CN109241392A - Recognition methods, device, system and the storage medium of target word - Google Patents
Recognition methods, device, system and the storage medium of target word Download PDFInfo
- Publication number
- CN109241392A CN109241392A CN201710538781.3A CN201710538781A CN109241392A CN 109241392 A CN109241392 A CN 109241392A CN 201710538781 A CN201710538781 A CN 201710538781A CN 109241392 A CN109241392 A CN 109241392A
- Authority
- CN
- China
- Prior art keywords
- candidate
- string
- characteristic value
- participle
- target word
- 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.)
- Pending
Links
Classifications
-
- 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
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to field of information processing, in particular to a kind of recognition methods of target word, device, system and storage medium.A kind of recognition methods of target word provided by one embodiment of the present invention carries out participle division based on candidate string of the text data of minimum particle size to acquisition, obtains word segmentation result;Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;At least one described specific characteristic value is inputted into default classifier, obtains the decision content of the string of candidate corresponding to the specific characteristic value;The string of candidate corresponding to the decision content for meeting preset condition is set as target word.This programme embodiment identifies target word by calculating at least one characteristic value and the combining classification device of candidate string, relative to directly artificial settings threshold value and simple linear analysis method, accuracy rate has with recall rate significantly to be promoted, artificial screening cost is largely reduced, the recognition efficiency of target word is improved.
Description
[technical field]
The present invention relates to field of information processing, in particular to a kind of recognition methods of target word, device, system and storage are situated between
Matter.
[background technique]
In recent years, the fast development with internet in the world, the information that people face are in exponential increase.In people
Have a large amount of neologisms, such as movie and television play name, trade name, network popular word in the information that is faced.Therefore, how accurate quick
Automatic discovery neologisms be just particularly important.
The method that existing new word discovery is mainly based upon rule, main thought are according to the word-building characteristic of neologisms or outer
Type feature establishes rule base, specialized dictionary or pattern base, then finds neologisms by rule match.Rule-based method is main
Disadvantage is the problem for being confined to some field, and needing to establish rule base etc., and have recall rate inadequate, cannot identify
Target neologisms under current complicated internet environment.
[summary of the invention]
One object of the present invention aims to solve the problem that at least one above-mentioned problem, provide a kind of target word recognition methods,
Device, system and storage medium.
To realize the purpose, the present invention adopts the following technical scheme:
An embodiment provides a kind of recognition methods of target word comprising:
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains word segmentation result;
Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;
At least one described specific characteristic value is inputted into default classifier, obtains the string of candidate corresponding to the specific characteristic value
Decision content;
The string of candidate corresponding to the decision content for meeting preset condition is set as target word.
Specifically, described carry out participle division based on candidate string of the text data of minimum particle size to acquisition, segmented
As a result the step of, comprising:
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains candidate participle;
At least one described candidate participle is combined, the corresponding word segmentation result of the candidate string is obtained.
Further, the specific characteristic value includes at least one of:
Foundation characteristic, statistics category feature, incidence coefficient category feature, context category feature.
Specifically, the foundation characteristic includes at least one of: candidate's string inquiry times, are waited at candidate string participle mode
The ratio of the mean search frequency, candidate string inquiry times that the maximum value of choosing participle length, candidate segment.
Specifically, the statistics category feature includes at least one of: tightness, joint probability, conditional probability, anti-condition
Probability, point mutual information, second order point mutual information, log-likelihood ratio, normalization expectation.
Specifically, the incidence coefficient category feature includes at least one of: probability ratio, increment, Jie Kade distance, Xin Pu
Gloomy distance.
Specifically, the context category feature includes at least one of: context entropy, context diversity, left information
Entropy, right comentropy, left adjacent diversity, right adjacent diversity.
Optionally, the default classifier includes the progressive decision tree classifier of gradient.
The decision content for meeting preset condition, comprising: meet the decision content of predetermined threshold value range, and the pre- gating
Limits are corresponding at least one described specific characteristic value.
Further, the recognition methods further includes being based on default extracting rule orientation by search engine to obtain candidate
String.
Further, the recognition methods further includes carrying out data purification pretreatment to the candidate string.
Another embodiment of the invention provides a kind of identification device of target word comprising:
Word segmentation module carries out participle division for candidate string of the text data based on minimum particle size to acquisition, is divided
Word result;
Characteristic value calculating module, for calculating at least one specific characteristic of each candidate participle in the word segmentation result
Value;
It is specified to obtain this at least one described specific characteristic value to be inputted default classifier for decision content computing module
The decision content of the string of candidate corresponding to characteristic value;
Selecting module is set as target word for will meet the string of candidate corresponding to the decision content of preset condition.
Specifically, the word segmentation module carries out participle division based on candidate string of the text data of minimum particle size to acquisition,
Obtain word segmentation result, comprising:
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains candidate participle;
At least one described candidate participle is combined, the corresponding word segmentation result of the candidate string is obtained.
Specifically, the specific characteristic value includes at least one of:
Foundation characteristic, statistics category feature, incidence coefficient category feature, context category feature.
Specifically, the foundation characteristic includes at least one of: candidate's string inquiry times, are waited at candidate string participle mode
The ratio of the mean search frequency, candidate string inquiry times that the maximum value of choosing participle length, candidate segment.
Specifically, the statistics category feature includes at least one of: tightness, joint probability, conditional probability, anti-condition
Probability, point mutual information, second order point mutual information, log-likelihood ratio, normalization expectation.
Specifically, the incidence coefficient category feature includes at least one of: probability ratio, increment, Jie Kade distance, Xin Pu
Gloomy distance.
Specifically, the context category feature includes at least one of: context entropy, context diversity, left information
Entropy, right comentropy, left adjacent diversity, right adjacent diversity.Optionally, the default classifier includes the progressive decision of gradient
Tree Classifier.
Specifically, the decision content for meeting preset condition, comprising: meet the decision content of predetermined threshold value range, and institute
It is corresponding at least one described specific characteristic value to state predetermined threshold value range.
Further, the identification device further includes that candidate string obtains module,
The candidate string obtains module, obtains candidate string for being based on default extracting rule orientation by search engine.
Further, the candidate string obtains module and is also used to: carrying out data purification pretreatment to the candidate string.
Yet another embodiment of the present invention provides a kind of identifying system of target word comprising:
The identification device of target word carries out participle for candidate string of the text data based on minimum particle size to acquisition and draws
Point, obtain word segmentation result;Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;By described at least
One specific characteristic value inputs default classifier, obtains the decision content of the string of candidate corresponding to the specific characteristic value;It will meet pre-
If the string of candidate corresponding to the decision content of condition is set as target word;
Training sample identification device is arranged needed for the parameter of classifier for providing to the identification device of the target word
Training sample word;
Target word collection device, the target word that the identification device for receiving the target word is identified.
Another embodiment of the present invention provides a kind of storage mediums, which is characterized in that the storage medium includes depositing
The program of storage, wherein the step of system where controlling the storage medium in described program operation executes any preceding method.
Compared with prior art, the present invention has following advantage:
The recognition methods of a kind of target word provided by one embodiment of the present invention, based on the text data of minimum particle size to obtaining
The candidate string taken carries out participle division, obtains word segmentation result;Calculate at least one each candidate segmented in the word segmentation result
Specific characteristic value;At least one described specific characteristic value is inputted into default classifier, obtains time corresponding to the specific characteristic value
Select the decision content of string;The string of candidate corresponding to the decision content for meeting preset condition is set as target word.This programme embodiment passes through
At least one characteristic value of candidate string and combining classification device are calculated to identify target word, relative to directly artificial settings threshold value and letter
Single linear analysis method, accuracy rate has with recall rate significantly to be promoted, and largely reduces artificial screening cost, improves target word
Recognition efficiency.
The additional aspect of the present invention and advantage will be set forth in part in the description, these will become from the following description
It obtains obviously, or recognized by the embodiment of this programme.
[Detailed description of the invention]
It to describe the technical solutions in the embodiments of the present invention more clearly, below will be to embodiment or description of the prior art
Needed in attached drawing be briefly described, it should be apparent that, the accompanying drawings in the following description is only some realities of the invention
Apply example, however, the present invention is not limited thereto.
Fig. 1 is the flow diagram in one embodiment of the recognition methods of target word of the present invention;
Fig. 2 is the flow diagram in one embodiment of the recognition methods of target word of the present invention;
Fig. 3 is the structural schematic diagram in one embodiment of the identification device of target word of the present invention;
Fig. 4 is the structural schematic diagram in one embodiment of the identification device of target word of the present invention;
Fig. 5 is the structural schematic diagram in one embodiment of the identifying system of target word of the present invention;
Fig. 6 is the structural schematic diagram in one embodiment of the identifying system of target word of the present invention.
[specific embodiment]
The present invention is further described with exemplary embodiment with reference to the accompanying drawing, the examples of the embodiments are attached
It is shown in figure, in which the same or similar labels are throughly indicated same or similar element or there is same or like function
Element.The embodiments described below with reference to the accompanying drawings are exemplary, for explaining only the invention, and cannot be construed to pair
Limitation of the invention.In addition, if the detailed description of known technology is for showing the invention is characterized in that unnecessary, then by it
It omits.
Those skilled in the art of the present technique are appreciated that unless expressly stated, singular " one " used herein, " one
It is a ", " described " and "the" may also comprise plural form.It is to be further understood that being arranged used in specification of the invention
Diction " comprising " refer to that there are the feature, integer, step, operation, element and/or component, but it is not excluded that in the presence of or addition
Other one or more features, integer, step, operation, element, component and/or their group.It should be understood that when we claim member
Part is " connected " or when " coupled " to another element, it can be directly connected or coupled to other elements, or there may also be
Intermediary element.In addition, " connection " used herein or " coupling " may include being wirelessly connected or wirelessly coupling.It is used herein to arrange
Diction "and/or" includes one or more associated wholes for listing item or any cell and all combinations.
Those skilled in the art of the present technique are appreciated that unless otherwise defined, all terms used herein (including technology art
Language and scientific term), there is meaning identical with the general understanding of those of ordinary skill in fields of the present invention.Should also
Understand, those terms such as defined in the general dictionary, it should be understood that have in the context of the prior art
The consistent meaning of meaning, and unless idealization or meaning too formal otherwise will not be used by specific definitions as here
To explain.
Those skilled in the art of the present technique are appreciated that server used herein above, cloud, remote network devices etc. are general
It reads, there is effects equivalent comprising but it is not limited to computer, network host, single network server, multiple network server collection
Or the cloud that multiple servers are constituted.Here, cloud is by a large amount of computers or network clothes based on cloud computing (Cloud Computing)
Business device is constituted, wherein cloud computing is one kind of distributed computing, and one consisting of a loosely coupled set of computers super
Virtual machine.In the embodiment of the present invention, it can lead between remote network devices, the identifying system equipment and WNS server
Cross any communication mode and realize communication, including but not limited to, mobile communication based on 3GPP, LTE, WIMAX, based on TCP/IP,
The computer network communication of udp protocol and low coverage wireless transmission method based on bluetooth, Infrared Transmission standard.
It should be noted that An embodiment provides a kind of recognition methods of target word, i.e., from server
Visual angle the recognition methods is described, the recognition methods of target word can be embodied as computer program in distal end net by programming
Realized in network equipment comprising but be not limited to computer, network host, single network server, multiple network server collection or
The cloud that multiple servers are constituted.
Referring to Fig. 1, an a kind of exemplary embodiments of the recognition methods of target word of the invention, specifically include following step
It is rapid:
S11 carries out participle division based on candidate string of the text data of minimum particle size to acquisition, obtains word segmentation result.
In an embodiment of the invention, the text data of minimum particle size can be single word, i.e., is with word by candidate's string
Unit carries out participle division;Can certainly individually segment, in a preferred embodiment of this programme, language can be expressed
Candidate's string is divided into multiple candidate participles by text data of the most succinct vocabulary of justice as minimum particle size.Wherein, participle division can
To use the segmenting method based on dictionary, the segmenting method based on statistics can also be used, since the accuracy of participle is to final
The accuracy rate of the target word identified has a certain impact, and therefore, it is necessary to select suitable segmenting method according to the actual situation.
In order to improve target word described in this programme recognition methods accuracy, it may be considered that calculate target word at least
One specific characteristic value.In one embodiment of this programme, based on candidate of the text data of minimum particle size to acquisition go here and there into
Row participle divides, and after obtaining candidate participle, then is combined at least one described candidate participle, obtains the candidate string and correspond to
Word segmentation result.It can be appreciated that including in word segmentation result: candidate's string itself is divided with the text data of minimum particle size
To candidate participle, at least two candidates divide contamination.
For example it is assumed that candidate string obtained is movie name " I is not Lady Pan Jinlian ", participle division is carried out to it, it is acquired
Candidate participle be " I ", "no", " Lady Pan Jinlian ", then its corresponding word segmentation result includes: " I ", "no", " Pan Jin
Lotus ", " I is not ", " not being Lady Pan Jinlian ", " my Lady Pan Jinlian ", " I is not Lady Pan Jinlian ".
Further, it can be appreciated that carrying out participle division in candidate string of the text data based on minimum particle size to acquisition,
Before obtaining word segmentation result, it is also necessary to obtain multiple candidate strings.Attached drawing 2 is referred to, it is described in one embodiment of this programme
The recognition methods of target word further comprises the steps of:
S101 is based on default extracting rule orientation by search engine and obtains candidate string.
It can be appreciated that the comparison that the desired candidate string obtained of the present embodiment occurs in certain internet corpus is frequent,
And occur in other internet corpus it is then fewer, therefore, current embodiment require that based on default extracting rule orientation
Obtain candidate string.The quality of candidate string can be improved in the mode that orientation obtains, and avoids since corpus is not smart caused some original
The vocabulary for being not belonging to candidate string is mixed into the ranks of candidate string.
For example, in one exemplary embodiment of the present invention, using oriented network spider from BBS, blog or use of trust
Candidate string is extracted in the dictionary of family, for example, the user's neologisms upload function etc. provided on search dog input method official homepage.Certainly, it stands
The selection of point can be specified sites, can be the classification point filtering based on extracting content on web pages.Further, of the invention
In one embodiment, the default extracting rule such as contextual features such as frequency, punctuate, length that can occur according to corpus is obtained
Candidate's string.Certainly, those skilled in that art should know, which is only exemplary, and cannot constitute to the present invention
The restriction of scheme.
In yet another embodiment of the present invention, it in order to improve the quality of selected candidate string, and then improves in the present embodiment
The accuracy rate and efficiency for identifying target word, after being based on default extracting rule orientation by search engine and obtaining candidate string, also
Comprising steps of
S102 carries out data purification pretreatment to the candidate string.
Specifically, in the exemplary embodiment of this programme, it can be from the html tag removed in advance on format in webpage
Etc. invalid informations, or from some fixed form information etc. removed in advance in content in BBS webpage.
Further, attached drawing 1 is referred to, the recognition methods of target word of the present invention further comprises the steps of:
S12 calculates at least one specific characteristic value of each candidate participle in the word segmentation result.
Specifically, using plurality of classes to improve the accuracy rate and recall rate of the recognition methods of target word in the present invention
Specific characteristic value to candidate string determine.The four category feature values that one embodiment of the invention considers, it is special based on distinguishing
Sign, statistics category feature, incidence coefficient category feature and context category feature.
In one embodiment of the application, when calculating candidate string participle mode, according to every in the word segmentation result of candidate's string
The number of words of a candidate participle carrys out partition mode.Wherein " 1+1+1 " indicates that word segmentation result is 3 words, 1 word of each word;" 3+2 " table
Show that word segmentation result is 2 words, respectively 3 words and 2 words, such as the candidate string participle mould of candidate string " I/be not/Lady Pan Jinlian "
Formula is 1-2-3;Its candidate's participle length maximum value is 3.
Specifically, wherein foundation characteristic, which includes at least, candidate string inquiry times, candidate string participle mode, candidate participle
The maximum value of length, the mean search frequency of candidate participle and ratio of candidate string inquiry times etc..
Specifically, statistics category feature indicates the statistical information of candidate string, including at least has tightness, joint probability, item
Part probability, anti-conditional probability, point mutual information, second order point mutual information, log-likelihood ratio, normalization expectation etc..Likewise, in this Shen
In one embodiment please, when calculating joint probability, the point statistics category feature such as mutual information and conspicuousness, expression is candidate string
Former and later two candidate participle between calculating, each specific characteristic can be obtained according to the calculation formula of each specific characteristic value
Value;If candidate string there are multiple candidate participles, after calculating the corresponding specific characteristic value of candidate participle two-by-two, calculation can be taken respectively
Art is average, geometric average, maximum value and minimum value calculate the corresponding specific characteristic value of candidate string.
Specifically, incidence coefficient category feature indicates the incidence relation between the candidate participle of candidate string, calculation method with
It is similar to count category feature comprising probability ratio, increment, Jie Kade distance (Jaccard Distance), Simpson's distance
(Simpson Distance), Coase measurement (Kosgen Measurement), Pi Tesiji Xia Piluo measurement
(PiaterskyShapiro Measurement) etc..For example, calculating Jie Kade distance in one embodiment of the application
When (Jaccard distance) and chance ratio (Odds ratio, OR), it is assumed that x, y respectively indicate the front and back two of a candidate string
A candidate participle;Xy indicates that two candidates segment while occurring,Indicate that x does not occur and y occurs,Indicate x occur and y not
Occur,It indicates that x and y does not occur, calculates separately the number of above four kinds of situations, further in accordance with Jie Kade distance or chance ratio
Calculation formula calculate.Certainly, those skilled in that art should know, which is only exemplary, and be unable to structure
The restriction of pairs of the present invention program.
Specifically, context category feature indicates the relationship between candidate string and context, include at least context entropy, on
Hereafter diversity, left comentropy, right comentropy, left adjacent diversity, right adjacent diversity, the candidate context gone here and there with the word
Cosine distance between TF-IDF vector etc..Wherein, context entropy indicates the variation institute of the context of candidate string in queries
The information content for including;Context diversity indicates that the different word of how many kinds of occurred in the context of candidate string in queries;Left neighbour
Connecing variation number indicates that the different word of how many kinds of occurred in the left adjacent word of candidate string in queries;It can be according to each specific characteristic value
Calculation formula obtain corresponding specific characteristic value.
Specifically, assuming in one embodiment of this programme, only with foundation characteristic, statistical nature, incidence coefficient class
It is calculated with several characteristic values corresponding to any one in contextual feature, without considering other three category features, although energy
Identify target word, but accuracy rate and recall rate are far below several features for selecting wherein at least two or more seed types
Value.
It can be appreciated that it is special to choose foundation characteristic, statistical nature, incidence coefficient class and context according to the actual situation
Several characteristic values in sign calculate.In an exemplary embodiment of this programme, it can be used including tightness, candidate string
Participle mode, joint probability, the maximum value of candidate participle length, context diversity, conspicuousness, candidate string inquiry times,
Left adjacent variation number, than 10 Jie Kade distance, chance specific characteristic values, have obtained preferably determining result.It can be appreciated that
In the other embodiments of this programme, one or more in above-mentioned 10 specific characteristic values can also be selected, is obtained preferably
Determine result.Certainly, those skilled in that art should know, which is only exemplary, and cannot constitute to this hair
The restriction of bright scheme.
Further, attached drawing 1 is referred to, the recognition methods of target word of the present invention further comprises the steps of:
The specific characteristic value is inputted default classifier, obtains sentencing for the string of candidate corresponding to the specific characteristic value by S13
Definite value.
It can be appreciated that in order to improve the accuracy rate of the recognition methods of target word described in this programme and recall rate, this programme
Nonlinear Classifier is used in one embodiment, rather than uses the simple linear classifier of direct given threshold.For example, can adopt
With the progressive decision tree of gradient (Gradient boosting and Decision tree, GBDT) sorting algorithm or depth nerve
Network (Deep Neural Network, DNN) sorting algorithm selects training sample word required for deep neural network certainly
It wants enough, can select which kind of non1inear classifying algorithm used according to the actual situation.In an exemplary implementation of the invention
In example, the present embodiment is realized using GBDT sorting algorithm, GBDT sorting algorithm is formed using several decision trees.
Further, it can be appreciated that first to adjust the parameter of GBDT.Specifically, in one embodiment of this programme, it will
The step of specific characteristic value inputs default classifier, obtains the decision content that candidate corresponding to the specific characteristic value goes here and there it
Before, further includes:
According to institute after the default classifier of at least one specific characteristic value input by the sample word in training sample set of words
The decision content of acquisition, to adjust the parameter of the default classifier;Wherein the training sample word set is combined into identified sample
The set of word.
Specifically, the parameter of the classifier includes minimum sample number required for peripheral node in decision tree, tree
Depth capacity, the quantity of decision tree, for classification specific characteristic value number, determine whether candidate string is the default of target word
Threshold value range etc..Wherein, in order to avoid there is the case where overfitting, minimum sample number cannot be too small, and the maximum set is deep
Degree also should not too deep and the specific characteristic value for classification number it is also unsuitable too many, certainly, excessive sample number may also can
There is poor fitting phenomenon.In an example embodiment of this programme, the number of the specific characteristic value for classification is preferably with referring to
The square root for determining the sum of characteristic value can at most be attempted certainly in practical application, should be attempted using different values with referring to
Determine the 30%-40% of the sum of characteristic value, the case where overfitting can be avoided the occurrence of.
Specifically, by decision content obtained after the default classifier of at least one specific characteristic value input of sample word and being somebody's turn to do
The recognition result of sample word is compared, if it is determined that value is different from the recognition result of the sample word, then adjusts the ginseng of the classifier
Number repeats above procedure, until the default classifier is consistent with recognition result.
Further, after the parameter for obtaining optimal GBDT classifier, by least one specific characteristic value of candidate's string
After inputting default classifier, the decision content of the string of candidate corresponding to specific characteristic value is calculated.
Further, attached drawing 1 is referred to, the recognition methods of target word of the present invention further comprises the steps of:
The string of candidate corresponding to the decision content for meeting preset condition is set as target word by S14.
Specifically, in one embodiment of the invention, preset condition can be the judgement for meeting predetermined threshold value range
Value.Wherein the predetermined threshold value range can be passes through what the training of great amount of samples word obtained in abovementioned steps.It can be appreciated that
After selecting different specific characteristic values corresponding to different characteristic types to input preset classifier, obtained decision content is not
Together, and predetermined threshold value range corresponding to the decision content is also corresponding different, i.e., predetermined threshold value range is specified with selected
Characteristic value is corresponding.Certainly, those skilled in that art should know, which is merely exemplary, and cannot constitute to this
The restriction of scheme of the invention.
For example, in an application scenarios of this programme, we are described by taking movie name " I/be not/Lady Pan Jinlian " as an example
The implementation of case embodiment.More than 10 a specific characteristics are extracted first, obtain " I/be not/Lady Pan Jinlian " this candidate more than 10 gone here and there
Specific characteristic value, as the candidate goes here and there inquiry times (taking logarithm): 17.835, tightness: 0.995, adjacent diversity: 1287 etc.
Deng;By these specific characteristic values by a GBDT classifier, obtain a candidate string whether be target word decision content
0.868783;It is 0.8 by the minimum threshold that the training of great amount of samples word obtains, then classifier score 0.868783 is greater than classification
Thresholding 0.8, then " I is not Lady Pan Jinlian " is judged as meeting the target word of preset condition.
Similarly, in another application scenarios of this programme, with another candidate's string " film/I/be not/Lady Pan Jinlian/"
For the implementation of this programme embodiment described.More than 10 a specific characteristics are extracted first, obtain " film/I/it is not/Lady Pan Jinlian "
The correspondence specific characteristic value of this candidate's string, such as candidate string inquiry times (taking logarithm): 11.644, tightness: 0.679, it is adjacent
Diversity: 81 etc.;By these specific characteristic values by a GBDT classifier, obtain whether a candidate string is target word
Decision content 0.008744;It is 0.8 by the minimum threshold that the training of great amount of samples word obtains, then classifier score 0.008744 is small
In classification thresholding 0.8, then " film I be not Lady Pan Jinlian " is judged as non-targeted word.
In conclusion a kind of recognition methods of target word provided by one embodiment of the present invention, the text based on minimum particle size
Notebook data carries out participle division to the candidate string of acquisition, obtains word segmentation result;Calculate each candidate participle in the word segmentation result
At least one specific characteristic value;At least one described specific characteristic value is inputted into default classifier, obtains the specific characteristic value
The decision content of corresponding candidate string;The string of candidate corresponding to the decision content for meeting preset condition is set as target word.This programme
Embodiment identifies target word by calculating at least one characteristic value and the combining classification device of candidate string, relative to directly artificially setting
Determine threshold value and simple linear analysis method, accuracy rate has with recall rate significantly to be promoted, artificial screening cost is largely reduced,
Improve the recognition efficiency of target word.
Further, according to the function modoularization thinking of computer software, one embodiment of the present of invention additionally provides one
The identification device of kind target word.Refer to attached drawing 3, the identification device include word segmentation module 11, characteristic value calculating module 12,
Decision content computing module 13 and selecting module 14 utilize the word segmentation module 11, characteristic value calculating module 12, decision content computing module
13 and selecting module 14 erect the principle framework of identification device, to realize modularization embodiment.It discloses in detail below
The concrete function that each module is realized.
The word segmentation module 11 carries out participle division for candidate string of the text data based on minimum particle size to acquisition,
Obtain word segmentation result.
In an embodiment of the invention, the text data of minimum particle size can be single word, i.e., is with word by candidate's string
Unit carries out participle division;Can certainly individually segment, in a preferred embodiment of this programme, language can be expressed
Candidate's string is divided into multiple candidate participles by text data of the most succinct vocabulary of justice as minimum particle size, the word segmentation module 11.
Wherein, participle, which divides, can use the segmenting method based on dictionary, the segmenting method based on statistics can also be used, due to participle
Accuracy have a certain impact to the accuracy rate of the target word finally identified, therefore, it is necessary to select to close according to the actual situation
Suitable segmenting method.
In order to improve target word described in this programme identification accuracy, it may be considered that calculate at least one of target word
Specific characteristic value.In one embodiment of this programme, the word segmentation module 11 is based on the text data of minimum particle size to acquisition
Candidate string carry out participle division, after obtaining candidate participle, then at least one described candidate participle is combined, is obtained described
The corresponding word segmentation result of candidate's string.It can be appreciated that including in word segmentation result: candidate goes here and there in itself, with the textual data of minimum particle size
According to the candidate divided segments, at least two candidates divide contamination.
For example it is assumed that candidate string obtained is movie name " I is not Lady Pan Jinlian ", the word segmentation module 11 carries out it
Participle divides, and obtained candidate participle is " I ", "no", " Lady Pan Jinlian ", then its corresponding word segmentation result includes:
" I ", "no", " Lady Pan Jinlian ", " I is not ", " not being Lady Pan Jinlian ", " my Lady Pan Jinlian ", " I is not Lady Pan Jinlian ".
Further, it can be appreciated that in the word segmentation module 11 based on the text data of minimum particle size to the candidate of acquisition
String carries out participle division, before obtaining word segmentation result, it is also necessary to obtain multiple candidate strings.Attached drawing 4 is referred to, the one of this programme
In a embodiment, the identification device of the target word further include:
Candidate's string obtains module 10, obtains candidate string for being based on default extracting rule orientation by search engine.
It can be appreciated that the comparison that the desired candidate string obtained of the present embodiment occurs in certain internet corpus is frequent,
And occur in other internet corpus it is then fewer, therefore, candidate described in the present embodiment go here and there obtain module 10 need
It is oriented based on default extracting rule and obtains candidate string.The quality of candidate string can be improved in the mode that orientation obtains, and avoids due to language
Expect that not smart caused some vocabulary for being not belonging to candidate string originally are mixed into the candidate ranks gone here and there.
For example, the candidate string obtains module 10 and uses oriented network spider in one exemplary embodiment of the present invention
Candidate string is extracted from BBS, blog or user thesaurus of trust, for example, the user provided on search dog input method official homepage is new
Word upload function etc..Certainly, the selection of website can be specified sites, can be the classification based on extracting content on web pages and put
Filter.Further, in one embodiment of the invention, the candidate string obtains the frequency that module 10 can occur according to corpus
The default extracting rule such as the contextual features such as rate, punctuate, length obtains candidate string.Certainly, those skilled in that art should
Know, which is only exemplary, and cannot constitute the restriction to the present invention program.
In yet another embodiment of the present invention, it in order to improve the quality of selected candidate string, and then improves in the present embodiment
It is fixed based on default extracting rule by search engine to obtain module 10 in the candidate string for the accuracy rate and efficiency of identification target word
To after obtaining candidate string, data purification pretreatment is carried out to the candidate string.
Specifically, the candidate string obtains module 10 can be from format in advance in the exemplary embodiment of this programme
Remove the invalid informations such as html tag in webpage, or from some fixed forms removed in advance in BBS webpage in content
Information etc..
Further, attached drawing 3 is referred to, the characteristic value calculating module 12 is each in the word segmentation result for calculating
At least one specific characteristic value of candidate's participle.
Specifically, in order to improve the recognition accuracy and recall rate of target word in the present invention, the characteristic value calculating module
12 determine candidate's string using the specific characteristic value of plurality of classes.It can be appreciated that common characteristic value be divided into foundation characteristic,
Count category feature, incidence coefficient category feature and context category feature.
In one embodiment of the application, the characteristic value calculating module 12 when calculating candidate string participle mode, according to
The number of words of each candidate participle carrys out partition mode in the word segmentation result of candidate's string.Wherein " 1+1+1 " indicates that word segmentation result is 3
Word, 1 word of each word;" 3+2 " indicates that word segmentation result is 2 words, respectively 3 words and 2 words, for example, candidate string " I/no
/ Lady Pan Jinlian " candidate string participle mode be 1-2-3;Its candidate's participle length maximum value is 3.
Specifically, wherein foundation characteristic, which includes at least, candidate string inquiry times, candidate string participle mode, candidate participle
The maximum value of length, the mean search frequency of candidate participle and ratio of candidate string inquiry times etc..
Specifically, statistics category feature indicates the statistical information of candidate string, including at least has tightness, joint probability, item
Part probability, anti-conditional probability, point mutual information, second order point mutual information, log-likelihood ratio, normalization expectation etc..Likewise, in this Shen
In one embodiment please, when calculating joint probability, the point statistics category feature such as mutual information and conspicuousness, expression is candidate string
Former and later two candidate participle between calculating, each specific characteristic can be obtained according to the calculation formula of each specific characteristic value
Value;If candidate string there are multiple candidate participles, after calculating the corresponding specific characteristic value of candidate participle two-by-two, calculation can be taken respectively
Art is average, geometric average, maximum value and minimum value calculate the corresponding specific characteristic value of candidate string.
Specifically, incidence coefficient category feature indicates the incidence relation between the candidate participle of candidate string, calculation method with
It is similar to count category feature comprising probability ratio, increment, Jie Kade distance (Jaccard Distance), Simpson's distance
(Simpson Distance), Coase measurement (Kosgen Measurement), Pi Tesiji Xia Piluo measurement
(PiaterskyShapiro Measurement) etc..For example, calculating Jie Kade distance in one embodiment of the application
When (Jaccard distance) and chance ratio (Odds ratio, OR), it is assumed that x, y respectively indicate the front and back two of a candidate string
A candidate participle;Xy indicates that two candidates segment while occurring,Indicate that x does not occur and y occurs,Indicate x occur and y not
Occur,It indicates that x and y does not occur, calculates separately the number of above four kinds of situations, further in accordance with Jie Kade distance or chance ratio
Calculation formula calculate.Certainly, those skilled in that art should know, which is only exemplary, and be unable to structure
The restriction of pairs of the present invention program.
Specifically, context category feature indicates the relationship between candidate string and context, include at least context entropy, on
Hereafter diversity, left comentropy, right comentropy, left adjacent diversity, right adjacent diversity, the candidate context gone here and there with the word
Cosine distance between TF-IDF vector etc..Wherein, context entropy indicates the variation institute of the context of candidate string in queries
The information content for including;Context diversity indicates that the different word of how many kinds of occurred in the context of candidate string in queries;Left neighbour
Connecing variation number indicates that the different word of how many kinds of occurred in the left adjacent word of candidate string in queries;It can be according to each specific characteristic value
Calculation formula obtain corresponding specific characteristic value.
Specifically, assuming in one embodiment of this programme, only with foundation characteristic, statistical nature, incidence coefficient class
It is calculated with several characteristic values corresponding to any one in contextual feature, without considering other three category features, although energy
Identify target word, but accuracy rate and recall rate are far below several characteristic values for selecting wherein at least two or more types.
It can be appreciated that it is special to choose foundation characteristic, statistical nature, incidence coefficient class and context according to the actual situation
Several characteristic values in sign calculate.In an exemplary embodiment of this programme, it can be used including tightness, candidate string
Participle mode, joint probability, the maximum value of candidate participle length, context diversity, conspicuousness, candidate string inquiry times,
Left adjacent variation number, than 10 Jie Kade distance, chance specific characteristic values, have obtained preferably determining result.It can be appreciated that
In the other embodiments of this programme, one or more in above-mentioned 10 specific characteristic values can also be selected, is obtained preferably
Determine result.Certainly, those skilled in that art should know, which is only exemplary, and cannot constitute to this hair
The restriction of bright scheme.
Further, attached drawing 3 is referred to, the decision content computing module 13 is pre- for inputting the specific characteristic value
If classifier, the decision content of the string of candidate corresponding to the specific characteristic value is obtained.
It can be appreciated that in order to improve the accuracy rate of the identification of target word described in this programme and recall rate, one of this programme
Decision content computing module described in embodiment 13 uses Nonlinear Classifier, rather than uses the simple linear of direct given threshold
Classifier.For example, the progressive decision tree of gradient (Gradient boosting and can be used in the decision content computing module 13
Decision tree, GBDT) sorting algorithm or deep neural network (Deep Neural Network, DNN) sorting algorithm, when
So training sample word required for deep neural network is selected to want enough, can selected according to the actual situation non-thread using which kind of
Property sorting algorithm.In an exemplary embodiment of the present invention, the decision content computing module 13 uses GBDT sorting algorithm
Realize the present embodiment, GBDT sorting algorithm formed using several decision trees.
Further, it can be appreciated that first to adjust the parameter of GBDT.Specifically, in one embodiment of this programme, also
It include classifier setup module, the classifier setup module, for the decision content computing module 13 at least one by described in
A specific characteristic value inputs default classifier, before obtaining the decision content of the string of candidate corresponding to the specific characteristic value, foundation
At least one specific characteristic value of sample word in training sample set of words is inputted to decision content obtained after presetting classifier,
To adjust the parameter of the default classifier;Wherein the training sample word set is combined into the set of identified sample word.
Specifically, the parameter of the classifier includes minimum sample number required for peripheral node in decision tree, tree
Depth capacity, the quantity of decision tree, for classification specific characteristic value number, determine whether candidate string is the default of target word
Threshold value range etc..Wherein, in order to avoid there is the case where overfitting, minimum sample number cannot be too small, and the maximum set is deep
Degree also should not too deep and the specific characteristic value for classification number it is also unsuitable too many, certainly, excessive sample number may also can
There is poor fitting phenomenon.In an example embodiment of this programme, finger of the classifier setup module setting for classification
The number for determining characteristic value is preferably the total square root of specific characteristic value, certainly in practical application, different values should be used
Come the case where attempting, can at most attempt the 30%-40% of the sum with specific characteristic value, overfitting can be avoided the occurrence of.
Specifically, after at least one specific characteristic value of sample word is inputted default classifier by the classifier setup module
Decision content obtained is compared with the recognition result of the sample word, if it is determined that value is different from the recognition result of the sample word,
The parameter of the classifier is then adjusted, above procedure is repeated, until the default classifier is consistent with recognition result.
Further, described to sentence after obtaining the parameter of optimal GBDT classifier by the classifier setup module
After at least one specific characteristic value of candidate's string is inputted default classifier by fixed value calculation module 13, specific characteristic value is calculated
The decision content of corresponding candidate string.
Further, attached drawing 3, the selecting module 14, for that will meet corresponding to the decision content of preset condition are referred to
Candidate string be set as target word.
Specifically, in one embodiment of the invention, preset condition can be the judgement for meeting predetermined threshold value range
Value.It can be appreciated that after selecting different specific characteristic values corresponding to different characteristic types to input preset classifier, gained
The decision content that arrives is different, and predetermined threshold value range corresponding to the decision content is also corresponding different, i.e., predetermined threshold value range with
Selected specific characteristic value is corresponding.Certainly, those skilled in that art should know, which is merely exemplary,
The restriction to the present invention program cannot be constituted.
For example, in an application scenarios of this programme, we are described by taking movie name " I/be not/Lady Pan Jinlian " as an example
The implementation of case embodiment.More than 10 a specific characteristics are extracted first, obtain " I/be not/Lady Pan Jinlian " this candidate more than 10 gone here and there
Specific characteristic value, as the candidate goes here and there inquiry times (taking logarithm): 17.835, tightness: 0.995, adjacent diversity: 1287 etc.
Deng;By these specific characteristic values by a GBDT classifier, obtain a candidate string whether be target word decision content
0.868783;It is 0.8 by the minimum threshold that the training of great amount of samples word obtains, then classifier score 0.868783 is greater than classification
Thresholding 0.8, then " I is not Lady Pan Jinlian " is judged as meeting the target word of preset condition.
Similarly, in another application scenarios of this programme, with another candidate's string " film/I/be not/Lady Pan Jinlian/"
For the implementation of this programme embodiment described.More than 10 a specific characteristics are extracted first, obtain " film/I/it is not/Lady Pan Jinlian "
The correspondence specific characteristic value of this candidate's string, such as candidate string inquiry times (taking logarithm): 11.644, tightness: 0.679, it is adjacent
Diversity: 81 etc.;By these specific characteristics by a GBDT classifier, obtain whether a candidate string is sentencing for target word
Definite value 0.008744;It is 0.8 by the minimum threshold that the training of great amount of samples word obtains, then classifier score 0.008744 is less than
Classify thresholding 0.8, then " film I be not Lady Pan Jinlian " is judged as non-targeted word.
In conclusion a kind of identification device of target word provided by one embodiment of the present invention, passes through 11 base of word segmentation module
Participle division is carried out to the candidate string of acquisition in the text data of minimum particle size, obtains word segmentation result;Characteristic value calculating module 12
Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;The general of decision content computing module 13 is described at least
One specific characteristic value inputs default classifier, obtains the decision content of the string of candidate corresponding to the specific characteristic value;Selecting module
The string of candidate corresponding to the decision content for meeting preset condition is set as target word by 14.This programme embodiment is by calculating candidate string
At least one characteristic value and combining classification device identify target word, relative to directly artificial settings threshold value and simple linear analysis
Method, accuracy rate has with recall rate significantly to be promoted, and artificial screening cost is largely reduced, and improves the recognition efficiency of target word.
Further, attached drawing 5 is referred to, one embodiment of this programme additionally provides a kind of identifying system of target word,
Comprising:
The identification device 100 of target word is segmented for candidate string of the text data based on minimum particle size to acquisition
It divides, obtains word segmentation result;Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;By described in extremely
A few specific characteristic value inputs default classifier, obtains the decision content of the string of candidate corresponding to the specific characteristic value;It will meet
The string of candidate corresponding to the decision content of preset condition is set as target word;
Training sample identification device 200, for providing the parameter institute of setting classifier to the identification device of the target word
The training sample word needed;
Target word collection device 300, the target word that the identification device for receiving the target word is identified.
For ease of description, it illustrate only part related to the embodiment of the present invention, it is disclosed by specific technical details,
Please refer to present invention method part.
It can be appreciated that above-mentioned module as illustrated by the separation member may or may not be physically separated,
Component shown as a unit may or may not be physical unit, it can and it is in one place, or can also divide
On cloth to multiple network units, can selecting some or all of the modules therein according to the actual needs, scheme is not real to realize
Apply the purpose of example.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage medium
In, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic
Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access
Memory, RAM) etc..Further, attached drawing 6 is referred to, it illustrates the identification systems of the target word of one embodiment of the present of invention
The structural block diagram of system, the system for realizing above-mentioned target word recognition methods.For ease of description, it illustrate only and this
The relevant part of inventive embodiments, it is disclosed by specific technical details, please refer to present invention method part.
The identifying system includes processor 40 and storage medium 50.Wherein, storage medium 40 can be used for storing software
Program and module, processor 50 is stored in the software program and module of storage medium 40 by operation, thereby executing described
The various function application and data processing of identifying system.Storage medium 40 can mainly include storing program area and storing data
Area, wherein storing program area can application program needed for storage program area, at least one function etc.;Storage data area can deposit
Storage uses created data (such as audio data, phone directory etc.) etc. according to the identifying system.In addition, storage medium 40
It may include high random access storage medium, can also include non-volatile memory medium, for example, at least a disk storage
Medium member, flush memory device or other volatile solid-state storage medium parts.
The processor 50 is the control centre of system, utilizes the entire identifying system of various interfaces and connection
Various pieces by running or execute the software program and/or module that are stored in storage medium 40, and are called and are stored in
Data in storage media 40 execute the various functions and processing data of the identifying system, to carry out to the identifying system
Integral monitoring.Optionally, processor 50 may include one or more processing units;Preferably, processor 50 can integrate at
Manage device and modem processor, wherein the main processing operation system of application processor, user interface and application program etc. are adjusted
Demodulation processor processed mainly handles wireless communication.It is understood that above-mentioned modem processor can not also integrate everywhere
It manages in device 50.
It can be appreciated that identifying system further includes the power supply powered to all parts, it is preferred that power supply can although being not shown
With by power-supply management system and processor 50 it is logically contiguous, thus by power-supply management system realize management charge, discharge, with
And the functions such as power managed;It can also include radio frequency (Radio Frequency, RF) circuit, input unit, display unit, biography
The components such as sensor, voicefrequency circuit, wireless module.
In one embodiment of this programme, a kind of storage medium 40 is provided, the storage medium 40 includes storage
Program, wherein the target identification system where controlling the storage medium 40 in described program operation executes any aforementioned mesh
The step of marking the recognition methods of word, to realize following functions:
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains word segmentation result;
Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;
At least one described specific characteristic value is inputted into default classifier, obtains the string of candidate corresponding to the specific characteristic value
Decision content;
The string of candidate corresponding to the decision content for meeting preset condition is set as target word.
Specifically, described carry out participle division based on candidate string of the text data of minimum particle size to acquisition, segmented
As a result the step of, comprising:
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains candidate participle;
At least one described candidate participle is combined, the corresponding word segmentation result of the candidate string is obtained.
Further, the specific characteristic value includes at least one of:
Foundation characteristic, statistics category feature, incidence coefficient category feature, context category feature.
Specifically, the foundation characteristic includes at least one of: candidate's string inquiry times, are waited at candidate string participle mode
The ratio of the mean search frequency, candidate string inquiry times that the maximum value of choosing participle length, candidate segment.
Specifically, the statistics category feature includes at least one of: tightness, joint probability, conditional probability, anti-condition
Probability, point mutual information, second order point mutual information, log-likelihood ratio, normalization expectation.
Specifically, the incidence coefficient category feature includes at least one of: probability ratio, increment, Jie Kade distance, Xin Pu
Gloomy distance.
Specifically, the context category feature includes at least one of: context entropy, context diversity, left information
Entropy, right comentropy, left adjacent diversity, right adjacent diversity.
Optionally, the default classifier includes the progressive decision tree classifier of gradient.
The decision content for meeting preset condition, comprising: meet the decision content of predetermined threshold value range, and the pre- gating
Limits are corresponding at least one described specific characteristic value.
Further, the recognition methods further includes being based on default extracting rule orientation by search engine to obtain candidate
String.
Further, the recognition methods further includes carrying out data purification pretreatment to the candidate string.
In the instructions provided here, although the description of a large amount of detail.It is to be appreciated, however, that of the invention
Embodiment can practice without these specific details.In some embodiments, it is not been shown in detail well known
Methods, structures and technologies, so as not to obscure the understanding of this specification.
Although having been illustrated with some exemplary embodiments of the invention above, those skilled in the art will be managed
Solution, in the case where not departing from the principle of the present invention or spirit, can make a change these exemplary embodiments, of the invention
Range is defined by the claims and their equivalents.
Claims (14)
1. a kind of recognition methods of target word characterized by comprising
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains word segmentation result;
Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;
At least one described specific characteristic value is inputted into default classifier, obtains sentencing for the string of candidate corresponding to the specific characteristic value
Definite value;
The string of candidate corresponding to the decision content for meeting preset condition is set as target word.
2. the method according to claim 1, wherein it is described based on the text data of minimum particle size to the time of acquisition
The step of choosing string carries out participle division, obtains word segmentation result, comprising:
Participle division is carried out based on candidate string of the text data of minimum particle size to acquisition, obtains candidate participle;
At least one described candidate participle is combined, the corresponding word segmentation result of the candidate string is obtained.
3. the method according to claim 1, wherein the specific characteristic value includes at least one of:
Foundation characteristic, statistics category feature, incidence coefficient category feature, context category feature.
4. according to the method described in claim 3, it is characterized in that, the foundation characteristic includes at least one of: candidate's string
Inquiry times, candidate string participle mode, the maximum value of candidate participle length, the mean search frequency of candidate participle, candidate string are looked into
Ask the ratio of number.
5. according to the method described in claim 3, it is characterized in that, the statistics category feature includes at least one of: close
Degree, joint probability, conditional probability, anti-conditional probability, point mutual information, second order point mutual information, log-likelihood ratio, normalization expectation.
6. according to the method described in claim 3, it is characterized in that, the incidence coefficient category feature includes at least one of:
Probability ratio, increment, Jie Kade distance, Simpson's distance.
7. according to the method described in claim 3, it is characterized in that, the context category feature includes at least one of: on
Hereafter entropy, context diversity, left comentropy, right comentropy, left adjacent diversity, right adjacent diversity.
8. according to the method described in claim 1, it is characterized by: the default classifier includes the progressive decision tree of gradient point
Class device.
9. the method according to claim 1, wherein the decision content for meeting preset condition, comprising: meet pre-
The decision content of gating limits, and the predetermined threshold value range is corresponding at least one described specific characteristic value.
10. the method according to claim 1, wherein further include:
Default extracting rule orientation, which is based on, by search engine obtains candidate string.
11. method according to claim 10, which is characterized in that described to be based on default extracting rule orientation by search engine
After the step of obtaining candidate string, further includes:
Data purification pretreatment is carried out to the candidate string.
12. a kind of identification device of target word characterized by comprising
Word segmentation module carries out participle division for candidate string of the text data based on minimum particle size to acquisition, obtains participle knot
Fruit;
Characteristic value calculating module, for calculating at least one specific characteristic value of each candidate participle in the word segmentation result;
Decision content computing module obtains the specific characteristic at least one described specific characteristic value to be inputted default classifier
The decision content of the corresponding candidate string of value;
Selecting module is set as target word for will meet the string of candidate corresponding to the decision content of preset condition.
13. a kind of identifying system of target word characterized by comprising
The identification device of target word carries out participle division for candidate string of the text data based on minimum particle size to acquisition, obtains
To word segmentation result;Calculate at least one specific characteristic value of each candidate participle in the word segmentation result;Will it is described at least one
Specific characteristic value inputs default classifier, obtains the decision content of the string of candidate corresponding to the specific characteristic value;Default item will be met
The string of candidate corresponding to the decision content of part is set as target word;
Training sample identification device, for training needed for providing the parameter that classifier is arranged to the identification device of the target word
Sample word;
Target word collection device, the target word that the identification device for receiving the target word is identified.
14. a kind of storage medium, which is characterized in that the storage medium includes the program of storage, wherein run in described program
When control the storage medium where system perform claim the step of requiring 1 to 11 any the method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710538781.3A CN109241392A (en) | 2017-07-04 | 2017-07-04 | Recognition methods, device, system and the storage medium of target word |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710538781.3A CN109241392A (en) | 2017-07-04 | 2017-07-04 | Recognition methods, device, system and the storage medium of target word |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109241392A true CN109241392A (en) | 2019-01-18 |
Family
ID=65083219
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710538781.3A Pending CN109241392A (en) | 2017-07-04 | 2017-07-04 | Recognition methods, device, system and the storage medium of target word |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109241392A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110175109A (en) * | 2019-05-31 | 2019-08-27 | 北京北信源软件股份有限公司 | A kind of determination method, determining device, equipment and the medium of user type |
CN110275938A (en) * | 2019-05-29 | 2019-09-24 | 广州伟宏智能科技有限公司 | Knowledge extraction method and system based on non-structured document |
CN110909725A (en) * | 2019-10-18 | 2020-03-24 | 平安科技(深圳)有限公司 | Method, device and equipment for recognizing text and storage medium |
CN113342936A (en) * | 2021-06-08 | 2021-09-03 | 北京明略软件***有限公司 | Word formation compactness determining method and device, electronic equipment and storage medium |
CN114363218A (en) * | 2022-01-07 | 2022-04-15 | 合肥工业大学 | Communication reachable rate detection method based on end-to-end learning |
CN114386407A (en) * | 2021-12-23 | 2022-04-22 | 北京金堤科技有限公司 | Word segmentation method and device for text |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101114298A (en) * | 2007-08-31 | 2008-01-30 | 北京搜狗科技发展有限公司 | Method for gaining oral vocabulary entry, device and input method system thereof |
CN102043843A (en) * | 2010-12-08 | 2011-05-04 | 百度在线网络技术(北京)有限公司 | Method and obtaining device for obtaining target entry based on target application |
CN102411563A (en) * | 2010-09-26 | 2012-04-11 | 阿里巴巴集团控股有限公司 | Method, device and system for identifying target words |
CN103092838A (en) * | 2011-10-28 | 2013-05-08 | 腾讯科技(深圳)有限公司 | Method and device for obtaining English words |
CN103593427A (en) * | 2013-11-07 | 2014-02-19 | 清华大学 | New word searching method and system |
CN105095196A (en) * | 2015-07-24 | 2015-11-25 | 北京京东尚科信息技术有限公司 | Method and device for finding new word in text |
CN105183923A (en) * | 2015-10-27 | 2015-12-23 | 上海智臻智能网络科技股份有限公司 | New word discovery method and device |
CN105389349A (en) * | 2015-10-27 | 2016-03-09 | 上海智臻智能网络科技股份有限公司 | Dictionary updating method and apparatus |
CN106445915A (en) * | 2016-09-14 | 2017-02-22 | 科大讯飞股份有限公司 | New word discovery method and device |
-
2017
- 2017-07-04 CN CN201710538781.3A patent/CN109241392A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101114298A (en) * | 2007-08-31 | 2008-01-30 | 北京搜狗科技发展有限公司 | Method for gaining oral vocabulary entry, device and input method system thereof |
CN102411563A (en) * | 2010-09-26 | 2012-04-11 | 阿里巴巴集团控股有限公司 | Method, device and system for identifying target words |
CN102043843A (en) * | 2010-12-08 | 2011-05-04 | 百度在线网络技术(北京)有限公司 | Method and obtaining device for obtaining target entry based on target application |
CN103092838A (en) * | 2011-10-28 | 2013-05-08 | 腾讯科技(深圳)有限公司 | Method and device for obtaining English words |
CN103593427A (en) * | 2013-11-07 | 2014-02-19 | 清华大学 | New word searching method and system |
CN105095196A (en) * | 2015-07-24 | 2015-11-25 | 北京京东尚科信息技术有限公司 | Method and device for finding new word in text |
CN105183923A (en) * | 2015-10-27 | 2015-12-23 | 上海智臻智能网络科技股份有限公司 | New word discovery method and device |
CN105389349A (en) * | 2015-10-27 | 2016-03-09 | 上海智臻智能网络科技股份有限公司 | Dictionary updating method and apparatus |
CN106445915A (en) * | 2016-09-14 | 2017-02-22 | 科大讯飞股份有限公司 | New word discovery method and device |
Non-Patent Citations (1)
Title |
---|
孙晓,孙重远,任福继: ""基于深层结构模型的新词发现与情感倾向判定"", 《计算机科学》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110275938A (en) * | 2019-05-29 | 2019-09-24 | 广州伟宏智能科技有限公司 | Knowledge extraction method and system based on non-structured document |
CN110275938B (en) * | 2019-05-29 | 2021-09-17 | 广州伟宏智能科技有限公司 | Knowledge extraction method and system based on unstructured document |
CN110175109A (en) * | 2019-05-31 | 2019-08-27 | 北京北信源软件股份有限公司 | A kind of determination method, determining device, equipment and the medium of user type |
CN110175109B (en) * | 2019-05-31 | 2023-05-26 | 北京北信源软件股份有限公司 | User type determining method, determining device, equipment and medium |
CN110909725A (en) * | 2019-10-18 | 2020-03-24 | 平安科技(深圳)有限公司 | Method, device and equipment for recognizing text and storage medium |
CN110909725B (en) * | 2019-10-18 | 2023-09-19 | 平安科技(深圳)有限公司 | Method, device, equipment and storage medium for recognizing text |
CN113342936A (en) * | 2021-06-08 | 2021-09-03 | 北京明略软件***有限公司 | Word formation compactness determining method and device, electronic equipment and storage medium |
CN113342936B (en) * | 2021-06-08 | 2024-03-22 | 北京明略软件***有限公司 | Word compactness determining method and device, electronic equipment and storage medium |
CN114386407A (en) * | 2021-12-23 | 2022-04-22 | 北京金堤科技有限公司 | Word segmentation method and device for text |
CN114363218A (en) * | 2022-01-07 | 2022-04-15 | 合肥工业大学 | Communication reachable rate detection method based on end-to-end learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109241392A (en) | Recognition methods, device, system and the storage medium of target word | |
CN110020122B (en) | Video recommendation method, system and computer readable storage medium | |
CN106874292B (en) | Topic processing method and device | |
CN109471944B (en) | Training method and device of text classification model and readable storage medium | |
CN108197144B (en) | Hot topic discovery method based on BTM and Single-pass | |
CN108021708B (en) | Content recommendation method and device and computer readable storage medium | |
CN105512676A (en) | Food recognition method at intelligent terminal | |
CN107209860A (en) | Optimize multiclass image classification using blocking characteristic | |
US20180046721A1 (en) | Systems and Methods for Automatic Customization of Content Filtering | |
CN107545038B (en) | Text classification method and equipment | |
CN113722438B (en) | Sentence vector generation method and device based on sentence vector model and computer equipment | |
CN111914159B (en) | Information recommendation method and terminal | |
CN104537341A (en) | Human face picture information obtaining method and device | |
CN103353881A (en) | Method and device for searching application | |
CN106294473B (en) | Entity word mining method, information recommendation method and device | |
CN108319672A (en) | Mobile terminal malicious information filtering method and system based on cloud computing | |
CN113821657A (en) | Artificial intelligence-based image processing model training method and image processing method | |
CN112348188B (en) | Model generation method and device, electronic device and storage medium | |
CN105512300A (en) | Information filtering method and system | |
CN108875050B (en) | Text-oriented digital evidence-obtaining analysis method and device and computer readable medium | |
CN112035685B (en) | Album video generating method, electronic device and storage medium | |
CN108287850A (en) | The optimization method and device of textual classification model | |
CN111324725B (en) | Topic acquisition method, terminal and computer readable storage medium | |
CN104991920A (en) | Label generation method and apparatus | |
Harakawa et al. | Extraction of hierarchical structure of Web communities including salient keyword estimation for Web video retrieval |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190118 |