CN109614541A - A kind of event recognition method, medium, device and calculate equipment - Google Patents
A kind of event recognition method, medium, device and calculate equipment Download PDFInfo
- Publication number
- CN109614541A CN109614541A CN201811472932.0A CN201811472932A CN109614541A CN 109614541 A CN109614541 A CN 109614541A CN 201811472932 A CN201811472932 A CN 201811472932A CN 109614541 A CN109614541 A CN 109614541A
- Authority
- CN
- China
- Prior art keywords
- event
- text data
- event information
- deep learning
- learning model
- 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
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Machine Translation (AREA)
Abstract
Embodiments of the present invention provide a kind of event recognition method, medium, device and calculate equipment.The event recognition method includes: text data to be obtained from network data and with the event information of matches text data as first matching result, and text data includes event to be identified；Using first matching result as the input of deep learning model, the fiducial probability of the matching relationship accuracy between text data and event information is used to indicate using the output of deep learning model；If fiducial probability meets preset condition, event to be identified is determined according to matching relationship.Embodiments of the present invention help to improve the portability of event recognition scheme, expand the application scenarios of event recognition scheme, and it additionally aids and avoids directly using over-fitting caused by deep learning model, the processing pressure for mitigating deep learning model, promotes the reliability, real-time and recognition effect of event recognition.
Description
Technical field
Embodiments of the present invention are related to software technology field, more specifically, embodiments of the present invention are related to a kind of thing
Part recognition methods, medium, device and calculating equipment.
Background technique
Background that this section is intended to provide an explanation of the embodiments of the present invention set forth in the claims or context.Herein
Description recognizes it is the prior art not because not being included in this section.
With the development of big data technology, the specific thing such as negative event, sensitive event is identified from network data
Part, and individual or company's machine is had become based on the analysis analyzed/assessed result/assessment result is carried out to these particular events
One of important decision foundation of structure.Currently, common event recognition scheme mainly has: the event recognition scheme based on pattern match
With the event recognition scheme based on machine learning.
First, the event recognition scheme based on pattern match is usually to depend on various pattern algorithms to certain class event type
Or Event element carries out matching to realize event recognition, but the domain knowledge that various pattern algorithms need largely to be manually set
The data manually marked, it is strong to the dependence of specific field, the portability of event recognition scheme is greatly reduced, is limited
The application scenarios of event recognition scheme.
Second, the event recognition scheme based on machine learning would generally utilize support vector machines, Bayesian model, decision
Event recognition problem is converted into sequence labelling problem or classification problem by the machine learning models such as tree, neural network.Existing base
Although the event recognition scheme in machine learning can extend the application scenarios of event recognition scheme with fitting data, it is easy
Existing over-fitting, the i.e. performance of event recognition scheme identification new data are caused far below the performance for the data that recognition training is crossed
Poor reliability, the real-time of event recognition are weak, effect is poor.
To sum up, there are portable poor, the scheme application scenarios of scheme more to limit to for existing event recognition scheme, and event is known
The problems such as other real-time is weak, effect is poor.
Summary of the invention
There are portable poor, the scheme application scenarios of scheme more to limit to for event recognition scheme at present, the reality of event recognition
The problems such as when property is weak, effect is poor.In the present context, embodiments of the present invention are intended to provide a kind of event recognition method, dress
Set, medium and calculate equipment.
In the first aspect of embodiment of the present invention, a kind of event recognition method is provided, comprising:
Text data is obtained from network data and with the event information of matches text data as first matching result, text
Notebook data includes event to be identified；
Using first matching result as the input of deep learning model, textual data is used to indicate using the output of deep learning model
According to the fiducial probability of the matching relationship accuracy between event information；
If the fiducial probability meets preset condition, event to be identified is determined according to the matching relationship.
In one embodiment of the invention, as follows using first matching result as the defeated of deep learning model
Enter, specifically include: obtaining the just corresponding term vector of matching result, term vector includes the corresponding text term vector of text data and thing
The corresponding event information vector of part information；Pass through time recurrent neural network (Long Short Term Memory, LSTM) mould
Type encodes text term vector, so that text term vector is converted to hidden state；Determine hidden state and event information vector
Input as next layer deep learning model.
In one embodiment of the invention, next layer deep learning model is attention mechanism (Attention) mould
Type.The confidence for being used to indicate the matching relationship accuracy between text data and event information using the output of deep learning model is general
Rate specifically includes: being weighted and averaged by Attention model to hidden state；Based on the hidden state after judgement weighted average
With the matching relationship between event information vector, the fiducial probability is exported using full articulamentum.
In one embodiment of the invention, hidden state is weighted and averaged by Attention model, comprising: when
It is the different weights of the hidden states configuration of difference in multiple hidden states based on Attention model when the quantity of hidden state is multiple, and
Multiple hidden states after configuration weight are weighted and averaged.
In one embodiment of the invention, event information include event to be identified keyword and event to be identified
Central person.
In the second aspect of embodiment of the present invention, a kind of event recognition device is provided, comprising:
Matching unit, for from network data obtain text data and with the event information conduct of matches text data
First matching result, text data include event to be identified；
Judging unit is exported for the input using first matching result as deep learning model using deep learning model
It is used to indicate the fiducial probability of the matching relationship accuracy between text data and event information；
Recognition unit, if meeting preset condition for fiducial probability, according between this article notebook data and event information
Matching relationship determines event to be identified.
In one embodiment of the invention, judging unit is in the input using first matching result as deep learning model
When, it is specifically used for: obtains the just corresponding term vector of matching result, term vector includes the corresponding text term vector of text data and thing
The corresponding event information vector of part information；Text term vector is encoded by time recurrent neural network LSTM model, from
And text term vector is converted into hidden state；Determine hidden state with event information vector as the defeated of next layer deep learning model
Enter.
In one embodiment of the invention, next layer deep learning model is Attention model.Judging unit is being adopted
When being used to indicate the fiducial probability of the matching relationship accuracy between text data and event information with the output of deep learning model,
It is specifically used for: hidden state is weighted and averaged by Attention model；Based on the hidden state and event letter after weighted average
The matching relationship between vector is ceased, fiducial probability is exported using full articulamentum.
In one embodiment of the invention, judging unit hidden state is weighted by Attention model it is flat
When equal, it is specifically used for: is different hidden shapes in multiple hidden states based on Attention model when the quantity of hidden state is multiple
State configures different weights, and is weighted and averaged to multiple hidden states after configuration weight.
In one embodiment of the invention, event information include event to be identified keyword and event to be identified
Central person.
In the third aspect of embodiment of the present invention, a kind of medium is provided, which has computer executable
Instruction, the method that computer executable instructions are used to that computer to be made to execute any embodiment in first aspect.
In the fourth aspect of embodiment of the present invention, provide a kind of calculating equipment, including processing unit, memory with
And input/output (In/Out, I/O) interface；Memory, the program or instruction executed for storage processing unit；Processing unit,
Program or instruction for being stored according to memory, the method for executing any embodiment in first aspect；I/O interface is used for
Data are received or sent under the control of processing unit.
Embodiments of the present invention provide technical solution, the available text data comprising event to be identified and this
The event information of matches text data obtains textual data as first matching result, and by first matching result input deep learning model
According to the fiducial probability of the matching relationship between event information, to filter out the matching relationship that fiducial probability meets preset condition
To determination event to be identified.Technical solution provided by the invention sieves first matching result by introducing deep learning model
Choosing gets rid of the data that event recognition scheme is marked to the domain knowledge of artificial settings and manually, to help to improve event
The portability of identifying schemes expands the application scenarios of event recognition scheme.Also, technical solution provided by the invention is sufficiently sharp
With domain knowledge combination attention mechanism that can be extensive, to help avoid directly using caused by deep learning model
Over-fitting mitigates the processing pressure of deep learning model, promotes the reliability, real-time and identification effect of event recognition
Fruit.
Detailed description of the invention
The following detailed description is read with reference to the accompanying drawings, above-mentioned and other mesh of exemplary embodiment of the invention
, feature and advantage will become prone to understand.In the accompanying drawings, if showing by way of example rather than limitation of the invention
Dry embodiment, in which:
Fig. 1 schematically show the present embodiments relate to a kind of event recognition method flow diagram；
Fig. 2 schematically shows the present embodiments relate to a kind of deep learning model structural schematic diagram；
Fig. 3 schematically show the present embodiments relate to a kind of event recognition device structural schematic diagram；
Fig. 4 schematically shows a kind of structural schematic diagrams of medium of the embodiment of the present invention；
Fig. 5 schematically shows a kind of structural schematic diagram of calculating equipment of the embodiment of the present invention.
In the accompanying drawings, identical or corresponding label indicates identical or corresponding part.
Specific embodiment
The principle and spirit of the invention are described below with reference to several illustrative embodiments.It should be appreciated that providing this
A little embodiments are used for the purpose of making those skilled in the art can better understand that realizing the present invention in turn, and be not with any
Mode limits the scope of the invention.On the contrary, these embodiments are provided so that this disclosure will be more thorough and complete, and energy
It is enough that the scope of the present disclosure is completely communicated to those skilled in the art.
One skilled in the art will appreciate that embodiments of the present invention can be implemented as a kind of system, device, equipment, method
Or computer program product.Therefore, the present disclosure may be embodied in the following forms, it may be assumed that complete hardware, complete software
The form that (including firmware, resident software, microcode etc.) or hardware and software combine.
Embodiment according to the present invention proposes a kind of event recognition method, medium, device and calculates equipment.
Herein, it is to be understood that the meaning of related several terms is respectively as follows:
Mode (matching) algorithm: for carrying out matching to realize event recognition to certain class event type or Event element,
It is commonly used in the event recognition scheme based on pattern match.For example, the general structure of the text engineering of commercial field event recognition
(General Architecture for Text Engineering, GATE) dictionary, or foot is extracted from news web page
Ball event ontology orientation event recognition system (Ontology Directed Event Extraction System,
ODEES)。
Time recurrent neural network (Long Short Term Memory, LSTM): i.e. shot and long term memory network is a kind of
The Recognition with Recurrent Neural Network of modified version is suitable for being spaced and postpone relatively long important thing in processing and predicted time sequence
Part.
Attention mechanism (Attention): for simulating the attention degree for giving different things, it can be understood as pay attention to
There is power certain weight to distinguish at any time.For example, the entirety of width picture would generally be paid close attention to when an ornamental width is drawn
Overall picture, and while deeply examining, would generally then focus on the regional area of whole picture picture.
Fiducial probability (Confidence Probability): being the probability for the statistical inference degree of reliability, confidence is general
Rate can be used for characterizing certain parameter when carrying out statistical inference in probability within a certain range.
In addition, any number of elements in attached drawing is used to example rather than limitation and any name are only used for distinguishing,
Without any restrictions meaning.
Below with reference to several representative embodiments of the invention, the principle and spirit of the present invention are explained in detail.
Summary of the invention
The inventors discovered that the event recognition scheme currently based on pattern match is strong to the dependence of specific field, significantly
The portability for reducing event recognition scheme limits the application scenarios of event recognition scheme；And the thing based on machine learning
Part identifying schemes are since the capability of fitting of machine learning model is stronger, although can extend event recognition scheme to a certain extent
Application scenarios, but it is easy to appear over-fitting, so that machine learning model only can recognize that in the text data trained
Event, and can not effectively identify the event in new text data, i.e., event recognition scheme identifies that the performance of new text data is far low
In the performance for the text data that recognition training is crossed, cause that the poor reliability of event recognition, real-time are weak, effect is poor.To sum up, existing
Event recognition scheme there are portable poor, the scheme application scenarios of scheme more to limit to, the real-time of event recognition is weak, effect
The problems such as poor.
In view of the above-mentioned problems, the present invention provides a kind of event recognition method, device, medium and calculating equipment.The present invention
In the technical solution of offer, the event information of the available text data comprising event to be identified and text Data Matching
The matching between text data and event information is obtained as first matching result, and by first matching result input deep learning model
The fiducial probability of relationship, so that filtering out fiducial probability meets the matching relationship of preset condition to determination event to be identified.This
It invents the technical solution provided to screen first matching result by introducing deep learning model, gets rid of event recognition scheme
Domain knowledge to artificial settings and the data manually marked are opened up to help to improve the portability of event recognition scheme
The application scenarios of exhibit-business part identifying schemes.Also, technical solution provided by the invention makes full use of domain knowledge knot that can be extensive
Attention mechanism is closed, to help avoid directly using over-fitting caused by deep learning model, mitigates depth
The processing pressure for practising model, promotes the reliability, real-time and recognition effect of event recognition.
After introduced the basic principles of the present invention, lower mask body introduces various non-limiting embodiment party of the invention
Formula.
Application scenarios overview
The embodiment of the present invention can be applied to the event recognition scene of event recognition scene, especially special object.This hair
The event recognition scene that bright embodiment is related to, which for example can be, bears mechanism/company/entertainment star's implementation based on network data
Face event recognition scene is also possible to the focus incident assessment scene for entertainment star, can also be other event recognition fields
Scape does not limit in the embodiment of the present invention.The present embodiments relate to network data include but be not limited to news report, media
Article, user's message, webpage information, video comments information.
Illustrative methods
It is described with reference to Figure 1 the method for event recognition of illustrative embodiments according to the present invention.It should be noted that
It is which is shown only for the purpose of facilitating an understanding of the spirit and principles of the present invention for above-mentioned application scenarios, embodiments of the present invention are herein
Aspect is unrestricted.On the contrary, embodiments of the present invention can be applied to applicable any scene.
The embodiment of the invention provides a kind of event recognition methods, as shown in Figure 1, this method comprises:
S101, text data is obtained from network data and is just matched with the conduct of the event information of text Data Matching
As a result, text data includes event to be identified；
S102, using first matching result as the input of deep learning model, using deep learning model output be used to indicate
The fiducial probability of matching relationship accuracy between text data and event information；
If S103, the fiducial probability meet preset condition, event to be identified is determined according to matching relationship.
By event recognition method shown in fig. 1, deep learning model can be introduced, first matching result is screened, put
The data that event identifying schemes are marked to the domain knowledge of artificial settings and manually are taken off, to help to improve event recognition side
The portability of case expands the application scenarios of event recognition scheme.Also, technical solution provided by the invention makes full use of can be general
The domain knowledge combination attention mechanism of change, to help avoid directly using over-fitting caused by deep learning model
Phenomenon mitigates the processing pressure of deep learning model, promotes the reliability, real-time and recognition effect of event recognition.
In the embodiment of the present invention, text data is the natural language data extracted from network data, herein natural language
Say that data include event to be identified.The present embodiments relate to text data can be the text data based on single languages,
Such as Chinese text data；The present embodiments relate to text data be also possible to based on multilingual mixed text data,
Such as the text data based on Chinese and English mixing.Event information with matches text data includes but is not limited to event to be identified
Type, time, central person.Optionally, it pre-establishes and safeguards the data list for storing event information；The data list
Such as can be antistop list for storing event type to be identified, it should if the antistop list is entertainment event keyword table
Antistop list covers typical event type and related saying in entertainment industry；The data list for example can be for storing
Entertainment star's table of the central person of entertainment media event, entertainment star Biao Biao cover mutually speaking on somebody's behalf for entertainment industry star
Method.By establishing and safeguard that tables of data arranges, help to identify entity (such as people from center of event to be identified in network data in advance
Object), it helps domain-specific knowledge is introduced, and then improves the ability of identification specific area event.
When being multiple with the event information of one text Data Matching, in first matching result this article notebook data can with it is more
A event information matches, i.e., there are a variety of matching relationships between this article notebook data and event information.Optionally, one text number
According to matching respectively with multiple event informations, this multiple event information just matches knot with one text data composition multiple groups respectively
Fruit, it is subsequent to judge respectively between every group of first matching result text data and the event information of text Data Matching
Whether matching relationship is correct.
The embodiment of the present invention, which does not limit, obtains text data and the method with the event information of text Data Matching.
Illustratively, it is assumed that event to be identified is negative event, and text data is the network text segment comprising the negative event, with text
The matched event information of notebook data is the type of negative event and the central person of the negative event, and one kind of S101 is possible
In implementation, filtered out from network data by match search algorithm comprising negative based on antistop list and entertainment star's table
The network text segment of event, and the corresponding negative thing of the network text segment is found out from antistop list and entertainment star's table
The type of part and multiple central persons, the network text segment filtered out is corresponding with the network text segment negative respectively
The type of event and multiple central persons form multiple groups just matching result, as shown in table 1 below.
The first matching result of table 1
| Central person | Negative event type | Network text segment |
| A | Scandal | B indicates that A may be true about the scandal of XX |
| B | Scandal | B indicates that A may be true about the scandal of XX |
| XX | Scandal | B indicates that A may be true about the scandal of XX |
Negative event class in table 1, for consolidated network text fragments there are 3 kinds of matching relationships, in this 3 kinds of matching relationships
Type is identical but central person is different.
The present embodiments relate to match search algorithm include be not limited to towards amusement object entity identification algorithms, mould
Paste matched Keywords matching algorithm.Further, fuzzy matching is passed through based on antistop list and entertainment star's table in S101
Keywords matching algorithm finds out candidate negative event type from network data, and is calculated by the Entity recognition towards amusement object
Method identifies the central person of negative event, so that the central person of negative event and negative event type are carried out maximization group
It closes.
There are many ways to S102 is realized in the embodiment of the present invention, will illustrate S102 by taking one of method as an example below
Specific implementation process.
In the embodiment of the present invention, deep learning model includes but is not limited to Attention model and LSTM model.It is optional
, the structure of deep learning model is multilayered structure.In a kind of possible multilayered structure, Attention model is LSTM model
Next layer deep learning model, the output of LSTM model is the part input or complete of Attention model in this configuration
Portion's input.It will illustrate the implementation of S102 by taking Attention model and LSTM model as an example below, the specific steps are as follows:
Step 1: the just corresponding term vector of matching result is obtained.
Term vector include but is not limited to the corresponding text term vector of text data and the corresponding event information of event information to
Amount.The form of term vector can be real number value vector, for example the form of term vector can be the real vector that dimension is 100 dimensions.
Illustratively, the just corresponding one group of text term vector of matching result (i.e. " other than fans' expression "), the group are obtained in step 1
The natural language vocabulary in text data that text term vector includes with first matching result matches one by one, as shown in Figure 2.
Step 2: text term vector is encoded by LSTM model, so that text term vector is converted to hidden state.
In order to control long term state, the LSTM model in step 2 uses three control switch (Gate) Lai Shixian.Its
In, three control switches that LSTM model uses be respectively input gate (Input Gate), forget door (Forget Gate) with
And out gate (Output Gate)；Input gate is responsible for saving long term state, that is, is responsible for replication processes and inputs information；It is negative to forget door
Immediate status input as long term state by duty, that is, is responsible for safeguarding inputting information, out gate is responsible for using long term state as LSTM mould
The input of type；The basic unit of LSTM can be called cell (cell).The forward calculation formula of these three control switches is as follows:
it=σ (Wixt+Uiht-1+bi) (1)
ft=σ (Wfxt+Ufht-1+bf) (2)
ot=σ (Woxt+Uoht-1+bo) (3)
Wherein, t moment is current time, and the t-1 moment is last moment, itIndicate the value of input gate, ftIt indicates to forget door
Value, otIndicate the value of out gate.W indicates the weight parameter being applied on x, and x indicates input value, and U expression is applied on h
Weight parameter, b indicate offset, ht-1Indicate the hidden state inscribed when t-1.
Further, Input transformation formula and hidden state htMore new formula it is as follows:
ct=ftOct-1+itOtanh(Wcxt+Ucht-1+bc) (4)
ht=otOtanh(cc) (5)
Wherein, ctIndicate cell state, htIndicate the hidden state of t moment.
In step 2, text term vector is input to LSTM model, text term vector turns by a time series coding
Be changed to multiple hidden states, wherein this multiple hidden state for the implicit multiple natural language vocabularies indicated in text data and this
Position of multiple natural language vocabularies in text data, i.e. this multiple hidden state are multiple natural language vocabularies in text data
Vectorization indicate.Illustratively, one group of text term vector is input to LSTM model, this group of text term vector warp in step 2
Crossing a time series code conversion is one group of hidden state (h1, h2..., hn-1, hn), as shown in Figure 2.
Step 3: hidden state and input of the event information vector as next layer deep learning model are determined.Wherein, next
Layer deep learning model can be Attention model.
Illustratively, by one group of hidden state (h in step 31, h2..., hn-1, hn), the event for indicating star
Information vector and input of the event information vector as Attention model for indicating event to be identified, such as Fig. 2 institute
Show.
First matching result can be converted to the input of next layer deep learning model by step 1 to three, this facilitates
Just matching result is implicitly introduced to deep learning model.
The standard of the matching relationship between text data and event information is used to indicate using the output of deep learning model in S102
The fiducial probability of exactness, specific steps are as follows:
Step 4: hidden state is weighted and averaged by Attention model.
It is that difference is hidden in multiple hidden states based on Attention model when the quantity of hidden state is multiple in step 4
State configures different weights, and is weighted and averaged to multiple hidden states after configuration weight.It illustratively, will be defeated in step 4
Hidden state (the h of one group entered1, h2... ..., hn-1, hn), the event information vector for indicating star and for indicating to be identified
The event information vector of event is weighted and averaged by Attention model, the calculated result (α after being weighted and averaged1,
α2..., αn), as shown in Figure 2.
In the embodiment of the present invention, attention mechanism is used to calculate the weight of hidden state and is weighted and averaged, so as to mould
Different attention levels is given to each vocabulary when quasi- reading text.In a kind of possible implementation, Attention model
Reckoning formula is as follows:
ut=tanh (Ww[ht；vstar；vevent]+bw) (6)
(6) formula is used to calculate the weight of the natural language vocabulary of each position in the weight and text data of hidden state, (7)
Formula is normalized for weight, and (8) formula is used for according to the weight after normalization to the natural language word of position each in text data
Corresponding hidden vector is weighted and averaged, to obtain the mathematical notation of text data.Wherein, the v of (6) formulastar；veventTable respectively
Show the central person of event to be identified and the text term vector for indicating event to be identified；(6) the hidden state h of formula and (8) formulat
With state h hidden in (5) formulatIt is identical.
Step 5: based on the matching relationship between the hidden state after weighted average and event information vector, using full connection
Layer output fiducial probability.
In step 5, a kind of possible implementation is, by after weighted average hidden state and event information vector input
The full articulamentum (dense layer) of deep learning model carries out linear transformation, is normalized to " just after softmax is calculated
Really "/" mistake " probability distribution, and export the probability distribution.It is used to indicate here by the probability distribution that normalization obtains
The fiducial probability of matching relationship accuracy between text data and event information.It illustratively, will weighted average in step 5
Calculated result (α afterwards1, α2... ..., αn) input deep learning model full articulamentum (dense layer) linearly become
It changes, the probability distribution of " correct "/" mistake " is normalized to after softmax is calculated, as shown in Fig. 2.
It is realized step 1 to five and is participated in a manner of Attention model calculates by term vector come to deep learning mould
Type implicitly input just matching result, i.e., to the implicit input text data of deep learning model and with the event of matches text data
Information, to obtain the fiducial probability for being used to indicate matching relationship accuracy in first matching result, so as in S103 according to screening
Matching relationship out completes event recognition.
Especially needed understanding, this input side that just matching result is implicitly inputted to deep learning model in S102
Formula can point out the position that event information is in text data indirectly, help avoid event information and directly participate in matching
In the judgement of relationship, to facilitate over-fitting caused by the position for preventing from being in text data by direct instruction event information
The generation of phenomenon.
In S103, judge that the confidence of the matching relationship between the text data and event information of deep learning model output is general
Whether rate meets preset condition, and wherein preset condition can be auxiliary judgement criterion, which is used for auxiliary judgment
The correctness of the matching relationship；If it is determined that fiducial probability meets preset auxiliary judgement criterion, then illustrate that the matching relationship is correct,
Namely the matching relationship event to be identified that matching result includes corresponding just exists, it can be according to the matching relationship in the case of this
Determine the corresponding event to be identified of the matching relationship.It optionally, is that first matching result marks " just according to the fiducial probability of output
Really "/" mistake " tag along sort.It should be noted that not limited in addition to auxiliary judgement criterion, in the embodiment of the present invention pre-
If condition is other conditions, such as preset condition can be preset fiducial probability threshold value.
By event recognition method shown in fig. 1, available text data and the text comprising event to be identified
The event information of Data Matching as first matching result, and by first matching result input deep learning model obtain text data with
The fiducial probability of matching relationship between event information, thus filter out fiducial probability meet the matching relationship of preset condition to
Determine event to be identified.Technical solution provided by the invention screens first matching result by introducing deep learning model,
The data that event recognition scheme is marked to the domain knowledge of artificial settings and manually are got rid of, to help to improve event recognition
The portability of scheme expands the application scenarios of event recognition scheme.Also, technical solution provided by the invention makes full use of can
Extensive domain knowledge combination attention mechanism, to help avoid directly using excessively quasi- caused by deep learning model
Phenomenon is closed, mitigates the processing pressure of deep learning model, promotes the reliability, real-time and recognition effect of event recognition.
Exemplary means
After describing the method for exemplary embodiment of the invention, next, introducing, the present invention provides exemplary
The device of implementation.
With reference to Fig. 3, the present invention provides a kind of event recognition devices, and it is corresponding which may be implemented Fig. 1
To the method in exemplary embodiment of the invention.As shown in fig.3, the event recognition device includes: matching unit, judgement list
Member and recognition unit.Wherein
Matching unit, for from network data obtain text data and with the event information conduct of matches text data
First matching result, text data include event to be identified；
Judging unit is exported for the input using first matching result as deep learning model using deep learning model
It is used to indicate the fiducial probability of the matching relationship accuracy between text data and event information；
Recognition unit determines event to be identified according to the matching relationship if meeting preset condition for fiducial probability.
Optionally, judging unit is specifically used in the input using first matching result as deep learning model: obtaining just
The corresponding term vector of matching result, term vector include the corresponding text term vector of text data and the corresponding event letter of event information
Cease vector；Text term vector is encoded by time recurrent neural network LSTM model, so that text term vector be converted
For hidden state；Determine hidden state and input of the event information vector as next layer deep learning model.
Optionally, next layer deep learning model is Attention model；Judging unit is defeated using deep learning model
When being used to indicate the fiducial probability of the matching relationship accuracy between text data and event information out, it is specifically used for: passes through
Attention model is weighted and averaged hidden state；Based between the hidden state after weighted average and event information vector
Matching relationship exports fiducial probability using full articulamentum.
Optionally, judging unit is specifically used for when being weighted and averaged by Attention model to hidden state: when
It is the different weights of the hidden states configuration of difference in multiple hidden states based on Attention model when the quantity of hidden state is multiple, and
Multiple hidden states after configuration weight are weighted and averaged.
Optionally, event information includes the keyword of event to be identified and the central person of event to be identified.
Exemplary media
After describing the method and apparatus of exemplary embodiment of the invention, next, the present invention mentions with reference to Fig. 4
A kind of exemplary media is supplied, which there are computer executable instructions, which can be used for making institute
It states computer and executes method described in any one of corresponding exemplary embodiment of the invention of Fig. 1.
Exemplary computer device
After method, medium and the device for describing exemplary embodiment of the invention, next, being introduced with reference to Fig. 5
A kind of exemplary computer device 50 provided by the invention, the calculating equipment 50 include processing unit 501, memory 502, bus
503, external equipment 504, I/O interface 505 and network adapter 506, the memory 502 include random access memory
(random access memory, RAM) 5021, cache memory 5022, read-only memory (Read-Only
Memory, ROM) 5023 and at least memory cell array 5025 that constitutes of a piece of storage unit 5024.The wherein memory
502, the program or instruction executed for storage processing unit 501；The processing unit 501, for being stored according to the memory 502
Program or instruction, execute method described in any one of corresponding exemplary embodiment of the invention of Fig. 1；The I/O interface
505, for receiving or sending data under the control of the processing unit 501.
Here, the exemplary computer device 50 its include but is not limited to user equipment, the network equipment or the network equipment with
User equipment is integrated constituted equipment by network；The user equipment includes but is not limited to that any one can be logical with user
Cross keyboard, remote controler, touch tablet or voice-operated device carry out the electronic product of human-computer interaction, such as computer, smart phone, common
Mobile phone, tablet computer etc.；The network equipment includes but is not limited to computer, network host, single network server, multiple nets
The cloud that network server set or multiple servers are constituted.
It should be noted that although being referred to several units/modules or subelement/module of device in the above detailed description,
But it is this division be only exemplary it is not enforceable.In fact, embodiment according to the present invention, above-described
The feature and function of two or more units/modules can embody in a units/modules.Conversely, above-described one
The feature and function of a units/modules can be to be embodied by multiple units/modules with further division.
In addition, although describing the operation of the method for the present invention in the accompanying drawings with particular order, this do not require that or
Hint must execute these operations in this particular order, or have to carry out shown in whole operation be just able to achieve it is desired
As a result.Additionally or alternatively, it is convenient to omit multiple steps are merged into a step and executed by certain steps, and/or by one
Step is decomposed into execution of multiple steps.
Although detailed description of the preferred embodimentsthe spirit and principles of the present invention are described by reference to several, it should be appreciated that, this
It is not limited to the specific embodiments disclosed for invention, does not also mean that the feature in these aspects cannot to the division of various aspects
Combination is benefited to carry out, this to divide the convenience merely to statement.The present invention is directed to cover appended claims spirit and
Included various modifications and equivalent arrangements in range.
Claims (12)
1. a kind of event recognition method characterized by comprising
Text data is obtained from network data and with the event information of the matches text data as first matching result, institute
Stating text data includes event to be identified；
Using the just matching result as the input of deep learning model, institute is used to indicate using deep learning model output
State the fiducial probability of the matching relationship accuracy between text data and the event information；
If the fiducial probability meets preset condition, the event to be identified is determined according to the matching relationship.
2. the method as described in claim 1, which is characterized in that as follows using the just matching result as depth
The input for practising model, specifically includes:
The just corresponding term vector of matching result is obtained, the term vector includes the corresponding text term vector of the text data
Event information vector corresponding with the event information；
The text term vector is encoded by time recurrent neural network LSTM model, thus by the text term vector
Be converted to hidden state；
Determine the hidden state and input of the event information vector as next layer deep learning model.
3. method according to claim 2, which is characterized in that next layer deep learning model is attention mechanism
Attention model；
It is described that matching between the text data and the event information is used to indicate using deep learning model output
The fiducial probability of relationship accuracy, specifically includes:
The hidden state is weighted and averaged by Attention model；
It is defeated using full articulamentum based on the hidden state after weighted average and the matching relationship between the event information vector
The fiducial probability out.
4. method as claimed in claim 3, which is characterized in that described to be carried out by Attention model to the hidden state
Weighted average, comprising:
It is different hidden state configurations in multiple hidden states based on Attention model when the quantity of the hidden state is multiple
Different weights, and the multiple hidden state after configuration weight is weighted and averaged.
5. the method as described in claim 2 to 4 is any, which is characterized in that the event information includes the event to be identified
Keyword and the event to be identified central person.
6. a kind of event recognition device characterized by comprising
Matching unit, for from network data obtain text data and with the event information conduct of the matches text data
First matching result, the text data include event to be identified；
Judging unit, for the input using the just matching result as deep learning model, using the deep learning model
Output is used to indicate the fiducial probability of the matching relationship accuracy between the text data and the event information；
Recognition unit determines the event to be identified according to the matching relationship if meeting preset condition for fiducial probability.
7. device as claimed in claim 6, which is characterized in that the judging unit is using the just matching result as depth
When the input of learning model, it is specifically used for:
The just corresponding term vector of matching result is obtained, the term vector includes the corresponding text term vector of the text data
Event information vector corresponding with the event information；
The text term vector is encoded by time recurrent neural network LSTM model, thus by the text term vector
Be converted to hidden state；
Determine the hidden state and input of the event information vector as next layer deep learning model.
8. device as claimed in claim 7, which is characterized in that next layer deep learning model is attention mechanism
Attention model；
The judging unit is being used to indicate the text data and the event information using deep learning model output
Between matching relationship accuracy fiducial probability when, be specifically used for:
The hidden state is weighted and averaged by Attention model；
It is defeated using full articulamentum based on the hidden state after weighted average and the matching relationship between the event information vector
The fiducial probability out.
9. device as claimed in claim 8, which is characterized in that the judging unit is passing through Attention model to described
When hidden state is weighted and averaged, it is specifically used for:
It is different hidden state configurations in multiple hidden states based on Attention model when the quantity of the hidden state is multiple
Different weights, and the multiple hidden state after configuration weight is weighted and averaged.
10. the device as described in claim 7 to 9 is any, which is characterized in that the event information includes the event to be identified
Keyword and the event to be identified central person.
11. a kind of calculating equipment, which is characterized in that including processor, memory and transceiver；
The memory, the program executed for storing the processor；
The processor, the program for being stored according to the memory, perform claim require 1 to 5 described in any item methods；
The transceiver, for receiving or sending data under the control of the processor.
12. a kind of medium, which is characterized in that the media storage has computer executable instructions, and the computer is executable to be referred to
It enables for making the computer perform claim require 1 to 5 described in any item methods.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811472932.0A CN109614541A (en) | 2018-12-04 | 2018-12-04 | A kind of event recognition method, medium, device and calculate equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811472932.0A CN109614541A (en) | 2018-12-04 | 2018-12-04 | A kind of event recognition method, medium, device and calculate equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109614541A true CN109614541A (en) | 2019-04-12 |
Family
ID=66007011
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811472932.0A Pending CN109614541A (en) | 2018-12-04 | 2018-12-04 | A kind of event recognition method, medium, device and calculate equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109614541A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110084373A (en) * | 2019-04-22 | 2019-08-02 | 腾讯科技（深圳）有限公司 | Information processing method, device, computer readable storage medium and computer equipment |
| CN113392213A (en) * | 2021-04-19 | 2021-09-14 | 合肥讯飞数码科技有限公司 | Event extraction method, electronic device and storage device |
| CN118070753A (en) * | 2024-04-17 | 2024-05-24 | 国网江苏省电力有限公司苏州供电分公司 | GIM model and entity ID matching method and system based on deep learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107622333A (en) * | 2017-11-02 | 2018-01-23 | 北京百分点信息科技有限公司 | A kind of event prediction method, apparatus and system |
| CN108446275A (en) * | 2018-03-21 | 2018-08-24 | 北京理工大学 | Long text emotional orientation analytical method based on attention bilayer LSTM |
-
2018
- 2018-12-04 CN CN201811472932.0A patent/CN109614541A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107622333A (en) * | 2017-11-02 | 2018-01-23 | 北京百分点信息科技有限公司 | A kind of event prediction method, apparatus and system |
| CN108446275A (en) * | 2018-03-21 | 2018-08-24 | 北京理工大学 | Long text emotional orientation analytical method based on attention bilayer LSTM |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110084373A (en) * | 2019-04-22 | 2019-08-02 | 腾讯科技（深圳）有限公司 | Information processing method, device, computer readable storage medium and computer equipment |
| CN110084373B (en) * | 2019-04-22 | 2021-08-24 | 腾讯科技（深圳）有限公司 | Information processing method, information processing device, computer-readable storage medium and computer equipment |
| CN113392213A (en) * | 2021-04-19 | 2021-09-14 | 合肥讯飞数码科技有限公司 | Event extraction method, electronic device and storage device |
| CN113392213B (en) * | 2021-04-19 | 2024-05-31 | 合肥讯飞数码科技有限公司 | Event extraction method, electronic equipment and storage device |
| CN118070753A (en) * | 2024-04-17 | 2024-05-24 | 国网江苏省电力有限公司苏州供电分公司 | GIM model and entity ID matching method and system based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Top-down tree long short-term memory networks | |
| Yang et al. | Cross-modal relationship inference for grounding referring expressions | |
| CN107992597B (en) | Text structuring method for power grid fault case | |
| Xiang et al. | A convolutional neural network-based linguistic steganalysis for synonym substitution steganography | |
| CN111143576A (en) | Event-oriented dynamic knowledge graph construction method and device | |
| CN110209806A (en) | File classification method, document sorting apparatus and computer readable storage medium | |
| CN108984530A (en) | A kind of detection method and detection system of network sensitive content | |
| CN107330011A (en) | The recognition methods of the name entity of many strategy fusions and device | |
| CN108255805A (en) | The analysis of public opinion method and device, storage medium, electronic equipment | |
| Marquez et al. | A machine learning approach to POS tagging | |
| CN109086269B (en) | Semantic bilingual recognition method based on semantic resource word representation and collocation relationship | |
| CN111460820A (en) | Network space security domain named entity recognition method and device based on pre-training model BERT | |
| Kim et al. | Multi-attention multimodal sentiment analysis | |
| CN110502742B (en) | Complex entity extraction method, device, medium and system | |
| CN111694937A (en) | Interviewing method and device based on artificial intelligence, computer equipment and storage medium | |
| CN109614541A (en) | A kind of event recognition method, medium, device and calculate equipment | |
| Sartakhti et al. | Persian language model based on BiLSTM model on COVID-19 corpus | |
| US20200364301A1 (en) | Affect-enriched vector representation of words for use in machine-learning models | |
| Zhang et al. | Making the relation matters: Relation of relation learning network for sentence semantic matching | |
| Sun et al. | Multi-channel CNN based inner-attention for compound sentence relation classification | |
| CN112395421A (en) | Course label generation method and device, computer equipment and medium | |
| Li et al. | Image describing based on bidirectional LSTM and improved sequence sampling | |
| Han et al. | An attention-based neural framework for uncertainty identification on social media texts | |
| Li et al. | MIA-Net: Multi-modal interactive attention network for multi-modal affective analysis | |
| CN116341519A (en) | Event causal relation extraction method, device and storage medium based on background knowledge |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190412 |
|
| WD01 | Invention patent application deemed withdrawn after publication |