CN108563790A - A kind of semantic understanding method and device, equipment, computer-readable medium - Google Patents
A kind of semantic understanding method and device, equipment, computer-readable medium Download PDFInfo
- Publication number
- CN108563790A CN108563790A CN201810401066.XA CN201810401066A CN108563790A CN 108563790 A CN108563790 A CN 108563790A CN 201810401066 A CN201810401066 A CN 201810401066A CN 108563790 A CN108563790 A CN 108563790A
- Authority
- CN
- China
- Prior art keywords
- word
- sentence
- generic clause
- information
- domain classification
- 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
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
Abstract
This application provides a kind of semantic understanding method and devices, equipment, according to preset domain classification, sentence to be understood is divided into the first generic clause, again according to preset intent classifier, each first generic clause is divided into the second generic clause, the finally extraction tank information from each second generic clause, wherein, each first generic clause belongs to a classification in domain classification, each second generic clause belongs to a classification in intent classifier, it can be seen that, semantic understanding method and device described herein, using different partition dimensions, in the way of dividing clause step by step, it can obtain the slot information under multi-field more intentions, therefore, the sentence more being intended to can be directed to, obtain the higher understanding results being intended to of accuracy more.
Description
Technical field
This application involves electronic information field more particularly to a kind of semantic understanding method and devices, equipment, computer-readable
Medium.
Background technology
With the continuous breakthrough of artificial intelligence technology and becoming increasingly popular for various intelligent terminals, human-computer interaction is in people
The frequency occurred in routine work and life is higher and higher.Interactive voice has become as most easily one of interactive mode
The important means of human-computer interaction.Human-computer interaction conversational system is used in various intelligent terminals, such as:TV, mobile phone, vehicle
Load, smart home and robot etc..And how to understand that being intended that for user is the most key in human-computer interaction conversational system
Technology.
Since human language has diversity, complexity and ambiguousness, a variety of intentions can be often in short expressed, but work as
Preceding human-computer interaction conversational system is merely able to provide a kind of semantic understanding as a result, (be intended to for the voice for being related to more being intended to more
Can be understood as at least two events, such as " open air-conditioning, close radio "), it may appear that no result only provides one kind
The problem of semantic understanding result, such as:" first open air-conditioning and be then shut off radio ", may be without output as a result, output result:
" opening air-conditioning ", in another example:" navigate to WanDa Plaza and select a not stifled route " may export result:" navigation is not blocked up
Route ".
As it can be seen that existing semantic understanding method, the understanding of the sentence for more being intended to are not accurate enough.
Invention content
This application provides a kind of semantic understanding method and device, equipment, computer-readable mediums, it is therefore intended that solution pair
In the problem that the understanding for the sentence more being intended to is not accurate enough.
To achieve the goals above, this application provides following technical schemes:
A kind of semantic understanding method, including:
According to preset domain classification, sentence to be understood is divided into the first generic clause, wherein each first kind
Clause belongs to a classification in the domain classification;
According to preset intent classifier, each first generic clause is divided into the second generic clause, wherein each described
Second generic clause belongs to a classification in the intent classifier;
The extraction tank information from each second generic clause.
Optionally, described according to preset domain classification, sentence to be understood, which is divided into the first generic clause, includes:
Using each word in the sentence to be understood as input, using field discrimination model, obtain described waiting understanding
Sentence in each word field label, the field label of any one word in the sentence to be understood includes:
The title and first position information of domain classification belonging to the word, the first position information are that the word is meeting first condition
All words in location information, the first condition is to belong to identical domain classification with the word in the sentence to be understood
Word, the position includes starting, intermediate and terminate.
Optionally, the training process of the field discrimination model includes:
For each classification in the preset domain classification, the first preset quantity sentence is collected respectively, and right
The sentence of collection marks the field label;
It collects the second preset quantity item and is not belonging to the sentence of any one of described preset domain classification classification, and mark
Note is not belonging to the label of the preset domain classification;
Using the object function of sentence training the first sequence marking model after mark, the field discrimination model is obtained.
Optionally, described according to preset intent classifier, each first generic clause is divided into the second generic clause packet
It includes:
Using each word with same area specific name in first generic clause as input, using with described the
The corresponding intention discrimination module of domain classification belonging to one generic clause, obtains the intention mark of each word in first generic clause
It signs, the intention labels of any one word in first generic clause include:The title of intent classifier belonging to the word with
And second position information, the second position information is location information of the word in all words for meeting second condition, described
Second condition is the word for belonging to identical domain classification and identical intent classifier in first generic clause with the word, described
Position includes starting, centre and terminates, wherein each domain classification corresponds to an intention discrimination module, different field point
The corresponding intention discrimination module of class is different.
Optionally, the corresponding training method for being intended to discrimination module of any one of domain classification includes:
The third preset quantity sentence for belonging to the domain classification is collected, and the intention is marked to every sentence of collection
Label;
The sentence that the 4th preset quantity is not belonging to the classification of any one of preset intent classifier is collected, and marks and does not belong to
In the label of the preset domain classification;
Using the object function of sentence training the second sequence marking model after mark, the corresponding institute of the domain classification is obtained
State intention discrimination module.
Optionally, the extraction tank information from each second generic clause includes:
Using the intention vector of each word and second generic clause in the second generic clause as input, using with it is described
The corresponding slot information discrimination model of domain classification belonging to second generic clause, obtains the slot of each word in second generic clause
The slot information labels of information labels, any one word in second generic clause include:The title of slot information belonging to the word
And the third place information, the third place information are location information of the word in all words for meeting third condition, institute
It is to belong to identical domain classification, identical intent classifier in second generic clause with the word and have phase to state third condition
The word of same slot information;
Wherein, each domain classification corresponds to a slot information discrimination module, the corresponding slot information of different domain classifications
Discrimination module is different.
Optionally, the training method of the corresponding slot information discrimination module of any one of domain classification includes:
The 5th preset quantity article sentence for belonging to the domain classification is collected, and slot information labels are marked to the sentence of collection;
The sentence that the 6th preset quantity is not belonging to any one of preset slot information slot information is collected, and marks and does not belong to
In the label of the preset slot information;
The object function for the marking model that sorted using the sentence training third after mark, obtains the corresponding institute of the domain classification
State slot information discrimination module.
Optionally, further include after extraction tank information in each second generic clause described:
By the institute in the word segmentation result of the sentence to be understood, slot information candidate path and the sentence to be understood
The intention vector of the second generic clause is stated as input, using preset model, is scored slot information candidate path, the slot
It is combined and is formed by least one slot information extracted from each second generic clause in information candidate path;
According to the scoring, slot information extraction result is determined.
A kind of semantic understanding device, including:
First division module, for according to preset domain classification, sentence to be understood to be divided into the first generic clause,
In, each first generic clause belongs to a classification in the domain classification;
Second division module, for according to preset intent classifier, each first generic clause to be divided into the second class
Clause, wherein each second generic clause belongs to a classification in the intent classifier;
Extraction module, for the extraction tank information from each second generic clause.
Optionally, first division module is used for according to preset domain classification, and sentence to be understood is divided into the
One generic clause includes:
First division module is specifically used for, and using each word in the sentence to be understood as input, uses neck
Domain discrimination model obtains the field label of each word in the sentence to be understood, arbitrary in the sentence to be understood
The field label of one word includes:The title and first position information of domain classification belonging to the word, described first
Confidence breath is location information of the word in all words for meeting first condition, and the first condition is the sentence to be understood
In belong to the word of identical domain classification with the word, the position includes starting, intermediate and terminate.
Optionally, first division module is additionally operable to, with the following method, the training field discrimination model:
For each classification in the preset domain classification, the first preset quantity sentence is collected respectively, and right
The sentence of collection marks the field label;
It collects the second preset quantity item and is not belonging to the sentence of any one of described preset domain classification classification, and mark
Note is not belonging to the label of the preset domain classification;
Using the object function of sentence training the first sequence marking model after mark, the field discrimination model is obtained.
Optionally, second division module is used for according to preset intent classifier, and each first generic clause is drawn
Being divided into the second generic clause includes:
Second division module is specifically used for, by each in first generic clause with same area specific name
Word as input, using intention discrimination module corresponding with the domain classification belonging to first generic clause, obtain described the
The intention labels of the intention labels of each word in one generic clause, any one word in first generic clause include:
The title and second position information of intent classifier belonging to the word, the second position information are that the word is meeting second condition
All words in location information, the second condition be first generic clause in the word belong to identical domain classification with
And the word of identical intent classifier, the position include starting, centre and terminate, wherein each domain classification corresponds to one
It is intended to discrimination module, the corresponding intention discrimination module of different domain classifications is different.
Optionally, second division module is additionally operable to, in accordance with the following methods, any one of domain classification pair of training
The intention discrimination module answered:
The third preset quantity sentence for belonging to the domain classification is collected, and the intention is marked to every sentence of collection
Label;
The sentence that the 4th preset quantity is not belonging to the classification of any one of preset intent classifier is collected, and marks and does not belong to
In the label of the preset domain classification;
Using the object function of sentence training the second sequence marking model after mark, the corresponding institute of the domain classification is obtained
State intention discrimination module.
Optionally, the extraction module be used for from each second generic clause extraction tank information include:
The extraction module is specifically used for, by the intention of each word and second generic clause in the second generic clause to
Amount is as input, using slot information discrimination model corresponding with the domain classification belonging to second generic clause, obtains described the
The slot information labels of the slot information labels of each word in two generic clauses, any one word in second generic clause include:
The title and the third place information of slot information belonging to the word, the third place information are that the word is meeting third condition
Location information in all words, the third condition are to belong to identical domain classification, phase with the word in second generic clause
The word of same intent classifier and slot information having the same;
Wherein, each domain classification corresponds to a slot information discrimination module, the corresponding slot information of different domain classifications
Discrimination module is different.
Optionally, the extraction module is additionally operable to, and in accordance with the following methods, any one of domain classification of training is corresponding
The slot information discrimination module:
The 5th preset quantity article sentence for belonging to the domain classification is collected, and slot information labels are marked to the sentence of collection;
The sentence that the 6th preset quantity is not belonging to any one of preset slot information slot information is collected, and marks and does not belong to
In the label of the preset slot information;
The object function for the marking model that sorted using the sentence training third after mark, obtains the corresponding institute of the domain classification
State slot information discrimination module.
Optionally, further include:
Screening module is used in the extraction module after extraction tank information in each second generic clause, by institute
State second generic clause in word segmentation result, slot information candidate path and the sentence to be understood of sentence to be understood
Intention vector as input, using preset model, score slot information candidate path, slot information candidate path
It is combined and is formed by least one slot information extracted from each second generic clause;And according to the scoring, determine that slot is believed
Breath extraction result.
A kind of semantic understanding equipment, including:
Memory and processor;
The memory is for storing one or more programs;
The processor is for executing one or more of programs, so that the semantic understanding equipment realizes the above institute
State semantic understanding method.
A kind of computer-readable medium is stored with instruction in the computer readable storage medium, when its on computers
When operation so that computer executes semantic understanding method described above.
Semantic understanding method and device described herein, equipment, according to preset domain classification, by sentence to be understood
It is divided into the first generic clause, then according to preset intent classifier, each first generic clause is divided into the second generic clause, most
The extraction tank information from each second generic clause afterwards, wherein each first generic clause belongs to a classification in domain classification, often
A second generic clause belongs to a classification in intent classifier, it is seen then that semantic understanding method and device described herein uses
Different partition dimensions can obtain the slot information under multi-field more intentions, therefore, energy in the way of dividing clause step by step
It is enough to be directed to the sentence more being intended to, obtain the higher understanding results being intended to of accuracy more.
Description of the drawings
In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below
There is attached drawing needed in technology description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of application for those of ordinary skill in the art without creative efforts, can be with
Obtain other attached drawings according to these attached drawings.
Fig. 1 is a kind of flow chart of semantic understanding method disclosed in the embodiment of the present application;
Fig. 2 is the flow chart of the specific implementation process of semantic understanding method disclosed in the embodiment of the present application;
Fig. 3 is to choose a road in semantic understanding method disclosed in the embodiment of the present application from a plurality of slot information candidate path
The flow chart of diameter;
Fig. 4 is a kind of structural schematic diagram of semantic understanding device disclosed in the embodiment of the present application.
Specific implementation mode
The embodiment of the present application discloses a kind of semantic understanding method and device, can apply with human-computer interaction function
In electronic equipment, such as the equipment such as TV, mobile phone, vehicle-mounted, smart home and robot.
Below in conjunction with the attached drawing in the embodiment of the present application, by taking intelligent vehicle-mounted system as an example, in the embodiment of the present application
Technical solution is clearly and completely described, it is clear that and described embodiments are only a part of embodiments of the present application, without
It is whole embodiment.Based on the embodiment in the application, those of ordinary skill in the art are not before making creative work
The every other embodiment obtained is put, shall fall in the protection scope of this application.
Fig. 1 is a kind of semantic understanding method disclosed in the embodiment of the present application, is included the following steps:
S101:According to preset domain classification, sentence to be understood is divided into clause, wherein each clause belongs to one
A domain classification.For the ease of subsequently illustrating, the clause divided in the step is known as the first generic clause.
Wherein, sentence to be understood can be the sentence gone out according to the speech recognition of collected user.
Table 1 is by taking onboard system as an example, and preset domain classification includes:Map, phone, radio, weather and order control
System.
Table 1
For example, sentence to be understood is:" open air-conditioning and play Anhui Traffic Announcement ", the first generic clause marked off
(belong to the radio neck in table 1 for " opening air-conditioning " (the order control field belonged in table 1) and " playing Anhui Traffic Announcement "
Domain).
In another example sentence to be understood is:" navigate to WanDa Plaza and select most fast route ", the first generic clause marked off
Still it is former sentence:" navigate to WanDa Plaza and select most fast route " (belonging to the map field in table 1).
S102:According to preset intent classifier, each first generic clause is divided into the second generic clause, wherein each
Second generic clause belongs to an intent classifier.
Table 2 shows the intent classifier under each domain classification shown in table 1 by taking onboard system as an example, for example, order control
Under domain classification processed, including intent classifier have:It opens, close and returns.
Table 2
Example is connected, sentence to be understood is:" open air-conditioning and play Anhui Traffic Announcement ", the first generic clause marked off is
" opening air-conditioning " and " playing Anhui Traffic Announcement ".Wherein, the second generic clause that clause's " opening air-conditioning " marks off still is " to open
Air-conditioning " (belong to opening be intended to) under the order control field in table 2, clause's " playing Anhui Traffic Announcement " mark off second
Generic clause is still " playing Anhui Traffic Announcement " (broadcasting belonged under the radio field in table 2 is intended to).
In another example sentence to be understood is:" navigate to WanDa Plaza and select most fast route ", the first generic clause marked off
For:" navigate to WanDa Plaza and select most fast route ", the second generic clause marked off are that " navigating to WanDa Plaza " (belongs to table 2
In map field under navigation purposes) and " selection most fast route " (route belonged under the map field in table 2 is intended to).
S103:The extraction tank information from each second generic clause.
Wherein, slot information can be understood as keyword.
Table 3 is the example of the slot information under domain classification and intent classifier shown in table 2.
Table 3
For example, the slot information of the second generic clause " navigating to WanDa Plaza " extraction is " WanDa Plaza " (endLoc).It needs
Illustrating, each is intended to the part that the lower slot information extracted can be pre-defined slot information, and not all, for example,
The slot information of " navigating to WanDa Plaza " extraction be " WanDa Plaza " (endLoc) rather than startLoc endLoc
Landmark whole in distance.
It should be noted that table shown in the present embodiment is only the citing for intelligent vehicle-mounted system, in practice,
Different domain classifications, intent classifier and slot information can be formulated according to different product demands.
The semantic understanding method described in the present embodiment is can be seen that from process shown in FIG. 1, first from the dimension of domain classification
Degree, is divided into the first generic clause, then from the dimension of intent classifier, each first generic clause is divided by sentence to be understood
Second generic clause finally extracts the slot information of each second generic clause, included therefore, it is possible to obtain in sentence to be understood
Different field classify and different intent classifier under slot information, that is to say, that the result of semantic understanding include different field not
Therefore for the sentence to be identified more being intended to, it can completely understand with the recognition result under being intended to and multiple intentions, improve reason
The accuracy of solution.
Also, applicant has found in the course of the study, existing semantic understanding method, why can not understand it is more
A intention is because if using existing semantic understanding algorithm, it is understood that the algorithm performs difficulty for going out multiple intentions is high:Although
Theoretically, it can understand the modes of multiple intentions from understanding that the mode for an intention expands to, it is contemplated that algorithm
The reasons such as complexity and the support of hardware, in fact, being appreciated that out that the algorithm of multiple intentions is difficult to be performed.And Fig. 1 institutes
The semantic understanding method shown, splits clause step by step, the object that each clause that upper level is split splits as next stage, it is this by
The mode that grade splits, understands step by step has higher enforceability.
Each step in Fig. 1, can use trained model realization, be described more detail below how training pattern
And how to use trained model realization step shown in FIG. 1.
Fig. 2 is the specific implementation flow of step shown in FIG. 1, is included the following steps:
S201:Using field discrimination model, the field label of each word in sentence to be understood is obtained.
The input of sentence that will be to be understood as field discrimination model, field discrimination model export in sentence to be understood
Each word field label.
The field label of any one word in sentence to be understood includes:The title of domain classification belonging to the word and
First position information, first position information are location information of the word in all words for meeting first condition, wherein first
Part is the word for belonging to identical domain classification in sentence to be understood with the word.Specifically, position includes starting (such as start information
Indicated with b), intermediate (such as average information is indicated with m) and terminate (such as ending message is indicated with e).It should be noted that if
One word is not belonging to the classification of any one of preset domain classification, then output is specific, can distinguish and preset field
The mark of bookmark name, such as specifically it is identified as s.It should be noted that domain classification is labeled as the word of s, it can be without
The identification of follow-up intention and slot information, to save resource.
Table 4 is the example of the field label of each word of output after the input by sentence field discrimination model understood.This reality
The example in example is applied, by taking the domain classification, intent classifier and slot information illustrated in table 3 as an example.
Table 4
Example 1:It opens air-conditioning and plays Anhui Traffic Announcement
pos | It beats | It opens | It is empty | It adjusts | It broadcasts | It puts | Peace | Emblem | It hands over | It is logical | Extensively | It broadcasts | pos |
pos | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | pos |
pos | cmd_b | cmd_m | cmd_m | cmd_e | radio_b | radio_m | radio_m | radio_m | radio_m | radio_m | raido_m | radio_e | pos |
Example 2:It navigates to WanDa Plaza and selects most fast route
pos | It leads | Boat | It arrives | Ten thousand | It reaches | Extensively | Choosing | It selects | Most | Soon | Road | Line | pos | |
pos | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | pos |
pos | map_b | map_m | map_m | map_m | map_m | map_m | map_m | map_m | map_m | map_m | map_m | map_m | map_e | pos |
Example 3:Inquire the price of No. 92 gasoline
pos | It looks into | It askes | One | Under | 92 | Number | Vapour | Oil | 's | Valence | Lattice | pos |
pos | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | ↓ | pos |
pos | s | s | s | s | s | s | s | s | s | s | s | pos |
By taking example 1 as an example, in " beating " corresponding field label cmd_b, cmd indicates the title of domain classification:Order control,
B indicates that " beating " position shared in all words for belonging to the fields cmd is initial position, i.e., " beating " is belong to the fields cmd the
One word.
In example 3, it is not determined by the domain classification of each word, therefore, is exported as s.
Based on first position information, it can be seen that after obtaining the field label of each word, that is, mark off the first generic clause.
In the present embodiment, each word can be used as a vectorial input field discrimination model, and vectorial length can be 100.
In above each exemplary table, POS indicates separator.
In the present embodiment, field discrimination model uses but is not limited to " the word insertion two-way long short-term memories of embedding+
Model lstm+ condition random field crf " neural network models (other sequence marking models).The training of field discrimination model
Method is:For each classification in preset domain classification, the first preset quantity such as 100,000 sentences are collected respectively, often
Sentence marks field label.It collects the second preset quantity such as 200,000 and is not belonging to any one of preset domain classification
The sentence of domain classification, and the field label for marking every sentence is s.Using the corpus data with label of above-mentioned collection,
The object function of training neural network model:.Wherein,For prediction field label it is general
Rate value.Specific training algorithm may refer to the prior art, and which is not described herein again.
S202:Using intention discrimination module corresponding with the domain classification belonging to the first generic clause, the meaning of each word is obtained
Icon label.Wherein, the domain classification belonging to the first generic clause is by the field tag representation that is obtained in S201.
In the present embodiment, each domain classification in preset domain classification, corresponding intention discrimination module.By
Word of each in one generic clause with same area specific name is as the corresponding input for being intended to discrimination module, it is intended that differentiates
Module exports the intention labels of each word.
The intention labels of any one word in first generic clause include:The title of intent classifier belonging to the word and
Two location informations, second position information are location information of the word in all words for meeting second condition, and second condition is institute
State the word for belonging to identical domain classification and identical intent classifier in the first generic clause with the word.Specifically, position includes
It originates (such as start information is indicated with b), intermediate (such as average information is indicated with m) and terminates (such as ending message is indicated with e).Such as
One word of fruit is not belonging to any one classification in preset intent classifier, then exports specific identifier, such as s.
Table 5 be belong to order control and map domain classification word input corresponding intention discrimination module respectively after, output
Each word intention labels example.
Table 5
Example 1:It opens air-conditioning and plays Anhui Traffic Announcement
Example 2:It navigates to WanDa Plaza and selects most fast route
For " the beating " of example 1 in table 4, it is intended that in label open_b, open indicates the opening under order control field
It is intended to, b indicates to be in the word that " beating " belongs to identical intent classifier in the clause's " opening air-conditioning " for belonging to order control
Beginning position.
Based on second position information, it can be seen that after obtaining the intention labels of each word, that is, mark off the second generic clause.
In the present embodiment, each word can be used as a vector input and be intended to domain discrimination model, and vectorial length can be
100。
In the present embodiment, it is intended that discrimination model uses but is not limited to " the two-way lstm+crf of embedding+ " neural network mould
Type (other sequence marking models).Each domain classification corresponds to an intention discrimination model.Any one domain classification pair
The training method for the intention discrimination model answered is:Third preset quantity such as 100,000 sentences for belonging to the domain classification are collected,
Every sentence marks intention labels.Collect such as 200,000 articles of the 4th preset quantity is not belonging in preset intent classifier any one
The sentence of a classification, and the intention labels for marking every sentence are s.Use the corpus data with label of above-mentioned collection, instruction
Practice the object function of neural network model: For the probability of the intention labels of prediction
Value.Specific training algorithm may refer to the prior art, and which is not described herein again.
S203:To each word (i.e. word in the second generic clause) with field label and intention labels, determine that slot is believed
Cease label.
The slot information labels of any one word in second generic clause include:Slot information belonging to the word title (such as
Shown in table 3) and the third place information, the third place information is position letter of the word in all words for meeting third condition
Breath, third condition are to belong to identical domain classification, identical intent classifier and having the same in the second generic clause with the word
The word of slot information.Specifically, position include starting (such as start information is indicated with b), intermediate (such as average information is indicated with m) and
Terminate (such as ending message is indicated with e).
Because different types of slot information has the characteristics that itself, therefore, in order to preferably extract slot information, the present embodiment
In, using different extraction modes:
1, grammar rule mode, which is suitable for enumerate but the extraction of regular strong slot information.Such as the time,
Amount of money etc. analyzes extraction tank information using the following abnf syntax, that is, regular expression:
$ digital=0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9;
$ date=$ digital<2,4>Year $ digital<1-2>Moon $ digital<1-2>Day;
$ money=$ digital<0,->[hundred million | ten thousand | thousand | hundred | ten] member.
2, dictionary mode, which are suitable for the extraction of ordinal type slot information.By taking city name as an example, dictionary list
(city)={ Beijing, Shanghai, Guangzhou, Shenzhen, Hangzhou ... }.By the way of dictionary matching, extraction tank information.
The specific implementation of both the above mode may refer to the prior art, and which is not described herein again.
3, it for that can not enumerate and the not strong slot information of regularity, in the present embodiment, is extracted using slot information discrimination model:
The input of slot information discrimination model is the intention vector of each word and the second generic clause in the second generic clause.Its
In, the example of the intention vector of any one the second generic clause is:The length for being intended to vector is the intention under a domain classification
The total quantity of classification, the intent classifier under the domain classification correspond an element in vector, the intention point described in the word
The corresponding element of class is 1, and other elements are 0.
Each word can be used as a vector, and vectorial length can be 100.
Table 6 is the slot Examples of information of clause under the navigation purposes of map field:
Table 6
In the present embodiment, slot information discrimination model uses but is not limited to " the two-way lstm+crf of embedding+ " neural network
Model (other sequence marking models).Each domain classification corresponds to a slot information discrimination model.Any one field point
The training method of the corresponding slot information discrimination model of class is:Collect such as 100,000 articles of the 5th preset quantity for belonging to the domain classification
Sentence, every sentence mark slot information labels.Such as 200,000 articles of the 6th preset quantity is collected to be not belonging in preset slot information
The sentence of any one slot information, and the slot information labels for marking every sentence are s.Use the language with label of above-mentioned collection
Expect data, the object function of training neural network model:Wherein,For the slot information of prediction
The probability value of label.Specific training algorithm may refer to the prior art, and which is not described herein again.
It should be noted that three of the above slot information extraction mode, can use parallel, i.e., each word is carried out with
The extraction of upper three kinds of modes, in practical applications, some possible modes are without output as a result, knot may also be exported there are many mode
Fruit extracts different slot information, it is also possible to which a kind of mode exports multiple slot information, for the final result of slot information
Selection, will subsequently illustrate.
It is of course also possible to train a slot information discrimination model, this case not to limit herein for each intent classifier.
In the case where each intent classifier corresponds to a slot information discrimination model respectively, the input data of slot information discrimination model
In, can not include the intention vector of the second generic clause.
So far, the domain classification, intent classifier and slot information that each word in sentence to be understood is determined respectively, because
This, completes for sentence comprehension to be understood.
Based on the mode for gradually understanding domain classification, intent classifier and slot information, each step is respectively trained accordingly
Model, so, the enforceability of model training is higher.Also, for different domain classifications, differentiated using different intentions
Module and slot information discrimination module so that for the accuracy higher for understanding result more being intended to.
It should be noted that it is shown in Fig. 2 in the way of model judgement, a kind of specific implementation of only Fig. 1
Citing can also use other manner, such as sentenced according to the mode that preset correspondence judges in addition to utilizing model
Fixed, which is not described herein again.
As previously mentioned, the slot information finally extracted may have it is multigroup, it is also possible to one group of slot information includes multiple slot information,
Multiple slot information can form different slot information paths, for example, slot information path 1 (path1)=<Slot_a, slot_b,
slot_c>, path2=<Slot_a, slot_d>.That is slot information a, b and c forms a kind of slot information candidate path, slot information a and
Another slot information candidate of d compositions path.Slot information candidate path described in the present embodiment from the second generic clause by extracting
At least one slot information combine to be formed, in order to improve efficiency of selection, optionally, different slot information candidate paths is at least
There are one different slot information for tool.
Therefore, method described herein further can choose a road from a plurality of slot information candidate path
Diameter.Detailed process is as shown in figure 3, include the following steps:
S301:Sentence to be understood is subjected to word segmentation processing, obtains word segmentation result.
S302:Using order models, the intention vector of word segmentation result, slot information candidate path and the second generic clause is made
For input, the scoring in each slot information candidate path is determined.
Wherein, word segmentation result can be the vector of each of word segmentation result participle composition, and vectorial length can be m*
100, wherein m are the number of participle.
Slot information candidate path can be the vector in each slot information candidate path and each participle composition, vectorial length
Can be n*100, n is the number of path candidate participle.
The intention vector of second generic clause is as previously mentioned, which is not described herein again.
It should be noted that three of the above vector might not be used as input, can at least select a kind of as sequence
The input of model, but vector as input is more, then and the result exported is more accurate.
In the present embodiment, order models use but are not limited to CNN models (other order models).For different necks
Domain is classified, and identical sorting module can be used.The training method of order models is:The training language of multiplexing field discrimination module
Sentence, and correct path pathT is marked to each sentence, erroneous path pathF is forged by program, finally by correct path and mistake
Accidentally path forms pair pairs<PathT, pathF>.Using the corpus data with label of above-mentioned collection, CNN models are trained
Object function:
S(θ)>thre
Wherein, PT (θ)iIndicate correct path score, PF (θ)iIndicate that erroneous path score, final score S (θ) are more than threshold
Value thre.
S303:According to the scoring in each slot information candidate path, final slot information path is determined.
In general, the highest slot information path that will score is as final slot information path.
For example, the slot information candidate path of clause " ten streets Li Miao for navigating to 3 kilometers " extraction includes:
path1:Navigate to the $ endLoc of $ distance
path2:Navigate to the mausoleums the $ distance street of $ distance
Executing the scoring obtained after process shown in Fig. 3 is respectively:Path1,0.95.Path2,0.2.Therefore, by path1
As final slot information path, the slot information in the path, for finally determining slot information.
Fig. 4 is a kind of semantic understanding device disclosed in the embodiment of the present application, including:First division module, second divide mould
Block and extraction module can also include optionally screening module.
Wherein, the first division module is used for according to preset domain classification, and sentence to be understood is divided into first kind
Sentence.Second division module is used to, according to preset intent classifier, each first generic clause is divided into the second generic clause.It carries
Modulus block is used for the extraction tank information from each second generic clause.Screening module is used in the extraction module from each institute
It states in the second generic clause after extraction tank information, scores slot information candidate path, and according to the scoring, determine that slot is believed
Breath extraction result.
The specific implementation process of the function of the above modules, may refer to above method part, which is not described herein again.
Semantic understanding device shown in Fig. 3, using different partition dimensions (i.e. field, intention, slot information), according to step by step
The mode for dividing clause can obtain the slot information under multi-field more intentions, therefore, it is possible to for the sentence more being intended to, obtain
The higher understanding results being intended to of accuracy more.
The embodiment of the present application also discloses a kind of semantic understanding equipment and computer-readable medium.
Specifically, semantic understanding equipment includes memory and processor.Memory is for storing one or more programs.Place
Reason device is for executing one or more of programs, so that the semantic understanding equipment realizes following functions previous embodiment institute
The semantic understanding function of stating.
Instruction is stored in computer readable storage medium, when run on a computer so that before computer executes
State the semantic understanding function described in embodiment.
If the function described in the embodiment of the present application method is realized in the form of SFU software functional unit and as independent production
Product are sold or in use, can be stored in a computing device read/write memory medium.Based on this understanding, the application is real
Applying the part of a part that contributes to existing technology or the technical solution can be expressed in the form of software products,
The software product is stored in a storage medium, including some instructions are used so that a computing device (can be personal meter
Calculation machine, server, mobile computing device or network equipment etc.) execute each embodiment the method for the application whole or portion
Step by step.And storage medium above-mentioned includes:USB flash disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), with
Machine accesses various Jie that can store program code such as memory (RAM, Random Access Memory), magnetic disc or CD
Matter.
Each embodiment is described by the way of progressive in this specification, the highlights of each of the examples are with it is other
The difference of embodiment, just to refer each other for same or similar part between each embodiment.
The foregoing description of the disclosed embodiments enables professional and technical personnel in the field to realize or use the application.
Various modifications to these embodiments will be apparent to those skilled in the art, as defined herein
General Principle can in other embodiments be realized in the case where not departing from spirit herein or range.Therefore, the application
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest range caused.
Claims (11)
1. a kind of semantic understanding method, which is characterized in that including:
According to preset domain classification, sentence to be understood is divided into the first generic clause, wherein each first generic clause
Belong to a classification in the domain classification;
According to preset intent classifier, each first generic clause is divided into the second generic clause, wherein each described second
Generic clause belongs to a classification in the intent classifier;
The extraction tank information from each second generic clause.
2. according to the method described in claim 1, it is characterized in that, described according to preset domain classification, by language to be understood
Sentence is divided into the first generic clause and includes:
Using each word in the sentence to be understood as input, using field discrimination model, the language to be understood is obtained
The field label of the field label of each word in sentence, any one word in the sentence to be understood includes:The word
The title and first position information of affiliated domain classification, the first position information are the word in the institute for meeting first condition
It is to belong to identical domain classification with the word in the sentence to be understood to have the location information in word, the first condition
Word, the position include starting, centre and terminate.
3. according to the method described in claim 2, it is characterized in that, the training process of the field discrimination model includes:
For each classification in the preset domain classification, the first preset quantity sentence is collected respectively, and to collecting
Sentence mark the field label;
The sentence that the second preset quantity item is not belonging to the classification of any one of described preset domain classification is collected, and is marked not
Belong to the label of the preset domain classification;
Using the object function of sentence training the first sequence marking model after mark, the field discrimination model is obtained.
It, will each described the 4. according to the method described in claim 1, it is characterized in that, described according to preset intent classifier
One generic clause is divided into the second generic clause:
Using each word with same area specific name in first generic clause as input, using with the first kind
The corresponding intention discrimination module of domain classification belonging to clause, obtains the intention labels of each word in first generic clause,
The intention labels of any one word in first generic clause include:The title of intent classifier belonging to the word and
Two location informations, the second position information be location information of the word in all words for meeting second condition, described second
Condition is to belong to the word of identical domain classification and identical intent classifier, the position in first generic clause with the word
Including starting, centre and end, wherein each domain classification corresponds to an intention discrimination module, different domain classifications pair
The intention discrimination module answered is different.
5. according to the method described in claim 4, it is characterized in that, the corresponding intention of any one of domain classification is sentenced
The training method of other module includes:
The third preset quantity sentence for belonging to the domain classification is collected, and the intention is marked to every sentence of collection and is marked
Label;
The sentence that the 4th preset quantity is not belonging to the classification of any one of preset intent classifier is collected, and marks and is not belonging to institute
State the label of preset domain classification;
Using the object function of sentence training the second sequence marking model after mark, the corresponding meaning of the domain classification is obtained
Figure discrimination module.
6. according to the method described in claim 1, it is characterized in that, the extraction tank information from each second generic clause
Including:
Using the intention vector of each word and second generic clause in second generic clause as input, using with it is described
The corresponding slot information discrimination model of domain classification belonging to second generic clause, obtains the slot of each word in second generic clause
The slot information labels of information labels, any one word in second generic clause include:The title of slot information belonging to the word
And the third place information, the third place information are location information of the word in all words for meeting third condition, institute
It is to belong to identical domain classification, identical intent classifier in second generic clause with the word and have phase to state third condition
The word of same slot information;
Wherein, each domain classification corresponds to a slot information discrimination module, and the corresponding slot information of different domain classifications differentiates
Module is different.
7. according to the method described in claim 6, it is characterized in that, the corresponding slot information of any one of domain classification
The training method of discrimination module includes:
The 5th preset quantity article sentence for belonging to the domain classification is collected, and slot information labels are marked to the sentence of collection;
The sentence that the 6th preset quantity is not belonging to any one of preset slot information slot information is collected, and marks and is not belonging to institute
State the label of preset slot information;
The object function for the marking model that sorted using the sentence training third after mark, obtains the corresponding slot of the domain classification
Information discrimination module.
8. according to claim 1-7 any one of them methods, which is characterized in that described from each second generic clause
After extraction tank information, further include:
By described in the word segmentation result of the sentence to be understood, slot information candidate path and the sentence to be understood
The intention vector of two generic clauses, using preset model, scores to slot information candidate path as input, the slot information
Path candidate is combined by least one slot information extracted from each second generic clause and is formed;
According to the scoring, slot information extraction result is determined.
9. a kind of semantic understanding device, which is characterized in that including:
First division module, for according to preset domain classification, sentence to be understood to be divided into the first generic clause, wherein
Each first generic clause belongs to a classification in the domain classification;
Second division module, for according to preset intent classifier, each first generic clause to be divided into the second generic clause,
Wherein, each second generic clause belongs to a classification in the intent classifier;
Extraction module, for the extraction tank information from each second generic clause.
10. a kind of semantic understanding equipment, which is characterized in that including:
Memory and processor;
The memory is for storing one or more programs;
The processor is for executing one or more of programs, so that the semantic understanding equipment realizes claim 1-
Semantic understanding method described in any one of 8.
11. a kind of computer-readable medium, which is characterized in that instruction is stored in the computer readable storage medium, when it
When running on computers so that computer perform claim requires the semantic understanding method described in any one of 1-8.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810401066.XA CN108563790B (en) | 2018-04-28 | 2018-04-28 | Semantic understanding method and device, equipment and computer readable medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810401066.XA CN108563790B (en) | 2018-04-28 | 2018-04-28 | Semantic understanding method and device, equipment and computer readable medium |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108563790A true CN108563790A (en) | 2018-09-21 |
CN108563790B CN108563790B (en) | 2021-10-08 |
Family
ID=63537574
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810401066.XA Active CN108563790B (en) | 2018-04-28 | 2018-04-28 | Semantic understanding method and device, equipment and computer readable medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108563790B (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109753565A (en) * | 2018-12-27 | 2019-05-14 | 厦门智融合科技有限公司 | Intellectual Property intelligent service method and system |
CN109783801A (en) * | 2018-12-14 | 2019-05-21 | 厦门快商通信息技术有限公司 | A kind of electronic device, multi-tag classification method and storage medium |
CN109783819A (en) * | 2019-01-18 | 2019-05-21 | 广东小天才科技有限公司 | A kind of generation method and system of regular expression |
CN110008325A (en) * | 2019-03-29 | 2019-07-12 | 海南中智信信息技术有限公司 | A kind of conversational language understanding and Improvement based on commercial conversational system |
CN110111766A (en) * | 2019-04-22 | 2019-08-09 | 南京硅基智能科技有限公司 | A kind of multi-field Task conversational system and terminal |
CN110837547A (en) * | 2019-10-16 | 2020-02-25 | 云知声智能科技股份有限公司 | Method and device for understanding multi-intention text in man-machine interaction |
CN111078844A (en) * | 2018-10-18 | 2020-04-28 | 上海交通大学 | Task-based dialog system and method for software crowdsourcing |
CN111292751A (en) * | 2018-11-21 | 2020-06-16 | 北京嘀嘀无限科技发展有限公司 | Semantic analysis method and device, voice interaction method and device, and electronic equipment |
CN111339760A (en) * | 2018-12-18 | 2020-06-26 | 北京京东尚科信息技术有限公司 | Method and device for training lexical analysis model, electronic equipment and storage medium |
WO2020147609A1 (en) * | 2019-01-18 | 2020-07-23 | 阿里巴巴集团控股有限公司 | Speech recognition method and apparatus |
CN111611793A (en) * | 2019-02-22 | 2020-09-01 | 北京猎户星空科技有限公司 | Data processing method, device, equipment and storage medium |
CN111625634A (en) * | 2020-05-25 | 2020-09-04 | 泰康保险集团股份有限公司 | Word slot recognition method and device, computer-readable storage medium and electronic device |
CN111639160A (en) * | 2020-05-29 | 2020-09-08 | 达闼机器人有限公司 | Domain identification method, interaction method, electronic device and storage medium |
CN112597310A (en) * | 2020-12-25 | 2021-04-02 | 深圳市声希科技有限公司 | Domain and intention hierarchical joint classification method, device, equipment and storage medium |
JP2021081712A (en) * | 2019-11-14 | 2021-05-27 | バイドゥ オンライン ネットワーク テクノロジー (ベイジン) カンパニー リミテッド | Method, device, electronic apparatus, and computer readable storage media for voice interaction |
CN113779964A (en) * | 2021-09-02 | 2021-12-10 | 中联国智科技管理(北京)有限公司 | Statement segmentation method and device |
CN115440200A (en) * | 2021-06-02 | 2022-12-06 | 上海擎感智能科技有限公司 | Control method and control system of vehicle-mounted machine system |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20150042533A (en) * | 2013-10-11 | 2015-04-21 | 에스케이텔레콤 주식회사 | Apparatus for analyzing complex sentence, and recording medium therefor |
US20170169013A1 (en) * | 2015-12-11 | 2017-06-15 | Microsoft Technology Licensing, Llc | Personalizing Natural Language Understanding Systems |
WO2017151400A1 (en) * | 2016-02-29 | 2017-09-08 | Microsoft Technology Licensing, Llc | Interpreting and resolving conditional natural language queries |
CN107209758A (en) * | 2015-01-28 | 2017-09-26 | 三菱电机株式会社 | It is intended to estimation unit and is intended to method of estimation |
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 |
CN107665188A (en) * | 2016-07-27 | 2018-02-06 | 科大讯飞股份有限公司 | A kind of semantic understanding method and device |
CN107750360A (en) * | 2015-06-15 | 2018-03-02 | 微软技术许可有限责任公司 | Generated by using the context language of language understanding |
CN107885874A (en) * | 2017-11-28 | 2018-04-06 | 上海智臻智能网络科技股份有限公司 | Data query method and apparatus, computer equipment and computer-readable recording medium |
-
2018
- 2018-04-28 CN CN201810401066.XA patent/CN108563790B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20150042533A (en) * | 2013-10-11 | 2015-04-21 | 에스케이텔레콤 주식회사 | Apparatus for analyzing complex sentence, and recording medium therefor |
CN107209758A (en) * | 2015-01-28 | 2017-09-26 | 三菱电机株式会社 | It is intended to estimation unit and is intended to method of estimation |
CN107750360A (en) * | 2015-06-15 | 2018-03-02 | 微软技术许可有限责任公司 | Generated by using the context language of language understanding |
US20170169013A1 (en) * | 2015-12-11 | 2017-06-15 | Microsoft Technology Licensing, Llc | Personalizing Natural Language Understanding Systems |
WO2017151400A1 (en) * | 2016-02-29 | 2017-09-08 | Microsoft Technology Licensing, Llc | Interpreting and resolving conditional natural language queries |
CN107665188A (en) * | 2016-07-27 | 2018-02-06 | 科大讯飞股份有限公司 | A kind of semantic understanding 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 |
CN107885874A (en) * | 2017-11-28 | 2018-04-06 | 上海智臻智能网络科技股份有限公司 | Data query method and apparatus, computer equipment and computer-readable recording medium |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111078844B (en) * | 2018-10-18 | 2023-03-14 | 上海交通大学 | Task-based dialog system and method for software crowdsourcing |
CN111078844A (en) * | 2018-10-18 | 2020-04-28 | 上海交通大学 | Task-based dialog system and method for software crowdsourcing |
CN111292751B (en) * | 2018-11-21 | 2023-02-28 | 北京嘀嘀无限科技发展有限公司 | Semantic analysis method and device, voice interaction method and device, and electronic equipment |
CN111292751A (en) * | 2018-11-21 | 2020-06-16 | 北京嘀嘀无限科技发展有限公司 | Semantic analysis method and device, voice interaction method and device, and electronic equipment |
CN109783801A (en) * | 2018-12-14 | 2019-05-21 | 厦门快商通信息技术有限公司 | A kind of electronic device, multi-tag classification method and storage medium |
CN109783801B (en) * | 2018-12-14 | 2023-08-25 | 厦门快商通信息技术有限公司 | Electronic device, multi-label classification method and storage medium |
CN111339760A (en) * | 2018-12-18 | 2020-06-26 | 北京京东尚科信息技术有限公司 | Method and device for training lexical analysis model, electronic equipment and storage medium |
CN109753565A (en) * | 2018-12-27 | 2019-05-14 | 厦门智融合科技有限公司 | Intellectual Property intelligent service method and system |
CN109783819A (en) * | 2019-01-18 | 2019-05-21 | 广东小天才科技有限公司 | A kind of generation method and system of regular expression |
CN109783819B (en) * | 2019-01-18 | 2023-10-20 | 广东小天才科技有限公司 | Regular expression generation method and system |
WO2020147609A1 (en) * | 2019-01-18 | 2020-07-23 | 阿里巴巴集团控股有限公司 | Speech recognition method and apparatus |
CN111611793B (en) * | 2019-02-22 | 2023-06-13 | 北京猎户星空科技有限公司 | Data processing method, device, equipment and storage medium |
CN111611793A (en) * | 2019-02-22 | 2020-09-01 | 北京猎户星空科技有限公司 | Data processing method, device, equipment and storage medium |
CN110008325A (en) * | 2019-03-29 | 2019-07-12 | 海南中智信信息技术有限公司 | A kind of conversational language understanding and Improvement based on commercial conversational system |
CN110008325B (en) * | 2019-03-29 | 2020-02-07 | 海南中智信信息技术有限公司 | Spoken language understanding and rewriting method based on commercial conversation system |
CN110111766A (en) * | 2019-04-22 | 2019-08-09 | 南京硅基智能科技有限公司 | A kind of multi-field Task conversational system and terminal |
CN110837547A (en) * | 2019-10-16 | 2020-02-25 | 云知声智能科技股份有限公司 | Method and device for understanding multi-intention text in man-machine interaction |
JP2021081712A (en) * | 2019-11-14 | 2021-05-27 | バイドゥ オンライン ネットワーク テクノロジー (ベイジン) カンパニー リミテッド | Method, device, electronic apparatus, and computer readable storage media for voice interaction |
JP7300435B2 (en) | 2019-11-14 | 2023-06-29 | バイドゥ オンライン ネットワーク テクノロジー(ペキン) カンパニー リミテッド | Methods, apparatus, electronics, and computer-readable storage media 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 |
CN111625634A (en) * | 2020-05-25 | 2020-09-04 | 泰康保险集团股份有限公司 | Word slot recognition method and device, computer-readable storage medium and electronic device |
CN111625634B (en) * | 2020-05-25 | 2023-08-22 | 泰康保险集团股份有限公司 | Word slot recognition method and device, computer readable storage medium and electronic equipment |
WO2021239078A1 (en) * | 2020-05-29 | 2021-12-02 | 达闼机器人有限公司 | Field recognition method, interaction method, electronic device, and storage medium |
CN111639160A (en) * | 2020-05-29 | 2020-09-08 | 达闼机器人有限公司 | Domain identification method, interaction method, electronic device and storage medium |
CN112597310A (en) * | 2020-12-25 | 2021-04-02 | 深圳市声希科技有限公司 | Domain and intention hierarchical joint classification method, device, equipment and storage medium |
CN115440200A (en) * | 2021-06-02 | 2022-12-06 | 上海擎感智能科技有限公司 | Control method and control system of vehicle-mounted machine system |
CN115440200B (en) * | 2021-06-02 | 2024-03-12 | 上海擎感智能科技有限公司 | Control method and control system of vehicle-mounted system |
CN113779964A (en) * | 2021-09-02 | 2021-12-10 | 中联国智科技管理(北京)有限公司 | Statement segmentation method and device |
Also Published As
Publication number | Publication date |
---|---|
CN108563790B (en) | 2021-10-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108563790A (en) | A kind of semantic understanding method and device, equipment, computer-readable medium | |
CN110717339B (en) | Semantic representation model processing method and device, electronic equipment and storage medium | |
CN109918673A (en) | Semantic referee method, device, electronic equipment and computer readable storage medium | |
Carvalho et al. | Learning to extract signature and reply lines from email | |
CN102262634B (en) | Automatic questioning and answering method and system | |
CN102289522B (en) | Method of intelligently classifying texts | |
CN108763510A (en) | Intension recognizing method, device, equipment and storage medium | |
CN106776538A (en) | The information extracting method of enterprise's noncanonical format document | |
CN105117429A (en) | Scenario image annotation method based on active learning and multi-label multi-instance learning | |
CN106294684A (en) | The file classification method of term vector and terminal unit | |
CN112836487B (en) | Automatic comment method and device, computer equipment and storage medium | |
CN112256845A (en) | Intention recognition method, device, electronic equipment and computer readable storage medium | |
CN111340034B (en) | Text detection and identification method and system for natural scene | |
CN107729468A (en) | Answer extracting method and system based on deep learning | |
CN101178714A (en) | Web page classification method and device | |
CN103778142A (en) | Conditional random fields (CRF) based acronym expansion explanation recognition method | |
CN103631874A (en) | UGC label classification determining method and device for social platform | |
CN110956044A (en) | Attention mechanism-based case input recognition and classification method for judicial scenes | |
CN108009248A (en) | A kind of data classification method and system | |
CN116049412A (en) | Text classification method, model training method, device and electronic equipment | |
CN116737922A (en) | Tourist online comment fine granularity emotion analysis method and system | |
CN110222139A (en) | Road solid data De-weight method, calculates equipment and medium at device | |
CN112148994B (en) | Information push effect evaluation method and device, electronic equipment and storage medium | |
CN115187910A (en) | Video classification model training method and device, electronic equipment and storage medium | |
CN111125387B (en) | Multimedia list generation and naming method and device, electronic equipment and storage 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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |