CN110309276A - Electric car dialogue state management method and system - Google Patents
Electric car dialogue state management method and system Download PDFInfo
- Publication number
- CN110309276A CN110309276A CN201810267049.1A CN201810267049A CN110309276A CN 110309276 A CN110309276 A CN 110309276A CN 201810267049 A CN201810267049 A CN 201810267049A CN 110309276 A CN110309276 A CN 110309276A
- Authority
- CN
- China
- Prior art keywords
- information
- semantic understanding
- retrieval
- dialogue state
- field
- 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
- G06F16/3329—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
- G06F16/35—Clustering; Classification
- G06F16/355—Class or cluster creation or modification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of electric car dialogue state management method and systems, and the method includes parsing the demand information of input, the demand information, which is converted to semantic understanding, to be indicated；It indicates to be retrieved in search library according to the semantic understanding, obtains multiple search results；Every data in the search library includes preset or passing semantic understanding expression and pending mission bit stream corresponding with described preset or passing semantic understanding expression；It is reordered using machine learning order models to the search result, defines a top ranked search result and attach most importance to ranking results, the pending mission bit stream in the result that reorders described in output.Using electric car dialogue state management method of the invention or system, human-computer interaction can be made easy to maintain, more flexible, pluggable and being capable of autonomous learning.
Description
Technical field
The present invention relates to human-computer dialogue fields, a kind of dialogue state management method more particularly to electric car and are
System.
Background technique
With the development of science and technology, the mankind have entered artificial intelligence epoch, wisdom and energy of the artificial intelligence for the mankind of extending
Power simulates the thought process and intelligent behavior of the mankind, enables the machine to the competent complexity that usually requires human intelligence and could complete
Work.
Human-computer dialogue is a sub- direction of artificial intelligence field, is to allow people by human language, with such as gesture, language
The interactive mode of sound, text carries out the process of information exchange with computer.
In general, complete interactive system is related to voice technology, natural language processing, and knowledge base talks with shape
State maintenance etc..Planning and reasoning interaction logic are responsible in dialogue state maintenance.User intentionally gets satisfaction spy with specific purpose
The information or service of definite limitation condition, such as: when making a phone call, navigate, inquiring weather, listen to music, dialogue state maintenance needs to mention
Produce the intention and satisfaction of user.Because the demand of user can be more complicated, it may be necessary to divide more wheels to be stated, Yong Huye
The demand of oneself constantly may be modified or improved in dialog procedure.
Existing interactive method, dialogue state are difficult to safeguard, can not quickly modify that error rate is high, are easy to happen mistake
Accidentally, not flexible, it can not autonomous learning.
Summary of the invention
It is a primary object of the present invention to, overcome defect existing for the interactive method of existing electric car, and mention
A kind of new dialogue state management method and system out, the technical problem to be solved is that make its height robustness, it is more flexible, can
Plug and can data-driven, autonomous learning.
The object of the invention to solve the technical problems adopts the following technical solutions to realize.It proposes according to the present invention
A kind of electric car dialogue state management method, comprising the following steps: the demand information for parsing input, by the demand information
Be converted to semantic understanding expression；It indicates to be retrieved in search library according to the semantic understanding, obtains multiple search results；Institute
State every data in search library include preset or passing semantic understanding indicate and with described preset or passing language
Reason and good sense solution indicates corresponding pending mission bit stream；It is reordered using machine learning order models to the search result,
It defines a top ranked search result to attach most importance to ranking results, the pending task letter in the result that reorders described in output
Breath.
It the purpose of the present invention and solves its technical problem and can also be further achieved by the following technical measures.
Dialogue state management method above-mentioned, wherein semantic understanding expression includes: field field, it is described for recording
Field belonging to demand information；It is intended to field, for recording the intention of the demand information；Language slot position information list field,
For recording each language slot position information of the demand information, each language slot position information includes extracting from demand
The information of one word；The word information includes the content of word, or content and classification including word.
Dialogue state management method above-mentioned, wherein further include confidence level in each field that the semantic understanding indicates,
For indicating the order of accuarcy for recording content in the field.
Dialogue state management method above-mentioned, wherein the pending mission bit stream includes: aiming field, for recording
State the classification of task；Action field, for recording the task instructions to be executed；Field is replied, for recording described appoint
The information to be fed back of business.
Dialogue state management method above-mentioned, wherein described indicate to be examined in search library according to the semantic understanding
Rope further comprises: being indicated to obtain semantic retrieval request according to semantic understanding；In search library using it is complete or by the way of to semanteme
Retrieval request is retrieved, and the retrieval relevance score of the search result and the search result is obtained.
Dialogue state management method above-mentioned, wherein the semantic retrieval request in further include vehicle running state information,
Vehicle position information, on-board status information, userspersonal information, one or some in user preference information.
Dialogue state management method above-mentioned, wherein described carry out the search result using machine learning order models
It reorders, further are as follows: choose the search result of the forward preset quantity of the retrieval relevance score rank or choose institute
Multiple search results that retrieval relevance score is greater than default score are stated, are reordered described in progress.
Dialogue state management method above-mentioned, wherein described carry out the search result using machine learning order models
Reorder, further comprise: extract sequencing feature, the range of extraction include the search result, the semantic understanding indicate and
The retrieval relevance score；Sorting data, every sequence number are extracted from preset or passing dialogue state data
It include a semantic understanding table, semantic retrieval corresponding with semantic understanding expression request, corresponding pending task in
And the markup information of the sorting data；Machine learning training is carried out according to the sequencing feature and the sorting data, is obtained
To the machine learning order models；The sequence for obtaining each search result using the machine learning order models is related
Property score, it is described sequence Relevance scores it is higher search result ranking it is more forward.
Dialogue state management method above-mentioned, wherein the markup information includes: to indicate described in the sorting data
The information of the executive condition of pending task.
Dialogue state management method above-mentioned, further includes: retrieval relevance threshold test is carried out to the result that reorders
With at least one of relevance threshold detection of reordering；If the retrieval relevance score of the result that reorders reaches
Preset first threshold then determines to pass through the retrieval relevance threshold test；If the sequence of the result that reorders
Relevance scores reach preset second threshold, then determine through the relevance threshold detection of reordering；If the rearrangement
Sequence result does not pass through carried out retrieval relevance threshold test or not by the relevance threshold detection of reordering carried out, then
Feedback can not identify the demand.
Dialogue state management method above-mentioned, the dialogue are more wheels, the dialogue state management method further include: incite somebody to action this
The semantic understanding of demand information described in wheel dialogue indicates to indicate to carry out with the semantic understanding of the demand information of each wheel input before
In conjunction with.
The object of the invention to solve the technical problems also uses following technical scheme to realize.It is proposed according to the present invention
A kind of electric car dialogue state management system, including semantic understanding module, for parsing the demand information of input, by the need
Information is asked to be converted to semantic understanding expression；Retrieval module, for indicating to be retrieved in search library according to the semantic understanding,
Obtain multiple search results；Every data in the search library include preset or passing semantic understanding indicate and with institute
Stating preset or passing semantic understanding indicates corresponding pending mission bit stream；Reorder module, for utilizing machine learning
Order models reorder to the search result, define a top ranked search result and attach most importance to ranking results,
Pending mission bit stream in the result that reorders described in output.
It the purpose of the present invention and solves its technical problem and can also be further achieved by the following technical measures.
Dialogue state management system above-mentioned, wherein semantic understanding expression includes: field field, it is described for recording
Field belonging to demand information；It is intended to field, for recording the intention of the demand information；Language slot position information list field,
For recording each language slot position information of the demand information, each language slot position information includes extracting from demand
The information of one word；The word information includes the content of word, or content and classification including word.
Dialogue state management system above-mentioned, wherein in each field that the semantic understanding indicates further include: confidence level,
For indicating the order of accuarcy for recording content in the field.
Dialogue state management system above-mentioned, wherein the pending mission bit stream includes: aiming field, for recording
State the classification of task；Action field, for recording the task instructions to be executed；Field is replied, for recording described appoint
The information to be fed back of business.
Dialogue state management system above-mentioned, wherein described be specifically used for according to the retrieval module: according to semantic understanding
Expression obtains semantic retrieval request；In search library using it is complete or by the way of semantic retrieval request is retrieved, obtain described
The retrieval relevance score of search result and the search result.
Dialogue state management system above-mentioned, wherein the semantic retrieval request in further include vehicle running state information,
Vehicle position information, on-board status information, userspersonal information, one or some in user preference information.
Dialogue state management system above-mentioned, wherein the module that reorders is specifically used for: choosing the retrieval relevance
The search result or the selection retrieval relevance score of the forward preset quantity of score rank are greater than the multiple of default score
Search result reorders described in progress.
Dialogue state management system above-mentioned, wherein the module that reorders further comprises: sequencing feature extracting unit,
For extracting sequencing feature, the range of extraction includes the search result, the semantic understanding indicates and the retrieval relevance
Score；Sorting data extracting unit, for extracting sorting data from preset or passing dialogue state data, described in every
Include in sorting data semantic understanding indicates, semantic retrieval corresponding with semantic understanding expression is requested, it is corresponding to
The markup information of execution task and the sorting data；Training unit, for according to the sequencing feature and the sequence number
According to machine learning training is carried out, the machine learning order models are obtained；Search result sequencing unit, for utilizing the machine
Study order models obtain the sequence Relevance scores of each search result, the higher retrieval of the sequence Relevance scores
As a result ranking is more forward.
Dialogue state management system above-mentioned, wherein the markup information includes: to indicate described in the sorting data
The information of the executive condition of pending task.
Dialogue state management system above-mentioned, further includes relevance threshold detection module, is used for: to the result that reorders
Carry out retrieval relevance threshold test and at least one of relevance threshold detection of reordering；If the result that reorders
The retrieval relevance score reaches preset first threshold, then determines to pass through the retrieval relevance threshold test；If institute
The sequence Relevance scores for stating the result that reorders reach preset second threshold, then determine through the correlation that reorders
Threshold test；If the result that reorders does not pass through carried out retrieval relevance threshold test or not by carried out weight
The relevance threshold that sorts detection, then feedback can not identify the demand.
Dialogue state management system above-mentioned, the dialogue are more wheels, and the dialogue state management system further includes semanteme
It updates and maintenance module, the semantic understanding for demand information described in talking with epicycle indicates the demand with each wheel input before
The semantic understanding expression of information is combined.
The object of the invention to solve the technical problems also uses following technical scheme to realize.It is proposed according to the present invention
A kind of controller comprising memory and processor, the memory are stored with program, and described program is held by the processor
The step of can be realized aforementioned any the method when row.
The object of the invention to solve the technical problems also uses following technical scheme to realize.It is proposed according to the present invention
A kind of computer readable storage medium, for storing computer instruction, described instruction is when by a computer or processor execution
The step of realizing aforementioned any the method.
The present invention has obvious advantages and beneficial effects compared with the existing technology.By above-mentioned technical proposal, the present invention
Dialogue state management method and system can reach comparable technical progress and practicability, and with the extensive exploitation value in industry
Value, at least has the advantage that
1, decouple: dialogue state management method of the invention and system will by the way of retrieving to search library
Dialogue state tracing module and man-machine interactive system decouple, and reduce the complexity of whole system；
2, height robustness: after semantic understanding, when the semantic incomplete or system Extracting Information of user's expression is not complete or has
In the case where accidentally, based on it is complete or mode retrieve, still correct user can be intended to be retrieved, greatly enhance the Shandong of system
Stick and adaptability can better meet user demand；
3, flexibly plug: dialogue state management method and system based on search engine are, it can be achieved that real-time or semireal time
Data are synchronous, increase some dialogue states newly, available data is deleted or modified, all can timely be reacted in system；
4, autonomous learning: data-driven after having accumulated certain data volume, is tied retrieval by the study of order models
Fruit carries out two minor sorts, to obtain better result.The acquisition of data can be by manually marking and a variety of sides such as system log
Formula obtains.Via the dialogue state management method of data-driven with the increase of data volume, the learning ability of system promoted,
User demand is satisfied.
The above description is only an overview of the technical scheme of the present invention, in order to better understand the technical means of the present invention,
And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects, features and advantages of the invention can
It is clearer and more comprehensible, it is special below to lift preferred embodiment, and cooperate attached drawing, detailed description are as follows.
Detailed description of the invention
Fig. 1 is the flow diagram of one embodiment of electric car dialogue state management method of the present invention.
Fig. 2 is the flow diagram of another embodiment of electric car dialogue state management method of the present invention.
Fig. 3 is the schematic illustration of one embodiment of electric car dialogue state management method of the present invention.
Fig. 4 is the schematic diagram for the data stored in search library of the invention.
Fig. 5 is the structural block diagram of one embodiment of electric car dialogue state management system of the present invention.
Fig. 6 is the structural block diagram of another embodiment of electric car dialogue state management system of the present invention.
Specific embodiment
It is of the invention to reach the technical means and efficacy that predetermined goal of the invention is taken further to illustrate, below in conjunction with
Attached drawing and preferred embodiment, to electric car dialogue state management method proposed according to the present invention and system its specific embodiment party
Formula, method, structure, feature and its effect, detailed description is as follows.
Fig. 1 is the schematic flow chart of electric car dialogue state management method one embodiment of the invention.It please refers to
Fig. 1, the exemplary electric car dialogue state management method of the present invention include:
Step S11 parses the demand of input, and the demand information of input, which is converted to dialogue state management system, to be utilized
Semantic understanding indicate.
Step S12 is indicated according to the semantic understanding parsed from the demand information, is arranged by search engine and machine learning
Sequence model obtains the corresponding pending task of the demand information.
Specifically, first can indicate to be retrieved in search library according to semantic understanding, therefrom obtain and the semantic understanding table
The every data retrieved might as well be defined as a search result by the forward a plurality of data of the relevance rank shown, be recycled
Machine learning order models reorder to these search results, obtain from multiple search results maximally related with the demand
One search result, and define this maximally related search result and attach most importance to ranking results, export the result that reorders；It should be noted that
, every data in the search library includes that preset or passing semantic understanding indicates and preset or passing with this
Semantic understanding indicates corresponding pending mission bit stream, thus by search library to the semantic understanding in every data indicate into
The search result and the result that reorders by reordering that row retrieval obtains can include pending mission bit stream.
Fig. 2 is the schematic flow chart of electric car dialogue state one specific embodiment of management method of the invention.Fig. 3
It is the schematic block diagram of electric car dialogue state management method one embodiment of the invention.Please refer to Fig. 2 and figure
3, the exemplary electric car dialogue state management method of the present invention includes:
The demand information (Query) of input is converted to the form of semantic understanding table (Belief table) by step S21.It needs to infuse
Meaning, the mode for inputting demand are not limited to voice input, are also possible to picture, the interactive modes such as text, touch, gesture, or more
Kind interactive mode is used in combination.
Specifically, semantic understanding table therein is made of the semantic understanding feature field of various dimensions, comprising:
Field field (Domain), for recording the field of demand information, such as navigation, call, weather lookup, music
Deng；In fact, each field can be considered as to a kind of classification of demand information；
It is intended to field (Intent), for recording the specifically intended information in demand information, it is in fact possible to by the intention
Information is considered as the specific descriptions of the information indicated to field field, or is considered as each field of demand information and further divides
Class, such as when field field is phone, it is intended that field can anticipate to dial number, receive calls, when field field is music
Figure field can be to search for and play music, pause music etc.；
Language slot position information list field (Slot List), for recording all slot position information of the demand information
(Slot), what each slot position information recorded is the information for the word released from the demand information of input；Word letter
Breath includes the content of word, can also include the classification of the word, specifically can be the classification of the word meaning of a word.In fact, can
It, will be infinite using the classification of word as the title of a slot position, and using the content of word as the design parameter of the slot position
Word is mapped as limited slot position, and in retrieval or sequence step later, can title to slot position without to slot position
Parameter is retrieved or is reordered, such as word is the situation of proper noun；Or it can be using the content of word as a slot position
Title, and the slot position do not have parameter, such as word be a part of verb situation.
Using by field field, be intended to field, the semantic understanding that language slot position information list field these three fields form
Table can completely express the demand information of input.
Optionally, semantic understanding table may also include problem types field (Question Type), for recording demand information
The problem of type, such as can be indicated with y/n the demand information be judge whether class problem, indicate the demand information with where
To inquire location category problem, indicate that the demand information is inquiry quantity class problem with how_many.
Further, in each semantic understanding feature field, it may include record the accurate of content in the field for indicating
The part of degree.For example, may include confidence level in field field, intention field, language slot position information list field
(Confidence), confidence level is real number of the range between 0 to 1, the bigger expression this feature of the confidence level of a feature field
The order of accuarcy of content recorded in field is higher.
Each semantic understanding feature field that demand information can be obtained by machine learning, obtains language by syntactic analysis
Say each slot position information in slot position information list field.
In a specific example, if the demand information of input is that { I wants to go to restaurant to Query=, gives me from five road junctions
Navigate to Wangjing }, then the semantic understanding table obtained may is that
Domain therein: " navigation ", confidence=0.8 are field field, indicate the field of the demand
For navigation (navigation), and it is 0.8 that field, which is the order of accuarcy of navigation,.Intent is intended to field, slot for indicating
List is used for representation language slot position information list field, and specific format is similar with field field.Particularly, " tag ": " restaurant ", "
Five road junction from_loc ": " ", " dest_loc ": " Wangjing " is slot position information (slot) respectively.Specifically, for example in " from_
Five road junction loc ": " " in, " five road junctions " this word in demand is had recorded in the form of slot position: " from_loc " has recorded this
The title of slot position, in fact and the classification at " five road junctions ", " five road junctions " have recorded the parameter of the slot position.
Step S22 is updated semantic understanding and maintains in mostly wheel human-computer dialogue.By sheet obtained in step S21
The semantic understanding of demand is combined with the semantic understanding for respectively taking turns demand before in wheel dialogue, so that in mostly wheel human-computer dialogues,
Semantic understanding update by wheel, addition, is deleted, and maintains the semantic understanding information of each round interaction, gradually clearly to need
It asks.
When user interacts with dialogue state management system progress T wheel, the T demand taken turns is understood into (Query
Belief (T)) with the aggregate demand respectively taken turns before T wheel understand that (Context (T-1)) is combined and updates, generate T wheel
Aggregate demand understand (Context (T)), be embodied formula are as follows: Context (T-1)+Query Belief (T)=Context
(T)。
It may be noted that the update and maintenance of semantic historical context can be carried out both for semantic understanding table (belief table)
, after user interacts with dialogue state management system, the semantic understanding table of demand will be updated, adds, delete.
Step S23, the semantic understanding after the update according to obtained in step S22 maintains indicates, by search engine and again
Order models, the result as a result, this reorders that reordered include pending task, which is sent to corresponding execution
Module or control system, to execute operation according to the task and reply user.The mission bit stream may include target (Goal)
Field, movement (Action) field and reply field.Target therein is the classification of the task, can be used for judging that the task should be sent out
Which execution module given；Movement is the task specific instruction or order to be executed, such as can be and change vehicle operation
The control command of state, or adjusted the instruction of the service in vehicle control system；Reply is that the task needs to feed back to use
The information at family, including but not limited to text reply, are based on spatial term (NLG, Natural Language
Generation reply), such as speech answering (TTS, Text to Speech), response (View) of interactive interface etc..
It should be noted that semantic understanding (Belief) is the combination of numerous elements, it is equivalent to variable X=(x1,x2,
x3...), and pending task is a limited operation class set, is equivalent to preferred set y.The effect of step S23 can be managed
Xie Wei is established by semantic understanding to the mapping of pending task, y=f (X), and the mapping is by search engine and sequence mould
What type was realized.In fact, establishing by semantic understanding to the mapping of pending task, it is understood that be divided semantic understanding
Class is mapped to a classification set via elements numerous in semantic understanding, and the element in this classification set is limited wait hold
The classification of row task, such as { make a phone call, listen song, navigate }.Also note that arriving, although pending set of tasks, which is one, to be had
Limit set, but in the mapped, finally obtained pending task can be with the information that semantic understanding hands down, for example, semantic reason
What is recorded in solution is the white windmill of Zhou Jielun, then in the pending task obtained other than including the demand of " listening song ", also wraps
Including singer is Zhou Jielun, and title of the song is white windmill.
The set of pending task can carry out different fractionations according to the difference of specific field, such as in onboard system,
Point of interest (POI, Point of Interest) inquiry, neighbouring search inquiry etc. can be split into navigation field, and in matchmaker
Inquiry song, broadcasting/pause song etc. can be divided into body stream field.
Optionally, the result that can will also reorder is recorded in history log, includes in addition semantic understanding in the result that reorders
The situation of information also carries out the semantic of step S22 using the result that reorders and updates and maintain, for can base in step S22
The semantic understanding for including in the result that reorders carries out the update and maintenance of next round.
In one embodiment, step S23 may include step in detail below:
Step S231 retrieves information to generative semantics retrieval request with other according to semantic understanding table.It does not limit therein
The type of other retrieval information, such as in one example, other retrieval information may include vehicle running state information (speed),
Vehicle position information, on-board status information, userspersonal information's (such as gender, age), (such as user likes walking user preference information
High speed still walk on path) etc. in one or some.Optionally, when in the feature field of semantic understanding table include confidence level
When, semantic retrieval request in can also include confidence level (can not also include confidence level).
Step S232, using search engine, to the semantic retrieval request obtained by step S231 using complete in search library
Or mode retrieved, obtain search result and obtain the retrieval relevance score of search result.Every number in search library
Information and pending mission bit stream are indicated according to comprising semantic understanding, also may include other corresponding retrieval information, such as can be
The semantic understanding of one preset or passing demand indicates and pending task corresponding with the preset or passing demand
Information.Include the information of pending task in each search result, also may include that corresponding semantic understanding indicates information and other
Retrieve information.The pending mission bit stream can be made of aiming field, action field and reply field.Retrieval relevance therein
Score is a kind of index of characterization search result order of accuarcy commonly obtained by search engine, can be according to any means known meter
It calculates and obtains.In one embodiment, can choose the forward preset quantity of retrieval relevance score rank search result or
Choose multiple search results that retrieval relevance score is greater than default score；In addition can to the search result that these are selected, according to
Ranking is carried out from high to low according to retrieval relevance score.
Above-mentioned complete or mode is meant: if semantic retrieval request content be domain: " call ", intent: "
Search number " }, then using it is complete or by the way of retrieved when, can be included all knots of { domain: " call " }
Fruit, and can also be included all results of { intent: " search number " }.
It is since semantic understanding state can be by field (Domain), intention by the way of being retrieved to search library
(Intent) it is composed with slot position information (Slot), the quantity for combining sum M reaches M=2|Domain|*2|Intent|*2|Slot|, wherein | Domain |, | Intent |, | Slot | field, intention, the respective set sizes of slot position information are respectively indicated, from
And the number of combinations of semantic understanding state reaches index rank, and all semantic understanding shapes can not be covered with common processing mode
State.Most related a few datas can be obtained from data acquisition system limited in search library by retrieving to search library, so as to
Problem complexity is effectively reduced.And with the continuous expansion of data acquisition system in search library and increase, the correlation of retrieval can obtain
It is improved to significant.
In search engine, the set of pending task can be added or modify at any time, at any time in search engine platform,
It increases some dialogue states newly, available data is deleted or modified, to realize that real-time or semireal time data are synchronous, make dialogue state
Management system property easy to maintain.
Dialogue state search engine is also used to first be indexed for inquiring to search library foundation for semantic understanding table.Establish index
Data source in history log or artificial mark, every data includes field, intention, slot position information, target, movement, reply
And other relevant fields.
Fig. 4 is the schematic diagram for the data stored in search library.Every a line in Fig. 4 represents a number in a search library
According to each column represent the parameter information that data include, in fact if the data source in search library is in human-computer dialogue history day
Will, then every data can be the dialogue state of a passing demand.Every data includes semantic understanding table information, other retrievals
Information and pending mission bit stream, semantic understanding table information therein specifically include field, intention, slot position information, it is therein its
It includes gender, age that he, which retrieves information, and pending mission bit stream therein includes target, movement, reply.Particularly, slot position is believed
Breath can only include input demand in each word classification, without the content including word, for example, in Fig. 4 the first row data
In slot position information list include two slot positions title: From_loc and dest_loc, and do not record the tool of the two slot positions
Hold " San Litun ", " Tsinghua University " in vivo.
It when being retrieved to search library, is requested according to semantic retrieval, to field, intention, slot position information, property in search library
Not, the age is retrieved；And return search result be whole data comprising target, movement, reply, so as to by pair
The retrieval of search library obtains corresponding pending mission bit stream.Thus the column and record of semantic understanding table information will might as well be recorded
The column of other retrieval information are referred to as retrieval column.
Although each retrieval column all may be used in some instances it should be noted that the example in figure does not include confidence level
With corresponding confidence level.In addition, retrieval column are not limited in these enumerated in figure, it can also include other column, and can
To add at any time.
Step S233 reorders to from the received search result of step S232, obtains using machine learning order models
To with maximally related one of semantic understanding (Belief) as a result, might as well be known as reordering as a result, the result that will reorder is sent to pair
The execution module or control system answered, to execute the pending task for including in the result that reorders.
In a kind of specific example of step S233, step S233 specifically includes the following steps:
Step S2331 extracts sequencing feature from all search results, semantic retrieval request, retrieval relevance score；Its
In sequencing feature, including but not limited to: field, intention, slot position information, confidence level, retrieval relevance score, vehicle-mounted state,
Speed, vehicle location, userspersonal information, user preference etc., which, which specifically extracts, as sequencing feature can be artificial select
, in general, the sequencing feature of extraction is The more the better；
Step S2332 extracts sorting data from preset or passing dialogue state data, wraps in every sorting data
It includes a semantic understanding expression, semantic retrieval corresponding with semantic understanding expression request, corresponding pending task and is somebody's turn to do
The markup information of sorting data, such as sorting data can be extracted from interactive dialogue state history log data；It should
The particular content of markup information can be different according to the difference of mark strategy, for example, when using user feedback formula mark strategy,
The markup information can be the specific feedback of user, and when tactful using heuristic mark, which can be expression row
The information of the executive condition of pending task of the ordinal number in, such as whether execute, whether complete, whether by user negate；It is optional
, cleaning data can be also carried out, for data to be examined and verified again, including check data consistency, processing is invalid
Value and missing values, it is therefore intended that delete infull invalid data；
Step S2333 carries out machine learning training according to sequencing feature and sorting data, obtains machine learning sequence mould
Type；
Step S2334, using by learning training the machine learning order models, obtain each search result based on
The sequence Relevance scores of the sequencing feature；The sequence Relevance scores are intended to indicate that the search result and this dialogue of input
Demand information be corresponding order of accuarcy index, the higher search result of sequence Relevance scores more with this dialogue
Demand information is corresponding, search result ranking is also more forward；It should be noted that with the increasing of dialogue state historical data
More, the learning ability of order models can be enhanced, and the accuracy of the sequence Relevance scores obtained by order models also can be with
Increase；
Step S2335 will sort one of Relevance scores highest (reordering top ranked) in all search results
The control system of corresponding execution module or vehicle is sent to as the result that reorders, to execute appointing in the result that reorders
Business.
It may be noted that the mark that a variety of strategies are ranked up data can be used, it is not limited to following manner:
Artificial mark strategy, the artificial demand for marking the result that dialogue state management system obtains and whether meeting input；
User feedback formula mark strategy, is labeled according to the feedback of a large number of users, specifically can be, set in human-computer interaction
It is standby upper, the button of " thumbing up " and " stepping on " is set, user is allowed to feed back, " thumbing up " mostly be considered as correlation, the view of " stepping on " mostly
It is uncorrelated；
Heuristic mark strategy, specifically can be, in daily record data, dialogue state management system returns to user one
Pending task, if the pending task that system obtains is negated in next interaction by user or this is pending
Task is not performed or is not completed, then the data is labeled as uncorrelated；If user does not negate in next round interaction
The pending task or default receives the pending task or the pending task is completed that system obtains, then the data
It is labeled as correlation.
Further, relevance threshold detection can be also carried out in step S23, detect whether the result that reorders passes through retrieval phase
Close property threshold test and/or whether pass through reorder relevance threshold detection.Wherein if the retrieval relevance for the result that reorders
Score reaches preset first threshold, then determines by the retrieval relevance threshold test, if the sequence phase for the result that reorders
Closing property score reaches preset second threshold, then determines relevance threshold detection of reordering by this.If the result that reorders is logical
The threshold test carried out has been crossed, then has exported the result that reorders；If the result that reorders does not pass through any one threshold test,
This is not exported then to reorder as a result, but feedback can not identify the demand of input.Optionally, the correlation that reorders can only be carried out
Threshold test.
In a specific example of step S23, the demand of input is that " I wants to go to restaurant, navigates to from five road junctions to me
Wangjing ".According to the demand, semantic retrieval request can be obtained, semantic retrieval request may is that
{ domain: " navigation ", confidence=0.8；
Intent: " navigate ", confidence=0.9；
Slot list={ " tag ": " restaurant ", confidence=0.9；" from_loc ": " five road junctions ",
Confidence=0.6；" dest_loc ": " Wangjing ", confidence=0.7 } }.
By semantic retrieval request in all feature field and corresponding confidence level (Confidence) be put into search library
By it is complete or in a manner of retrieved, obtain the search result of N item before retrieval relevance score rank.In this example, retrieval knot
Fruit may is that
Search result 1 (retrieval relevance score=0.8): target=" searching restaurant nearby ", movement=" adjusted and purchased by group
App searches restaurant nearby ", reply=" finding for you to go to restaurant "；
Search result 2 (retrieval relevance score=0.75): target=" navigation ", movement=" have adjusted navigation app, guide
App injection of navigating originates and purpose parameter ", reply=" $ dest_loc is being navigate to for you, be your programme path ... "；
…
Search result N (retrieval relevance score=0.70) ....
It extracts sequencing feature, extract sorting data, order models are obtained by machine learning training, then by each retrieval
As a result the order models are brought into obtain sequence Relevance scores.The feature wherein extracted include field, intention, slot position information,
Vehicle-mounted state, speed, vehicle location, userspersonal information, user preference, the confidence level of these features and this N item retrieval knot
Retrieval relevance score of fruit etc..Finally, if sequence Relevance scores highest that search result 2 obtains and being greater than manually is set
Fixed relevance threshold (for example score section is 0~1.0 point, threshold value may be set to 0.7 point), then select search result 2 as
The output for the demand that dialogue state management system talks with epicycle.After completing more wheel dialogues, onboard system will execute final
One takes turns the search result of selection.
Include two demands in the original demands information of input it may be noted that in this example, one be navigation the other is
Restaurant is looked for, and navigation therein is only the real demand of user.But in the search result obtained by search engine, really use
The retrieval relevance score of family demand (search result 2) may not be highest.To solve this problem, the present invention is added to
The step of reordering carries out sequence once based on machine learning to search result again, to obtain the real demand of user.
Fig. 5 is the schematic block diagram of electric car dialogue state management system one embodiment of the invention.It please join
Fig. 5 is read, the exemplary electric car dialogue state management system of the present invention includes semantic understanding (Query Belief) module 100,
Dialogue state tracks (DST, Dialog State Track) module 200；Semantic understanding module 100 therein is for parsing input
Demand, the demand information of input, which is converted to the semantic understanding that dialogue state management system can utilize, to be indicated；Dialogue state
Tracing module 200 is used to be indicated according to the semantic understanding parsed from the demand information, be arranged by search engine and machine learning
Sequence model obtains the corresponding pending task of the demand information.
Dialogue state tracing module 200 therein specifically includes retrieval submodule 210 and the submodule 220 that reorders.Retrieval
Submodule is used to indicate to be retrieved in search library according to semantic understanding, therefrom obtains the correlation indicated with the semantic understanding
The every data retrieved might as well be defined as a search result by a plurality of data in the top.The submodule that reorders is used for
It is reordered, is obtained from multiple search results and the demand most phase to these search results using machine learning order models
The search result closed, and define this maximally related search result and attach most importance to ranking results, export the result that reorders.It needs
It is noted that every data in the search library include preset or passing semantic understanding indicate and with the preset or mistake
Past semantic understanding indicates corresponding pending mission bit stream, thus by search library to the semantic understanding table in every data
Show that the search result retrieved and the result that reorders by reordering can include pending mission bit stream.
Fig. 6 is the schematic block diagram of electric car dialogue state one specific embodiment of management system of the invention.
Fig. 6 and Fig. 3 are please referred to, exemplary electric car dialogue state management system includes semantic understanding (Query according to the present invention
Belief) module 100, semantic update track (DST) module with (Update Belief) module 300 and dialogue state is maintained
200。
Semantic understanding module 100 therein, for demand (Query) information of input to be converted to semantic understanding table
The form of (Belief table), and the semantic understanding table is sent to semantic update and maintenance module 300.It may be noted that input demand
Mode be not limited to voice input, be also possible to picture, the interactive modes such as text, touch, gesture or a variety of interactive modes
It is used in combination.
Specifically, semantic understanding table therein is made of the semantic understanding feature field of various dimensions, comprising:
Field field (Domain), for recording the field of demand information, such as navigation, call, weather lookup, music
Deng；In fact, each field can be considered as to a kind of classification of demand information；
It is intended to field (Intent), for recording the specifically intended information in demand information, it is in fact possible to by the intention
Information is considered as the specific descriptions of the information indicated to field field or is considered as the further classification in each field of demand information,
Such as field field is when being phone, it is intended that field can be to dial number, receive calls, when field field is music, it is intended that
Field can be to search for and play music, pause music etc.；
Language slot position information list field (Slot List), for recording all slot position information of the demand information
(Slot), what each slot position information recorded is the information for the word released from the demand information of input；Word letter
Breath includes the content of word, can also include the classification of the word, specifically can be the classification of the word meaning of a word.In fact, can
It, will be infinite using the classification of word as the title of a slot position, and using the content of word as the design parameter of the slot position
Word is mapped as limited slot position, and in dialogue state tracing module 200, can title to slot position without the ginseng to slot position
Number is retrieved or is sorted, such as word is the situation of proper noun；Or it can be using the content of word as the name of a slot position
Claim, and the slot position does not have parameter, such as word is the situation of a part of verb.
Using by field field, be intended to field, the semantic understanding that language slot position information list field these three fields form
Table can completely express the demand information of input.
Optionally, semantic understanding table may also include problem types field (Question Type), for indicating demand information
The problem of type, such as can be indicated with y/n the demand information be judge whether class problem, indicate the demand information with where
To inquire location category problem, indicate that the demand information is inquiry quantity class problem with how_many.
Further, in each semantic understanding feature field, it may include record the accurate of content in the field for indicating
The part of degree.For example, may include confidence level in field field, intention field, language slot position information list field
(Confidence), confidence level is real number of the range between 0 to 1, the bigger expression this feature of the confidence level of a feature field
The order of accuarcy of content recorded in field is higher.
Each semantic understanding feature field that demand information can be obtained by machine learning, obtains language by syntactic analysis
Say each slot position information in slot position information list field.
In a specific example, if the demand information of input is that { I wants to go to restaurant to Query=, gives me from five road junctions
Navigate to Wangjing }, then the semantic understanding table that semantic understanding module 100 obtains may is that
Domain therein: " navigation ", confidence=0.8 are field field, indicate the field of the demand
For navigation (navigation), and it is 0.8 that field, which is the order of accuarcy of navigation,.Intent is intended to field, slot for indicating
List is used for representation language slot position information list field, and specific format is similar with field field.Particularly, " tag ": " restaurant ", "
Five road junction from_loc ": " ", " dest_loc ": " Wangjing " is slot position information (slot) respectively.Specifically, for example in " from_
Five road junction loc ": " " in, " five road junctions " this word in demand is had recorded in the form of slot position: " from_loc " has recorded this
The title of slot position, in fact and the classification at " five road junctions ", " five road junctions " have recorded the parameter of the slot position.
Semantic update therein and maintenance (Update Belief) module 300, for taking turns in human-computer dialogue, to language more
Reason and good sense solution is updated and maintains, by epicycle talk in demand semantic understanding and before respectively wheel demand semantic understanding tie
It closes, so that update by wheel to semantic understanding, addition, deleting, and maintain the language of each round interaction in mostly wheel human-computer dialogue
Reason and good sense solution information, with gradually clear demand.
When user interacts with dialogue state management system progress T wheel, the T demand taken turns is understood into (Query
Belief (T)) understand that (Context (T-1)) is combined and updates with the aggregate demand respectively taken turns before T wheel, generate T wheel
Aggregate demand understands (Context (T)), formula is embodied are as follows: Context (T-1)+Query Belief (T)=Context
(T)。
It may be noted that the update and maintenance of semantic historical context can be carried out both for semantic understanding table (belief table)
, after user interacts with dialogue state management system, the semantic understanding table of demand will be updated, adds, delete.
Dialogue state tracing module 200 therein, for passing through search engine according to the semantic understanding table updated after maintaining
With the model that reorders, the result as a result, this reorders that reordered includes pending task, which is sent to corresponding
Execution module or control system, to execute operation according to the task and reply user.The mission bit stream may include target
(Goal) field, movement (Action) field and reply field.Target therein is the classification of the task, can be used for judging this
Which execution module business should be sent to；Movement is the task specific instruction or order to be executed, such as can be change vehicle
The control command of operating status, or adjusted the instruction of the service in vehicle control system；Reply is that the task needs instead
It feeds the information of user, including but not limited to text replys, is based on spatial term (NLG, Natural Language
Generation reply), such as speech answering (TTS, Text to Speech), response (View) of interactive interface etc..
It should be noted that semantic understanding (Belief) is the combination of numerous elements, it is equivalent to variable X=(x1,x2,
x3...), and pending task is a limited operation class set, is equivalent to preferred set y.Dialogue state can be tracked into mould
The effect of block 200 is interpreted as, and is established by semantic understanding to the mapping of pending task, y=f (X), and the mapping is to pass through retrieval
What engine and order models were realized.In fact, establishing by semantic understanding to the mapping of pending task, it is understood that be by language
Reason and good sense solution is classified, and is mapped to a classification set via elements numerous in semantic understanding, the element in this classification set
It is limited the classification of pending task, such as { make a phone call, listen song, navigate }.Also note that arriving, although pending task
Set is a finite aggregate, but in the mapped, and finally obtained pending task can have the information that semantic understanding hands down,
For example, what is recorded in semantic understanding is the white windmill of Zhou Jielun, then in addition to including " listening song " in the pending task obtained
It further include singer is Zhou Jielun outside demand, title of the song is white windmill.
The set of pending task can carry out different fractionations according to the difference of specific field, such as in onboard system,
Point of interest (POI, Point of Interest) inquiry, neighbouring search inquiry etc. can be split into navigation field, and in matchmaker
Inquiry song, broadcasting/pause song etc. can be divided into body stream field.
Optionally, the result that can will reorder of dialogue state tracing module 200 is recorded in history log, is in addition being reordered
It as a result include the situation of semantic understanding information, the result that can will also reorder is sent to semantic update and maintenance module 300, for language
Justice updates the update and maintenance for carrying out next round based on the semantic understanding for including in the result that reorders with maintenance module 300.
In some embodiments, dialogue state tracing module 200 may particularly include: retrieval request generates submodule 230, inspection
Large rope module 210 and the submodule 220 that reorders.
Retrieval request therein generates submodule 230, generates language for retrieving information to according to semantic understanding table with other
Adopted retrieval request, and semantic retrieval request is sent to retrieval submodule 210.The kind of other retrieval information therein is not limited
Class, for example, other retrieval information may include vehicle running state information (speed), vehicle position information, on-board status information, use
One in family personal information (such as gender, age), user preference information (such as user likes walking still walking on path at a high speed)
Or it is some.It optionally, can also include setting in semantic retrieval request in the feature field of semantic understanding table when including confidence level
Reliability (can not also include confidence level).
Retrieval submodule 210 therein, including dialogue state search engine, for requesting in search library semantic retrieval
Using it is complete or by the way of retrieved, obtain search result and obtain the retrieval relevance score of search result, by search result
It is sent to the submodule 220 that reorders.Every data in search library includes that semantic understanding indicates information and pending mission bit stream,
May also comprise it is corresponding other retrieval information, such as can be a preset or passing demand semantic understanding indicate and with
The corresponding pending mission bit stream of the preset or passing demand.It include the information of pending task in each search result,
May also comprise corresponding semantic understanding indicates information and other retrieval information.The pending mission bit stream can be by aiming field, dynamic
Make field and replys field composition.Retrieval relevance score therein is a kind of characterization retrieval commonly obtained by search engine
As a result the index of order of accuarcy can be calculated according to any means known and be obtained.In one embodiment, retrieval relevance can be chosen
The search result of the forward preset quantity of score rank chooses multiple retrievals that retrieval relevance score is greater than default score
As a result, being sent to the submodule 220 that reorders；In addition can to the search result that these are selected, according to retrieval relevance score by
High to Low carry out ranking.
Above-mentioned complete or mode is meant: if semantic retrieval request content be domain: " call ", intent: "
Search number " }, then using it is complete or by the way of retrieved when, can be included all knots of { domain: " call " }
Fruit, and can also be included all results of { intent: " search number " }.
It is since semantic understanding state can be by field (Domain), intention by the way of being retrieved to search library
(Intent) it is composed with slot position information (Slot), the quantity for combining sum M reaches M=2|Domain|*2|Intent|*2|Slot|, wherein | Domain |, | Intent |, | Slot | field, intention, the respective set sizes of slot position information are respectively indicated, from
And the number of combinations of semantic understanding state reaches index rank, and all semantic understanding shapes can not be covered with common processing mode
State.Most related a few datas can be obtained from data acquisition system limited in search library by retrieving to search library, so as to
Problem complexity is effectively reduced.And with the continuous expansion of data acquisition system in search library and increase, the correlation of retrieval can obtain
It is improved to significant.
In retrieval submodule 210, the set of pending task can be added or modify at any time, it is flat in search engine at any time
It in platform, increases some dialogue states newly, available data is deleted or modified, to realize in real time or the data of semireal time are synchronous, make pair
Speech phase management system property easy to maintain.
Dialogue state search engine is also used to first be indexed for inquiring to search library foundation for semantic understanding table.Establish index
Data source in history log or artificial mark, every data includes field, intention, slot position information, target, movement, reply
And other relevant fields.
Fig. 4 is the schematic diagram for the data stored in search library.Every a line in Fig. 4 represents a number in a search library
According to each column represent the parameter information that data include, in fact if the data source in search library is in human-computer dialogue history day
Will, then every data can be the dialogue state of a passing demand.Every data includes semantic understanding table information, other retrievals
Information and pending mission bit stream, semantic understanding table information therein specifically include field, intention, slot position information, it is therein its
It includes gender, age that he, which retrieves information, and pending mission bit stream therein includes target, movement, reply.Particularly, slot position is believed
Breath can only include input demand in each word classification, without the content including word, for example, in Fig. 4 the first row data
In slot position information list include two slot positions title: From_loc and dest_loc, and do not record the tool of the two slot positions
Hold " San Litun ", " Tsinghua University " in vivo.
It when being retrieved to search library, is requested according to semantic retrieval, to field, intention, slot position information, property in search library
Not, the age is retrieved；And return search result be whole data comprising target, movement, reply, so as to by pair
The retrieval of search library obtains corresponding pending mission bit stream.Thus the column and record of semantic understanding table information will might as well be recorded
The column of other retrieval information are referred to as retrieval column.
Although each retrieval column all may be used in some instances it should be noted that the example in figure does not include confidence level
With corresponding confidence level.In addition, retrieval column are not limited in these enumerated in figure, it can also include other column, and can
To add at any time.
The submodule 220 therein that reorders, using machine learning order models to from retrieval the received retrieval of submodule 210
As a result it reorders, obtains with maximally related one of semantic understanding (Belief) as a result, might as well be known as reordering as a result, will weigh
Ranking results are sent to corresponding execution module or control system, to execute task of including in reordering result.
In a kind of example for reordering submodule 220, the submodule 220 that reorders is specifically included:
Sequencing feature extracting unit, for being extracted from all search results, semantic retrieval request, retrieval relevance score
Sequencing feature；Sequencing feature therein, including but not limited to: field, intention, slot position information, confidence level, retrieval relevance obtain
Point, vehicle-mounted state, speed, vehicle location, userspersonal information, user preference etc., which is specifically extracted can as sequencing feature
Be it is artificially selected, in general, the sequencing feature of extraction is The more the better；
Sorting data extracting unit, for extracting sorting data from preset or passing dialogue state data, every
It include that a semantic understanding indicates, semantic retrieval corresponding with semantic understanding expression requests, is corresponding wait hold in sorting data
The markup information of row task and the sorting data, such as can be extracted from interactive dialogue state history log data
Sorting data；The particular content of the markup information can be different according to the difference of mark strategy, for example, using user feedback formula
When mark strategy, which can be the specific feedback of user, and when tactful using heuristic mark, which can
To be the information for indicating the executive condition of the pending task in sorting data, such as whether executing, whether completing, whether by user
Negative；Optionally, cleaning data can be also carried out, for data to be examined and verified again, including check data consistency,
Handle invalid value and missing values, it is therefore intended that delete infull invalid data；
Training unit obtains machine learning sequence for carrying out machine learning training according to sequencing feature and sorting data
Model；
Search result sequencing unit, for obtaining each inspection using the machine learning order models for passing through learning training
The sequence Relevance scores based on the sequencing feature of hitch fruit；The sequence Relevance scores are intended to indicate that the retrieval knot of input
Fruit is the index of corresponding order of accuarcy with the demand information that this is talked with, and the higher search result of sequence Relevance scores is more
With this dialogue demand information is corresponding, search result ranking is also more forward；It should be noted that as dialogue state is gone through
History data increase, and the learning ability of order models can be enhanced, by the standard for the sequence Relevance scores that order models obtain
Exactness can also increase with it；
Output unit, by the work of Relevance scores highest (reordering top ranked) that sorts in all search results
Ranking results of attaching most importance to are sent to the control system of corresponding execution module or vehicle, to execute the task in the result that reorders.
It may be noted that the mark that a variety of strategies are ranked up data can be used, it is not limited to following manner:
Artificial mark strategy, the artificial demand for marking the result that dialogue state management system obtains and whether meeting input；
User feedback formula mark strategy, is labeled according to the feedback of a large number of users, specifically can be, set in human-computer interaction
It is standby upper, the button of " thumbing up " and " stepping on " is set, user is allowed to feed back, " thumbing up " mostly be considered as correlation, the view of " stepping on " mostly
It is uncorrelated；
Heuristic mark strategy, specifically can be, in daily record data, dialogue state management system returns to user one
Pending task, if the pending task that system obtains is negated in next interaction by user or this is pending
Task is not performed or is not completed, then the data is labeled as uncorrelated；If user does not negate in next round interaction
The pending task or default receives the pending task or the pending task is completed that system obtains, then the data
It is labeled as correlation.
Further, dialogue state tracing module 200 may also include relevance threshold detection sub-module 240, for detecting
Whether whether the result that reorders pass through retrieval relevance threshold test and/or by relevance threshold detection of reordering.Wherein such as
The reorder retrieval relevance score of result of fruit reaches preset first threshold, then determines to examine by the retrieval relevance threshold value
It surveys, if the sequence Relevance scores for the result that reorders reach preset second threshold, determines to pass through the correlation that reorders
Threshold test.If the result that reorders has passed through the threshold test carried out, the result that reorders is exported；If reordering knot
Fruit does not pass through any one threshold test that relevance threshold submodule carries out, then does not export this and reorder as a result, but feeding back
It can not identify the demand of input.Optionally, relevance threshold detection sub-module can only carry out relevance threshold detection of reordering.
In a specific example, the demand of input is " I wants to go to restaurant, navigates to Wangjing from five road junctions to me ".Retrieval
Request generates submodule 230 and is used to obtain semantic retrieval request according to the demand, and semantic retrieval request may is that
{ domain: " navigation ", confidence=0.8；
Intent: " navigate ", confidence=0.9；
Slot list={ " tag ": " restaurant ", confidence=0.9；" from_loc ": " five road junctions ",
Confidence=0.6；" dest_loc ": " Wangjing ", confidence=0.7 } }.
Retrieval submodule 210 is for feature field and corresponding confidence level all in requesting semantic retrieval
(Confidence) be put into search library by it is complete or in a manner of retrieved, obtain the inspection of N item before retrieval relevance score rank
Hitch fruit.In this example, search result may is that
Search result 1 (retrieval relevance score=0.8): target=" searching restaurant nearby ", movement=" adjusted and purchased by group
App searches restaurant nearby ", reply=" finding for you to go to restaurant "；
Search result 2 (retrieval relevance score=0.75): target=" navigation ", movement=" have adjusted navigation app, guide
App injection of navigating originates and purpose parameter ", reply=" $ dest_loc is being navigate to for you, be your programme path ... "；
…
Search result N (retrieval relevance score=0.70) ....
The submodule 220 that reorders is sorted for extracting sequencing feature, extracting sorting data by machine learning training
Then model brings each search result into the order models to obtain sequence Relevance scores.The feature wherein extracted includes
Field, intention, slot position information, vehicle-mounted state, speed, vehicle location, userspersonal information, user preference, these features are set
Reliability and the retrieval relevance score of this N search result etc..Finally, if the sequence correlation that search result 2 obtains
(for example score section is 0~1.0 point to the relevance threshold that highest scoring and being greater than manually is set, and threshold value may be set to 0.7
Point), then the output for the demand for selecting search result 2 to talk with as dialogue state management system to epicycle.Talk with when completing more wheels
Later, onboard system will execute final one search result for taking turns selection.
Include two demands in the original demands information of input it may be noted that in this example, one be navigation the other is
Restaurant is looked for, and navigation therein is only the real demand of user.But in the search result obtained by retrieval submodule 210, very
The retrieval relevance score of positive user demand (search result 2) may not be highest.To solve this problem, of the invention
The submodule 220 that reorders is added in dialogue state tracing module 200, for the retrieval to being obtained by retrieval submodule 210
As a result sequence once based on machine learning is carried out, again to obtain the real demand of user.
To better illustrate electric car dialogue state management system of the invention, corresponds to different field, enumerate following
Example:
1. navigation field is illustrated
The demand of input is query=" I to navigate Tsinghua University "；
Semantic understanding that semantic understanding module 100 obtains indicates can be with are as follows:
{domain:"navigation",intent:"navigate",
Slot list:[" navigate ", " dest_loc: Tsinghua University "] }；
Result that dialogue state tracing module 200 obtained reorder can be with are as follows:
{goal:"navigation",
Action: " nav_poi service " (having adjusted map navigation service),
TTS: " planned for you and remove $ dest_loc route " }.
2. music field is illustrated
The demand of input is query=" me is helped to play the blue and white porcelain of Zhou Jielun "；
Semantic understanding that semantic understanding module 100 obtains indicates can be with are as follows:
{domain:"music",intent:"search and play music",
Slot list:[" play_music ", " music_artist: Zhou Jielun ", " music_name: blue and white porcelain "] }；
Result that dialogue state tracing module 200 obtained reorder can be with are as follows:
{goal:"seach music",
Action: " play music service " (having adjusted music service),
TTS: " searched for for you and play $ music_name song, please later " }.
3. telephone field is illustrated
The demand of input is query=" me please be helped to make a phone call to the connection number of Li Si "；
Semantic understanding that semantic understanding module 100 obtains indicates can be with are as follows:
{domain:"call",intent:"call someone",
Slot list:[" person_name: Li Si ", " phone_operator: connection ", " call "] }；
Result that dialogue state tracing module 200 obtained reorder can be with are as follows:
{goal:"call",
Action:"call service",
TTS: " the $ phone_operator number of $ person_name is called for you " }.
It may be noted that in practice, based on the difference of data in search library, the obtained result that reorders can be exemplary with this
As a result difference；When in search library, data volume is sufficiently large in suitable and library, the available result similar with this example.
Further, the embodiment of the present invention also proposed a kind of controller comprising memory and processor, the memory
It is stored with computer program, described program can be realized any of the above-described kind of electric car dialogue shape when being executed by the processor
The step of state management method.It should be appreciated that the instruction stored in memory is and it can be realized when being executed by processor
Electric car dialogue state management method specific example the step of it is corresponding.
Further, the embodiment of the present invention also proposed a kind of computer readable storage medium, for storing computer instruction,
Described instruction realizes the step of any of the above-described kind of electric car dialogue state management method when by a computer or processor execution
Suddenly.It should be appreciated that the instruction stored in computer readable storage medium is the electronic vapour that can be realized when executed with it
The step of specific example of vehicle dialogue state management method, is corresponding.
The above described is only a preferred embodiment of the present invention, be not intended to limit the present invention in any form, though
So the present invention has been disclosed as a preferred embodiment, and however, it is not intended to limit the invention, any technology people for being familiar with this profession
Member, without departing from the scope of the present invention, when the technology contents using the disclosure above make a little change or modification
For the equivalent embodiment of equivalent variations, but anything that does not depart from the technical scheme of the invention content, according to the technical essence of the invention
Any simple modification, equivalent change and modification to the above embodiments, all of which are still within the scope of the technical scheme of the invention.
Claims (12)
1. a kind of electric car dialogue state management method, comprising the following steps:
The demand information for parsing input, the demand information, which is converted to semantic understanding, to be indicated；
It indicates to be retrieved in search library according to the semantic understanding, obtains multiple search results；It is every in the search library
Data includes that preset or passing semantic understanding indicates and corresponding with described preset or passing semantic understanding expression
Pending mission bit stream；
It is reordered using machine learning order models to the search result, defines the retrieval knot of top ranked one
Fruit is attached most importance to ranking results, the pending mission bit stream in the result that reorders described in output.
2. dialogue state management method according to claim 1, wherein semantic understanding expression includes:
Field field, for recording field belonging to the demand information；
It is intended to field, for recording the intention of the demand information；
Language slot position information list field, for recording each language slot position information of the demand information, each language
Slot position information includes the information of the word extracted from demand；The word information includes the content of word, or including
The content and classification of word.
3. dialogue state management method according to claim 2, wherein also being wrapped in each field that the semantic understanding indicates
Confidence level is included, for indicating the order of accuarcy for recording content in the field.
4. dialogue state management method according to claim 1, wherein the pending mission bit stream includes:
Aiming field, for recording the classification of the task；
Action field, for recording the task instructions to be executed；
Reply field, the information to be fed back for recording the task.
5. dialogue state management method according to claim 1, wherein described indicate according to the semantic understanding in search library
In retrieved, further comprise:
It is indicated to obtain semantic retrieval request according to semantic understanding；
In search library using it is complete or by the way of semantic retrieval request is retrieved, obtain the search result and the retrieval
As a result retrieval relevance score.
6. dialogue state management method according to claim 5, wherein further including vehicle driving in semantic retrieval request
Status information, vehicle position information, on-board status information, userspersonal information, one or some in user preference information.
7. dialogue state management method according to claim 5, wherein described utilize machine learning order models to the inspection
Hitch fruit is reordered, further are as follows:
It chooses the search result of the forward preset quantity of the retrieval relevance score rank or chooses the retrieval relevance
Score is greater than multiple search results of default score, reorders described in progress.
8. dialogue state management method according to claim 7, wherein described utilize machine learning order models to the inspection
Hitch fruit is reordered, and further comprises:
Sequencing feature is extracted, the range of extraction includes the search result, the semantic understanding indicates and the retrieval relevance
Score；
Sorting data is extracted from preset or passing dialogue state data, includes a semanteme in every sorting data
Understand table, semantic retrieval corresponding with semantic understanding expression request, corresponding pending task and the sorting data
Markup information；
Machine learning training is carried out according to the sequencing feature and the sorting data, obtains the machine learning order models；
The sequence Relevance scores of each search result are obtained using the machine learning order models, the sequence is related
The property higher search result ranking of score is more forward.
9. dialogue state management method according to claim 8, wherein the markup information includes: to indicate the sorting data
In the pending task executive condition information.
10. dialogue state management method according to claim 8, further includes:
Retrieval relevance threshold test is carried out to the result that reorders and at least one of relevance threshold detection of reordering；
If the retrieval relevance score of the result that reorders reaches preset first threshold, judgement passes through the retrieval phase
Closing property threshold test；If the sequence Relevance scores of the result that reorders reach preset second threshold, determine
Pass through the relevance threshold detection of reordering；
If the result that reorders does not pass through carried out retrieval relevance threshold test or is not reordered by what is carried out
Relevance threshold detection, then feedback can not identify the demand.
11. dialogue state management method according to claim 1, the dialogue is more wheels, the dialogue state management method
Further include: by epicycle talk with described in the semantic understanding of demand information indicate to manage with the semantic of demand information of each wheel input before
Solution indicates to be combined.
12. a kind of electric car dialogue state management system, comprising:
Semantic understanding module, for parsing the demand information of input, the demand information, which is converted to semantic understanding, to be indicated；
Retrieval module obtains multiple search results for indicating to be retrieved in search library according to the semantic understanding；It is described
Every data in search library include preset or passing semantic understanding indicate and with described preset or passing semanteme
Understanding indicates corresponding pending mission bit stream；
Reorder module, for being reordered using machine learning order models to the search result, defines top ranked
A search result attach most importance to ranking results, the pending mission bit stream in the result that reorders described in output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810267049.1A CN110309276B (en) | 2018-03-28 | 2018-03-28 | Electric vehicle conversation state management method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810267049.1A CN110309276B (en) | 2018-03-28 | 2018-03-28 | Electric vehicle conversation state management method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110309276A true CN110309276A (en) | 2019-10-08 |
| CN110309276B CN110309276B (en) | 2022-11-29 |
Family
ID=68073762
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810267049.1A Active CN110309276B (en) | 2018-03-28 | 2018-03-28 | Electric vehicle conversation state management method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110309276B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022057712A1 (en) * | 2020-09-15 | 2022-03-24 | 华为技术有限公司 | Electronic device and semantic parsing method therefor, medium, and human-machine dialog system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140108321A1 (en) * | 2012-10-12 | 2014-04-17 | International Business Machines Corporation | Text-based inference chaining |
| CN105183850A (en) * | 2015-09-07 | 2015-12-23 | 百度在线网络技术（北京）有限公司 | Information querying method and device based on artificial intelligence |
| CN107316643A (en) * | 2017-07-04 | 2017-11-03 | 科大讯飞股份有限公司 | Voice interactive method and device |
| CN107315737A (en) * | 2017-07-04 | 2017-11-03 | 北京奇艺世纪科技有限公司 | A kind of semantic logic processing method and system |
| CN107785018A (en) * | 2016-08-31 | 2018-03-09 | 科大讯飞股份有限公司 | More wheel interaction semantics understanding methods and device |
-
2018
- 2018-03-28 CN CN201810267049.1A patent/CN110309276B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140108321A1 (en) * | 2012-10-12 | 2014-04-17 | International Business Machines Corporation | Text-based inference chaining |
| CN105183850A (en) * | 2015-09-07 | 2015-12-23 | 百度在线网络技术（北京）有限公司 | Information querying method and device based on artificial intelligence |
| CN107785018A (en) * | 2016-08-31 | 2018-03-09 | 科大讯飞股份有限公司 | More wheel interaction semantics understanding methods and device |
| CN107316643A (en) * | 2017-07-04 | 2017-11-03 | 科大讯飞股份有限公司 | Voice interactive method and device |
| CN107315737A (en) * | 2017-07-04 | 2017-11-03 | 北京奇艺世纪科技有限公司 | A kind of semantic logic processing method and system |
Non-Patent Citations (2)
| Title |
|---|
| LIU B等: "attention based recurrent neural network models for joint intent detection and slot filling", 《ARXIV PREPRINT ARXIV:1609.01454》 * |
| 徐梓翔: "任务型对话系统平台的设计与实现", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2022057712A1 (en) * | 2020-09-15 | 2022-03-24 | 华为技术有限公司 | Electronic device and semantic parsing method therefor, medium, and human-machine dialog system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110309276B (en) | 2022-11-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110427463B (en) | Search statement response method and device, server and storage medium | |
| CN108345690B (en) | Intelligent question and answer method and system | |
| CN109885672B (en) | Question-answering type intelligent retrieval system and method for online education | |
| CN107818781B (en) | Intelligent interaction method, equipment and storage medium | |
| US8756245B2 (en) | Systems and methods for answering user questions | |
| CN107832439B (en) | Method, system and the terminal device of more wheel state trackings | |
| CN104331449B (en) | Query statement and determination method, device, terminal and the server of webpage similarity | |
| CN109165380A (en) | A kind of neural network model training method and device, text label determine method and device | |
| US20210018332A1 (en) | Poi name matching method, apparatus, device and storage medium | |
| CN108280155A (en) | The problem of based on short-sighted frequency, retrieves feedback method, device and its equipment | |
| CN109299245B (en) | Method and device for recalling knowledge points | |
| CN110096567A (en) | Selection method, system are replied in more wheels dialogue based on QA Analysis of Knowledge Bases Reasoning | |
| CN114020924A (en) | Optimizing graph traversal | |
| CN109829052A (en) | A kind of open dialogue method and system based on human-computer interaction | |
| CN109189959A (en) | A kind of method and device constructing image data base | |
| CN108664599A (en) | Intelligent answer method, apparatus, intelligent answer server and storage medium | |
| CN112115252B (en) | Intelligent auxiliary writing processing method and device, electronic equipment and storage medium | |
| CN106951503A (en) | Information providing method, device, equipment and storage medium | |
| CN107193882A (en) | Why not query answer methods based on figure matching on RDF data | |
| CN112115242A (en) | Intelligent customer service question-answering system based on naive Bayes classification algorithm | |
| CN108959529A (en) | Determination method, apparatus, equipment and the storage medium of problem answers type | |
| CN116959433A (en) | Text processing method, device, electronic equipment and storage medium | |
| CN113515616B (en) | Task driving system based on natural language | |
| CN110795544A (en) | Content search method, device, equipment and storage medium | |
| CN110795547B (en) | Text recognition method and related product |
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 | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200727 Address after: Susong Road West and Shenzhen Road North, Hefei Economic and Technological Development Zone, Anhui Province Applicant after: Weilai (Anhui) Holding Co.,Ltd. Address before: China Hong Kong Applicant before: NIO NEXTEV Ltd. |
|
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: 230601 Susong Road West, Shenzhen Road North, Hefei Economic and Technological Development Zone, Anhui Province Applicant after: Weilai Holdings Ltd. Address before: 230601 Susong Road West, Shenzhen Road North, Hefei Economic and Technological Development Zone, Anhui Province Applicant before: Weilai (Anhui) Holding Co.,Ltd. |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |