CN105868181B - The automatic identifying method of natural language parallel construction based on new neural network - Google Patents

The automatic identifying method of natural language parallel construction based on new neural network Download PDF

Info

Publication number
CN105868181B
CN105868181B CN201610250258.6A CN201610250258A CN105868181B CN 105868181 B CN105868181 B CN 105868181B CN 201610250258 A CN201610250258 A CN 201610250258A CN 105868181 B CN105868181 B CN 105868181B
Authority
CN
China
Prior art keywords
parallel construction
neural network
phrase
ingredient
word
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201610250258.6A
Other languages
Chinese (zh)
Other versions
CN105868181A (en
Inventor
黄书剑
周逸初
戴新宇
陈家骏
张建兵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN201610250258.6A priority Critical patent/CN105868181B/en
Publication of CN105868181A publication Critical patent/CN105868181A/en
Application granted granted Critical
Publication of CN105868181B publication Critical patent/CN105868181B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
    • 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)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Machine Translation (AREA)

Abstract

The present invention proposes the automatic identifying method of the natural language parallel construction based on new neural network, including：Syntactic analysis processing first is carried out to sentence to be analyzed, the candidate collection of a parallel construction is obtained, is then given a mark to the parallel construction in candidate collection using novel neural network learning device, to select final output of the best parallel construction as system.This method has considered the similitude between the phrase independence of parallel construction and phrase, improves parallel construction accuracy of identification.Existing other technologies are compared, this method protrusion can automatically identify arbitrary parallel construction, and other technologies can only identify certain types of parallel construction, the parallel construction that such as only noun forms.Method proposes a kind of more effective parallel construction recognition methods, improve identification quality in practical applications.

Description

The automatic identifying method of natural language parallel construction based on new neural network
Technical field
The present invention relates to a kind of methods using Computer Automatic Recognition parallel construction, are based particularly on new neural network Natural language parallel construction automatic identifying method.
Background technology
Syntactic analysis technology has developed very rapidly since the nineties in last century, has made significant headway, had become Research hotspot in natural language processing field.
Although syntactic analysis treatment technology has had a significant progress, the practicability of current syntactic analysis technology and Availability is not very high, also undesirable to the handling result of complicated sentence, especially to the sentence comprising labyrinth, such as simultaneously The quality of array structure, syntactic analysis is also improved.According to statistics, there are about 10% mistakes to come from knot arranged side by side in syntactic analysis Structure.Therefore, in the case where the difficulty for promoting syntactic analysis ability is increasing, how by focusing special construction, than Such as parallel construction, the quality to promote syntactic analysis becomes a major issue.
In syntactic analysis technology, a kind of very efficient mode is the syntactic analysis technology based on state transition method, Its course of work is as follows：Sentence to be analyzed is inputted, system is unit according to word, and sequence reads in word one by one from left to right Language often reads in a word, it is possible to carry out stipulations operation to the sequence of terms read in, and when carry out stipulations and what is carried out Kind stipulations operation will be determined by the trained scoring model finished.As the word in sentence is read in one by one, tree The longer structure will be the bigger, and when entire sentence is all read in, syntax tree also completes with regard to analysis.From the foregoing, it will be observed that when carry out It is an important factor for influencing syntactic analysis ability that stipulations, which operate, carry out which kind of stipulations operation all,.Knot arranged side by side is automatically identified in advance This partial information is simultaneously input in syntactic analysis system by structure, it will help system correctly sentences above-mentioned two factor It is disconnected, have greatly improved effect for the syntactic analysis ability of entire sentence, the present invention focuses on the automatic identification of parallel construction Research, will improve the quality of syntactic analysis in actual use.
It is to carry out automatic identifications for some special parallel constructions, such as only by teasing in existing inventive technique Number parallel construction separated, only the parallel construction etc. that is made of noun, these methods and techniques can not all accomplish to automatically identify The parallel construction being likely to occur in any one natural language.Therefore, it in order to continue the ability of raising syntactic analysis, needs to seek It can identify the new method of arbitrary parallel construction.
Invention content
Goal of the invention：The technical problem to be solved by the present invention is to focus only on identification for the identification of current parallel construction Special parallel construction, there is no enough generalization abilities, do not play the role of the weakness of raising to syntax analyzing processing, propose A method of utilizing arbitrary parallel construction in neural network learning device automatic identification natural language.
In order to solve the above-mentioned technical problem, the invention discloses the natural language parallel constructions based on new neural network Automatic identifying method.
In the automatic identification natural language sentences of the present invention using neural network structure the method for parallel construction include with Lower step：
Step 1, it includes natural language sentences text file to be analyzed that computer, which reads one, is carried out to the sentence of reading For the syntactic analysis of parallel construction, obtains the merging of parallel construction syntax tree Candidate Set and be input in neural network learning device；
Step 2, neural network learning device gives a mark to parallel construction all in parallel construction syntax tree candidate collection, Therefrom select best parallel construction.
Step 1 includes the following steps：
Step 1-1 is successively read each word in natural language sentences, using based on state according to sequence from left to right The syntactic analysis technology of transfer techniques carries out the syntactic analysis just for parallel construction to the sentence of input, is obtained after analysis side by side Structure syntax tree candidate collection.
Step 1-2, extract parallel construction syntax tree candidate collection in all parallel constructions left ingredient phrase and it is right at Divide phrase and tentatively given a mark, the left ingredient phrase of all parallel constructions and right ingredient phrase are input to neural network learning In device.
The neural network learning device is made of two Recognition with Recurrent Neural Network and a neural networks with single hidden layer, two cycles Neural network shares identical parameter setting, and its hidden layer is directly connected to the input layer of neural networks with single hidden layer, and two are followed There is individual output layer to be independent of each other independently of each other for ring neural network and neural networks with single hidden layer.
Step 1-2 includes the following steps：
Step 1-2-1 extracts parallel construction to each parallel construction in parallel construction syntax tree candidate collection Left ingredient phrase SleftWith right ingredient phrase Sright, Sleft=w0w1…wn1, Sright=w '0w′1…w′m1, wherein wn1Indicate left N-th in ingredient phrase1A word, w 'm1Indicate the m in right ingredient phrase1A word；
Step 1-2-2, using following formula by left ingredient phrase SleftWith right ingredient phrase SrightIt is input to identical In two Recognition with Recurrent Neural Network of parameter setting：
Y (t)=g (Vs (t)),
S (t)=f (U0w(t)+U1O (t)+Ps (t-1)),
Wherein, y (t) is the final output of Recognition with Recurrent Neural Network, and w indicates that the word in sentence, o indicate the word of corresponding word Property label, t indicate currently processed to t-th word；W (t) indicates that t-th of word, o (t) indicate the part of speech label of t-th of word；s (t), s (t-1) indicates that the vector of t-th of word indicates and the vector of the t-1 word indicates respectively； U0、U1, V and P be trained Good model parameter, the typically form of matrix, each element in matrix can arbitrary real number value, concrete numerical value is by being The automatic study of system obtains；F () and g () is activation primitive and normalized function in Recognition with Recurrent Neural Network, Vs (t), U respectively0w (t),U1O (t), Ps (t-1) are matrix multiple operations.
Using Recognition with Recurrent Neural Network respectively to SleftAnd SrightGive a mark, using the final output of Recognition with Recurrent Neural Network as The score of left and right phrase, is denoted as Score respectivelyleftAnd Scoreright。
Step 2 includes the following steps：
Step 2-1, by left ingredient phrase Sleft, right ingredient phrase SrightAnd their common contextual information c are defeated simultaneously Enter into the neural network of single hidden layer, is integrally given a mark to parallel construction according to following formula:
H=f (Rc),
Y=g (Q0s0(n2)+Q1s1(m2)+Th),
Wherein, h is that the vector of contextual information indicates, the final output of the neural network of the single hidden layer of y expressions, wherein R, Q0、Q1It is trained model parameter, the typically form of matrix with T, each element in matrix can be arbitrarily real Numerical value, concrete numerical value are learnt to obtain automatically by system.n2、m2The length of the length and right ingredient phrase of left ingredient phrase is indicated respectively Degree, s0(n2) and s1(m2) left ingredient phrase S is indicated respectivelyleftIndicated by the vector obtained after neural circuitry network and it is right at Divide phrase SrightIt is indicated by the vector obtained after neural circuitry network；The final output of the neural network of the list hidden layer is just made For the score of current parallel construction, it is denoted as Score； Rc,Q0s0(n2),Q1s1(m2), Th is matrix multiple operation；
Step 2-2 considers the marking in step 1-2-2 and step 2-1, to Scoreleft,Scoreright,Score Average value is calculated, selects the highest parallel construction of average mark as best parallel construction.
Wherein, f (z) and g (z) is common activation primitive and normalized function, specific shape in Recognition with Recurrent Neural Network respectively Formula is：
Wherein, z is the input parameter of activation primitive and normalized function, and e indicates that natural logrithm, x indicate the dimension of vector, K is a counting to vector element.
Advantageous effect：The present invention has considered the local message and global information of phrase simultaneously, is selected in this, as basis Best parallel construction is selected, the recognition capability of parallel construction is improved.
Description of the drawings
The present invention is done with reference to the accompanying drawings and detailed description and is further illustrated, of the invention is above-mentioned And/or otherwise advantage will become apparent.
Fig. 1 and Fig. 2 indicates two different syntax trees being likely to occur in syntax tree analytic process in embodiment 1.
Fig. 3 is the flow chart of the present invention.
Fig. 4 and Fig. 5 indicates two different syntax trees being likely to occur in syntax tree analytic process in embodiment 2.
Specific implementation mode
The present invention proposes the automatic identifying method of the natural language parallel construction based on new neural network.First with Syntactic analysis technology finds out possible candidate collection, then using neural network learning device found out from candidate collection it is best and Array structure.In existing system can only identification division parallel construction, such as only by separated by commas parallel construction, only by noun phrase At parallel construction etc., these methods and techniques can not all accomplish to automatically identify to be likely to occur in any one natural language Parallel construction.
As shown in figure 3, the invention discloses one kind based on arranged side by side in new neural network structure automatic identification natural language The method of structure, the system based on the present invention have considered the local message and Global Information of parallel construction as a whole, identify best Parallel construction.Fig. 3 describes the new neural network structure proposed in the present invention.
The process of parallel construction includes the following steps in identification natural language of the present invention:
Step 11, it includes natural language sentences text file to be analyzed that computer, which reads one, is turned using based on state The syntactic analysis technology of shifting method carries out syntactic analysis to the sentence of input, syntactic analysis herein by the corresponding syntax about Beam, can only carry out parallel construction syntactic analysis, and analysis obtains the candidate collection of a parallel construction syntax tree.
Step 12, in the candidate collection of parallel construction syntax tree, all possible parallel construction candidate is extracted, by this A little candidate parallel constructions are input in new neural network proposed by the present invention.
It is as follows that process is identified in new neural network learning device of the present invention:
Step 21, system receives candidate parallel construction set, therefrom extracts the left ingredient phrase S of parallel constructionleft: w0w1…wnWith right ingredient phrase Sright:w′0w′1…w′m。
Step 22, the left and right ingredient phrase of parallel construction is input to two cycle nerve nets with identical parameters simultaneously In network structure, as shown in structure in box in Fig. 2.By the neural network structure of two shared parameters, system is according to following public Formula is to SleftAnd SrightMarking：
Y (t)=g (Vs (t))
S (t)=f (U0w(t)+U1o(t)+Ps(t-1))
Wherein, y (t) is the final output score of neural network, and w indicates that the word in sentence, o indicate the word of corresponding word Property label, t indicate currently processed to t-th word；W (t), o (t) indicate that t-th of word and its part of speech mark respectively； s(t), S (t-1) indicates that t-th of word and the vector of the t-1 word indicate respectively,；U0、U1, V and P be trained model parameter, The typically form of matrix, each element in matrix can arbitrary real number value, concrete numerical value learns automatically by system Go out；F and g is activation primitive and normalized function in Recognition with Recurrent Neural Network respectively.Using the network respectively to SleftAnd Sright It gives a mark, the score by the final output of network as left and right phrase is denoted as respectively：ScoreleftAnd Scoreright。Vs (t),U0w(t),U1O (t), Ps (t-1) are matrix multiple operations.
Step 23, by left ingredient phrase Sleft, right ingredient phrase SrightAnd its common contextual information c is input to simultaneously In the neural network of one single hidden layer, following formula is utilized:
H=f (Rc)
Y=g (Q0s0(n)+Q1s1(m)+Th)
It integrally gives a mark to parallel construction.
Wherein, h is that the vector of contextual information indicates that y indicates the final output of model, wherein R, Q0、Q1It is with T Trained model parameter, the typically form of matrix, each element in matrix can arbitrary real number value, concrete numerical value Learnt to obtain automatically by system.N, m indicates the length of the length and right ingredient phrase of left ingredient phrase respectively, and S0(n) and S1 (m) left ingredient phrase S is indicated respectivelyleftIt is indicated and right ingredient phrase S by the vector obtained after neural circuitry networkrightThrough The vector obtained after neural circuitry network is crossed to indicate；The output of the network is denoted as with regard to the score as current parallel construction Score。Rc,Q0s0(n),Q1s1(m), Th is matrix multiple operation.
Step 24, after giving a mark respectively to left and right phrase and overall structure, to this three (Scoreleft,Scoreright, Score score) calculates average mark, the final score as current parallel construction.
Step 25, the operation that step 21 arrives step 24 is executed to the parallel construction of all candidates, therefrom selects highest scoring Parallel construction, as best parallel construction.
Embodiment 1
The present embodiment utilizes parallel construction operational process in new neural network structure recognition natural language as follows:
1. input natural language sentences to be analyzed:" Pudong, Shanghai exploitation is synchronous with legal construction ", wherein really simultaneously Array structure is " exploitation and legal construction ".
2. system starts the natural language sentences progress to input just for the syntactic analysis of parallel construction, obtain possible Parsing tree, as depicted in figs. 1 and 2:
3. pair parallel construction syntax tree being likely to occur, system extracts its parallel construction, for Fig. 1, extraction and Array structure is " exploitation and legal system "；For Fig. 2, the parallel construction of extraction is " exploitation and legal construction ".
4. the parallel construction S that will be extracted1=" exploitation and legal system " and S2=" exploitation and legal construction " is input to this hair In new neural network in bright.
5. after neural network receives the parallel construction set of input, it is short to extract its left and right to each parallel construction Language, for S1For, its left phraseRight phrase isAnd for S2For, its left side PhraseRight phrase is
6. willWithIt is input in Recognition with Recurrent Neural Network, is given a mark to it using Recognition with Recurrent Neural Network simultaneously, Score is respectivelyWithIt willWithIt is input in Recognition with Recurrent Neural Network simultaneously, It is given a mark to it using Recognition with Recurrent Neural Network, score is respectivelyWith
7. by S1And S2It is input in neural networks with single hidden layer, is given a mark to parallel construction using neural networks with single hidden layer, S1It is scored at Score1=0.7, S2It is scored at Score2=0.9.
8. pairAnd Score1Calculate average valueIt is right And Score2Calculate average valueThus judge, S2Highest scoring, therefore " exploitation built with legal system If " will be as the final output of system.
Embodiment 2
The present embodiment utilizes parallel construction operational process in new neural network structure recognition natural language as follows:
1. input natural language sentences to be analyzed:" new situation, the new problem that encountered ", wherein real parallel construction It is " new situation, new problem ".
2. system starts the natural language sentences progress to input just for the syntactic analysis of parallel construction, obtain possible Parsing tree, as shown in Figure 4 and Figure 5:
3. pair parallel construction syntax tree being likely to occur, system extracts its parallel construction, for Fig. 4, extraction and Array structure is " new situation, new problem "；For Fig. 2, the parallel construction of extraction is " situation, new problem ".
4. the parallel construction S that will be extracted1=" new situation, new problem " and S2=" situation, new problem " is input to this hair In new neural network in bright.
5. after neural network receives the parallel construction set of input, it is short to extract its left and right to each parallel construction Language, for S1For, its left phraseRight phrase isAnd for S2For, it Left phraseRight phrase is
6. willWithIt is input in Recognition with Recurrent Neural Network, is given a mark to it using Recognition with Recurrent Neural Network simultaneously, Score is respectivelyWithIt willWithIt is input to Recognition with Recurrent Neural Network simultaneously In, it is given a mark to it using Recognition with Recurrent Neural Network, score is respectivelyWith
7. by S1And S2It is input in neural networks with single hidden layer, is given a mark to parallel construction using neural networks with single hidden layer, S1It is scored at Score1=0.95, S2It is scored at Score2=0.6.
8. pairAnd Score1Calculate average valueIt is right And Score2Calculate average valueThus judge, S1Highest scoring, therefore " new situation and new problem " will Final output as system.
The present invention provides the automatic identifying method of the natural language parallel construction based on new neural network, specific implementations The method and approach of the technical solution have very much, and the above is the preferred embodiment of the present invention.The present invention is based on a kind of new The neural network structure of type is given a mark between each ingredient of parallel construction and integrally so that be respectively using neural network System can automatically identify any type of parallel construction.In concrete practice, method proposed by the present invention and other manner phase Than being not limited to special parallel construction, such as the parallel construction by separated by commas, the parallel construction etc. that only noun forms, energy Enough automatically identify the parallel construction of arbitrary structures.It should be pointed out that for those skilled in the art, not Under the premise of being detached from the principle of the invention, several improvements and modifications can also be made, these improvements and modifications also should be regarded as the present invention Protection domain.The available prior art of each component part being not known in the present invention is realized.

Claims (1)

1. the automatic identifying method of the natural language parallel construction based on neural network, which is characterized in that include the following steps：
Step 1, it includes natural language sentences text file to be analyzed that computer, which reads one, is directed to the sentence of reading The syntactic analysis of parallel construction obtains the merging of parallel construction syntax tree Candidate Set and is input in neural network learning device；
Step 2, neural network learning device gives a mark to parallel construction all in parallel construction syntax tree candidate collection, therefrom Select best parallel construction；
Step 1 includes the following steps：
Step 1-1 is successively read each word in natural language sentences according to sequence from left to right, is shifted using based on state The syntactic analysis technology of technology carries out the syntactic analysis just for parallel construction to the sentence of input, and parallel construction is obtained after analysis Syntax tree candidate collection；
Step 1-2, the left ingredient phrase and right ingredient for extracting all parallel constructions in parallel construction syntax tree candidate collection are short Language is simultaneously tentatively given a mark, and the left ingredient phrase of all parallel constructions and right ingredient phrase are input to neural network learning device In；
The neural network learning device is made of two Recognition with Recurrent Neural Network and a neural networks with single hidden layer, two cycle nerves The identical parameter setting of network share, and its hidden layer is directly connected to the input layer of neural networks with single hidden layer, two cycle god Through network and neural networks with single hidden layer there is individual output layer to be independent of each other independently of each other；
Step 1-2 includes the following steps：
Step 1-2-1, to each parallel construction in parallel construction syntax tree candidate collection extract parallel construction it is left at Divide phrase SleftWith right ingredient phrase Sright, Sleft=w0w1...wn1, Sright=w '0w′1...w′m1, wherein wn1Indicate it is left at Divide n-th in phrase1A word, w 'm1Indicate the m in right ingredient phrase1A word；
Step 1-2-2, using following formula by left ingredient phrase SleftWith right ingredient phrase SrightIt is input to identical parameters In two Recognition with Recurrent Neural Network being arranged：
Y (t)=g (Vs (t)),
S (t)=f (U0w(t)+U1O (t)+Ps (t-1)),
Wherein, y (t) is the final output of Recognition with Recurrent Neural Network, and w indicates that the word in sentence, o indicate the part of speech mark of corresponding word Note, t indicate currently processed to t-th word；W (t) indicates that t-th of word, o (t) indicate the part of speech label of t-th of word；s(t)、s (t-1) indicate that the vector of t-th of word indicates and the vector of the t-1 word indicates respectively；U0、U1, V and P be trained mould Shape parameter；F () and g () is activation primitive and normalized function in Recognition with Recurrent Neural Network, Vs (t), U respectively0W (t), U1o (t), Ps (t-1) is matrix multiple operation；
Using Recognition with Recurrent Neural Network respectively to SleftAnd SrightGive a mark, using the final output of Recognition with Recurrent Neural Network as it is left, The score of right phrase, is denoted as Score respectivelyleftAnd Scoreright；
Step 2 includes the following steps：
Step 2-1, by left ingredient phrase Sleft, right ingredient phrase SrightAnd their common contextual information c are input to simultaneously In the neural network of single hidden layer, integrally given a mark to parallel construction according to following formula：
H=f (Rc),
Y=g (Q0s0(n2)+Q1s1(m2)+Th),
Wherein, h is that the vector of contextual information indicates, y indicates the final output of the neural network of single hidden layer, wherein R, Q0、Q1With T is trained model parameter；n2、m2The length of the length and right ingredient phrase of left ingredient phrase, s are indicated respectively0(n2) And s1(m2) left ingredient phrase S is indicated respectivelyleftIt is indicated and right ingredient phrase by the vector obtained after neural circuitry network SrightIt is indicated by the vector obtained after neural circuitry network；The final output of the neural network of the list hidden layer is just used as current The score of parallel construction, is denoted as Score；Rc, Q0s0(n2), Q1s1(m2), Th is matrix multiple operation；
Step 2-2 considers the marking in step 1-2-2 and step 2-1, calculates average value, it is highest simultaneously to select average mark Array structure is as best parallel construction；
F (z) and g (z) is activation primitive and normalized function in Recognition with Recurrent Neural Network respectively, and concrete form is：
Wherein, z is the input parameter of activation primitive and normalized function, and e indicates that natural logrithm, x indicate that the dimension of vector, k are A counting to vector element.
CN201610250258.6A 2016-04-21 2016-04-21 The automatic identifying method of natural language parallel construction based on new neural network Active CN105868181B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610250258.6A CN105868181B (en) 2016-04-21 2016-04-21 The automatic identifying method of natural language parallel construction based on new neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610250258.6A CN105868181B (en) 2016-04-21 2016-04-21 The automatic identifying method of natural language parallel construction based on new neural network

Publications (2)

Publication Number Publication Date
CN105868181A CN105868181A (en) 2016-08-17
CN105868181B true CN105868181B (en) 2018-08-21

Family

ID=56632710

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610250258.6A Active CN105868181B (en) 2016-04-21 2016-04-21 The automatic identifying method of natural language parallel construction based on new neural network

Country Status (1)

Country Link
CN (1) CN105868181B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP6663881B2 (en) * 2017-04-13 2020-03-13 日本電信電話株式会社 Parallel phrase analysis device, parallel phrase analysis model learning device, method, and program
CN110046338B (en) * 2018-01-15 2022-11-11 深圳市腾讯计算机系统有限公司 Context selection method and device, electronic equipment and storage medium
CN114722774B (en) * 2022-04-07 2024-01-30 平安科技（深圳）有限公司 Data compression method, device, electronic equipment and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103810999A (en) * 2014-02-27 2014-05-21 清华大学 Linguistic model training method and system based on distributed neural networks
CN104102630A (en) * 2014-07-16 2014-10-15 复旦大学 Method for standardizing Chinese and English hybrid texts in Chinese social networks
CN104463324A (en) * 2014-11-21 2015-03-25 长沙马沙电子科技有限公司 Convolution neural network parallel processing method based on large-scale high-performance cluster
CN104572892A (en) * 2014-12-24 2015-04-29 中国科学院自动化研究所 Text classification method based on cyclic convolution network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103810999A (en) * 2014-02-27 2014-05-21 清华大学 Linguistic model training method and system based on distributed neural networks
CN104102630A (en) * 2014-07-16 2014-10-15 复旦大学 Method for standardizing Chinese and English hybrid texts in Chinese social networks
CN104463324A (en) * 2014-11-21 2015-03-25 长沙马沙电子科技有限公司 Convolution neural network parallel processing method based on large-scale high-performance cluster
CN104572892A (en) * 2014-12-24 2015-04-29 中国科学院自动化研究所 Text classification method based on cyclic convolution network

Also Published As

Publication number Publication date
CN105868181A (en) 2016-08-17

Similar Documents

Publication Publication Date Title
CN110795543B (en) Unstructured data extraction method, device and storage medium based on deep learning
Abdul-Mageed et al. NADI 2021: The second nuanced Arabic dialect identification shared task
Forster et al. RWTH-PHOENIX-weather: A large vocabulary sign language recognition and translation corpus.
Dalal et al. A survey of extractive and abstractive text summarization techniques
CN110110330B (en) Keyword extraction method based on text and computer equipment
CN106407235B (en) A kind of semantic dictionary construction method based on comment data
CN107766371A (en) A kind of text message sorting technique and its device
CN107679224B (en) Intelligent question and answer method and system for unstructured text
CN109543181B (en) Named entity model and system based on combination of active learning and deep learning
CN105868181B (en) The automatic identifying method of natural language parallel construction based on new neural network
Sandaruwan et al. Sinhala hate speech detection in social media using text mining and machine learning
CN107247751B (en) LDA topic model-based content recommendation method
CN108121702A (en) Mathematics subjective item reads and appraises method and system
CN106227719B (en) Chinese word segmentation disambiguation method and system
CN109062894A (en) The automatic identification algorithm of Chinese natural language Entity Semantics relationship
Younes et al. Constructing linguistic resources for the Tunisian dialect using textual user-generated contents on the social web
CN109213856A (en) A kind of method for recognizing semantics and system
CN109783624A (en) Answer generation method, device and the intelligent conversational system in knowledge based library
Sifa et al. Towards contradiction detection in german: a translation-driven approach
CN104317882B (en) Decision-based Chinese word segmentation and fusion method
CN106528694A (en) Artificial intelligence-based semantic judgment processing method and apparatus
CN106777336A (en) A kind of exabyte composition extraction system and method based on deep learning
Cornia et al. Universal captioner: Inducing content-style separation in vision-and-language model training
Zhou et al. Icdar 2015 text reading in the wild competition
JP2018026098A (en) Identification control method and identification control device

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant