CN110309277A - Human-computer dialogue semanteme parsing method and system - Google Patents

Human-computer dialogue semanteme parsing method and system Download PDF

Info

Publication number
CN110309277A
CN110309277A CN201810267052.3A CN201810267052A CN110309277A CN 110309277 A CN110309277 A CN 110309277A CN 201810267052 A CN201810267052 A CN 201810267052A CN 110309277 A CN110309277 A CN 110309277A
Authority
CN
China
Prior art keywords
information
demand information
slot position
word
type
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
Application number
CN201810267052.3A
Other languages
Chinese (zh)
Other versions
CN110309277B (en
Inventor
夏妍
温泉
林锋
徐龙生
马天泽
赵浩天
葛斯函
马英财
卢瑶琪
陈盛
陈功
芮锐
芮元勋
庄莉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
NIO Holding Co Ltd
Original Assignee
NIO Nextev Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by NIO Nextev Ltd filed Critical NIO Nextev Ltd
Priority to CN201810267052.3A priority Critical patent/CN110309277B/en
Publication of CN110309277A publication Critical patent/CN110309277A/en
Application granted granted Critical
Publication of CN110309277B publication Critical patent/CN110309277B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Human Computer Interaction (AREA)
  • Machine Translation (AREA)

Abstract

The present invention relates to a kind of human-computer dialogue semanteme parsing method and systems, which comprises carries out the first parsing to the demand information of input, obtains the first slot position of the demand information and one or more classification informations of the demand information；First slot position is used to record the attribute of short sentence, phrase or word and the short sentence, phrase or word in the demand information；The classification information is for recording classification belonging to the demand information；The semantic expressiveness of the demand information is obtained according to first slot position and the classification information.Using human-computer dialogue semanteme parsing method and system of the invention, the semanteme of the demand information of input can be understood, is accurately obtained.

Description

Human-computer dialogue semanteme parsing method and system
Technical field
The present invention relates to human-computer dialogue fields, more particularly to a kind of semantic parsing method and 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.
Complete interactive system is related to the sides such as voice technology, natural language processing, knowledge base, dialogue state maintenance Face.The request for the natural language form for inputting user is responsible in semanteme parsing, judges the intention of user, extracts relevant element User demand information is converted to the internal representation of interactive system by (slot position), so that interactive system asks user It asks and is further processed.
Existing semanteme analytic method semanteme parsing result is not accurate enough, leads to conversational system during human-computer dialogue It replys unreasonable, occurs not meeting the situations such as the reply of dialogue habit of people.
Summary of the invention
It is a primary object of the present invention to, overcome defect existing for existing semantic analytic method, and propose a kind of new Human-computer dialogue semanteme parsing method and system, the technical problem to be solved is that provide a kind of semantic parsing it is more acurrate become apparent from, It is more in line with human conversation habit, intelligent higher human-computer dialogue semanteme parsing method and system.
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 human-computer dialogue semanteme analytic method, comprising the following steps: the first parsing is carried out to the demand information of input, is obtained described First slot position of demand information and one or more classification informations of the demand information；First slot position is for recording institute State the attribute of short sentence, phrase or the word and the short sentence, phrase or word in demand information；The classification information is for recording institute State classification belonging to demand information；The semantic table of the demand information is obtained according to first slot position and the classification information Show.
It the purpose of the present invention and solves its technical problem and can also be further achieved by the following technical measures.
Human-computer dialogue semanteme analytic method above-mentioned, wherein the classification information includes: issue type information, field letter One of breath, intent information are some；Described problem type information, for recording the interrogative sentence classification of the demand information； The realm information, for recording the field classification of the demand information；The intent information, for recording the demand information Intention classification.
Human-computer dialogue semanteme analytic method above-mentioned further comprises: preset described problem type information type, The type of the type of the realm information and the intent information, and one or more meanings are correspondingly arranged for the field of a type Figure, or one or more fields are correspondingly arranged for the intention of a type.
Human-computer dialogue semanteme analytic method above-mentioned further comprises: presetting the type of first slot position, and is The field of one type is correspondingly arranged one or more first slot positions, or for a type intention be correspondingly arranged one or Multiple first slot positions, or type is correspondingly arranged one or more first slot positions the problem of be a type.
Human-computer dialogue semanteme analytic method above-mentioned, wherein first parsing are as follows: utilize bidirectional circulating neural network In conjunction with the learning algorithm and the common learning algorithm of associated losses function of attention mechanism, obtains the described of the demand information and ask Inscribe type information, the realm information, the intent information and first slot position.
Human-computer dialogue semanteme analytic method above-mentioned, wherein described utilizes bidirectional circulating neural network combination attention machine The learning algorithm and the common learning algorithm of associated losses function of system obtain the described problem type information of the demand information, institute State realm information, the intent information and first slot position, comprising the following steps: learn to obtain based on large-scale corpus pre- If the pre-training term vector of phrase or default word indicates；Each phrase or the word in the demand information are obtained by participle, it will The each phrase or word of the demand information correspond on the pre-training term vector, obtain the corresponding word of the demand information to Moment matrix；The study that bidirectional circulating neural network combination attention mechanism is carried out to the demand information term vector matrix, obtains Expression, the global expression of the demand information, attention machine of the demand information at bidirectional circulating neural network each moment The output of system and the global of attention mechanism indicate；Using the demand information at bidirectional circulating neural network each moment It indicates the label for predicting first slot position of the demand information with attention mechanism output, the demand is utilized to believe The global intent information for indicating to indicate with the attention mechanism overall situation to predict the demand information of breath, the field Information and described problem type information, so that the demand information is indicated by field identification mission, intention assessment task, problem Type identification task and slot filling task are shared；Use the common learning areas identification mission of associated losses function, intention assessment Task, problem types identification mission and slot fill task, respectively obtain field identification mission, intention assessment task, problem types The class label of identification mission and slot filling task.
Human-computer dialogue semanteme analytic method above-mentioned further comprises: carrying out the second parsing to first slot position, obtains The each word for including in first slot position, and the relationship between the attribute and/or each word of each word is obtained, According to the relationship between the attribute and each word of each word, each word is filled into the second slot position In；The semanteme of the demand information is obtained according to second slot position and described problem type information, realm information, intent information It indicates.
Human-computer dialogue semanteme analytic method above-mentioned, wherein described carry out the second parsing including following to first slot position Step: participle and interdependent syntactic analysis are carried out to first slot position, the syntax tree of first slot position is obtained, according to the sentence Method tree determines the core word and determiner in the first slot position；It marks to obtain the attribute of the core word by part-of-speech tagging, entity With the attribute of the determiner, pass through a word in the attribute of each word in first slot position and/or first slot position With the relationship of other words, corresponding second slot position of predicate to determine.
Human-computer dialogue semanteme analytic method above-mentioned further comprises: by epicycle talk with described in demand information semanteme It indicates to be combined with the semantic expressiveness of the demand information of each wheel input before.
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 human-computer dialogue semanteme resolution system, including the first parsing module are obtained for carrying out the first parsing to the demand information of input To the first slot position of the demand information and one or more classification informations of the demand information；First slot position is used for Record the attribute of short sentence, phrase or the word and the short sentence, phrase or word in the demand information；The classification information is used for Record classification belonging to the demand information；Semantic expressiveness determining module, for being believed according to first slot position and the classification Breath obtains the semantic expressiveness of the demand information.
It the purpose of the present invention and solves its technical problem and can also be further achieved by the following technical measures.
Human-computer dialogue semanteme resolution system above-mentioned, wherein the classification information includes: issue type information, field letter One of breath, intent information are some；Described problem type information, for recording the interrogative sentence classification of the demand information； The realm information, for recording the field classification of the demand information；The intent information, for recording the demand information Intention classification.
Human-computer dialogue semanteme resolution system above-mentioned further comprises: classification information classification presetting module, for setting in advance The type of the type of described problem type information, the type of the realm information and the intent information is set, and is a type Field be correspondingly arranged one or more intentions, or be correspondingly arranged one or more fields for the intention of a type.
Human-computer dialogue semanteme resolution system above-mentioned further comprises: slot position classification presetting module, for presetting The type of the first slot position is stated, and is correspondingly arranged one or more first slot positions for the field of a type, or be one kind The intention of class is correspondingly arranged one or more first slot positions, or type is correspondingly arranged one or more the problem of be a type A first slot position.
Human-computer dialogue semanteme resolution system above-mentioned, wherein first parsing module includes machine learning submodule, For utilizing the learning algorithm and the common learning algorithm of associated losses function of bidirectional circulating neural network combination attention mechanism, Obtain described problem type information, the realm information, the intent information and first slot position of the demand information.
Human-computer dialogue semanteme resolution system above-mentioned is used for wherein the machine learning submodule includes: first unit The pre-training term vector for learning to obtain default phrase or default word based on large-scale corpus indicates；Second unit, for by dividing Word obtains each phrase or word in the demand information, and each phrase or word of the demand information are corresponded to the pre- instruction Practice on term vector, obtains the corresponding term vector matrix of the demand information；Third unit, for the demand information term vector Matrix carries out the study of bidirectional circulating neural network combination attention mechanism, obtains the demand information in bidirectional circulating nerve net The expression at network each moment, the demand information it is global indicate, the overall situation of the output of attention mechanism and attention mechanism It indicates；Unit the 4th, for utilizing the demand information in the expression and the attention at bidirectional circulating neural network each moment The label of first slot position of the demand information is predicted in the output of power mechanism, using the demand information it is global indicate and The attention mechanism overall situation indicates the intent information, the realm information and described problem to predict the demand information Type information so that the demand information indicate by field identification mission, intention assessment task, problem types identification mission and Slot filling task is shared；Unit the 5th, for being appointed using the common learning areas identification mission of associated losses function, intention assessment Business, problem types identification mission and slot fill task, respectively obtain field identification mission, intention assessment task, problem types and know The class label of other task and slot filling task.
Human-computer dialogue semanteme resolution system above-mentioned further comprises: the second parsing module, for first slot position The second parsing is carried out, each word for including in first slot position is obtained, and obtains the attribute of each word and/or described each Relationship between a word, according to the relationship between the attribute and each word of each word, by each word It is filled into the second slot position；The semantic expressiveness determining module, for according to second slot position and described problem type information, Realm information, intent information obtain the semantic expressiveness of the demand information.
Human-computer dialogue semanteme resolution system above-mentioned, wherein second parsing module includes: Unit the 6th, for institute It states the first slot position and carries out participle and interdependent syntactic analysis, obtain the syntax tree of first slot position, determined according to the syntax tree Core word and determiner in first slot position；Unit the 7th obtains the core word for marking by part-of-speech tagging, entity The attribute of attribute and the determiner passes through in the attribute of each word in first slot position and/or first slot position one The relationship of a word and other words, corresponding second slot position of predicate to determine.
Human-computer dialogue semanteme resolution system above-mentioned further comprises semantic update module, for institute in talking with epicycle The semantic expressiveness for stating the semantic expressiveness and the demand information of each wheel input before of demand 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 vehicle-mounted interactive method includes the steps that human-computer dialogue semanteme analytic method described in aforementioned any one.
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 vehicle-mounted interactive system, including human-computer dialogue semanteme resolution system described in aforementioned any one.
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.
By above-mentioned technical proposal, human-computer dialogue semanteme parsing method and system of the invention at least have following advantages and The utility model has the advantages that
(1) by parsing from the demand information that user inputs problem types, field, intention, slot position four dimensions Information, can understand, uniquely determine demand information semanteme；Pass through type the problem of parsing demand information, human-computer dialogue System can provide different answers according to the difference of problem types, be more in line with human conversation and be accustomed to, be intelligent higher；Pass through Parse the domain type of demand information, can to only by being intended to, the obtained semanteme of slot position correct, and can be can not explicit semantic meaning When, the field of information is reasonably replied according to demand；
(2) by being arranged a large amount of slot position classification, and by the division of slot position classification and realm information and/or intent information And/or issue type information is associated, so that the coverage area of slot position is bigger, more acurrate；
(3) it is learned jointly using the learning algorithm of bidirectional circulating neural network combination attention mechanism and associated losses function Practise algorithm, the inner link when parsing problem types, field, intention, slot position between consideration, to improve the accurate of parsing Rate；
It (4), can more accurately by the way that in more wheel dialogues, this parsing result is combined with previous parsing result Obtain the semanteme of demand information.
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 semantic analytic method of the invention.
Fig. 2 is the flow diagram of another embodiment of semantic analytic method of the invention.
Fig. 3 is the structural schematic diagram for the syntax tree that one embodiment of semantic analytic method of the invention provides.
Fig. 4 is the structural block diagram of one embodiment of semantic resolution system of the invention.
Fig. 5 is the structural block diagram of another embodiment of semantic resolution system of the 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 human-computer dialogue semanteme parsing method and system proposed according to the present invention its specific embodiment, Method, structure, feature and its effect, detailed description is as follows.
Fig. 1 is the schematic flow chart of human-computer dialogue semanteme analytic method one embodiment of the invention.Referring to Fig. 1, The exemplary human-computer dialogue semanteme analytic method of the present invention includes:
Step S110 carries out the first parsing to the demand information (query) of user's input, obtains the first parsing result, should First parsing result includes the characteristic information of the demand information.This feature information is used to record the content and category of the demand information Property, the classification information of slot position (slot) and one or more demand informations including the demand information.Wherein, slot position is used for The attribute of short sentence, phrase or the word and the short sentence, phrase or word in demand information is recorded, classification information is for recording demand Classification belonging to information.
Step S120 obtains the semantic expressiveness of demand information according to the first parsing result, which includes the demand The slot position and classification information of information.
Further, can classify in various ways to demand information, such as according to problem types (ques_type) Classified, classified according to field (domain), classified according to intention (intent), thus in step s 110 The classification information parsed may include problem types (ques_type) information, field (domain) information, intention (intent) information.The problem of determining demand information type information, realm information or intent information, seek to judge this Classification, the classification of affiliated intention of the classification of problem type, affiliated field belonging to demand information.
Issue type information therein, for characterizing interrogative sentence type belonging to demand information.The problem types include: non- Query type (None), ask whether type (YN), inquiry how/how type (How), inquiry venue type (Where), Query time type (When), inquiry quantity type (How_many), interrogating range type (How_far) etc..It should be noted that , issue type information is not limited to example enumerated above, in fact can be according to the actual situation to default problem class The type of type information is added or changes.
Realm information therein, for characterizing the type in field belonging to demand information.In fact, realm information can be regarded Classify for the summary of intent information.As an example, human-computer interaction scene in the car, realm information may include navigation, charging, Media, telephone relation, short message, information, specific equipment, search and interactive system chat etc..The field of should be noted that Information is not limited to example enumerated above, in fact can the type according to the actual situation to preset realm information add Add or changes.
Intent information therein for characterizing the specifically intended type of demand information is retouched to the refinement of realm information It states.
As an example, human-computer interaction scene in the car, it is intended that information can include:
From somewhere to somewhere (from_to)；Go down (to)；
It finds and plays music (search_and_play_music)；Volume adjustment (turn_volumn)；
Call someone (call_someone)；
It is (yes)；No (no)；
Movement (action/action_xxx) is executed, in general, execution movement may particularly include execution and start, execute Stop, executing pause, execute and continue；
Certain equipment (action_device) is operated, for different equipment intent information different froms, such as vehicle Lamp, the intent information of certain equipment class of the operation include opening/closing car light etc.；
About the movement (list_xxx) of inventory, such as goes to page up, goes to lower one page；
Inquire the status information (show_state) of in-vehicle device.
It should be noted that intent information is not limited to example enumerated above, it in fact can be right according to the actual situation The type of preset intent information is added or changes.
Slot position therein, the specific element for recording the extraction from user demand information are (such as specific short sentence, short Language, phrase, word or word) and the specific element belonging to classification.And (alternatively referred to as slot position mentions for the parsing of slot position (slots) Take, slot position identification or slot position filling) be equivalent to extracted from the demand information of input relevant element (short sentence, phrase, phrase, Word, word etc.) it is filled into slot position according to the attribute of element, or be equivalent to and label is filled out to each element of demand information, it can It is regarded as sequence labelling.
In fact, can will have using the classification of specific element as a slot position classification (the alternatively referred to as title of slot position) The content of element of volume is filled into the slot position as the design parameter of the slot position, infinite specific element is mapped as limited Slot position, such as in the situation that specific element is proper noun；Or can using the content of specific element as a slot position classification, and The slot position does not have parameter, such as in the situation that specific element is a part of verb.As an example, the demand of an input Information includes " from Beijing to Hangzhou " this sentence, which includes " Beijing ", " Hangzhou " two elements, wherein the class in " Beijing " Not Wei " departure place " (from_loc), the classification in " Hangzhou " is " destination " (to_loc), should " departure place ", " destination " just It is the slot position in this example, and " Beijing " is the content of " departure place " this slot position record, " Hangzhou " is " destination " this slot The content of position record.
Optionally, slot position information can be recorded with the form of " slot position classification: slot position parameter ", wherein slot position classification is for remembering The classification of the element (word, phrase) in demand information sentence is recorded, actually a kind of classification of slot position can be used as this slot position Title, and slot position parameter is used to record the content of element (word, phrase) in demand information sentence.It is interior man-machine below Some examples of the slot position of interaction scenarios:
" poi: point of interest ", which is a geographical entity, including generally qualifier (such as attribute, the adverbial modifier) One phrase, such as " coffee shop within 1 kilometer of Wangjing subway station "；
" around_poi: geographical location ", for recording geographical location locating for a point of interest, such as " Wangjing subway station 1 " Wangjing subway station " in coffee shop within kilometer "；
" tag: a kind of things entity ", for indicating the entity tag of a kind of things, such as " tag: coffee shop ", " tag: meal Shop ", " tag: parking lot "；
" loc: specified place ", which is the word or expression of specified place title, such as BoTai Building, the village Dong Wang Cell；
" fee: charge situation ", which includes whether the phrases such as charge, toll amount, expense；
" parking: Parking situation "；
" operation: on-unit ", the on-unit are the concrete operations needed to be implemented, such as " operation: search ", " operation: playing "；
" language: category of language ", the category of language are specific languages, such as English, western class's language, Guangdong language；
" volume: volume ", the volume are the phrase about volume；
" music_name: song title "；
" music_artist: singer "；
" person_name: name "；
" phone_digit: telephone number "；
" distance: the vocabulary about distance ", such as " distance: how far ", " within distance: one kilometer "；
" duration: the vocabulary about the time " etc..
It should be noted that slot position information is not limited to example enumerated above, it in fact can be right according to the actual situation Preset slot position classification is added or changes.
This man-machine dialog semantics analytic method may also include the kind of the type, realm information that preset issue type information The step of type of class and intent information, and the step of presetting the type of slot position.It further, can be according to problem class Type, field, intention, connecting each other the type of these characteristic informations is arranged between slot position.It is specific:
For realm information and intent information:
Since realm information is intended to the summary classification of information, it is intended that information is the refinement description to realm information, can be with The type of field, intention is arranged to connect each other:
Certain intentions only belong to one or more specific fields, for example, from somewhere to somewhere (from_to), go down (to), navigation field is generally corresponded only to；
In addition, same class is intended to included specifically intended information can be different, example for different specific areas Such as executing movement (action/action_xxx) this kind of intention, in field of media, it may include execution start, execute stopping, Execute that pause, to execute continuation, search song, switching song, change play mode etc. specifically intended, and in air-conditioning equipment field Execution acts this kind of be intended in addition to equally including that other than execution starts, executes stopping, executing pause, executes continuation, may also include It changes air conditioning mode, change set temperature, change wind direction etc. to be intended to, but can not include search song, switching song, change These intentions of play mode.
For slot position:
The corresponding relationship of slot position classification (slot position title) and problem types, field or intention can be set, such as can set Set problem type, field or intention corresponding to a slot position classification, or one problem types of setting or a field or one It is a to be intended to included slot position classification, so that slot position information is more acurrate.For example, " volume: volume " this slot position can be with " sound Amount adjusts (turn_volumn) " it is intended to correspond to, can be corresponding with " media " field, or may also be corresponding with " call " field, and It is general not corresponding with " go down (to) " intention, it is not corresponding with " navigation " field.
According to the association rule between above-mentioned problem type, field, intention, slot position, can be parsed by judgement Whether each characteristic information meets the association rule to determine whether the semantic parsing of this time succeeds, moreover it is possible to can not identify some spy When reference ceases, (or prediction) unrecognized information is supplemented using the association rule.
It can use machine learning model and carry out first parsing.For this reason, it may be necessary in advance to largely having marked great question The data of the characteristic informations such as type, field, intention and slot position carry out machine learning training, machine learning model are obtained, to one When demand information carries out the first parsing, which is input to this by the machine learning model of learning training, it should Machine learning model will export problem type, field, intention and slot position corresponding to this demand information.
Further, using the learning algorithm and associated losses letter of bidirectional circulating neural network combination attention mechanism The common learning algorithms of number carry out first parsing, when identifying problem types, field, intention, slot position between consideration in It is contacting, to improve the accuracy rate of the first parsing.Specifically, term vector square first can be converted by the demand information that user inputs Battle array, is learnt by bidirectional circulating neural network combination attention mechanism, so that the demand information is expressed as being identified by field Task, it is intended that identification mission, the term vector matrix of problem types identification mission and slot filling task sharing reuse associated losses Function carrys out common learning areas identification mission, it is intended that identification mission, problem types identification mission and slot fill task.In some realities Apply in example, step S110 specifically includes the following steps:
Step S111, the pre-training term vector for learning to obtain preset word (or word, phrase) based on large-scale corpus indicate.
Step S112, by segmenting each word (or word, phrase) in the demand information that is inputted, then by the demand Each word in information corresponds on the term vector of the pre-training obtained in step S111, corresponding to obtain this demand information Term vector matrix.
Step S113 obtains the term vector square of the demand information using bidirectional circulating neural network learning in step S112 Battle array obtains the demand information in the expression ht_forward at bidirectional circulating neural network each moment and (corresponds to forward circulation mind Through network), ht_backward (correspond to recycled back neural network), the demand information it is global indicate hu_forward, Hu_backward, the output Ct of attention mechanism, the global of attention mechanism indicate Cu.
Step S114, by the demand information obtained in step S113 bidirectional circulating neural network each moment table Show that the output Ct of ht_forward, ht_backward and attention mechanism is used to predict the label of each slot position, it will be in step The global of the demand information obtained in S113 indicates that hu_forward, hu_backward and the global of attention mechanism are indicated Cu is used to predict intention, the field, problem types of the demand information, so that the demand information expression of input is identified by field Task, intention assessment task, problem types identification mission and slot filling task are shared.
Step S115 uses the common learning areas identification mission of associated losses function, it is intended that identification mission, problem types are known Other task and slot fill task, respectively obtain field identification mission, it is intended that identification mission, problem types identification mission and slot filling The class label of task.
Fig. 2 is the schematic flow chart of another embodiment of human-computer dialogue semanteme analytic method of the invention.Due to step Slot position obtained in S110 may be short sentence or phrase comprising one or more qualifiers (such as attribute, the adverbial modifier), and there is no harm in will be by The slot position that step S110 is obtained is known as the first slot position, to obtain more accurate semantic expressiveness, referring to Fig. 2, in some embodiments In, human-computer dialogue semanteme analytic method of the invention includes:
Step S210 carries out the first parsing to the demand information (query) of user's input, obtains the first parsing result, should The problem of first parsing result includes demand information type information, realm information, intent information and the demand information first Slot position.
Step S220 carries out the second parsing to the first slot position of the demand information (query) obtained by step S210, obtains Second parsing result, second parsing result include that the first slot position each word (or phrase), each word that are included is (or short Language) attribute and each word (or phrase) between relationship, with from the short sentence of this first slot position for decomposing insufficient or short It is decomposed in language and extracts core word and qualifier, according to the attribute and each word of each word in the short sentence or phrase of the first slot position Between relationship, each word of the first slot position is filled into the second slot position more specifically refined, with more accurate earth's surface Up to the demand information of input.Specifically, one of syntactic analysis, participle, part-of-speech tagging, entity mark mode or more can be passed through The composition of kind mode carries out second parsing.
Step S230 obtains the semantic expressiveness of demand information in conjunction with the result of the first parsing and the result of the second parsing, should The problem of semantic expressiveness includes demand information type information, realm information, intent information and the second slot position.
As a specific example, for a demand information sentence, " coffee shop within 1 kilometer of Wangjing subway station has several Family ", extracts to have obtained first slot position " poi: the coffee shop within 1 kilometer of Wangjing subway station " by step S210.To this First slot position information carry out step S220 specifically includes the following steps:
Step S221 is obtained in " coffee shop within 1 kilometer of Wangjing subway station " by participle and interdependent syntactic analysis The syntax tree of specific element (generally word or expression), to determine core word and the restriction in specific element according to the syntax tree Word.Fig. 3 is the schematic diagram for the syntax tree that human-computer dialogue semanteme analytic method one embodiment of the invention provides.It please join Fig. 3 is read, the word node that the tree root in syntax tree directly links is exactly core word, and determiner can be in the child of core word in syntax tree In child node." coffee shop " is core word in this example, and " Wangjing subway station " and " within one kilometer " is all " coffee shop " Determiner.
Step S222 marks to obtain the attribute of each core word and each determiner by part-of-speech tagging, entity, to pass through The attribute of one specific element and/or the relationship of a specific element and other specific elements, to determine belonging to the specific element The second slot position classification.In this example, " coffee shop " part of speech is noun, and " Wangjing " is place name entity, to obtain core word " coffee shop " is limited the limitation of word place name entity " Wangjing ".
As an example, the demand information of input is " several coffee shops in one kilometer range of Wangjing subway station ", this The semantic parsing result that the human-computer dialogue semanteme analytic method of invention obtains may is that
{
ques_type:How_many；
domain:Nav；
intent:poi_lookup；
Slots: " around_poi: Wangjing subway station ", " within distance: one kilometer ", " tag: coffee shop "；
}
Wherein, the problem of " ques_type:How_many " is demand information type, the problem of indicating the demand information class Type is inquiry quantity type；" domain:Nav " is the domain type of demand information, indicates the field of the demand information for navigation； " intent:poi_lookup " is the intention type of demand information, and indicate the demand information is intended to inquiry point of interest (point of interest, poi)；" slots:around_poi: Wangjing subway station, within distance: one kilometer, tag: Coffee shop " is the language slot position of demand information, which includes three language slot positions.
It, can be more clear according to the semantic meaning representation comprising issue type information (ques_type), realm information (domain) The actual demand of Chu, accurately expression user, and can more accurately be judged in interactive subsequent process Finer processing.
Specifically, the demand information in user's input is showing for " several coffee shops in one kilometer range of Wangjing subway station " In example, if obtained intent information (intent) is " poi_lookup " (inquiry poi), obtained slot position (slot) is " around_poi: Wangjing subway station ", " within distance: one kilometer " and " tag: coffee shop " these three slot positions, in conjunction with this Intent information and the demand of the available user of slot position are: user wants to search for coffee in the range in one kilometer of Wangjing subway station The Room.But the demand information " several coffee shops in one kilometer range of Wangjing subway station " with " have in one kilometer of Wangjing subway station There is no coffee shop " it compares, the intention of the two user demands is almost the same, but in fact, according to the difference of issue details, it should The two demand informations are carried out with different answers, to provide properer reply.Therefore, human-computer dialogue semanteme solution of the invention Analysis method, can be similar by intent information, slot position by type information (ques_type) the problem of parsing user demand information User demand distinguish, such as type is How_many the problem of " several coffee shops in one kilometer range of Wangjing subway station " The problem of type, " either with or without coffee shop in one kilometer of Wangjing subway station ", type was YN type, thus during human-computer dialogue, Different answers can be provided according to the difference of problem types, kept interactive system more intelligence more humanized, allowed user Perceptual image is engaged in the dialogue with people, rather than and machine.
The realm information for the user demand information that human-computer dialogue semanteme analytic method of the invention is parsed (domain), can play the role of correcting and revealing all the details during human-computer dialogue.When intent information and slot position form contradiction, nothing It, can be in conjunction between realm information above-mentioned, slot position and intent information when method determines user demand during human-computer dialogue Association rule carries out human-computer dialogue according to the semantic expressiveness after correction, supplement to supplement or predict the intent information of the demand Subsequent judgement and answer in the process；Or revealing all the details back for some unifications based on realm information (domain) can also be designed It answers, such as " less understanding your meaning, may I ask will be where " revealing all the details back as default for navigation field, can be used It answers, and for music field, " less understanding your meaning, what music may I ask will listen " can be set and reveal all the details back as default It answers, so that interactive system is more intelligent, more meets the habit of people.
In some embodiments, human-computer dialogue semanteme analytic method of the invention further comprises the steps of: in more wheel dialogues, will The semantic understanding of demand information described in epicycle dialogue indicate to indicate with the semantic understanding of the demand information of each wheel input before into Row combines, so that update by wheel to semantic understanding, addition, deleting, and maintain each round interaction in mostly wheel human-computer dialogue Semantic understanding information, with gradually clear demand.
It is exemplified below the scene of the demand information of some inputs, and is obtained by human-computer dialogue semanteme analytic method of the invention The semantic parsing result arrived.
Scene one, demand information are as follows: with removing B；
Semantic parsing result are as follows: ques_type:None, domain: navigation, slots:{ to_loc:B }, intent: to。
Scene two, demand information are as follows: with onboard from A navigating to B；
Semantic parsing result are as follows: ques_type:None, domain: navigation, slots:{ from_loc:A, dest_ Loc:B }, intent:from-to.
Scene three, demand information are as follows: whether have traffic lights；
Semantic parsing result are as follows: ques_type:yn, domain: navigation, intent:not_support, slots: { route_attribute: traffic lights }.
It is worth noting that, being possible to the case where can not parsing some characteristic information occur, at this moment can be united with one One mark expression does not parse this feature information, such as " intent:not_support " in scene three this example is exactly Intent information for indicating not parsing the demand information.The language obtained using human-computer dialogue semanteme analytic method of the invention Reason and good sense solution indicates, when lacking some characteristic information, still can carry out human-computer dialogue based on other characteristic informations parsed Judge in journey, such as in the example of scene three, although not parsing intent information, can still be by the problem of parsing type Relationship between information, realm information, slot position and preset intent information and realm information, slot position, to judge demand information General idea and carry out other human-computer dialogue processes.
It should be noted that human-computer dialogue semanteme analytic method of the invention is not limited to be applied to above-mentioned scene.
Further, the embodiment of the present invention also proposed a kind of vehicle-mounted interactive method comprising any of the above-described kind man-machine The step of dialog semantics analytic method.
Fig. 4 is the schematic diagram of human-computer dialogue semanteme resolution system one embodiment of the invention.Referring to Fig. 4, The exemplary human-computer dialogue semanteme resolution system of the present invention includes:
First parsing module 310, the demand information (query) for inputting to user carry out the first parsing, obtain first Parsing result, first parsing result include the characteristic information of the demand information.This feature information is for recording the demand letter The content and attribute of breath, the classification information of slot position (slot) and one or more demand informations including the demand information. Wherein, slot position is used to record the attribute of short sentence, phrase or word and the short sentence, phrase or word in demand information, classification letter Breath is for recording classification belonging to demand information.
Semantic expressiveness determining module 320, for obtaining the semantic expressiveness of demand information according to the first parsing result, the semanteme Indicate slot position and classification information comprising the demand information.
Further, can classify in various ways to demand information, such as according to problem types (ques_type) Classified, classified according to field (domain), classified according to intention (intent), thus the first parsing module 301 classification informations parsed may include problem types (ques_type) information, field (domain) information, intention (intent) information.The problem of determining demand information type information, realm information or intent information, seek to judge this Classification, the classification of affiliated intention of the classification of problem type, affiliated field belonging to demand information.
Issue type information therein, for characterizing interrogative sentence type belonging to demand information.The problem types include: non- Query type (None), ask whether type (YN), inquiry how/how type (How), inquiry venue type (Where), Query time type (When), inquiry quantity type (How_many), interrogating range type (How_far) etc..It should be noted that , issue type information is not limited to example enumerated above, in fact can be according to the actual situation to default problem class The type of type information is added or changes.
Realm information therein, for characterizing the type in field belonging to demand information.In fact, realm information can be regarded Classify for the summary of intent information.As an example, human-computer interaction scene in the car, realm information may include navigation, charging, Media, telephone relation, short message, information, specific equipment, search and interactive system chat etc..The field of should be noted that Information is not limited to example enumerated above, in fact can the type according to the actual situation to preset realm information add Add or changes.
Intent information therein for characterizing the specifically intended type of demand information is retouched to the refinement of realm information It states.It should be noted that intent information is not limited to example enumerated above, it in fact can be according to the actual situation to preset The type of intent information is added or changes.
Slot position therein, the specific element for recording the extraction from user demand information are (such as specific short sentence, short Language, phrase, word or word) and the specific element belonging to classification.And (alternatively referred to as slot position mentions for the parsing of slot position (slots) Take, slot position identification or slot position filling) be equivalent to extracted from the demand information of input relevant element (short sentence, phrase, phrase, Word, word etc.) it is filled into slot position according to the attribute of element, or be equivalent to and label is filled out to each element of demand information, it can It is regarded as sequence labelling.
In fact, can will have using the classification of specific element as a slot position classification (the alternatively referred to as title of slot position) The content of element of volume is filled into the slot position as the design parameter of the slot position, infinite specific element is mapped as limited Slot position, such as in the situation that specific element is proper noun；Or can using the content of specific element as a slot position classification, and The slot position does not have parameter, such as in the situation that specific element is a part of verb.As an example, the demand of an input Information includes " from Beijing to Hangzhou " this sentence, which includes " Beijing ", " Hangzhou " two elements, wherein the class in " Beijing " Not Wei " departure place " (from_loc), the classification in " Hangzhou " is " destination " (to_loc), should " departure place ", " destination " just It is the slot position in this example, and " Beijing " is the content of " departure place " this slot position record, " Hangzhou " is " destination " this slot The content of position record.
Optionally, slot position information can be recorded with the form of " slot position classification: slot position parameter ", wherein slot position classification is for remembering The classification of the element (word, phrase) in demand information sentence is recorded, actually a kind of classification of slot position can be used as this slot position Title, and slot position parameter is used to record the content of element (word, phrase) in demand information sentence.It should be noted that Slot position information is not limited to example enumerated above, in fact can be added according to the actual situation to preset slot position classification Or change.
Human-computer dialogue semanteme resolution system of the invention may also include classification information classification presetting module and slot position classification is pre- If module (is not drawn into) in figure, classification information classification presetting module is used to preset the type of issue type information, field letter The type of breath and the type of intent information, slot position classification presetting module are used to preset the type of slot position.Further, classify Information category presetting module and slot position classification presetting module can be according to the mutually interconnections between problem types, field, intention, slot position It is the type these characteristic informations are arranged.It is specific:
For realm information and intent information:
Since realm information is intended to the summary classification of information, it is intended that information is the refinement description to realm information, can be with The type of field, intention is arranged to connect each other:
Certain intentions only belong to one or more specific fields, for example, from somewhere to somewhere (from_to), go down (to), navigation field is generally corresponded only to；
In addition, same class is intended to included specifically intended information can be different, example for different specific areas Such as executing movement (action/action_xxx) this kind of intention, in field of media, it may include execution start, execute stopping, Execute that pause, to execute continuation, search song, switching song, change play mode etc. specifically intended, and in air-conditioning equipment field Execution acts this kind of be intended in addition to equally including that other than execution starts, executes stopping, executing pause, executes continuation, may also include It changes air conditioning mode, change set temperature, change wind direction etc. to be intended to, but can not include search song, switching song, change These intentions of play mode.
For slot position:
The corresponding relationship of slot position classification (slot position title) and problem types, field or intention can be set, such as can set Set problem type, field or intention corresponding to a slot position classification, or one problem types of setting or a field or one It is a to be intended to included slot position classification, so that slot position information is more acurrate.For example, " volume: volume " this slot position can be with " sound Amount adjusts (turn_volumn) " it is intended to correspond to, can be corresponding with " media " field, or may also be corresponding with " call " field, and It is general not corresponding with " go down (to) " intention, it is not corresponding with " navigation " field.
According to the association rule between above-mentioned problem type, field, intention, slot position, can be parsed by judgement Whether each characteristic information meets the association rule to determine whether the semantic parsing of this time succeeds, moreover it is possible to can not identify some spy When reference ceases, (or prediction) unrecognized information is supplemented using the association rule.
First parsing module 310 may include machine learning submodule 311, for being somebody's turn to do using machine learning model First parsing.For this reason, it may be necessary in advance to the number for largely having marked the characteristic informations such as great question type, field, intention and slot position According to machine learning training is carried out, machine learning model is obtained, when carrying out the first parsing to a demand information, which is believed Breath is input to this by the machine learning model of learning training, which will be exported corresponding to this demand information The problem of type, field, intention and slot position.
Further, machine learning submodule 311 can be specifically used for utilizing bidirectional circulating neural network combination attention machine The learning algorithm and the common learning algorithm of associated losses function of system carry out first parsing, in identification problem types, field, meaning Inner link when figure, slot position between consideration, to improve the accuracy rate of the first parsing.Specifically, machine learning submodule 311 first can convert term vector matrix for the demand information that user inputs, and pass through bidirectional circulating neural network combination attention machine System learnt so that the demand information is expressed as by field identification mission, it is intended that identification mission, problem types identification mission and Slot fills the term vector matrix of task sharing, reuses associated losses function and carrys out common learning areas identification mission, it is intended that identification Task, problem types identification mission and slot fill task.In some embodiments, machine learning submodule 311 specifically includes:
First unit, for learning to obtain the pre-training term vector of preset word (or word, phrase) based on large-scale corpus It indicates.
Second unit, each word (or word, phrase) in demand information for being inputted by participle, then this is needed Each word in information is asked to correspond on the term vector of the pre-training obtained by first unit, it is corresponding to obtain this demand information Term vector matrix.
Third unit, for obtaining the term vector of the demand information by second unit using bidirectional circulating neural network learning Matrix obtains the demand information in the expression ht_forward at bidirectional circulating neural network each moment and (corresponds to forward circulation Neural network), ht_backward (correspond to recycled back neural network), the demand information it is global indicate hu_forward, Hu_backward, the output Ct of attention mechanism, the global of attention mechanism indicate Cu.
Unit the 4th, the demand information for will be obtained by third unit is at bidirectional circulating neural network each moment Expression ht_forward, ht_backward and output Ct of attention mechanism is used to predict the label of each slot position, will be by third The global of the demand information that unit obtains indicates that hu_forward, hu_backward and the global of attention mechanism indicate Cu For predicting intention, the field, problem types of the demand information, so that the demand information of input indicates to be appointed by field identification Business, intention assessment task, problem types identification mission and slot filling task are shared.
Unit the 5th, for using the common learning areas identification mission of associated losses function, it is intended that identification mission, problem class Type identification mission and slot fill task, respectively obtain field identification mission, it is intended that identification mission, problem types identification mission and slot The class label of filling task.
Fig. 5 is the schematic diagram of another embodiment of human-computer dialogue semanteme resolution system of the invention.Due to the first solution The slot position that analysis module 310 obtains may be short sentence or phrase comprising one or more qualifiers (such as attribute, the adverbial modifier), might as well incite somebody to action First slot position is known as by the slot position that the first parsing module 310 obtains, to obtain more accurate semantic expressiveness, referring to Fig. 5, one In a little embodiments, human-computer dialogue semanteme resolution system of the invention includes:
First parsing module 310 above-mentioned, the demand information (query) for inputting to user carry out the first parsing, obtain To the first parsing result, the problem of which includes the demand information type information, realm information, intent information and First slot position of the demand information；
Second parsing module 330, to the first slot position of the demand information (query) obtained by the first parsing module 310 into Row second parses, and obtains the second parsing result, which includes that first slot position each word for being included is (or short Language), the relationship between the attribute of each word (or phrase) and each word (or phrase), with from this decompose insufficient first It is decomposed in the short sentence or phrase of slot position and extracts core word and qualifier, according to each word in the short sentence or phrase of the first slot position Attribute and each word between relationship, each word of the first slot position is filled into the second slot position more specifically refined, To express the demand information of input more accurately.Specifically, can be by syntactic analysis, participle, part-of-speech tagging, entity mark A kind of mode or the compositions of various ways carry out second parsing；
Semantic expressiveness determining module 320 above-mentioned, the result that result and second for combining the first parsing parse obtain The problem of semantic expressiveness of demand information, which includes the demand information type information, realm information, intent information and Second slot position.
As a specific example, for a demand information sentence, " coffee shop within 1 kilometer of Wangjing subway station has several Family ", the first parsing module 310 has obtained first slot position " poi: the coffee shop within 1 kilometer of Wangjing subway station ".Second solution Analysis module 330 specifically includes:
Unit the 6th, for obtaining " coffee shop within 1 kilometer of Wangjing subway station " by participle and interdependent syntactic analysis In specific element (generally word or expression) syntax tree, with determined according to the syntax tree core word in specific element and Determiner.Fig. 3 is the schematic diagram for the syntax tree that one embodiment of the present of invention provides.Referring to Fig. 3, in syntax tree The word node that tree root directly links is exactly core word, and determiner can be in the child nodes of core word in syntax tree.Originally showing " coffee shop " is core word in example, and " Wangjing subway station " and " within one kilometer " is all the determiner of " coffee shop ".
Unit the 7th obtains the attribute of each core word and each determiner for marking by part-of-speech tagging, entity, with By the attribute of specific element and/or the relationship of a specific element and other specific elements, to determine the specific element The second affiliated slot position classification.In this example, " coffee shop " part of speech is noun, and " Wangjing " is place name entity, to obtain core Heart word " coffee shop " is limited the limitation of word place name entity " Wangjing ".
As an example, the demand information of input is " several coffee shops in one kilometer range of Wangjing subway station ", this The semantic parsing result that the human-computer dialogue semanteme resolution system of invention obtains may is that
{
ques_type:How_many；
domain:Nav；
intent:poi_lookup；
Slots: " around_poi: Wangjing subway station ", " within distance: one kilometer ", " tag: coffee shop "；
}
Wherein, the problem of " ques_type:How_many " is demand information type, the problem of indicating the demand information class Type is inquiry quantity type；" domain:Nav " is the domain type of demand information, indicates the field of the demand information for navigation； " intent:poi_lookup " is the intention type of demand information, and indicate the demand information is intended to inquiry point of interest (point of interest, poi)；" slots:around_poi: Wangjing subway station, within distance: one kilometer, tag: Coffee shop " is the language slot position of demand information, which includes three language slot positions.
It, can be more clear according to the semantic meaning representation comprising issue type information (ques_type), realm information (domain) Chu, accurately expression user actual demand, and can make interactive system carry out more accurately judgement and it is more smart Thin subsequent processing.
Specifically, the demand information in user's input is showing for " several coffee shops in one kilometer range of Wangjing subway station " In example, if obtained intent information (intent) is " poi_lookup " (inquiry poi), obtained slot position (slot) is " around_poi: Wangjing subway station ", " within distance: one kilometer " and " tag: coffee shop " these three slot positions, in conjunction with this Intent information and the demand of the available user of slot position are: user wants to search for coffee in the range in one kilometer of Wangjing subway station The Room.But the demand information " several coffee shops in one kilometer range of Wangjing subway station " with " have in one kilometer of Wangjing subway station There is no coffee shop " it compares, the intention of the two user demands is almost the same, but in fact, according to the difference of issue details, it should The two demand informations are carried out with different answers, to provide properer reply.Therefore, human-computer dialogue semanteme solution of the invention Analysis system, can be similar by intent information, slot position by type information (ques_type) the problem of parsing user demand information User demand distinguish, such as type is How_many the problem of " several coffee shops in one kilometer range of Wangjing subway station " The problem of type, " either with or without coffee shop in one kilometer of Wangjing subway station ", type was YN type, so that interactive system can be with Different answers is provided according to the difference of problem types, keeps interactive system more intelligence more humanized, user is allowed to feel It seem to engage in the dialogue with people, rather than and machine.
The realm information for the user demand information that human-computer dialogue semanteme resolution system of the invention is parsed (domain), can play the role of correcting and revealing all the details during human-computer dialogue.When intent information and slot position form contradiction, people When machine conversational system can not determine user demand, contacting between realm information above-mentioned, slot position and intent information can be combined Rule carries out interactive system according to the semantic expressiveness after correction, supplement to supplement or predict the intent information of the demand Subsequent judgement and answer；Or the answer of revealing all the details of some unifications based on realm information (domain) can also be designed, such as For navigation field, " less understanding your meaning, may I ask will be where " can be used as the answer of revealing all the details defaulted, and for " less understanding your meaning, what music may I ask will listen " can be set as answer of revealing all the details is defaulted, so that man-machine in music field Conversational system is more intelligent, more meets the habit of people.
In some embodiments, human-computer dialogue semanteme resolution system of the invention further includes semantic update module, is used for In more wheel dialogues, by epicycle talk with described in the semantic understanding of demand information indicate language with the demand information of each wheel input before The expression of reason and good sense solution is combined, so that update by wheel to semantic understanding, addition, deleting, and tie up in mostly wheel human-computer dialogue The semantic understanding information of each round interaction is held, with gradually clear demand.
Further, the embodiment of the present invention also proposed a kind of vehicle-mounted interactive system comprising any of the above-described kind man-machine Dialog semantics resolution system.
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 human-computer dialogue semanteme solution when being executed by the processor The step of analysis method.It should be appreciated that the people that the instruction stored in memory is and it can be realized when being executed by processor The step of specific example of machine dialog semantics analytic method, is corresponding.
Further, the embodiment of the present invention also proposed a kind of computer readable storage medium, for storing computer instruction, The step of described instruction realizes any of the above-described kind of human-computer dialogue semanteme analytic method when by a computer or processor execution.It answers This is it is understood that the instruction stored in computer readable storage medium is the human-computer dialogue semanteme that can be realized when executed with it The step of specific example of analytic 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 human-computer dialogue semanteme analytic method, comprising the following steps:
First parsing is carried out to the demand information of input, obtain the demand information the first slot position and the demand information One or more classification informations；First slot position is used to record short sentence, phrase or word in the demand information and described The attribute of short sentence, phrase or word；The classification information is for recording classification belonging to the demand information；
The semantic expressiveness of the demand information is obtained according to first slot position and the classification information.
2. human-computer dialogue semanteme analytic method according to claim 1, wherein the classification information includes: problem types letter One of breath, realm information, intent information are some；
Described problem type information, for recording the interrogative sentence classification of the demand information；
The realm information, for recording the field classification of the demand information；
The intent information, for recording the intention classification of the demand information.
3. human-computer dialogue semanteme analytic method according to claim 2, further comprises: presetting described problem type letter The type of the type of breath, the type of the realm information and the intent information, and one is correspondingly arranged for the field of a type A or multiple intentions, or one or more fields are correspondingly arranged for the intention of a type.
4. human-computer dialogue semanteme analytic method according to claim 3, further comprises: presetting first slot position Type, and one or more first slot positions are correspondingly arranged for the field of a type, or corresponding for the intention of a type One or more first slot positions of setting, or type is correspondingly arranged one or more first slots the problem of be a type Position.
5. human-computer dialogue semanteme analytic method according to claim 2, wherein first parsing are as follows: utilize bidirectional circulating The learning algorithm and the common learning algorithm of associated losses function of neural network combination attention mechanism, obtain the demand information Described problem type information, the realm information, the intent information and first slot position.
6. human-computer dialogue semanteme analytic method according to claim 5, wherein described utilizes bidirectional circulating neural network knot The learning algorithm and the common learning algorithm of associated losses function for closing attention mechanism obtain the described problem of the demand information Type information, the realm information, the intent information and first slot position, comprising the following steps:
The pre-training term vector for learning to obtain default phrase or default word based on large-scale corpus indicates；
Each phrase or the word in the demand information are obtained by participle, each phrase or word of the demand information is corresponding Onto the pre-training term vector, the corresponding term vector matrix of the demand information is obtained；
The study that bidirectional circulating neural network combination attention mechanism is carried out to the demand information term vector matrix obtains described Demand information the expression at bidirectional circulating neural network each moment, the demand information it is global indicate, attention mechanism Output and the global of attention mechanism indicate；
Using the demand information in the expression at bidirectional circulating neural network each moment and attention mechanism output come pre- The label for surveying first slot position of the demand information, global using the demand information indicate and the attention mechanism The overall situation indicates the intent information, the realm information and described problem type information to predict the demand information, so that Obtaining the demand information indicates to fill task institute by field identification mission, intention assessment task, problem types identification mission and slot It is shared；
It is filled out using the common learning areas identification mission of associated losses function, intention assessment task, problem types identification mission and slot Fill the post of business, respectively obtains the classification mark of field identification mission, intention assessment task, problem types identification mission and slot filling task Label.
7. human-computer dialogue semanteme analytic method according to claim 2, further includes steps of
Second parsing is carried out to first slot position, obtains each word for including in first slot position, and obtain described each Relationship between the attribute of word and/or each word, according between the attribute and each word of each word Each word is filled into the second slot position by relationship；
The language of the demand information is obtained according to second slot position and described problem type information, realm information, intent information Justice indicates.
8. human-computer dialogue semanteme analytic method according to claim 7, wherein described carry out the second solution to first slot position Analysis the following steps are included:
Participle and interdependent syntactic analysis are carried out to first slot position, the syntax tree of first slot position is obtained, according to the sentence Method tree determines the core word and determiner in the first slot position；
It marks to obtain the attribute of the core word and the attribute of the determiner by part-of-speech tagging, entity, passes through described first The relationship of a word and other words in the attribute of each word in slot position and/or first slot position, to determine, predicate is corresponding The second slot position.
9. human-computer dialogue semanteme analytic method according to claim 1, further comprises: by epicycle talk with described in demand believe The semantic expressiveness of breath and the semantic expressiveness of the demand information of each wheel input before are combined.
10. a kind of human-computer dialogue semanteme resolution system, comprising:
First parsing module obtains the first slot position of the demand information for carrying out the first parsing to the demand information of input And one or more classification informations of the demand information；First slot position is short in the demand information for recording The attribute of sentence, phrase or word and the short sentence, phrase or word；The classification information is for recording belonging to the demand information Classification；
Semantic expressiveness determining module, for obtaining the semanteme of the demand information according to first slot position and the classification information It indicates.
11. a kind of vehicle-mounted interactive method is parsed including human-computer dialogue semanteme described in any one of claims 1 to 9 The step of method.
12. a kind of vehicle-mounted interactive system, including human-computer dialogue semanteme resolution system described in any one of claim 10.
CN201810267052.3A 2018-03-28 2018-03-28 Man-machine conversation semantic analysis method and system, vehicle-mounted man-machine conversation method and system, controller and storage medium Active CN110309277B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810267052.3A CN110309277B (en) 2018-03-28 2018-03-28 Man-machine conversation semantic analysis method and system, vehicle-mounted man-machine conversation method and system, controller and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810267052.3A CN110309277B (en) 2018-03-28 2018-03-28 Man-machine conversation semantic analysis method and system, vehicle-mounted man-machine conversation method and system, controller and storage medium

Publications (2)

Publication Number Publication Date
CN110309277A true CN110309277A (en) 2019-10-08
CN110309277B CN110309277B (en) 2023-08-18

Family

ID=68073813

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810267052.3A Active CN110309277B (en) 2018-03-28 2018-03-28 Man-machine conversation semantic analysis method and system, vehicle-mounted man-machine conversation method and system, controller and storage medium

Country Status (1)

Country Link
CN (1) CN110309277B (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866090A (en) * 2019-11-14 2020-03-06 百度在线网络技术（北京）有限公司 Method, apparatus, electronic device and computer storage medium for voice interaction
CN110931009A (en) * 2019-12-12 2020-03-27 贵州电力交易中心有限责任公司 System for rapidly improving conversation capacity of reception robot in electric power transaction hall
CN111177310A (en) * 2019-12-06 2020-05-19 广西电网有限责任公司 Intelligent scene conversation method and device for power service robot
WO2021147041A1 (en) * 2020-01-22 2021-07-29 华为技术有限公司 Semantic analysis method and apparatus, device, and storage medium
CN113254610A (en) * 2021-05-14 2021-08-13 廖伟智 Multi-round conversation generation method for patent consultation
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 (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004045900A (en) * 2002-07-12 2004-02-12 Toyota Central Res & Dev Lab Inc Voice interaction device and program
CN106156003A (en) * 2016-06-30 2016-11-23 北京大学 A kind of question sentence understanding method in question answering system
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
US20180032082A1 (en) * 2016-01-05 2018-02-01 Mobileye Vision Technologies Ltd. Machine learning navigational engine with imposed constraints

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004045900A (en) * 2002-07-12 2004-02-12 Toyota Central Res & Dev Lab Inc Voice interaction device and program
US20180032082A1 (en) * 2016-01-05 2018-02-01 Mobileye Vision Technologies Ltd. Machine learning navigational engine with imposed constraints
CN106156003A (en) * 2016-06-30 2016-11-23 北京大学 A kind of question sentence understanding method in question answering system
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

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110866090A (en) * 2019-11-14 2020-03-06 百度在线网络技术（北京）有限公司 Method, apparatus, electronic device and computer storage medium for voice interaction
US11830482B2 (en) 2019-11-14 2023-11-28 Baidu Online Network Technology (Beijing) Co., Ltd Method and apparatus for speech interaction, and computer storage medium
CN111177310A (en) * 2019-12-06 2020-05-19 广西电网有限责任公司 Intelligent scene conversation method and device for power service robot
CN111177310B (en) * 2019-12-06 2023-08-18 广西电网有限责任公司 Intelligent scene conversation method and device for power service robot
CN110931009A (en) * 2019-12-12 2020-03-27 贵州电力交易中心有限责任公司 System for rapidly improving conversation capacity of reception robot in electric power transaction hall
WO2021147041A1 (en) * 2020-01-22 2021-07-29 华为技术有限公司 Semantic analysis method and apparatus, device, and storage medium
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
CN113254610A (en) * 2021-05-14 2021-08-13 廖伟智 Multi-round conversation generation method for patent consultation

Also Published As

Publication number Publication date
CN110309277B (en) 2023-08-18

Similar Documents

Publication Publication Date Title
CN110309277A (en) Human-computer dialogue semanteme parsing method and system
CN108763510B (en) Intention recognition method, device, equipment and storage medium
CN109670024B (en) Logic expression determination method, device, equipment and medium
CN111708869B (en) Processing method and device for man-machine conversation
CN111159385B (en) Template-free general intelligent question-answering method based on dynamic knowledge graph
CN111081220B (en) Vehicle-mounted voice interaction method, full-duplex dialogue system, server and storage medium
CN111062217B (en) Language information processing method and device, storage medium and electronic equipment
CN111738016A (en) Multi-intention recognition method and related equipment
CN106057200A (en) Semantic-based interaction system and interaction method
CN116959433B (en) Text processing method, device, electronic equipment and storage medium
CN109933792A (en) Viewpoint type problem based on multi-layer biaxially oriented LSTM and verifying model reads understanding method
CN115525753A (en) Task-oriented multi-turn dialogue method and system based on 1+ N
CN108364646A (en) Embedded speech operating method, device and system
CN111966809B (en) Knowledge question and answer method and device based on multiple rounds of conversations
CN115630146A (en) Method and device for automatically generating demand document based on human-computer interaction and storage medium
CN109189882A (en) Answer type recognition methods, device, server and the storage medium of sequence content
CN110969276B (en) Decision prediction method, decision prediction model obtaining method and device
CN110377691A (en) Method, apparatus, equipment and the storage medium of text classification
CN112036186A (en) Corpus labeling method and device, computer storage medium and electronic equipment
CN117407507A (en) Event processing method, device, equipment and medium based on large language model
CN115689603A (en) User feedback information collection method and device and user feedback system
CN114036268A (en) Task type multi-turn dialogue method and system based on intention gate
CN111914538B (en) Channel notification information intelligent space matching method and system
CN111681670B (en) Information identification method, device, electronic equipment and storage medium
CN111767465A (en) Information recommendation method and device, computing equipment and medium

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

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.

TA01 Transfer of patent application right
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.

CB02 Change of applicant information
GR01 Patent grant
GR01 Patent grant