US8156060B2 (en) | Systems and methods for generating and implementing an interactive man-machine web interface based on natural language processing and avatar virtual agent based character | |
CN114429133A (en) | Relying on speech analysis to answer complex questions through neuro-machine reading understanding | |
Bernard et al. | Cognitive interaction with virtual assistants: From philosophical foundations to illustrative examples in aeronautics | |
JP2023156447A (en) | Natural language solution | |
Burgan | Dialogue systems and dialogue management | |
Huang et al. | Developing context-aware dialoguing services for a cloud-based robotic system | |
CN110059164A (en) | Semantic expressiveness and realization for conversational system | |
CN115392217A (en) | Techniques for preserving pruning flows | |
Finch et al. | An approach to inference-driven dialogue management within a social chatbot | |
KR20230130580A (en) | Autonomous generation, deployment, and personalization of real-time interactive digital agents | |
NL2025235B1 (en) | Updating constraints for computerized assistant actions | |
Mutis et al. | The Bimbot: Mediating technology for enacting coordination in teamwork collaboration. | |
Keizer et al. | Reasoning under uncertainty in natural language dialogue using Bayesian Networks | |
Goschnick | Shadowboard: an Agent Architecture for enacting a sophisticated Digital Self | |
Schüssel | Multimodal input fusion for companion technology | |
Geraci | Design and implementation of embodied conversational agents | |
Tumuluri et al. | Standardized representations and markup languages for multimodal interaction | |
Marriott et al. | gUI: Specifying complete user interaction | |
Goertzel et al. | OpenCog Hyperon: A Framework for AGI at the Human Level and Beyond |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |