CN106844765A - Notable information detecting method and device based on convolutional neural networks - Google Patents
Notable information detecting method and device based on convolutional neural networks Download PDFInfo
- Publication number
- CN106844765A CN106844765A CN201710098500.7A CN201710098500A CN106844765A CN 106844765 A CN106844765 A CN 106844765A CN 201710098500 A CN201710098500 A CN 201710098500A CN 106844765 A CN106844765 A CN 106844765A
- Authority
- CN
- China
- Prior art keywords
- event
- paragraph
- information
- expression
- unsupervised
- 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
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 title claims abstract description 25
- 230000014509 gene expression Effects 0.000 claims abstract description 34
- 238000009826 distribution Methods 0.000 claims abstract description 20
- 239000013604 expression vector Substances 0.000 claims abstract description 15
- 238000013480 data collection Methods 0.000 claims abstract description 10
- 239000012141 concentrate Substances 0.000 claims abstract description 5
- 238000000605 extraction Methods 0.000 claims abstract 2
- 238000001514 detection method Methods 0.000 claims description 34
- 239000011159 matrix material Substances 0.000 claims description 16
- 238000012549 training Methods 0.000 claims description 14
- 239000013598 vector Substances 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 5
- 238000003062 neural network model Methods 0.000 claims description 3
- 208000001970 congenital sucrase-isomaltase deficiency Diseases 0.000 description 6
- 238000002474 experimental method Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 244000097202 Rathbunia alamosensis Species 0.000 description 3
- 235000009776 Rathbunia alamosensis Nutrition 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000007689 inspection Methods 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000007935 neutral effect Effects 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 229940050561 matrix product Drugs 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000006116 polymerization reaction Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/951—Indexing; Web crawling techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
Abstract
The invention discloses a kind of notable information detecting method and device based on convolutional neural networks.Methods described includes：For the data set for being crawled, determine that each event develops the Annual distribution in each stage, and determine timing node；For each event, the corresponding all of event information of the event sample is divided into by several pieces according to identified timing node, the content of text of event information in each time phase is spliced into a paragraph, generate paragraph data collection；Paragraph data described in distribution and expression Algorithm Learning according to paragraph concentrates the unsupervised expression vector of each paragraph；For an event, the unsupervised expression vector of each paragraph is input to depth convolutional neural networks model, the low layer in each stage of event to the expression of high level is obtained using multilayer convolution operation, the key feature in extraction each stage of event is operated by k maximum pondizations, the classification of unreal information is carried out to the information being input into finally by a full articulamentum.
Description
Technical field
The present invention relates to computer processing technology field, more particularly to a kind of notable information inspection based on convolutional neural networks
Survey method and device.
Background technology
The fast development of social media network, is widely used and is easily obtained, and on the one hand largely facilitates user
Life, enrich the experience of user, but simultaneously, the unreal Information Communication on social media network can also upset the normal of people
Life, misleads public sentiment, endangers public security and social stability.Therefore identified from the social media network information of magnanimity unreal
The task of information becomes more and more important and urgent, and the getting up early detection of unreal information also becomes more practical and effective.
The method that existing unreal information mirror method for distinguishing is mainly some Feature Engineerings, the manual feature extracted can be with
From the following aspects, User reliability, the content of microblogging level, the content of event level and from microblogging level to event
The polymerization of level.The manual feature extracted can substantially be divided into following a few classes, the conflict viewpoint in microblogging, microblogging forwarding quantity
Changed with time feature, and microblogging is replied and comprising the signal microblogging for suspecting attitude etc..But these are based on the side of manual feature
Method is all difficult to be related to emerging situation, and social media be it is dynamic, it is variable, it is complicated, can produce many special by hand
Levy the new situation for being difficult to be designed into.
CSID models can detect some significant letters according to user-generated content above social media and generation time
Breath, the including but not limited to discriminating of rumour information and early detection.Usually, microblogging event can include the micro- of thousands of correlations
It is rich, and the temperature difference of microblogging is huge.Firstly for unreal information and real information above data set, count they when
Between characteristic, herein refer to power-law distribution feature of the microblogging with the time.Then the microblogging that be related to for event by model according to it is corresponding when
Between characteristic packet transaction.Microblogging text for different groups, model introduces representative learning method (representation
Learning method), with the distributed expression learning algorithm (paragraph vector) of paragraph, learn each group of microblogging
The expression of text.Finally use deep layer convolutional neural networks, the high-order interaction between modeling each group microblogging, carry out from low order feature to
There is the implicit expression (latent representation) in each stage in the process of high-order feature learning, study event, and
Extract important factor.Examined based on these implicit expressions, the final expression of model event and in the detection of unreal information and early stage
Survey and be made that innovation contribution above.
The content of the invention
In view of there is technological deficiency in method of the tradition based on manual features, for more preferable detection information confidence level, the present invention
A kind of notable information detecting method and device based on convolutional neural networks are provided.
According to an aspect of the present invention, there is provided a kind of notable information detecting method based on convolutional neural networks, including with
Lower step：
Step S1, for the data set including multiple event informations for being crawled, determines the event in the data set
Corresponding each event of information develops the Annual distribution in each stage, and determines each time period corresponding timing node；It is described
Event information in data set includes unreal event information and real time information, and event information correspondence multiple event,
The multiple unreal event informations of each event correspondence or multiple real event information；
Step S2, it is according to identified timing node that the event sample is corresponding all of for each event
Event information is divided into several pieces, and the content of text of event information in each time phase is spliced into a paragraph, generates section
Fall data set；
Step S3, paragraph data described in the distribution and expression Algorithm Learning according to paragraph concentrates the unsupervised expression of each paragraph
Vector；
Step S4, for an event, depth convolutional neural networks is input to by the unsupervised expression vector of each paragraph
Model, the bottom in each stage of event is obtained to high-rise expression using multilayer convolution operation, is operated by k maximum pondizations abundant
Ground extracts the key feature in each stage of event, and dividing for unreal information is carried out to the information being input into finally by a full articulamentum
Class；After carrying out the above-mentioned training of step S4 to the depth convolutional neural networks model using all events, significantly believed
Breath detection model；
Step S5, treating detection information using the notable infomation detection model carries out classification and Detection.
Step S1 includes：
Determine the timestamp of the corresponding all event informations of the event；
For each event, the timestamp is ranked up according to time order and function order；
Earliest time stamp and latest time stamp corresponding time are divided into multiple time periods；
Determine corresponding timing node of the multiple time period.
Step S2 includes：
For each event, according to the multiple time periods and the corresponding event information of the event determined in step S1 when
Between stab, the corresponding event information of the event is divided into the different time periods；
The content of text of the event information in each time period is spliced into a paragraph, multiple time period correspondences are obtained
Multiple paragraphs, constitute paragraph data set.
Step S3 includes：
The paragraph data collection is regarded as a corpus, respectively in word rank and paragraph rank, with unsupervised word
Distributed expression learning algorithm study with paragraph obtains the unsupervised expression vector of each paragraph.
Step S4 includes：
For each event, the unsupervised vector expression of all paragraphs is spliced into a matrix；
The Input matrix to depth convolutional neural networks model is trained.
According to a second aspect of the present invention, there is provided a kind of notable information detector based on convolutional neural networks, including
Following steps：
Timing node determining module, is configured as the data set including multiple event informations for being crawled, it is determined that
Corresponding each event of event information in the data set develops the Annual distribution in each stage, and determines each time period pair
The timing node answered；Event information in the data set includes unreal event information and real time information, and the event
Information correspondence multiple event, the multiple unreal event informations of each event correspondence or multiple real event information；
Paragraph generation module, is configured as each event, according to identified timing node by the event sample
This corresponding all of event information is divided into several pieces, and the content of text of event information in each time phase is spliced into one
Individual paragraph, generates paragraph data collection；
Vector generation module, is configured as the paragraph data according to the distribution and expression Algorithm Learning of paragraph and concentrates each section
The unsupervised expression vector for falling；
Model training module, is configured as an event, and the unsupervised expression vector of each paragraph is input into depth
Degree convolutional neural networks model, obtains the bottom in each stage of event to the expression of high level, using multilayer convolution operation by k most
The key feature in each stage of event is fully extracted in great Chiization operation, and the information being input into is entered finally by a full articulamentum
The classification of the unreal information of row；Using all events the depth convolutional neural networks model is carried out step S4 above-mentioned training it
Afterwards, notable infomation detection model is obtained；
Detection module, is configured to, with the notable infomation detection model and treats detection information carrying out classification and Detection.
The timing node determining module：
First determination sub-module, is configured to determine that the timestamp of the corresponding all event informations of the event；
Sorting sub-module, is configured as, for each event, being ranked up the timestamp according to time order and function order；
Decile submodule, is configured as earliest time stamp and latest time stamp corresponding time being divided into multiple times
Section；
Second determination sub-module, is configured to determine that corresponding timing node of the multiple time period.
The paragraph generation module includes：
Time period divides submodule, is configured as each event, according to the multiple time periods and the event that determine
The timestamp of corresponding event information, the different time periods are divided into by the corresponding event information of the event；
Paragraph generates submodule, is configured as the content of text of the event information in each time period being spliced into one
Paragraph, obtains corresponding multiple paragraphs of multiple time periods, constitutes paragraph data set.
The vector generation module includes：
Unsupervised learning submodule, is configured as regarding the paragraph data collection as a corpus, respectively in word rank
In paragraph rank, the unsupervised table of each paragraph is obtained with the distributed expression learning algorithm study of unsupervised word and paragraph
Up to vector.
The model training module includes：
Splicing submodule, is configured as each event, and the unsupervised vector expression of all paragraphs is spliced into
One matrix；
Training submodule, is configured as being trained the Input matrix to depth convolutional neural networks model.
Brief description of the drawings
Fig. 1 is the schematic diagram of the notable infomation detection MODEL C SID based on convolutional neural networks in the present invention；
Fig. 2 is the power-law distribution diagram of unreal information and real information on microblog data collection in the present invention；
Fig. 3 is different control methods early detection Contrast on effect schematic diagram on microblog data collection.
Specific embodiment
To make the object, technical solutions and advantages of the present invention become more apparent, below in conjunction with specific embodiment, and reference
Accompanying drawing, the present invention is described in further detail.
The present invention discloses a kind of notable infomation detection model (Convolutional based on convolutional neural networks
Salient Information Detection, abbreviation CSID) training method, can be used for unreal in social media network
In information discriminating and early detection task.Model can learn the event comprising varying number level microblogging integrally expresses.Simultaneously
The temporal characteristics modeling event that CSID can also develop according to event develops each stage, from bottom to high-rise semantic meaning representation,
And the crucial feature of selection is operated by neatly k maximum pondizations, being transported to last full articulamentum carries out social media network
The classification learning of information.In a model, if all microbloggings that each event package contains are divided into according to the time phase that event develops
Dry group, every group of microblogging is sent to a depth convolutional neural networks after learning an expression, finally exports this time and belongs to unreal
The probability of information.CSID models are set up：1) for the substantial amounts of unreal information and the data set of real information that are crawled, from entirety
Upper research event develops the Annual distribution in each stage, and each time period corresponding timing node is determined according to this；2) for each
Individual event sample, several pieces are divided into according to the timing node for determining before by all of microblogging, will be micro- in each time phase
Rich content of text is spliced into a paragraph；3) it is whole data set generation paragraph is every according to the distribution and expression Algorithm Learning of paragraph
The unsupervised expression vector of individual paragraph；4) for an event sample, the expression vector in each stage is input to depth convolution
Neural network model, obtains the bottom in each stage of event to the expression of high level, using multilayer convolution operation by neatly k most
The key feature in each stage of event is fully extracted in great Chiization operation, and the information being input into is entered finally by a full articulamentum
The classification of the unreal information of row；5) on test set, by gradient anti-pass, visualized experiment has been carried out to convolution kernel and gradient, it is right
Model learning to notable information carried out deep analysis and demonstration.In Sina weibo data set and the experiment for pushing away special data set
On, obtain than other existing models more accurately prediction effect.
As shown in figure 1, a kind of notable information detecting method based on convolutional neural networks is the embodiment of the invention provides, should
Method includes：
Receive information to be sorted；
By the information input to be sorted to the good notable infomation detection model of training in advance；
It is real information or the result of unreal information that the notable infomation detection model exports the information to be sorted.
In one embodiment, the notable information monitoring model is first good by model training according to data with existing, obtains
After trained good model, for emerging information, also pass through similar operation and fresh information is input in model, then
Model can export a probable value, represent that input information belongs to the probability of unreal information, and output valve is bigger, and input information has been got over can
Can be unreal information.
Below in conjunction with each detailed problem involved in accompanying drawing detailed description technical solution of the present invention.It should be understood that
It is that described embodiment is intended merely to facilitate understanding, any restriction effect is not risen to the present invention.
In order to more fully understand that CSID models are acted in unreal infomation detection, and implementation result of the invention is verified,
Next illustrated by taking experiment as an example, this example uses Sina weibo database.Experimental data set is divided into 60% training set,
30% test set and 10% checking collection.
Experiment includes four evaluation index accuracys rate (Accuracy), accurate rate (Precision), recall rate (Recall)
And F1-score.Research object calculates Precision and Recall to show respectively when being respectively unreal information and real information
Representation model detects two kinds of abilities of information.The value of four kinds of evaluation indexes is bigger, and the detection performance of the unreal information of model is higher.
As shown in figure 1, specific experiment step is as follows on Sina weibo data set：
Step S1, for the substantial amounts of unreal information and the data set of real information that are crawled, includes multiple event E=
{ei, (for an event, the event can be described to that there should be multiple information, such as, for a significant time, had very
The information such as a plurality of microblogging or news describes the event), research event develops the Annual distribution in each stage on the whole, first
Collect timestamp (i.e. information of the corresponding all microbloggings of all events (this is sentenced as a example by microblogging, or other information)
The time point of issue), arranged according to time order and function order, the earliest and latest time stamp corresponding time period is then divided into M
Part (such as M=20 parts), and each time period corresponding timing node is determined according to this,
Ti=[ti-1, ti), i=1,2 ..., 20.
Wherein TiRepresent i-th time phase, ti-1And tiWhen representing i-th time phase initial time stamp respectively and terminating
Between stab.In addition it is also necessary to be normalized operation to each timing node, the resulting corresponding timestamp of timing node is returned
One changes to 0-1 intervals.
Step S2, a plurality of microblogging is included for eachEvent sampleFirst by this event package
The timestamp t of all microbloggings for containingj0-1 intervals are normalized to, be divided into for all of microblogging by the timing node determined further according to S1
Several pieces, a paragraph is spliced into by the content of text of microblogging in each time phase, i.e., the timestamp of microblogging at i-th
Time phase TiIn the content of all microbloggings be all spliced into a paragraph.
Step S3, regards all of content of microblog text data set in S2 as a corpus, respectively in word rank and section
Fall in rank, with unsupervised word and distributed expression the learning algorithm word2vec and para2vec of paragraph, study obtains every
The expression vector of individual word and each paragraph, separately constitutes matrix W, D.Each row in matrix W and D correspond to a word and section respectively
The expression vector for falling.
Wherein N represents the word number in paragraph, and the window width of context is 2k, that is, each k word is used as upper before and after selecting current word
Hereafter, algorithm to be mainly and express the recall info in vector by the word and paragraph of context and maximize all words in paragraph
Combination condition distribution probability p, Probability p calculated by softmax.yiI-th output response of word is represented, can be by following formula
Draw,
Y=b+UTh(pj, wn-k... wn+k；D, W)
Wherein pjFor the vector table of a paragraph reaches, wnRepresent that the vector table of n-th word in paragraph reaches, pjAnd wnRespectively
It is a certain row in matrix D and W.B and U is the parameter of softmax, and h is to average or concatenation.
Step S4, for an event sample, by the paragraph expression vector p in S3jIt is spliced into a matrixWherein the dimension of d and n representing matrixs P, is input to depth convolutional neural networks model, using multilayer convolution operation
The bottom in each stage of event is obtained to high-rise expression, is referred to as in the output result of a certain layer in deep neural network model
One characteristic pattern, the output result of the low layer of neutral net is referred to as low order characteristic pattern, and the high-rise output result of neutral net claims
It is high-order characteristic pattern, some element of characteristic pattern can be obtained by following convolution operation,
F [i]=tanh (<P[:,i:i+ω-1],C>F)
Wherein P [:,i:I+w-1] representing matrix E i-th to (i+ ω -1) row, ω represents the width of convolution kernel, and C is represented
Convolution weight matrix.Asking the operation of the mark after matrix product can be expressed as Frobenius inner product operations, as follows：
<X,Y>F=Tr (XYT)
The key feature for fully extracting each stage of event is operated by neatly k maximum pondizations, that is, extracts characteristic pattern
Middle k maximum elementAs new characteristic pattern.Unreal information is carried out to the information being input into finally by a full articulamentum
Classification.
Depth convolutional neural networks model can first random initializtion, then constantly trained by S4, the parameter of more new model.
Step S5, on test set, by gradient anti-pass, obtains gradient matrix of the output label for input, for defeated
Enter matrix and do the content of microblog that significance analysis obtain being played during correspondence is input into remarkable effect.In addition to first volume of convolutional layer
Product core has carried out deep visual analyzing, obtains the characteristic distributions of significant content of microblog in event.
Fig. 2 is the power-law distribution diagram of unreal information and real information on microblog data collection in the present invention；As shown in Figure 2
In data set, for real information and unreal information, the ratio shared by different phase microblogging number changes with time situation, reaction
Microblogging number with the time power-law distribution situation.Fig. 3 illustrates the experimental result of the early detection of unreal information.
Table 1 illustrates the statistics of attributes information in Twitter and Weibo data sets
Table 2：Unreal information differentiates (M:Unreal information, T:Real information)
Table 2 illustrates the Comparison of experiment results of proposed CSID methods and existing other method
Above-mentioned model proposed by the present invention discloses the power of microblogging quantity that event package in social media network contains with the time
The rule regularity of distribution, and according to this rule using overall etc. point of timing node in establishment each stage of event, then according to these
Time phase carries out cutting to each event, so not only ensures there is the microblogging of approximately the same number in each time interval
Number, and can on the whole ensure that all events share a time scale.It is more real that model can learn event
Expression, and can fully excavate and utilize the temporal regularity that information is distributed.Event each rank is obtained using multilayer convolution operation
The bottom of section can fully model the high-order interaction and Deep Semantics expression in each stage of event to high-rise expression；By spirit
The key feature in each stage of event is fully extracted in ground living k maximum pondization operations, model is more can adapt in DYNAMIC COMPLEX
Social medium scene.
Notable infomation detection task the present invention relates to be based on convolutional neural networks, it is big particular for information content scale,
Substantially, semantic scene is complicated for time span difference, the user behavior dynamically real social activity medium occasion of changeable grade, notable information inspection
Survey can obtain more accurately Detection results.
Particular embodiments described above, has been carried out further in detail to the purpose of the present invention, technical scheme and beneficial effect
Describe in detail bright, it should be understood that the foregoing is only specific embodiment of the invention, be not intended to limit the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements done etc. should be included in protection of the invention
Within the scope of.
Claims (10)
1. a kind of notable information detecting method based on convolutional neural networks, comprises the following steps：
Step S1, for the data set including multiple event informations for being crawled, determines the event information pair in the data set
Each event answered develops the Annual distribution in each stage, and determines each time period corresponding timing node；The data set
In event information include unreal event information and real event information, and the data set correspondence multiple events, each event
Correspondence at least one unreal event information and/or at least one real event information；
Step S2, for each event, according to identified timing node by the corresponding all of event of the event sample
Information is divided into several pieces, and the content of text of event information in each time phase is spliced into a paragraph, generates paragraph number
According to collection；
Step S3, paragraph data described in the distribution and expression Algorithm Learning according to paragraph concentrate the unsupervised expression of each paragraph to
Amount；
Step S4, for an event, depth convolutional neural networks model is input to by the unsupervised expression vector of each paragraph,
The bottom in each stage of event is obtained to high-rise expression using multilayer convolution operation, and extraction event is operated by k maximum pondizations
The key feature in each stage, the classification of unreal information is carried out finally by a full articulamentum to the information being input into；Using institute
After having event and carrying out the above-mentioned training of step S4 to the depth convolutional neural networks model, notable infomation detection mould is obtained
Type；
Step S5, treating detection information using the notable infomation detection model carries out classification and Detection.
2. method according to claim 1, it is characterised in that step S1 includes：
Determine the timestamp of the corresponding all event informations of the event；
For each event, the timestamp is ranked up according to time order and function order；
Earliest time stamp and latest time stamp corresponding time are divided into multiple time periods；
Determine corresponding timing node of the multiple time period.
3. method according to claim 1, it is characterised in that step S2 includes：
For each event, according to the multiple time periods determined in step S1 and the time of the corresponding event information of the event
Stamp, the different time periods are divided into by the corresponding event information of the event；
The content of text of the event information in each time period is spliced into a paragraph, multiple time periods is obtained corresponding many
Individual paragraph, constitutes paragraph data set.
4. method according to claim 1, it is characterised in that step S3 includes：
The paragraph data collection is regarded as a corpus, respectively in word rank and paragraph rank, with unsupervised word and section
The distributed expression learning algorithm study for falling obtains the unsupervised expression vector of each paragraph.
5. method according to claim 1, it is characterised in that step S4 includes：
For each event, the unsupervised vector expression of all paragraphs is spliced into a matrix；
The Input matrix to depth convolutional neural networks model is trained.
6. a kind of notable information detector based on convolutional neural networks, comprises the following steps：
Timing node determining module, is configured as the data set including multiple event informations for being crawled, it is determined that described
Corresponding each event of event information in data set develops the Annual distribution in each stage, and determines that each time period is corresponding
Timing node；Event information in the data set includes unreal event information and real time information, and the event information
Correspondence multiple event, the multiple unreal event informations of each event correspondence or multiple real event information；
Paragraph generation module, is configured as each event, according to identified timing node by the event sample pair
The all of event information answered is divided into several pieces, and the content of text of event information in each time phase is spliced into a section
Fall, generate paragraph data collection；
Vector generation module, is configured as the paragraph data according to the distribution and expression Algorithm Learning of paragraph and concentrates each paragraph
Unsupervised expression vector；
Model training module, is configured as an event, and the unsupervised expression vector of each paragraph is input into depth volume
Product neural network model, obtains the bottom in each stage of event to high-rise expression, by k maximums pond using multilayer convolution operation
Change the key feature that each stage of event is fully extracted in operation, the information being input into is carried out not finally by a full articulamentum
The classification of real information；After carrying out above-mentioned training to the depth convolutional neural networks model using all events, obtain significantly
Infomation detection model；
Detection module, is configured to, with the notable infomation detection model and treats detection information carrying out classification and Detection.
7. device according to claim 6, it is characterised in that timing node determining module：
First determination sub-module, is configured to determine that the timestamp of the corresponding all event informations of the event；
Sorting sub-module, is configured as, for each event, being ranked up the timestamp according to time order and function order；
Decile submodule, is configured as earliest time stamp and latest time stamp corresponding time being divided into multiple time periods；
Second determination sub-module, is configured to determine that corresponding timing node of the multiple time period.
8. device according to claim 6, it is characterised in that the paragraph generation module includes：
Time period divides submodule, is configured as each event, according to multiple time periods and the event correspondence for determining
Event information timestamp, the corresponding event information of the event is divided into the different time periods；
Paragraph generates submodule, is configured as the content of text of the event information in each time period being spliced into a section
Fall, obtain corresponding multiple paragraphs of multiple time periods, constitute paragraph data set.
9. device according to claim 6, it is characterised in that the vector generation module includes：
Unsupervised learning submodule, is configured as regarding the paragraph data collection as a corpus, respectively in word rank and section
Fall in rank, with the distributed expression learning algorithm study of unsupervised word and paragraph obtain the unsupervised expression of each paragraph to
Amount.
10. device according to claim 6, it is characterised in that the model training module includes：
Splicing submodule, is configured as each event, and the unsupervised vector expression of all paragraphs is spliced into one
Matrix；
Training submodule, is configured as being trained the Input matrix to depth convolutional neural networks model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710098500.7A CN106844765B (en) | 2017-02-22 | 2017-02-22 | Significant information detection method and device based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710098500.7A CN106844765B (en) | 2017-02-22 | 2017-02-22 | Significant information detection method and device based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106844765A true CN106844765A (en) | 2017-06-13 |
| CN106844765B CN106844765B (en) | 2019-12-20 |
Family
ID=59134861
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710098500.7A Active CN106844765B (en) | 2017-02-22 | 2017-02-22 | Significant information detection method and device based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106844765B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107688870A (en) * | 2017-08-15 | 2018-02-13 | 中国科学院软件研究所 | A kind of the classification factor visual analysis method and device of the deep neural network based on text flow input |
| WO2018126325A1 (en) * | 2017-01-06 | 2018-07-12 | The Toronto-Dominion Bank | Learning document embeddings with convolutional neural network architectures |
| CN108491480A (en) * | 2018-03-12 | 2018-09-04 | 义语智能科技（上海）有限公司 | Rumour detection method and equipment |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102411611A (en) * | 2011-10-15 | 2012-04-11 | 西安交通大学 | Instant interactive text oriented event identifying and tracking method |
| CN104834747A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Short text classification method based on convolution neutral network |
| CN104915448A (en) * | 2015-06-30 | 2015-09-16 | 中国科学院自动化研究所 | Substance and paragraph linking method based on hierarchical convolutional network |
| CN105608200A (en) * | 2015-12-28 | 2016-05-25 | 湖南蚁坊软件有限公司 | Network public opinion tendency prediction analysis method |
| CN105740349A (en) * | 2016-01-25 | 2016-07-06 | 重庆邮电大学 | Sentiment classification method capable of combining Doc2vce with convolutional neural network |
| CN105975497A (en) * | 2016-04-27 | 2016-09-28 | 清华大学 | Automatic microblog topic recommendation method and device |
| CN106202211A (en) * | 2016-06-27 | 2016-12-07 | 四川大学 | A kind of integrated microblogging rumour recognition methods based on microblogging type |
-
2017
- 2017-02-22 CN CN201710098500.7A patent/CN106844765B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102411611A (en) * | 2011-10-15 | 2012-04-11 | 西安交通大学 | Instant interactive text oriented event identifying and tracking method |
| CN104834747A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Short text classification method based on convolution neutral network |
| CN104915448A (en) * | 2015-06-30 | 2015-09-16 | 中国科学院自动化研究所 | Substance and paragraph linking method based on hierarchical convolutional network |
| CN105608200A (en) * | 2015-12-28 | 2016-05-25 | 湖南蚁坊软件有限公司 | Network public opinion tendency prediction analysis method |
| CN105740349A (en) * | 2016-01-25 | 2016-07-06 | 重庆邮电大学 | Sentiment classification method capable of combining Doc2vce with convolutional neural network |
| CN105975497A (en) * | 2016-04-27 | 2016-09-28 | 清华大学 | Automatic microblog topic recommendation method and device |
| CN106202211A (en) * | 2016-06-27 | 2016-12-07 | 四川大学 | A kind of integrated microblogging rumour recognition methods based on microblogging type |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018126325A1 (en) * | 2017-01-06 | 2018-07-12 | The Toronto-Dominion Bank | Learning document embeddings with convolutional neural network architectures |
| US10360303B2 (en) | 2017-01-06 | 2019-07-23 | The Toronto-Dominion Bank | Learning document embeddings with convolutional neural network architectures |
| US11030415B2 (en) | 2017-01-06 | 2021-06-08 | The Toronto-Dominion Bank | Learning document embeddings with convolutional neural network architectures |
| CN107688870A (en) * | 2017-08-15 | 2018-02-13 | 中国科学院软件研究所 | A kind of the classification factor visual analysis method and device of the deep neural network based on text flow input |
| CN107688870B (en) * | 2017-08-15 | 2020-07-24 | 中国科学院软件研究所 | Text stream input-based hierarchical factor visualization analysis method and device for deep neural network |
| CN108491480A (en) * | 2018-03-12 | 2018-09-04 | 义语智能科技（上海）有限公司 | Rumour detection method and equipment |
| CN108491480B (en) * | 2018-03-12 | 2021-05-11 | 义语智能科技（上海）有限公司 | Rumor detection method and apparatus |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106844765B (en) | 2019-12-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104216954B (en) | The prediction meanss and Forecasting Methodology of accident topic state | |
| CN104182517B (en) | The method and device of data processing | |
| CN103500175B (en) | A kind of method based on sentiment analysis on-line checking microblog hot event | |
| CN107193801A (en) | A kind of short text characteristic optimization and sentiment analysis method based on depth belief network | |
| CN106855853A (en) | Entity relation extraction system based on deep neural network | |
| CN108038725A (en) | A kind of electric business Customer Satisfaction for Product analysis method based on machine learning | |
| CN108920456A (en) | A kind of keyword Automatic method | |
| CN104820629A (en) | Intelligent system and method for emergently processing public sentiment emergency | |
| CN103324666A (en) | Topic tracing method and device based on micro-blog data | |
| CN105893609A (en) | Mobile APP recommendation method based on weighted mixing | |
| CN105608200A (en) | Network public opinion tendency prediction analysis method | |
| CN103136358B (en) | A kind of method of Automatic Extraction forum data | |
| CN108038205A (en) | For the viewpoint analysis prototype system of Chinese microblogging | |
| WO2015061046A2 (en) | Method and apparatus for performing topic-relevance highlighting of electronic text | |
| CN108647800A (en) | A kind of online social network user missing attribute forecast method based on node insertion | |
| CN104731958A (en) | User-demand-oriented cloud manufacturing service recommendation method | |
| CN109685153A (en) | A kind of social networks rumour discrimination method based on characteristic aggregation | |
| CN104699767B (en) | A kind of extensive Ontology Mapping Method towards Chinese language | |
| CN107679135A (en) | The topic detection of network-oriented text big data and tracking, device | |
| CN103823890A (en) | Microblog hot topic detection method and device aiming at specific group | |
| CN109918648B (en) | Rumor depth detection method based on dynamic sliding window feature score | |
| CN106844765A (en) | Notable information detecting method and device based on convolutional neural networks | |
| CN105869058B (en) | A kind of method that multilayer latent variable model user portrait extracts | |
| CN108304479A (en) | A kind of fast density cluster double-layer network recommendation method based on graph structure filtering | |
| Kumar et al. | Battling fake news: A survey on mitigation techniques and identification |
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 |