CN109543039A - A kind of natural language sentiment analysis method based on depth network - Google Patents
A kind of natural language sentiment analysis method based on depth network Download PDFInfo
- Publication number
- CN109543039A CN109543039A CN201811409537.8A CN201811409537A CN109543039A CN 109543039 A CN109543039 A CN 109543039A CN 201811409537 A CN201811409537 A CN 201811409537A CN 109543039 A CN109543039 A CN 109543039A
- Authority
- CN
- China
- Prior art keywords
- square
- context
- memory
- module
- loss
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Machine Translation (AREA)
Abstract
Natural language sentiment analysis method provided by the invention based on depth network, on the basis of memory network, semantic dependency information introduces the execution to guide attention mechanism, and the context square information for containing sentence entirety emotion information is also utilized for present analysis subject word and provides background information.Entire model includes insertion module, memory sequences construct module, semantic dependency mask pays attention to power module, context square emotion learning module and output module.In a model, the semantic dependency relations information of the subject word and context that are obtained according to interdependent syntax tree will be introduced into memory network, so that the memory sequences of each layer were dynamically generated, to guide the execution of the attention mechanism in the multilayer module of memory network.In addition, we have proposed the learning tasks based on context square in order to introduce the whole emotion information of sentence, i.e. relation information in same sentence between all subject words, the sentiment analysis of special object word is assisted by multi-task learning.
Description
Technical field
The present invention relates to the sentiment analysis fields in Computer Natural Language Processing, more particularly, to one kind based on deep
Spend the natural language sentiment analysis method of network.
Background technique
The main target of object level sentiment analysis task be for one or more evaluation goal present in given sentence,
Provide feeling polarities of each evaluation goal in sentence respectively (as positive, passive or neutral).For example, given sentence " this family
The price in restaurant is very cheap, but services very poor " and evaluation goal " price " and " service ", for evaluation object " price ", emotion
Polarity is positive, and for evaluation object " service ", feeling polarities are then passive.It is obvious that in the same sentence not
Same subject word, sentiment analysis result may be different.
As attention mechanism and memory network take in multiple natural language processing tasks such as machine translation, reading understanding
It obtains and well shows, the method for having merged attention mechanism and memory network also becomes the master for solving object level sentiment analysis task
Want method.The exemplary process of this respect has MemNet (Memory Network, memory network) and RAM (Recurrent
Attention Network on Memory acts on the recurrence attention network on memory).
MemNet model utilizes the location information of context and subject word using the word embeded matrix of sentence as memory sequences
Implement attention mechanism on memory sequences with content information, finally obtains the sentence feeling polarities for subject word.RAM model
Then on the basis of MemNet, square is embedded in word using LSTM (Long Short-Term Memory, long memory unit in short-term)
Battle array is operated, and the memory sequences comprising sentence internal structural information are obtained, then is gone to combine recurrence with a kind of nonlinear mode
Each layer of output in attention network.
Subject word is utilized in existing method and the positional relationship of context carries out tax power to memory sequences, so that being directed to
Different subject words can generate different memory sequences.However it goes to be weighted simultaneously memory sequences just with positional relationship
Contacting between subject word and context cannot be made full use of, including the semantic relation between subject word and context and is currently commented
Emotional relationship in valence subject word and sentence between other subject words is made a concrete analysis of as follows:
The semantic relation of subject word and context is lacked, has utilized location information merely, can make on semantic dependent tree
It is different from subject word distance, but attention of the text apart from identical word by equal extent, this is clearly unreasonable.In addition,
In complicated sentence, often there is semantic relation by force but the context words of text distance relatively far away from, if merely used
Position weighting can make model that can not capture and determine the highly important, context of distance relatively far away to subject word emotion
Word.
Considering for subject word relationship is lacked, so that with judging that the operation of different subject words is independent, nothing in sentence
Method considers influence of other subject words of same sentence for existing object word Judgment by emotion.It is some comprising comparing, it is arranged side by side
In the complicated sentence of multipair elephant, if not accounting for these relationships, individually go to judge each that there are the objects in sentence
The feeling polarities of word, it is clear that not than considering these relationships, allow the emotion recognition task of each subject word by same sentence
The assistance of other subject words is more helpful.
Summary of the invention
It only considered for the existing main stream approach based on attention mechanism and memory network when generating memory sequences
Positional relationship between subject word and context, the problem of without in view of semantic relation between subject word and context,
The present invention proposes a kind of natural language sentiment analysis method based on depth network, the technical solution adopted by the present invention is that:
A kind of natural language sentiment analysis method based on depth network, including insertion module, memory sequences building module,
Semantic dependency mask pays attention to power module, context square emotion learning module and output module;
The insertion look-up table that the insertion module is obtained using one by unsupervised approaches pre-training, will be in corpus
Word is converted to corresponding term vector;For the non-dictionary word being not present in look-up table, Gaussian Profile random initializtion is used
Its random transition is embedded at the word of a low dimensional;
The memory sequences building module passes through the two-way length insertion sequence that memory unit obtains insertion module in short-term
Memory sequences are converted to, the memory sequences after conversion can indicateWherein n is sequence length;
The semantic dependency mask notices that power module according to the interdependent syntax tree of sentence, extracts semantic dependency information, so
The execution of attention mechanism is guided, object is obtained come the different piece of dynamic select memory sequences according to semantic dependency information afterwards
The loss of word emotional semantic classification;The context square emotion learning module passes through the Cooperative Study to context square recurrence task simultaneously
Context memory sequence is constructed, and calculates context square and returns loss;
The output module by simultaneously minimize subject word emotional semantic classification loss and context square return loss come into
Row training, to predict the feeling polarities of subject word.
Described, the semantic mask notices that the workflow of power module includes the following steps:
Step 1: in l layers of mask memory sequences, each memory unit will carry out mask behaviour according to semantic dependency information
Make, i.e. up and down cliction corresponding memory unit of the selection semantic distance less than current layer number l, specific formula is as follows:
Wherein dist (wi,wt) referring to existing object word to the semantic distance of context, l is the layer of multilayer profound memory network
Number, l is positive integer;
Step 2: being in the mask memory sequences that each computation layer l is generatedRemembered by mask
Recall sequence, available in l layers of mask memory sequences, the attention score of each memory unit are as follows:
WhereindALIndicate weight dimension used in attention mechanism,rl-1And vaRespectively indicate memory
One layer of output and subject word indicate on unit, memory network;
Step 3: the score value of each memory unit then being obtained by softmax function normalizationTo gain attention
The final output of power mechanism layer:
WhereinIt is score value,Indicate corresponding memory unit
Step 4: a change-over gate and a carrying door is added, controls one layer of output r respectivelyl-1How many is noted power
Next layer is brought into after system conversion, and how many does not pass through conversion, is directly carried to next layer, to obtain this layer of output
rl;The sentence expression for being directed to special object word is obtained by the multilayer attention mechanism nonlinear iteration of profound memory network,
To obtain the emotional semantic classification of subject word as a result, using this as a result, being trained step obtains the loss of subject word emotional semantic classification.
Described, the workflow of the context emotion learning module is as follows:
Step 1: use -1,0 and 1 respectively indicate passive, neutral and positive three feeling polarities describe one with square in model
The feeling polarities distribution of all subject words, square are defined as follows in a sentence:
μk=E ((X- μ)k)
Wherein X is sample value, and E () indicates expectation function, and k indicates the order of square, when order is odd number, μk∈[-1,
1];When order is even number, μk∈[0,1];All squares are normalized into [0,1];
Step 2: using first moment μ1With second moment μ2As the target of global square study, obtained using attention mechanism
Sentence expression vs, then the estimated value of square, such as μ are obtained with different full articulamentums1Estimated value μ '1It can obtain as follows:
WhereinFor the weight of full articulamentum, vsIt is sentence expression
Step 3: and then each sample x is defined in global first moment μ1On loss it is as follows:
Step 4: obtaining global second moment using identical calculation and loseAnd then obtain global loss:
Step 5: the subject word emotion of a sentence is divided into two parts, i.e. left-half and right half part;If
Subject word number is odd number, then median can be divided into left-half, calculate separately left-half and right half part respectively
First order and second order moments;According to the calculating that global square loses, local moment loss is obtained:
Step 6: to lgloble(x) and llocal(x) it is weighted summation, to obtain context square study total losses.
It is described, for each sample x, one context square study loss l of context square emotion learning module definitionm
(x) carry out the optimization of auxiliary object word emotional semantic classification task;This loss loses l by the overall situationgloble(x) and local losses llocal(x)
Two parts composition:
Wherein naIndicate the number of subject word in a sentence, lgloble(x) for the loss of context square in entire sentence it
With, and llocal(x) be the respective context square of left-half and right half part in sentence the sum of loss.
Described, the described output module, specific formula is as follows:
Wherein C is emotional category collection, and D is training set;y∈R|C|It is an only hot vector, i.e., it is only on correct label
Component be 0, fc(x, θ) is model prediction as a result, λ is L2The weight of regular terms, and λmIt is context square recurrence learning loss lm
(x) weight;
In the training process, training will be realized by minimizing loss L, obtain the parameter that model can be made to optimize;?
In test process, prediction result f is obtained by the most optimized parameter that training process obtainsc(x, θ), wherein maximum point of score value
The corresponding classification of amount is exactly the classification predicted.
Compared with prior art, the beneficial effect of technical solution of the present invention is:
For confirmatory experiment effect, the Restaurant and Laptop two that we provide in 2014 Task 4 of SemEval
It is tested on a data set, and assesses experiment effect with accuracy rate, final experiment effect is as follows:
Model | Restaurant data set | Laptop data set |
MemNet | 78.2% | 70.3% |
RAM | 80.0% | 74.1% |
DMMN-SDCM | 81.9% | 75.1% |
Table 1.DMMN-SDCM can from the result of table compared with current main-stream method MemNet and RAM experiment effect
Out, performance of the DMMN-SDCM model in Restaurant and Laptop two datasets will be better than current main method
MemNet and RAM, wherein than the experiment accuracy rate of the RAM model that behaves oneself best at present on both data sets respectively than be higher by about
2 and 1 percentage point, this demonstrate that we are meaningful to the improvement of current method.
Detailed description of the invention
Fig. 1 is the flow chart of the natural language sentiment analysis method provided by the invention based on depth network.
Fig. 2 is that the semantic dependency mask of the natural language sentiment analysis method based on depth network pays attention to the work of power module
Flow chart.
Fig. 3 is the workflow of the context square emotion learning module of the natural language sentiment analysis method based on depth network
Cheng Tu.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, only for illustration, Bu Nengli
Solution is the limitation to this patent.Based on the embodiments of the present invention, those of ordinary skill in the art are not making creative labor
Every other embodiment obtained under the premise of dynamic, shall fall within the protection scope of the present invention.
The following further describes the technical solution of the present invention with reference to the accompanying drawings and examples.
Embodiment 1
Shown in Fig. 1~3, a kind of natural language sentiment analysis method based on depth network, including insertion module, memory sequence
Column building module, semantic dependency mask pay attention to power module, context square emotion learning module and output module;
The insertion look-up table that the insertion module is obtained using one by unsupervised approaches pre-training, will be in corpus
Word is converted to corresponding term vector;For the non-dictionary word being not present in look-up table, Gaussian Profile random initializtion is used
Its random transition is embedded at the word of a low dimensional;
The memory sequences building module passes through the two-way length insertion sequence that memory unit obtains insertion module in short-term
Memory sequences are converted to, the memory sequences after conversion can indicateWherein n is sequence length;
The semantic dependency mask notices that power module according to the interdependent syntax tree of sentence, extracts semantic dependency information, so
The execution of attention mechanism is guided, object is obtained come the different piece of dynamic select memory sequences according to semantic dependency information afterwards
The loss of word emotional semantic classification;The context square emotion learning module passes through the Cooperative Study to context square recurrence task simultaneously
Context memory sequence is constructed, and calculates context square and returns loss;
The output module by simultaneously minimize subject word emotional semantic classification loss and context square return loss come into
Row training, to predict the feeling polarities of subject word.
Described, the semantic mask notices that the workflow of power module includes the following steps:
Step 1: in l layers of mask memory sequences, each memory unit will carry out mask behaviour according to semantic dependency information
Make, i.e. up and down cliction corresponding memory unit of the selection semantic distance less than current layer number l, specific formula is as follows:
Wherein dist (wi,wt) referring to existing object word to the semantic distance of context, l is the layer of multilayer profound memory network
Number, l is positive integer;
Step 2: being in the mask memory sequences that each computation layer l is generatedRemembered by mask
Recall sequence, available in l layers of mask memory sequences, the attention score of each memory unit are as follows:
WhereindALIndicate weight dimension used in attention mechanism,rl-1And vaRespectively indicate memory
One layer of output and subject word indicate on unit, memory network;
Step 3: the score value of each memory unit then being obtained by softmax function normalizationTo gain attention
The final output of power mechanism layer:
WhereinIt is score value,Indicate corresponding memory unit
Step 4: a change-over gate and a carrying door is added, controls one layer of output r respectivelyl-1How many is noted power
Next layer is brought into after system conversion, and how many does not pass through conversion, is directly carried to next layer, to obtain this layer of output
rl;The sentence expression for being directed to special object word is obtained by the multilayer attention mechanism nonlinear iteration of profound memory network,
To obtain the emotional semantic classification of subject word as a result, using this as a result, being trained step obtains the loss of subject word emotional semantic classification.
Described, the workflow of the context emotion learning module is as follows:
Step 1: use -1,0 and 1 respectively indicate passive, neutral and positive three feeling polarities describe one with square in model
The feeling polarities distribution of all subject words, square are defined as follows in a sentence:
μk=E ((X- μ)k)
Wherein X is sample value, and E () indicates expectation function, and k indicates the order of square, when order is odd number, μk∈[-1,
1];When order is even number, μk∈[0,1];All squares are normalized into [0,1];
Step 2: using first moment μ1With second moment μ2As the target of global square study, obtained using attention mechanism
Sentence expression vs, then the estimated value of square, such as μ are obtained with different full articulamentums1Estimated value μ '1It can obtain as follows:
WhereinFor the weight of full articulamentum, vsIt is sentence expression
Step 3: and then each sample x is defined in global first moment μ1On loss it is as follows:
Step 4: obtaining global second moment using identical calculation and loseAnd then obtain global loss:
Step 5: the subject word emotion of a sentence is divided into two parts, i.e. left-half and right half part;If
Subject word number is odd number, then median can be divided into left-half, calculate separately left-half and right half part respectively
First order and second order moments;According to the calculating that global square loses, local moment loss is obtained:
Step 6: to lgloble(x) and llocal(x) it is weighted summation, to obtain context square study total losses.
It is described, for each sample x, one context square study loss l of context square emotion learning module definitionm
(x) carry out the optimization of auxiliary object word emotional semantic classification task;This loss loses l by the overall situationgloble(x) and local losses llocal(x)
Two parts composition:
Wherein naIndicate the number of subject word in a sentence, lgloble(x) for the loss of context square in entire sentence it
With, and llocal(x) be the respective context square of left-half and right half part in sentence the sum of loss.
Described, the described output module, specific formula is as follows:
Wherein C is emotional category collection, and D is training set;y∈R|C|It is an only hot vector, i.e., it is only on correct label
Component be 0, fc(x, θ) is model prediction as a result, λ is L2The weight of regular terms, and λmIt is context square recurrence learning loss lm
(x) weight;
In the training process, training will be realized by minimizing loss L, obtain the parameter that model can be made to optimize;?
In test process, prediction result f is obtained by the most optimized parameter that training process obtainsc(x, θ), wherein maximum point of score value
The corresponding classification of amount is exactly the classification predicted.
Embodiment 2
In the present embodiment,
Given sentence " Great food but the service was dreadful!" and subject word " food " and
“service”。
As a result: in RAM model, two subject words in this sentence are all judged as positive emotion, and DMMN-SDCM
Model then successfully identifies two respective feeling polarities of subject word.
Analysis: since DMMN-SDCM model replaces traditional text range information with semantic dependency information, so that
Model can judge that influence of the upper and lower cliction " dreadful " to subject word " service " is more compared to upper and lower cliction " Great "
It is deep, the feeling polarities judgement of this subject word is helped bigger.In addition, due to the learning tasks for introducing context square, so mould
Pair type can be gone while the relationship of learning object word " food " and " service " when constructing context memory sequence, i.e.,
Than relationship, so that the memory sequences that building is more scientific, assist the feeling polarities of two subject words to judge.
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention.For those of ordinary skill in the art, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
Made any modifications, equivalent replacements, and improvements etc., should be included in the claims in the present invention within the spirit and principle of invention
Protection scope within.
Claims (5)
1. a kind of natural language sentiment analysis method based on depth network, which is characterized in that including being embedded in module, memory sequences
Building module, semantic dependency mask pay attention to power module, context square emotion learning module and output module;
The insertion look-up table that the insertion module is obtained using one by unsupervised approaches pre-training turns the word in corpus
It is changed to corresponding term vector;For the non-dictionary word being not present in look-up table, use Gaussian Profile random initializtion by its
Random transition is embedded at the word of a low dimensional;
The memory sequences building module passes through the insertion sequence conversion that memory unit obtains insertion module in short-term of two-way length
For memory sequences, the memory sequences after conversion can be indicatedWherein n is sequence length;
The semantic dependency mask notices that power module according to the interdependent syntax tree of sentence, extracts semantic dependency information, then root
Carry out the different piece of dynamic select memory sequences according to semantic dependency information, guides the execution of attention mechanism, obtain subject word feelings
Feel Classification Loss;The context square emotion learning module is by returning the Cooperative Study of task to context square come structure simultaneously
Context memory sequence is built, and calculates context square and returns loss;
The output module is instructed by minimizing the loss of subject word emotional semantic classification and context square recurrence loss simultaneously
Practice, to predict the feeling polarities of subject word.
2. the natural language sentiment analysis method according to claim 1 based on depth network, which is characterized in that described
Semantic mask notices that the workflow of power module includes the following steps:
Step 1: in l layers of mask memory sequences, each memory unit will carry out mask operation according to semantic dependency information, i.e.,
Semantic distance is selected to be less than the corresponding memory unit of cliction up and down of current layer number l, specific formula is as follows:
Wherein dist (wi,wt) referring to existing object word to the semantic distance of context, l is the number of plies of multilayer profound memory network, l
For positive integer;
Step 2: being in the mask memory sequences that each computation layer l is generatedSequence is remembered by mask
Column, available in l layers of mask memory sequences, the attention score of each memory unit are as follows:
WhereindALIndicate weight dimension used in attention mechanism,rl-1And vaIt is single to respectively indicate memory
One layer of output and subject word indicate in member, memory network;
Step 3: the score value of each memory unit then being obtained by softmax function normalizationTo the power machine that gains attention
The final output of preparative layer:
WhereinIt is score value,Indicate corresponding memory unit
Step 4: a change-over gate and a carrying door is added, controls one layer of output r respectivelyl-1How many is noted power system
It brings next layer after conversion into, and how many does not pass through conversion, is directly carried to next layer, to obtain this layer of output rl;It is logical
The multilayer attention mechanism nonlinear iteration for crossing depth memory network obtains the sentence expression for being directed to special object word, thus
To subject word emotional semantic classification as a result, using this as a result, be trained step obtain subject word emotional semantic classification loss.
3. the natural language sentiment analysis method according to claim 2 based on depth network, which is characterized in that described
The workflow of context emotion learning module is as follows:
Step 1: use -1,0 and 1 respectively indicate passive, neutral and positive three feeling polarities describe a sentence with square in model
The feeling polarities distribution of all subject words, square are defined as follows in son:
μk=E ((X- μ)k)
Wherein X is sample value, and E () indicates expectation function, and k indicates the order of square, when order is odd number, μk∈[-1,1];When
When order is even number, μk∈[0,1];All squares are normalized into [0,1];
Step 2: using first moment μ1With second moment μ2As the target of global square study, sentence is obtained using attention mechanism
Indicate vs, then the estimated value of square, such as μ are obtained with different full articulamentums1Estimated value μ '1It can obtain as follows:
WhereinFor the weight of full articulamentum, vsIt is sentence expression
Step 3: and then each sample x is defined in global first moment μ1On loss it is as follows:
lμ1(x)=(μ '1-μ1)2
Step 4: obtaining global second moment using identical calculation and loseAnd then obtain global loss:
Step 5: the subject word emotion of a sentence is divided into two parts, i.e. left-half and right half part;If object
Word number is odd number, then median can be divided into left-half, calculate separately left-half and right half part respective one
Rank square and second moment;According to the calculating that global square loses, local moment loss is obtained:
Step 6: to lgloble(x) and llocal(x) it is weighted summation, to obtain context square study total losses.
4. the natural language sentiment analysis method according to claim 3 based on depth network, which is characterized in that for every
A sample x, one context square study loss l of context square emotion learning module definitionm(x) carry out auxiliary object word emotion point
The optimization of generic task;This loss loses l by the overall situationgloble(x) and local losses llocal(x) two parts form:
Wherein naIndicate the number of subject word in a sentence, lglobleIt (x) is the sum of the loss of context square in entire sentence, and
llocal(x) be the respective context square of left-half and right half part in sentence the sum of loss.
5. the natural language sentiment analysis method according to claim 3 based on depth network, which is characterized in that described
Output module, specific formula is as follows:
Wherein C is emotional category collection, and D is training set;y∈R|C|An only hot vector, i.e., its only on correct label point
Amount is 0, fc(x, θ) is model prediction as a result, λ is L2The weight of regular terms, and λmIt is context square recurrence learning loss lm(x)
Weight;
In the training process, training will be realized by minimizing loss L, obtain the parameter that model can be made to optimize;It is testing
In the process, prediction result f is obtained by the most optimized parameter that training process obtainsc(x, θ), the wherein maximum component institute of score value
Corresponding classification is exactly the classification predicted.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811409537.8A CN109543039B (en) | 2018-11-23 | 2018-11-23 | Natural language emotion analysis method based on deep network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811409537.8A CN109543039B (en) | 2018-11-23 | 2018-11-23 | Natural language emotion analysis method based on deep network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109543039A true CN109543039A (en) | 2019-03-29 |
CN109543039B CN109543039B (en) | 2022-04-08 |
Family
ID=65850342
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811409537.8A Active CN109543039B (en) | 2018-11-23 | 2018-11-23 | Natural language emotion analysis method based on deep network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109543039B (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110110323A (en) * | 2019-04-10 | 2019-08-09 | 北京明略软件系统有限公司 | A kind of text sentiment classification method and device, computer readable storage medium |
CN110134757A (en) * | 2019-04-19 | 2019-08-16 | 杭州电子科技大学 | A kind of event argument roles abstracting method based on bull attention mechanism |
CN110222827A (en) * | 2019-06-11 | 2019-09-10 | 苏州思必驰信息科技有限公司 | The training method of text based depression judgement network model |
CN110321563A (en) * | 2019-06-28 | 2019-10-11 | 浙江大学 | Text emotion analysis method based on mixing monitor model |
CN110349620A (en) * | 2019-06-28 | 2019-10-18 | 广州序科码生物技术有限责任公司 | One kind accurately identifying interaction of molecules and its polarity and directionality method from PubMed document |
CN110457480A (en) * | 2019-08-16 | 2019-11-15 | 国网天津市电力公司 | The construction method of fine granularity sentiment classification model based on interactive attention mechanism |
CN110490136A (en) * | 2019-08-20 | 2019-11-22 | 电子科技大学 | A kind of human body behavior prediction method of knowledge based distillation |
CN110990552A (en) * | 2019-12-18 | 2020-04-10 | 北京声智科技有限公司 | Method and device for determining operation sequence of natural language formula |
CN111160037A (en) * | 2019-12-02 | 2020-05-15 | 广州大学 | Fine-grained emotion analysis method supporting cross-language migration |
CN111259663A (en) * | 2020-01-14 | 2020-06-09 | 北京百度网讯科技有限公司 | Information processing method and device |
CN111488734A (en) * | 2020-04-14 | 2020-08-04 | 西安交通大学 | Emotional feature representation learning system and method based on global interaction and syntactic dependency |
CN111625652A (en) * | 2019-07-12 | 2020-09-04 | 杭州电子科技大学 | Attention neural network method based on multi-path dynamic mask |
CN111914185A (en) * | 2020-07-06 | 2020-11-10 | 华中科技大学 | Graph attention network-based text emotion analysis method in social network |
CN112035661A (en) * | 2020-08-24 | 2020-12-04 | 北京大学深圳研究生院 | Text emotion analysis method and system based on graph convolution network and electronic device |
WO2021012183A1 (en) * | 2019-07-23 | 2021-01-28 | 中山大学 | Deducible machine learning reading comprehension system, and storage medium |
CN112784532A (en) * | 2021-01-29 | 2021-05-11 | 电子科技大学 | Multi-head attention memory network for short text sentiment classification |
CN112784573A (en) * | 2021-01-25 | 2021-05-11 | 中南民族大学 | Text emotion content analysis method, device and equipment and storage medium |
CN113065331A (en) * | 2021-04-15 | 2021-07-02 | 上海金融期货信息技术有限公司 | Entity emotion recognition method and system based on entity context discrimination |
CN113157919A (en) * | 2021-04-07 | 2021-07-23 | 山东师范大学 | Sentence text aspect level emotion classification method and system |
CN113571097A (en) * | 2021-09-28 | 2021-10-29 | 之江实验室 | Speaker self-adaptive multi-view dialogue emotion recognition method and system |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107092596A (en) * | 2017-04-24 | 2017-08-25 | 重庆邮电大学 | Text emotion analysis method based on attention CNNs and CCR |
CN107578785A (en) * | 2017-09-05 | 2018-01-12 | 哈尔滨工业大学 | The continuous emotional feature analysis evaluation method of music based on Gamma distributional analysis |
CN107578092A (en) * | 2017-09-01 | 2018-01-12 | 广州智慧城市发展研究院 | A kind of emotion compounding analysis method and system based on mood and opinion mining |
CN108133038A (en) * | 2018-01-10 | 2018-06-08 | 重庆邮电大学 | A kind of entity level emotional semantic classification system and method based on dynamic memory network |
CN108399158A (en) * | 2018-02-05 | 2018-08-14 | 华南理工大学 | Attribute sensibility classification method based on dependency tree and attention mechanism |
CN108460009A (en) * | 2017-12-14 | 2018-08-28 | 中山大学 | The attention mechanism Recognition with Recurrent Neural Network text emotion analytic approach of embedded sentiment dictionary |
US10133791B1 (en) * | 2014-09-07 | 2018-11-20 | DataNovo, Inc. | Data mining and analysis system and method for legal documents |
-
2018
- 2018-11-23 CN CN201811409537.8A patent/CN109543039B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10133791B1 (en) * | 2014-09-07 | 2018-11-20 | DataNovo, Inc. | Data mining and analysis system and method for legal documents |
CN107092596A (en) * | 2017-04-24 | 2017-08-25 | 重庆邮电大学 | Text emotion analysis method based on attention CNNs and CCR |
CN107578092A (en) * | 2017-09-01 | 2018-01-12 | 广州智慧城市发展研究院 | A kind of emotion compounding analysis method and system based on mood and opinion mining |
CN107578785A (en) * | 2017-09-05 | 2018-01-12 | 哈尔滨工业大学 | The continuous emotional feature analysis evaluation method of music based on Gamma distributional analysis |
CN108460009A (en) * | 2017-12-14 | 2018-08-28 | 中山大学 | The attention mechanism Recognition with Recurrent Neural Network text emotion analytic approach of embedded sentiment dictionary |
CN108133038A (en) * | 2018-01-10 | 2018-06-08 | 重庆邮电大学 | A kind of entity level emotional semantic classification system and method based on dynamic memory network |
CN108399158A (en) * | 2018-02-05 | 2018-08-14 | 华南理工大学 | Attribute sensibility classification method based on dependency tree and attention mechanism |
Non-Patent Citations (1)
Title |
---|
梁斌等: "基于多注意力卷积神经网络的特定目标情感分析", 《计算机研究与发展》 * |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110110323A (en) * | 2019-04-10 | 2019-08-09 | 北京明略软件系统有限公司 | A kind of text sentiment classification method and device, computer readable storage medium |
CN110110323B (en) * | 2019-04-10 | 2022-11-11 | 北京明略软件系统有限公司 | Text emotion classification method and device and computer readable storage medium |
CN110134757A (en) * | 2019-04-19 | 2019-08-16 | 杭州电子科技大学 | A kind of event argument roles abstracting method based on bull attention mechanism |
CN110222827A (en) * | 2019-06-11 | 2019-09-10 | 苏州思必驰信息科技有限公司 | The training method of text based depression judgement network model |
CN110349620A (en) * | 2019-06-28 | 2019-10-18 | 广州序科码生物技术有限责任公司 | One kind accurately identifying interaction of molecules and its polarity and directionality method from PubMed document |
CN110321563A (en) * | 2019-06-28 | 2019-10-11 | 浙江大学 | Text emotion analysis method based on mixing monitor model |
CN111625652A (en) * | 2019-07-12 | 2020-09-04 | 杭州电子科技大学 | Attention neural network method based on multi-path dynamic mask |
WO2021012183A1 (en) * | 2019-07-23 | 2021-01-28 | 中山大学 | Deducible machine learning reading comprehension system, and storage medium |
CN110457480A (en) * | 2019-08-16 | 2019-11-15 | 国网天津市电力公司 | The construction method of fine granularity sentiment classification model based on interactive attention mechanism |
CN110490136A (en) * | 2019-08-20 | 2019-11-22 | 电子科技大学 | A kind of human body behavior prediction method of knowledge based distillation |
CN111160037A (en) * | 2019-12-02 | 2020-05-15 | 广州大学 | Fine-grained emotion analysis method supporting cross-language migration |
CN110990552A (en) * | 2019-12-18 | 2020-04-10 | 北京声智科技有限公司 | Method and device for determining operation sequence of natural language formula |
CN111259663A (en) * | 2020-01-14 | 2020-06-09 | 北京百度网讯科技有限公司 | Information processing method and device |
US11775776B2 (en) | 2020-01-14 | 2023-10-03 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for processing information |
CN111259663B (en) * | 2020-01-14 | 2023-05-26 | 北京百度网讯科技有限公司 | Information processing method and device |
CN111488734A (en) * | 2020-04-14 | 2020-08-04 | 西安交通大学 | Emotional feature representation learning system and method based on global interaction and syntactic dependency |
CN111488734B (en) * | 2020-04-14 | 2022-02-22 | 西安交通大学 | Emotional feature representation learning system and method based on global interaction and syntactic dependency |
CN111914185A (en) * | 2020-07-06 | 2020-11-10 | 华中科技大学 | Graph attention network-based text emotion analysis method in social network |
CN111914185B (en) * | 2020-07-06 | 2024-03-22 | 华中科技大学 | Text emotion analysis method in social network based on graph attention network |
CN112035661A (en) * | 2020-08-24 | 2020-12-04 | 北京大学深圳研究生院 | Text emotion analysis method and system based on graph convolution network and electronic device |
CN112784573A (en) * | 2021-01-25 | 2021-05-11 | 中南民族大学 | Text emotion content analysis method, device and equipment and storage medium |
CN112784573B (en) * | 2021-01-25 | 2023-12-19 | 中南民族大学 | Text emotion content analysis method, device, equipment and storage medium |
CN112784532B (en) * | 2021-01-29 | 2022-09-02 | 电子科技大学 | Multi-head attention memory system for short text sentiment classification |
CN112784532A (en) * | 2021-01-29 | 2021-05-11 | 电子科技大学 | Multi-head attention memory network for short text sentiment classification |
CN113157919A (en) * | 2021-04-07 | 2021-07-23 | 山东师范大学 | Sentence text aspect level emotion classification method and system |
CN113065331A (en) * | 2021-04-15 | 2021-07-02 | 上海金融期货信息技术有限公司 | Entity emotion recognition method and system based on entity context discrimination |
CN113571097A (en) * | 2021-09-28 | 2021-10-29 | 之江实验室 | Speaker self-adaptive multi-view dialogue emotion recognition method and system |
Also Published As
Publication number | Publication date |
---|---|
CN109543039B (en) | 2022-04-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109543039A (en) | A kind of natural language sentiment analysis method based on depth network | |
CN107239446B (en) | A kind of intelligence relationship extracting method based on neural network Yu attention mechanism | |
CN103207855B (en) | For the fine granularity sentiment analysis system and method for product review information | |
Tay et al. | Densely connected attention propagation for reading comprehension | |
CN106649275A (en) | Relation extraction method based on part-of-speech information and convolutional neural network | |
CN106855853A (en) | Entity relation extraction system based on deep neural network | |
CN105976056A (en) | Information extraction system based on bidirectional RNN | |
CN105975555A (en) | Bidirectional recursive neural network-based enterprise abbreviation extraction method | |
CN106598950A (en) | Method for recognizing named entity based on mixing stacking model | |
CN106202032A (en) | A kind of sentiment analysis method towards microblogging short text and system thereof | |
CN110110754B (en) | Method for classifying imbalance problems based on cost local generalization errors | |
CN109446331A (en) | A kind of text mood disaggregated model method for building up and text mood classification method | |
CN112561718A (en) | Case microblog evaluation object emotion tendency analysis method based on BilSTM weight sharing | |
CN110415071A (en) | A kind of competing product control methods of automobile based on opining mining analysis | |
Chen et al. | cs@ DravidianLangTech-EACL2021: Offensive language identification based on multilingual BERT model | |
CN108427865A (en) | A method of prediction LncRNA and environmental factor incidence relation | |
CN113177417A (en) | Trigger word recognition method based on hybrid neural network and multi-stage attention mechanism | |
Luu et al. | A Multiple Choices Reading Comprehension Corpus for Vietnamese Language Education | |
Qi et al. | MS-transformer: introduce multiple structural priors into a unified transformer for encoding sentences | |
Li | Application of an Internet of Things Oriented Network Education Platform in English Language Teaching | |
Xu et al. | AHRNN: Attention‐Based Hybrid Robust Neural Network for emotion recognition | |
Zhao et al. | Detection of Chinese Grammatical Errors with Context Representation | |
Xu et al. | BiRNN-DKT: transfer bi-directional LSTM RNN for knowledge tracing | |
Chandrasekaran et al. | Automating Transfer Credit Assessment in Student Mobility--A Natural Language Processing-based Approach | |
Umapathy et al. | Segmentation of Floorplans and Heritage Sites: An Approach to Unbalanced Dataset |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |