CN108124065A - A kind of method junk call content being identified with disposal - Google Patents
A kind of method junk call content being identified with disposal Download PDFInfo
- Publication number
- CN108124065A CN108124065A CN201711266910.4A CN201711266910A CN108124065A CN 108124065 A CN108124065 A CN 108124065A CN 201711266910 A CN201711266910 A CN 201711266910A CN 108124065 A CN108124065 A CN 108124065A
- Authority
- CN
- China
- Prior art keywords
- disposal
- identified
- call
- word
- junk call
- 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04M—TELEPHONIC COMMUNICATION
- H04M3/00—Automatic or semi-automatic exchanges
- H04M3/42—Systems providing special services or facilities to subscribers
- H04M3/436—Arrangements for screening incoming calls, i.e. evaluating the characteristics of a call before deciding whether to answer it
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/26—Speech to text systems
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique
- G10L25/30—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique using neural networks
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
- G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
Abstract
The invention discloses a kind of methods junk call content being identified with disposal.Specific steps of the present invention include S1, acquisition dialog context;S2, the dialog context is converted into text message;S3, using stammering, participle instrument does word segmentation processing to the text message;S4, phone examination model is obtained according to LSTM algorithms and DNN algorithms by the word of the word segmentation processing;S5, the telephone class obtained according to softmax graders, output by phone examination model;S6, call of the softmax graders output for junk call is interrupted.Present invention identification and the method for disposal junk call, can analyze the content in call in real time and block instantly, realize the real-time blocking to junk call.
Description
Technical field
The invention belongs to technical field of telecommunications more particularly to it is a kind of junk call content is identified with disposal
Method.
Background technology
With the continuous development of the worldwide communication technology, people are increasing to the dependence of mobile communication.It is mobile
The rapid development of communication is while bringing convenient, but also some are declared for commercial object using mobile communication
It passes advertisement, promote the sale of products or telecommunication fraud, should in time be found with control method by effectively monitoring and filter interception.
Therefore junk call is automatically identified, and junk call is interrupted in time, protect the life of people and property peace
Entirely, it is purpose of the present invention place.
As disclosed in the patent of invention of Publication No. CN103731832A a kind of anti-phone, short message fraud system and
The audio file content of upload can be converted into text file, changed by method, sound identification module, the sound identification module
Text file can be analyzed and processed by intelligent analysis module, intelligent analysis module passes through natural language processing pair first
Above- mentioned information is analyzed, and extracts keyword message, such as telephone number, account No., prize-winning content are identified, and with
Information in fraud information database is matched, and the fraud information of various word classes is mainly stored in fraud information database;
Audio fingerprint database, for having collected the audio-frequency fingerprint information of various swindle recording.The scheme of the invention is put forth effort on will be in call
Appearance is matched with database information, and fraud information is regarded as if successful match, but the style for nowadays swindling molecule is more, the hair
Database in bright scheme only preserves swindle molecule used fraud information, cause the information that the database preserves compared with
To be narrow, it is impossible to more comprehensively defend various harassing calls.
In view of in above-mentioned existing technology anti-telephone fraud scheme there are the defects of, the present inventor, which is based on being engaged in such product, to be set
Meter manufacture abundant for many years practical experience and professional knowledge, and coordinate the utilization of scientific principle, actively it is subject to research and innovation, to create
If a kind of method junk call content being identified with disposal, can improve it is general it is existing to junk call content into
The method of row identification and disposal, makes it have more practicability.By constantly studying, designing, and by study repeatedly sample and
After improvement, the present invention having practical value is created finally.
The content of the invention
For above-mentioned technical problem existing in the prior art, junk call content is known the present invention provides a kind of
Not with the method for disposal, the voice content of call is analyzed the present invention is based on depth learning technology, so as to identify rubbish electricity
Words, and junk call is interrupted in time using the technology of taking out stitches of gateway exchange, analysis intercepting system in real time is built, it is existing to solve
There is technology to lack and find the problem of junk call variation characteristic and improve the energy accurately identified to junk call in time
Power.
To reach above-mentioned technical purpose, the present invention adopts the following technical scheme that:
The method with disposal is identified in a kind of content to junk call, including:
S1, acquisition dialog context;
S2, the dialog context is converted into text message;
S3, using stammering, participle instrument does word segmentation processing to the text message;
S4, phone examination model is obtained according to LSTM algorithms and DNN algorithms by the word of the word segmentation processing;
S5, the telephone class obtained according to softmax graders, output by phone examination model;
S6, call of the softmax graders output for junk call is interrupted.
As a preference of the present invention, it is further included between step S2 and S3:The text envelope is removed using the method for regularization
Non-textual portions in breath.
As a preference of the present invention, prepending non-significant word, if the word is matched with the invalid word, deletes institute's predicate
Language.
As a preference of the present invention, the word is converted by term vector using Word2Vec methods, using LSTM algorithms
The term vector is converted into sentence vector, obtaining the phone using the input vector of the sentence vector as DNN algorithms screens mould
Type.
As a preference of the present invention, the text message is divided into training sample and test sample, with the training sample
It predicts that the phone screens model, verifies that the phone screens model with the test sample.
As a preference of the present invention, model tuning is screened to the phone using the self-learning capability of neutral net.
As a preference of the present invention, adding in activation primitive in the neutral net, and weights are adjusted to stablizing constant or reach
To specified threshold.
As a preference of the present invention, the softmax graders classify to phone type, rubbish electricity is specifically divided into
Words and normal telephone.
As a preference of the present invention, the dialog context is that the voice generated in communication process is gathered using recording device
Data.
Technical solution provided by the invention can include the following benefits:
1st, the present invention is preset with invalid word, removes normal vocabulary, only selects key vocabulary and carries out examination, shortens
The ability accurately identified to junk call is improved while recognition time;2nd, with reference to stammerer participle instrument and
Word2Vec methods process the text message of call, and the word to screen model judgement by phone is prepared;3rd, this hair
It is bright that phone examination model is established according to LSTM algorithms and DNN algorithms, examination is carried out to phone automatically;4th, neutral net is passed through
Self-learning capability screens model to obtained preliminary phone and carries out tuning, if required precision is also once not achieved after training,
Can be with repetition training, until meeting training precision, the accuracy that the present invention judges junk call is high;It is if the 5th, described
Softmax graders output phone be junk call, can in real time to phone carry out interrupt operation, the present invention call when
Can judge whether be junk call in short time, and recording processing can just need not be judged to lead to after hanging up the telephone
The classification of words realizes and automatically identifies junk call, and to purpose that junk call interrupts in time.
Description of the drawings
Fig. 1 is the operating diagram of the present invention;
Fig. 2 is deep learning algorithm schematic diagram of the present invention;
Fig. 3 is Word2Vec algorithm arrangements schematic diagram of the present invention;
Fig. 4 is LSTM algorithm arrangements schematic diagram of the present invention;
Fig. 5 is 5 operating diagram of the embodiment of the present invention;
Fig. 6 is 6 operating diagram of the embodiment of the present invention.
Specific embodiment
Below in conjunction with the attached drawing in the embodiment of the present invention, the technical solution in the embodiment of the present invention is carried out clear, complete
Site preparation describes, it is clear that described embodiment is part of the embodiment of the present invention, instead of all the embodiments.Based on this hair
Embodiment in bright, the every other reality that those of ordinary skill in the art are obtained without making creative work
Example is applied, belongs to the scope of protection of the invention.
Embodiment 1:
Step A1:Dialog context is gathered using recording device, speech recognition equipment identifies the recording data, and this is recorded
Sound data conversion is into text data;
Step A2:Non-textual portions in text data is removed using regularization method;
Step A3:By sample according to 3:1 ratio cut partition is training sample and test sample;
Step A4:Text segments:Word segmentation processing is done to short message text using stammerer participle instrument;
Step A5:Prepending non-significant word removes the invalid matched word of word in text;
Step A6:The word segmented is converted by term vector using Word2Vec technologies, word is done at vectorization
Reason;
Step A7:Term vector is converted by sentence vector using LSTM algorithms;
Step A8:Using sentence vector as the input vector of DNN disaggregated models;
Step A9:The result of probability value maximum is chosen as output classification;
Step A10:Counting loss function carries out backpropagation;
Step A11:Weighed value adjusting, until weights are stablized constant or reach specified threshold;
Step A12:Model prediction and assessment, by test sample input model, the accuracy rate of computation model, recall rate, F
Value;
Step A13:Model tuning, using neutral net self-learning capability to model carry out tuning, reduce actual value with
Gap between ideal value;
Step A14:Telephone class exports, and finally junk call is known according to the output result of softmax graders
Not.
The result of softmax graders output is junk call or normal telephone, if the knot of softmax graders output
Fruit is junk call, then the call is interrupted in time.
Embodiment 2
As shown in Fig. 2, the present embodiment is with " I is public security bureau, your account is accused of washing dirty money, you is asked to coordinate investigation "
Example, design of the invention is the hybrid production style based on LSTM and DNN.
In fig. 2, model is divided into 3 layers, first layer by using Word2Vec by the word in text be converted to word to
Amount;The second layer is LSTM layers, and term vector caused by first layer is input to LSTM layers, using LSTM algorithm structures, before calculating
Influence of the word to current word afterwards, most each individual term vector is converted into sentence vector at last;Third layer is DNN layers, and the second layer is given birth to
Into sentence vector as input layer, by hidden layer, using softmax activation primitives, telephone class is exported in output layer.
Embodiment 3
As shown in figure 3, the present embodiment specifically introduces the effect of Word2Vec algorithms in the present invention.
The problem of natural language understanding, will be converted into the problem of machine can be handled, and the first step must be by these symbolic numbers
The expression of text is mapped in the vector space of k dimensions by word.Word2Vec algorithms will be in the corpus that segmented
Cliction language is converted to term vector, and the term vector trained by Word2Vec is as follows:
vi=(a0,a1,L,ad) (1)
(1) in formula, d is the dimension of term vector.
Word2Vec algorithms implement process:
Step A61:Keyword in phone text feature storehouse is counted, it is assumed that have m keyword;
Step A62:One word first with one-hot-vector is converted into the vector x of a n dimension, is with " arrearage "
Example:
" arrearage " → [0,0,0,0,1..., 0,0]
Step A63:There is m neuron in hidden layer, it is known that input layer is a n-dimensional vector and is connected entirely with hidden layer,
So it needs in the hidden neuron that the weight matrix w of a n*m is 1*m the DUAL PROBLEMS OF VECTOR MAPPING that n is tieed up to latitude;
Step A64:Also with full connection from hidden layer to output layer, added in when output unit is calculated
Softmax graders, can be in the hope of final vectorial w by back transfer, can by being multiplied i.e. x*w with initial term vector
In the hope of the vectorial W (i) of final term vector, that is, 1*m;
X*w=W (i)=[Wi1 Wi2… Wim]
Step A65:The corresponding term vector of rubbish keyword of each appearance of taking on the telephone is added, obtains belonging to this
The text vector d to take on the telephone.
Embodiment 4
As shown in figure 4, the present embodiment specifically introduces the effect of LSTM algorithms in the present invention.
LSTM is a kind of special RNN (Recurrent Neural Network), and a LSTM unit is by a cell
With three doors (forgeing door, input gate, out gate) composition.This special structure is exactly based on, which LSTM could select believe
Breath passes into silence, which information is remembered.Certain moment each component of t, LSTM unit does following update:
ft=σ (Wfht-1+Ufxt+bf)
it=σ (Wiht-1+Uixt+bi)
at=tanh (Waht-1+Uaxt+ba)
Ct=Ct-1e ft+it e at
ot=σ (Woht-1+Uoxt+bo)
ht=ot e tanh(Ct)
Wherein, σ represents sigmoid activation primitives, and e accumulates for Hadamard, xtFor the input vector of t moment, htIt is hiding
State, Uf,Ui,Ua,UoRespectively xtThe not weight matrix of fellow disciple, and Wf,Wi,Wa,WoFor htThe not weight matrix of fellow disciple, bf,bi,
ba,boFor the biasing of each door, ft,it,Ct,otIt represents respectively and forgets door, input gate, mnemon state and out gate.
LSTM layers of input is Word2Vec layers of output term vector, and Word2Vec layers of an output is corresponding to one
The LSTM inputs of moment t.The input layer of DNN disaggregated models is sent into LSTM layers of output, is activated in output layer using softmax
Function calculates phone and belongs to probability of all categories, maximum probability is selected to be exported as telephone class.
Phone text term vector is converted into sentence vector and is as follows by LSTM algorithms:
Step A71:The term vector for taking on the telephone text by one arranges in sequence, it is assumed that one takes on the telephone by m term vector structure
Into i.e. x1,x2,L,xm;
Step A72:Initialization model parameter Wf,Uf,bf,Wa,Ua,Wi,Ui,bi,Wo,Uo,bo;
Step A73:By x1It is incoming to forget door f2, f2=σ (W2h1+U2x2+b2), door weights W is forgotten in update2,U2,b2;
Step A74:Update input gate parameter, it=σ (Wiht-1+Uixt+bi), at=tanh (Waht-1+Uaxt+ba),
Wherein Wi,Ui,bi,Wa,Ua,baFor the coefficient and bias of linear relationship, σ is sigmoid activation primitives;
Step A75:Model output state updates, Ct=Ct-1e ft+it e at, wherein e is Hadamard products;
Step A76:Update out gate parameter, ot=σ (Woht-1+Uoxt+bo), ht=ot e tanh(Ct), output is current
Sequence index predicted value:
Step A77:Step A72 to A76 is repeated, finally exports predicted value
Embodiment 5
As shown in figure 5, the present embodiment specifically introduces the effect of DNN disaggregated models in the present invention.
The basic structure of DNN models includes input layer, several hidden layers and output layer.
DNN models using Softmax graders, classify to phone type in output layer, be divided into junk call or
Normal telephone.Softmax formula are as follows:
Wherein P (y=i | x, θ) it is the probability that sample x belongs to i classes.
The process of DNN neural network algorithms can be divided into two stages:First stage is successively calculated by input layer
Each layer neuron is output and input, until output layer.Second stage is that each layer nerve is successively calculated by output layer
The output error of member, and principle is declined to adjust the connection weight of each layer and Node B threshold according to error gradient, make amended
The final output of network can be close to desired value.It, can be with repetition training, directly if required precision is also once not achieved after training
Until training precision is met.
Embodiment 6
As shown in fig. 6, simply introduce the network weight regulation mechanism of the application use
If input vector X=x1, x2..., xm)T, be each evaluation index value, hidden layer output vector h=h1, h2...,
hL)T, y is the reality output of network, is effect assessment value.The weights of input layer i to hidden layer node j are Wij, it is hidden
Weights containing node layer to output node layer are Vj, θjWithThe threshold value of hidden layer and output layer is represented respectively.Then
Wherein f (x) is activation primitive, and activation primitive is chosen to be sigmoid functions here, i.e.,
Sigmoid functions are by variable mappings to 0, between 1.
(1) calculating network reality output and the error of preferable output
In t moment, by the reality output y of networki(t) the target output d provided with samplei(t) it is compared, output production
Raw error εi(t) it is defined as follows:
εi(t)=di(t)-yi(t)
Control of the generated error signal drives to learning algorithm, the purpose is to the input weight of neuron into
A series of calibrations of row are adjusted, and the purpose for calibrating adjustment is by iteration step by step, makes output signal yi(t) become closer to
Target exports di(t), which can minimize to realize by cost function E (t).
(2) adjustment amount of calculating network weights
The adjustment amplitude of weight is
ΔWij(t)=η εi(t)xi(t)
ΔVj(t)=η εi(t)hj(t)
Wherein η is that a numerical value is positive constant, represents learning rate.
Weight after adjustment is
Wij(t+1)=α Wij(t)+ΔWij(t)
Vj(t+1)=α Vj(t)+ΔVj(t)
α be momentum item, Δ Wij(t) for by the weighed value adjusting amplitude of input layer to hidden layer, Δ Vj(t) it is by hidden layer
To the weighed value adjusting amplitude of output layer.
In the present invention, the overall of DNN neural network algorithms realizes that flow is as follows:
Step B1:Network connection weights, Node B threshold are initialized, including input layer to hidden layer, hidden layer to output layer
Connection weight and deviation;
Step B2:Taking a sample, each desired value of sample is the input value of network as input;
Step B3:Calculate hidden layer node output, export as the cumulative of each input layer and connection weight and, and band
Enter activation primitive to be exported;
Step B4:Calculate output node layer output, export as hidden layer node input and the cumulative of connection weight and, and
Activation primitive is brought into be exported;
Step B5:Calculate hidden layer and output layer error, error for model real output value and desired output it
Difference;
Step B6:Connection weight and Node B threshold are updated, by error back propagation, each connection weight is adjusted;
Step B7:Judge whether to take out whole samples, if not provided, return to step B2, is cycled, if so, performing
Following step;
Step B8:Whether error in judgement is less than specified threshold, if not, return to step B3, continues cycling through, if so, knot
Shu Xunhuan.
Those skilled in the art will readily occur to the disclosure after considering specification and putting into practice invention disclosed herein
Other embodiments.This application is intended to cover any variations, uses, or adaptations of the disclosure, these modifications, purposes
Or adaptive change follow the disclosure general principle and including the disclosure it is undocumented in the art it is known often
Knowledge or conventional techniques.Description and embodiments are considered only as illustratively, and the true scope and spirit of the disclosure are by following
Claim point out.
Claims (9)
1. the method with disposal is identified in a kind of content to junk call, which is characterized in that including:
S1, acquisition dialog context;
S2, the dialog context is converted into text message;
S3, using stammering, participle instrument does word segmentation processing to the text message;
S4, phone examination model is obtained according to LSTM algorithms and DNN algorithms by the word of the word segmentation processing;
S5, the telephone class obtained according to softmax graders, output by phone examination model;
S6, call of the softmax graders output for junk call is interrupted.
2. the method with disposal is identified in the content according to claim 1 to junk call, which is characterized in that step
It is further included between S2 and S3:Non-textual portions in the text message is removed using the method for regularization.
3. the method with disposal is identified in the content according to claim 1 to junk call, which is characterized in that default
Invalid word if the word is matched with the invalid word, deletes the word.
4. the method with disposal is identified in the content according to claim 1 to junk call, which is characterized in that uses
The word is converted into term vector by Word2Vec methods, the term vector is converted into sentence vector using LSTM algorithms, with institute
It states sentence vector and obtains the phone examination model as the input vector of DNN algorithms.
5. the method with disposal is identified in the content according to claim 2 to junk call, which is characterized in that by institute
It states text message and is divided into training sample and test sample, predict that the phone screens model with the training sample, with the survey
This verification of sample phone screens model.
6. the method with disposal is identified in the content according to claim 5 to junk call, which is characterized in that utilizes
The self-learning capability of neutral net screens model tuning to the phone.
7. the method with disposal is identified in the content according to claim 6 to junk call, which is characterized in that in institute
It states neutral net and adds in activation primitive, and adjust weights to stablizing constant or reach specified threshold.
8. the method with disposal is identified in the content according to claim 1 to junk call, which is characterized in that described
Softmax graders classify to phone type, are specifically divided into junk call and normal telephone.
9. the method with disposal is identified in the content according to claim 1 to junk call, which is characterized in that described
Dialog context is that the voice data generated in communication process is gathered using recording device.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711266910.4A CN108124065A (en) | 2017-12-05 | 2017-12-05 | A kind of method junk call content being identified with disposal |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711266910.4A CN108124065A (en) | 2017-12-05 | 2017-12-05 | A kind of method junk call content being identified with disposal |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108124065A true CN108124065A (en) | 2018-06-05 |
Family
ID=62228925
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711266910.4A Pending CN108124065A (en) | 2017-12-05 | 2017-12-05 | A kind of method junk call content being identified with disposal |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108124065A (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108924333A (en) * | 2018-06-12 | 2018-11-30 | 阿里巴巴集团控股有限公司 | Fraudulent call recognition methods, device and system |
CN109377983A (en) * | 2018-10-18 | 2019-02-22 | 深圳壹账通智能科技有限公司 | A kind of harassing call hold-up interception method and relevant device based on interactive voice |
CN109410496A (en) * | 2018-10-25 | 2019-03-01 | 北京交通大学 | Attack early warning method, apparatus and electronic equipment |
CN109658939A (en) * | 2019-01-26 | 2019-04-19 | 北京灵伴即时智能科技有限公司 | A kind of telephonograph access failure reason recognition methods |
CN109688275A (en) * | 2018-12-27 | 2019-04-26 | 中国联合网络通信集团有限公司 | Harassing call recognition methods, device and storage medium |
CN109905282A (en) * | 2019-04-09 | 2019-06-18 | 国家计算机网络与信息安全管理中心 | Fraudulent call prediction technique and forecasting system based on LSTM |
CN110334110A (en) * | 2019-05-28 | 2019-10-15 | 平安科技(深圳)有限公司 | Natural language classification method, device, computer equipment and storage medium |
CN110525821A (en) * | 2019-08-20 | 2019-12-03 | 北京精英系统科技有限公司 | A kind of system promoting separate waste collection site operation efficiency |
CN110929506A (en) * | 2019-12-04 | 2020-03-27 | 杭州安恒信息技术股份有限公司 | Junk information detection method, device and equipment and readable storage medium |
CN112193959A (en) * | 2020-09-25 | 2021-01-08 | 浙江新再灵科技股份有限公司 | Method and system for detecting abnormal sound of elevator |
CN112478975A (en) * | 2020-12-09 | 2021-03-12 | 浙江新再灵科技股份有限公司 | Elevator door fault detection method based on audio features |
CN112784038A (en) * | 2019-10-23 | 2021-05-11 | 阿里巴巴集团控股有限公司 | Information identification method, system, computing device and storage medium |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140278402A1 (en) * | 2013-03-14 | 2014-09-18 | Kent S. Charugundla | Automatic Channel Selective Transcription Engine |
US20150019226A1 (en) * | 1999-06-10 | 2015-01-15 | West View Research, Llc | Computerized information apparatus |
CN106504768A (en) * | 2016-10-21 | 2017-03-15 | 百度在线网络技术(北京)有限公司 | Phone testing audio frequency classification method and device based on artificial intelligence |
CN106599933A (en) * | 2016-12-26 | 2017-04-26 | 哈尔滨工业大学 | Text emotion classification method based on the joint deep learning model |
CN107222865A (en) * | 2017-04-28 | 2017-09-29 | 北京大学 | The communication swindle real-time detection method and system recognized based on suspicious actions |
CN107306306A (en) * | 2016-04-25 | 2017-10-31 | 腾讯科技(深圳)有限公司 | Communicating number processing method and processing device |
-
2017
- 2017-12-05 CN CN201711266910.4A patent/CN108124065A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150019226A1 (en) * | 1999-06-10 | 2015-01-15 | West View Research, Llc | Computerized information apparatus |
US20140278402A1 (en) * | 2013-03-14 | 2014-09-18 | Kent S. Charugundla | Automatic Channel Selective Transcription Engine |
CN107306306A (en) * | 2016-04-25 | 2017-10-31 | 腾讯科技(深圳)有限公司 | Communicating number processing method and processing device |
CN106504768A (en) * | 2016-10-21 | 2017-03-15 | 百度在线网络技术(北京)有限公司 | Phone testing audio frequency classification method and device based on artificial intelligence |
CN106599933A (en) * | 2016-12-26 | 2017-04-26 | 哈尔滨工业大学 | Text emotion classification method based on the joint deep learning model |
CN107222865A (en) * | 2017-04-28 | 2017-09-29 | 北京大学 | The communication swindle real-time detection method and system recognized based on suspicious actions |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108924333A (en) * | 2018-06-12 | 2018-11-30 | 阿里巴巴集团控股有限公司 | Fraudulent call recognition methods, device and system |
CN109377983A (en) * | 2018-10-18 | 2019-02-22 | 深圳壹账通智能科技有限公司 | A kind of harassing call hold-up interception method and relevant device based on interactive voice |
CN109410496A (en) * | 2018-10-25 | 2019-03-01 | 北京交通大学 | Attack early warning method, apparatus and electronic equipment |
CN109688275A (en) * | 2018-12-27 | 2019-04-26 | 中国联合网络通信集团有限公司 | Harassing call recognition methods, device and storage medium |
CN109658939A (en) * | 2019-01-26 | 2019-04-19 | 北京灵伴即时智能科技有限公司 | A kind of telephonograph access failure reason recognition methods |
CN109905282A (en) * | 2019-04-09 | 2019-06-18 | 国家计算机网络与信息安全管理中心 | Fraudulent call prediction technique and forecasting system based on LSTM |
CN110334110A (en) * | 2019-05-28 | 2019-10-15 | 平安科技(深圳)有限公司 | Natural language classification method, device, computer equipment and storage medium |
CN110525821A (en) * | 2019-08-20 | 2019-12-03 | 北京精英系统科技有限公司 | A kind of system promoting separate waste collection site operation efficiency |
CN112784038A (en) * | 2019-10-23 | 2021-05-11 | 阿里巴巴集团控股有限公司 | Information identification method, system, computing device and storage medium |
CN110929506A (en) * | 2019-12-04 | 2020-03-27 | 杭州安恒信息技术股份有限公司 | Junk information detection method, device and equipment and readable storage medium |
CN112193959A (en) * | 2020-09-25 | 2021-01-08 | 浙江新再灵科技股份有限公司 | Method and system for detecting abnormal sound of elevator |
CN112478975A (en) * | 2020-12-09 | 2021-03-12 | 浙江新再灵科技股份有限公司 | Elevator door fault detection method based on audio features |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108124065A (en) | A kind of method junk call content being identified with disposal | |
CN110084610B (en) | Network transaction fraud detection system based on twin neural network | |
CN108566627A (en) | A kind of method and system identifying fraud text message using deep learning | |
Schaaf et al. | Enhancing decision tree based interpretation of deep neural networks through l1-orthogonal regularization | |
CN110853680A (en) | double-BiLSTM structure with multi-input multi-fusion strategy for speech emotion recognition | |
CN105955959B (en) | A kind of sensibility classification method and system | |
CN111309909B (en) | Text emotion classification method based on hybrid model | |
CN109583565A (en) | Forecasting Flood method based on the long memory network in short-term of attention model | |
CN109756632B (en) | Fraud telephone analysis method based on multidimensional time sequence | |
CN109243494A (en) | Childhood emotional recognition methods based on the long memory network in short-term of multiple attention mechanism | |
Partila et al. | Pattern recognition methods and features selection for speech emotion recognition system | |
CN111222992A (en) | Stock price prediction method of long-short term memory neural network based on attention mechanism | |
CN113488073A (en) | Multi-feature fusion based counterfeit voice detection method and device | |
CN109308903A (en) | Speech imitation method, terminal device and computer readable storage medium | |
CN115689008A (en) | CNN-BilSTM short-term photovoltaic power prediction method and system based on ensemble empirical mode decomposition | |
CN114861835B (en) | Noise hearing loss prediction system based on asymmetric convolution | |
Hu et al. | pRNN: A recurrent neural network based approach for customer churn prediction in telecommunication sector | |
Kamalov et al. | Deep learning regularization in imbalanced data | |
Hu et al. | An efficient Long Short-Term Memory model based on Laplacian Eigenmap in artificial neural networks | |
Maqsudur Rahman et al. | The prediction of coronavirus disease 2019 outbreak on Bangladesh perspective using machine learning: a comparative study | |
Chen et al. | A modified HME architecture for text-dependent speaker identification | |
Kimura et al. | New perspective of interpretability of deep neural networks | |
Kachwala et al. | Predicting Rainfall from Historical Data Trends | |
Wang et al. | Deep learning model for human activity recognition and prediction in smart Homes | |
Kajornrit et al. | A comparative analysis of soft computing techniques used to estimate missing precipitation records |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180605 |