CN111325027A - Sparse data-oriented personalized emotion analysis method and device - Google Patents

Sparse data-oriented personalized emotion analysis method and device Download PDF

Info

Publication number
CN111325027A
CN111325027A CN202010102417.4A CN202010102417A CN111325027A CN 111325027 A CN111325027 A CN 111325027A CN 202010102417 A CN202010102417 A CN 202010102417A CN 111325027 A CN111325027 A CN 111325027A
Authority
CN
China
Prior art keywords
document
emotion
sentence
user
representation
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
Application number
CN202010102417.4A
Other languages
Chinese (zh)
Other versions
CN111325027B (en
Inventor
周德宇
张朦
张林海
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.)
Southeast University
Original Assignee
Southeast 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 Southeast University filed Critical Southeast University
Priority to CN202010102417.4A priority Critical patent/CN111325027B/en
Publication of CN111325027A publication Critical patent/CN111325027A/en
Application granted granted Critical
Publication of CN111325027B publication Critical patent/CN111325027B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

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

The invention discloses a sparse data-oriented personalized emotion analysis method and device, which are used for grouping users with similar scoring habits and enhancing user representation by utilizing grouping information to realize personalized emotion analysis. The method comprises the following steps: preprocessing a document; calculating to obtain an emotion scoring basis by using a basic emotion analysis model based on a deep neural network; calculating to obtain emotion scoring offset and fluctuation by using a group-based personalized analysis model; and calculating the final emotion scoring by combining the emotion scoring basis and the emotion scoring offset. Compared with the prior personalized emotion analysis method, the method can learn to obtain good user representation under the condition that the text data of the user is sparse, can effectively model the user in the personalized emotion analysis, and can more accurately perform the personalized emotion analysis.

Description

Sparse data-oriented personalized emotion analysis method and device
Technical Field
The invention relates to emotion analysis of a text by using user text data under the condition of sparse data, and belongs to the technical field of machine learning.
Background
The user generated text sentiment analysis is intended to calculate a corresponding sentiment score (e.g., satisfaction) based on a text (e.g., a Twitter or a shopping comment) composed by the user. Traditional emotion analysis methods consider the mapping between text and emotion scores to be the same for all users, i.e., do not distinguish individual differences between users. However, such an assumption is not in accordance with the actual situation. Because the emotion expression modes of users are different due to different educational backgrounds, social experiences and the like, the personalized emotion analysis for the users is very necessary. However, some existing personalized emotion analysis methods generally use a fixed-dimension user vector to represent each user, the user vector is generally randomly initialized and then learned by the network, and the user representation mode has strong dependence on data and the network. According to network statistics, most users of Twitter send Twitter rarely, and nearly 80% of Twitter is sent by 10% of active users. This means that in real life, the situation that user data is sparse often exists, so it has very important social meaning to solve the problem of personalized emotion analysis under the sparse environment of data.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides a sparse data-oriented personalized emotion analysis method and device, which can solve the problem of data sparsity in current personalized emotion analysis.
The technical scheme is as follows: in order to achieve the above purpose, the method for analyzing personalized emotion facing to sparse data, provided by the invention, comprises the following steps:
(1) preprocessing a document;
(2) using a basic emotion analysis model based on a deep neural network, taking words of a document as input, respectively calculating semantic representation of each sentence in the document and semantic representation of the document through sentence-level semantic representation learning and document-level semantic representation learning, and taking a numerical value obtained by mapping the semantic representation of the document as an emotion scoring basis;
(3) using a group-based personalized emotion analysis model, taking semantic representation, user vectors and global group vectors of a document obtained by a deep neural network-based basic emotion analysis model as input, respectively calculating and obtaining user representation of each sentence and user representation of the document in the document through sentence-level user representation learning and document-level user representation learning, cascading the user representation of the document and the semantic representation obtained by the deep neural network-based basic emotion analysis model as final representation of the document, and mapping the final representation of the document to two values to be used as emotion scoring offset and fluctuation respectively; the emotion scoring offset is used for final scoring calculation, and emotion scoring fluctuation is used for network optimization;
(4) and adding the emotion scoring basis and the emotion scoring offset to obtain a final emotion scoring.
Further, the document preprocessing in the step (1) comprises: the method comprises the steps of segmenting words of a document, and filtering out stop words in the document and words which only appear once in a processed data set.
Further, the step (2) of calculating the emotion scoring basis by using the basic emotion analysis model based on the deep neural network comprises the following steps:
(2.1) aiming at each word in the sentence, mapping the word into a word vector trained in advance, and then coding each word in the sentence by using a bidirectional long and short memory network Bi-LSTM to obtain a corresponding hidden state of each word; calculating a weight for each word using an attention mechanism; finally, weighting and summing each word to obtain semantic representation of each sentence;
(2.2) regarding each sentence in the document, taking semantic representation of the sentence as input, and coding each sentence in the document by utilizing Bi-LSTM to obtain a corresponding hidden state of each sentence; calculating a weight for each sentence using an attention mechanism; finally, weighting and summing each sentence to obtain semantic representation of the document;
(2.3) mapping the semantic representation of the document level to a numerical value, namely emotion scoring basis, by using a multi-layer perceptron.
Further, the calculating of the emotion scoring offset and fluctuation by using the group-based personalized emotion analysis model in the step (3) comprises:
(3.1) calculating to obtain the user hidden state of each word on the basis of the hidden state of each word in the Bi-LSTM, the group global vector and the user vector corresponding to the document; calculating the weight of the user hidden state corresponding to each word by using an attention mechanism; finally, weighting and summing the hidden states of the users corresponding to each word to obtain the user representation of the sentence;
(3.2) calculating to obtain the user hidden state of each sentence on the basis of the hidden state of each sentence, the group global vector and the sentence user representation in the Bi-LSTM; calculating a weight of the hidden state of each sentence user using an attention mechanism; finally, weighting and summing the hidden states of the users of each sentence to obtain the user representation of the document;
(3.3) concatenating the semantic representation of the document and the user representation as a final representation of the document;
and (3.4) respectively mapping the final document representation to two numerical values, namely emotion scoring offset calculation and emotion scoring fluctuation by using two multilayer perceptrons.
Further, the user of the sentence is represented as:
Figure BDA0002387303180000021
wherein ,
Figure BDA0002387303180000031
Figure BDA0002387303180000032
Figure BDA0002387303180000033
Figure BDA0002387303180000034
Figure BDA0002387303180000035
Figure BDA0002387303180000036
ekis the global vector for the k-th group,
Figure BDA0002387303180000037
is the word wijThe corresponding hidden state, u is the user vector corresponding to the document,
Figure BDA0002387303180000038
and
Figure BDA0002387303180000039
is a model parameter, softmax (·) is a normalized logistic regression function, and tanh (·) is a hyperbolic tangent activation function.
Further, the user of the document is represented as:
Figure BDA00023873031800000310
wherein ,
Figure BDA00023873031800000311
Figure BDA00023873031800000312
Figure BDA00023873031800000313
Figure BDA00023873031800000314
Figure BDA00023873031800000315
Figure BDA00023873031800000316
Figure BDA00023873031800000317
Figure BDA00023873031800000318
is the sentence siThe corresponding hidden state is set to be in a hidden state,
Figure BDA00023873031800000319
and
Figure BDA00023873031800000320
are the model parameters.
Further, using the joint loss to optimize the network includes: using mean square error loss for a basic emotion analysis model based on a deep neural network; the group-based personalized emotion analysis model is subjected to Gaussian penalty loss, so that emotion fluctuation is learned from a loss function, and the influence of samples with overlarge fluctuation on a network is reduced; adding a group vector-based penalty term to enable the group vector obtained by learning to have discriminability; l joining network parameters2The regularization term avoids overfitting.
Further, the loss function of the network is defined as:
Figure BDA0002387303180000041
wherein ,
Figure BDA0002387303180000042
Figure BDA0002387303180000043
Figure BDA0002387303180000044
λ‖Θ‖2is a network parameter L2The regularization term, T is the number of samples,
Figure BDA0002387303180000045
is the true sentiment score for the tth document,
Figure BDA0002387303180000046
is the output result of the basic emotion analysis model based on the deep neural network of the t document, ytThe group-based personalized emotion analysis model for the tth document outputs a final emotion score,
Figure BDA0002387303180000047
the emotion scoring fluctuation of the output of the t document based on the personalized emotion analysis model of the group, wherein I is an identity matrix, and E is { E ═ E }1,…,eKIs a matrix of group vectors, | - |FIs the Frobenius norm of the matrix.
Based on the same inventive concept, the individualized emotion analysis device facing to sparse data comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein the computer program realizes the individualized emotion analysis method facing to sparse data when being loaded to the processor.
Has the advantages that: compared with the existing personalized emotion analysis method, the method can effectively solve the problem of sparse user data in the personalized emotion analysis which generally exists in real life, and the personalized emotion analysis performance can be improved by establishing the emotion score of one user as Gaussian distribution and considering the offset and fluctuation of the emotion score of the user.
Drawings
FIG. 1 is a flow chart of a method in an embodiment of the invention.
Fig. 2 is a schematic diagram of calculation in user representation learning at a sentence level of a user Encoder (U-Encoder) in the embodiment of the present invention.
Detailed Description
The present invention is further illustrated by the following description in conjunction with the accompanying drawings and the specific embodiments, it is to be understood that these examples are given solely for the purpose of illustration and are not intended as a definition of the limits of the invention, since various equivalent modifications will occur to those skilled in the art upon reading the present invention and fall within the limits of the appended claims.
The problem can be described as follows: for a document and a user u thereof, the personalized emotion analysis task is to predict an emotion score y corresponding to the document (such as a satisfaction score, for example, 1-5 scores, 1 score indicates dissatisfaction, and 5 scores indicates satisfaction). According to the observation, since the user has a generalized emotion expression mode, an emotion scoring basis y can be obtainedbThe users have the deviation y of sentiment scoring due to individual differencesAnd a fluctuation σ2Therefore, the predicted sentiment score can be modeled as a value that follows a Gaussian distribution
Figure BDA0002387303180000051
So the personalized sentiment analysis problem can be converted into yb，ys and σ2The prediction problem of (1).
The individualized emotion analysis method facing to sparse data, disclosed by the embodiment of the invention, as shown in fig. 1, mainly comprises the following steps: .
S1: document preprocessing: segmenting the document, removing stop words in the document and words appearing only once in the data set to obtain a processed document d, wherein d comprises M sentences, and each sentence ciComprising NiA word.
S2: calculating to obtain an emotion scoring basis by using a basic emotion analysis model based on a deep neural network, and specifically comprising the following steps of:
(1) semantic representation learning at sentence level. Firstly, each word in a sentence is mapped into a word vector trained in advance, and then a sentence can be represented as
Figure BDA0002387303180000052
Each word in the sentence is then sequence encoded using Bi-LSTM:
Figure BDA0002387303180000053
Figure BDA0002387303180000054
then will be
Figure BDA0002387303180000055
And
Figure BDA0002387303180000056
splicing to obtain a word wijHidden state of
Figure BDA0002387303180000057
Since each word contributes differently to the semantic representation of the sentence, the attention mechanism is used so that important words in the sentence have higher weights, which is calculated as follows:
Figure BDA0002387303180000058
Figure BDA0002387303180000059
wherein
Figure BDA00023873031800000510
Is the word wijThe weight of (a) is determined,
Figure BDA00023873031800000511
and
Figure BDA00023873031800000512
are the model parameters. So the semantics s of the sentenceiThe representation can be obtained by weighted summation of the hidden states of the words:
Figure BDA0002387303180000061
(2) semantic representation learning at the document level. Bi-LSTM encodes each sentence representation in the document into a corresponding hidden state for each sentence:
Figure BDA0002387303180000062
Figure BDA0002387303180000063
then will be
Figure BDA0002387303180000064
And
Figure BDA0002387303180000065
splicing to obtain the hidden state of the sentence
Figure BDA0002387303180000066
Since different sentences contribute differently to the final semantic representation of the document, the weight of each sentence is further calculated using an attention mechanism, as follows:
Figure BDA0002387303180000067
Figure BDA0002387303180000068
wherein
Figure BDA0002387303180000069
Is the sentence siThe weight of (a) is determined,
Figure BDA00023873031800000610
and
Figure BDA00023873031800000611
are the model parameters. So the document semantic representation rbThe hidden state of a sentence can be weighted and summed to obtain:
Figure BDA00023873031800000612
(3) and calculating the emotion scoring basis. A semantic representation at the document level is mapped to a numerical value using a multi-layered perceptron,
namely emotion scoring basis:
yb＝MLP(rb)
s3: and calculating the emotion scoring offset and fluctuation by using a group-based personalized emotion analysis model. Assume that there are a total of K groups, each group having a corresponding global vector ekThis vector is initialized randomly and learned by network adaptation. The method specifically comprises the following steps:
(1) sentence-level users represent learning. With each word
Figure BDA00023873031800000613
Hidden state of
Figure BDA00023873031800000614
Based on the user vector U corresponding to the document, a user Encoder (U-Encoder) is designed to enhance the user representation, as shown in fig. 2, and includes: calculating the word wijCorresponding group representation
Figure BDA00023873031800000615
And computing an enhanced user representation
Figure BDA00023873031800000616
First to compute the relationship between words and groups, an attention mechanism is used to compute the words wijCorresponding group representation
Figure BDA00023873031800000617
If the word is associated with the groupIf the habit is more similar, the word has a higher weight, which is specifically calculated as follows:
Figure BDA0002387303180000071
Figure BDA0002387303180000072
then, since the learned user representation may be unreliable under the condition of sparse data, a group-representation-based mechanism is used to enhance the user representation, and the specific calculation manner is as follows:
Figure BDA0002387303180000073
Figure BDA0002387303180000074
wherein
Figure BDA0002387303180000075
Is an enhanced user representation for each word,
Figure BDA0002387303180000076
control of
Figure BDA0002387303180000077
To pijThe influence of (a) on the performance of the device,
Figure BDA0002387303180000078
and
Figure BDA0002387303180000079
are the model parameters. It should be noted that the conventional recurrent neural network usually calculates the hidden state of each word sequentially, but the U-Encoder is not a sequential structure, so that the user representations can be simultaneously calculated in parallel, and the calculation performance can be effectively improved.
Since different words contribute differently to the user representation of the sentence, the attention mechanism is further used to calculate the weight of the user hidden state corresponding to each word:
Figure BDA00023873031800000710
Figure BDA00023873031800000711
wherein
Figure BDA00023873031800000712
Is a user representation
Figure BDA00023873031800000713
The weight of (a) is determined,
Figure BDA00023873031800000714
and
Figure BDA00023873031800000715
are the model parameters. Finally, weighting and summing the hidden states of the users corresponding to each word to obtain the user representation v of the sentencei：
Figure BDA00023873031800000716
(2) The users at the document level represent learning. In the hidden state of each sentence
Figure BDA00023873031800000717
And corresponding user representation viOn the basis, the U-Encoder is used for coding, and the method comprises the following steps: calculating a sentence siCorresponding group representation
Figure BDA00023873031800000718
And computing an enhanced user representation
Figure BDA00023873031800000719
First to calculate the relationship between words and groups, attention is usedMechanism calculation sentence siCorresponding group representation
Figure BDA00023873031800000720
The specific calculation is as follows:
Figure BDA00023873031800000721
Figure BDA00023873031800000722
then, since the learned user representation may be unreliable under the condition of sparse data, a group-representation-based mechanism is used to enhance the user representation, and the specific calculation manner is as follows:
Figure BDA0002387303180000081
Figure BDA0002387303180000082
wherein
Figure BDA0002387303180000083
Is an enhanced user representation for each sentence,
Figure BDA0002387303180000084
control of
Figure BDA0002387303180000085
To pair
Figure BDA0002387303180000086
The influence of (a) on the performance of the device,
Figure BDA0002387303180000087
and
Figure BDA0002387303180000088
are the model parameters.
Since different sentences contribute differently to the user representation of the document, an attention mechanism is further used to calculate the weight of the user's hidden state for each sentence:
Figure BDA0002387303180000089
Figure BDA00023873031800000810
wherein
Figure BDA00023873031800000811
Is a user representation
Figure BDA00023873031800000812
The weight of (a) is determined,
Figure BDA00023873031800000813
and
Figure BDA00023873031800000814
are the model parameters. Finally, weighting and summing the hidden states of the users corresponding to each sentence to obtain the user representation r of the documentu i：
Figure BDA00023873031800000815
(3) And calculating emotion scoring offset and fluctuation. Representing a user of a document ruAnd semantic representation r obtained by basic emotion analysis model based on deep neural networkbAnd (3) the cascaded final representations are respectively mapped to the final representation of the document level by using two multi-layer perceptrons:
ys＝MLP([ru,rb])
σ2＝MLP([ru,rb])
s4: base y combined with sentiment scoringbSentiment score offset ysAnd emotional score fluctuation sigma2The final prediction result can be obtained
Figure BDA00023873031800000816
Due to yb+ysIs an unbiased estimate of y, so will yb+ysAs the value of y, i.e. y ═ yb+ys. And the emotional expression fluctuation is further used as a constraint in the optimization of the model.
Since the present invention can be considered a multitasking architecture, we use joint loss to optimize the network. For a traditional generalized emotion analysis model, we use the mean square error as a loss function:
Figure BDA00023873031800000817
where T is the number of samples and,
Figure BDA00023873031800000818
is the true sentiment score for the tth document,
Figure BDA00023873031800000819
is the output result of the basic emotion analysis model based on the deep neural network of the t document.
For the group-based personalized emotion analysis model, a Gaussian penalty function is used as a loss function, and the loss function comprises two parts: (1) residual regression based on the emotion fluctuation is used for learning from the loss function to obtain the emotion fluctuation, and the influence of a sample with overlarge fluctuation on a network is reduced; (2) and a regularization term of emotion fluctuation to avoid the predicted emotion fluctuation from being too large. The specific definition is as follows:
Figure BDA0002387303180000091
wherein ytThe group-based personalized emotion analysis model for the tth document outputs a final emotion score,
Figure BDA0002387303180000092
and fluctuating the emotion scoring output by the group-based personalized emotion analysis model of the tth document.
In addition, in order to make the global vectors of different groups have a certain distinctiveness, a group vector-based penalty term is further introduced:
Figure BDA0002387303180000093
where I is the identity matrix, E ═ E1,…,eKIs a matrix of group vectors, | - |FIs the Frobenius norm of the matrix.
The loss function of the network is defined as follows:
Figure BDA0002387303180000094
wherein λ‖Θ‖2Is a network parameter L2A regularization term.
In the experimental process, the experimental parameters are set as follows: pre-trained GloVe was used as word embedding, with the hidden state dimensions of word embedding, user and word all being 300. Dropout rate of 0.5, learning rate of 0.001, L2The weight lost is 0.01. The network was optimized using Adam with a number of groups of 6. The RMSE obtained on the user data sparse data set of the Yelp 2013 is 0.7166, the MAE is 0.5501, and the performance is superior to that of the existing personalized emotion analysis method.
Compared with the existing method, the individualized emotion analysis method facing sparse data considers the problem of sparse user data in individualized emotion analysis, which is ubiquitous in real life, and the method further enhances user representation by grouping users with similar emotion expression modes and utilizing group information. Meanwhile, the emotion score corresponding to the user text is generally considered to be a single numerical value by the original method, the emotion score of one user is established to be uniform Gaussian distribution by the method, and the deviation and fluctuation of the emotion score of the user are considered, so that the personalized emotion analysis performance is favorably improved.
Based on the same inventive concept, the individualized emotion analysis device for sparse data disclosed by the embodiment of the invention comprises a memory, a processor and a computer program which is stored on the memory and can run on the processor, wherein the computer program realizes the individualized emotion analysis method for sparse data when being loaded to the processor.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.

Claims (9)

1. A sparse data-oriented personalized emotion analysis method is characterized by comprising the following steps:
(1) preprocessing a document;
(2) using a basic emotion analysis model based on a deep neural network, taking words of a document as input, respectively calculating semantic representation of each sentence in the document and semantic representation of the document through sentence-level semantic representation learning and document-level semantic representation learning, and taking a numerical value obtained by mapping the semantic representation of the document as an emotion scoring basis;
(3) using a group-based personalized emotion analysis model, taking semantic representation, user vectors and global group vectors of a document obtained by a deep neural network-based basic emotion analysis model as input, respectively calculating and obtaining user representation of each sentence and user representation of the document in the document through sentence-level user representation learning and document-level user representation learning, cascading the user representation of the document and the semantic representation obtained by the deep neural network-based basic emotion analysis model as final representation of the document, and mapping the final representation of the document to two values to be used as emotion scoring offset and fluctuation respectively; the emotion scoring offset is used for final scoring calculation, and emotion scoring fluctuation is used for network optimization;
(4) and adding the emotion scoring basis and the emotion scoring offset to obtain a final emotion scoring.
2. The sparse-data-oriented personalized emotion analysis method of claim 1, wherein the document preprocessing in step (1) comprises: the method comprises the steps of segmenting words of a document, and filtering out stop words in the document and words which only appear once in a processed data set.
3. The sparse data-oriented personalized emotion analysis method of claim 1, wherein the step (2) of calculating the emotion scoring basis by using the deep neural network-based basic emotion analysis model comprises:
(2.1) aiming at each word in the sentence, mapping the word into a word vector trained in advance, and then coding each word in the sentence by using a bidirectional long and short memory network Bi-LSTM to obtain a corresponding hidden state of each word; calculating a weight for each word using an attention mechanism; finally, weighting and summing each word to obtain semantic representation of each sentence;
(2.2) regarding each sentence in the document, taking semantic representation of the sentence as input, and coding each sentence in the document by utilizing Bi-LSTM to obtain a corresponding hidden state of each sentence; calculating a weight for each sentence using an attention mechanism; finally, weighting and summing each sentence to obtain semantic representation of the document;
(2.3) mapping the semantic representation of the document level to a numerical value, namely emotion scoring basis, by using a multi-layer perceptron.
4. The sparse data-oriented personalized emotion analysis method of claim 1, wherein the step (3) of calculating emotion scoring offsets and fluctuations by using the group-based personalized emotion analysis model comprises:
(3.1) calculating to obtain the user hidden state of each word on the basis of the hidden state of each word in the Bi-LSTM, the group global vector and the user vector corresponding to the document; calculating the weight of the user hidden state corresponding to each word by using an attention mechanism; finally, weighting and summing the hidden states of the users corresponding to each word to obtain the user representation of the sentence;
(3.2) calculating to obtain the user hidden state of each sentence on the basis of the hidden state of each sentence, the group global vector and the sentence user representation in the Bi-LSTM; calculating a weight of the hidden state of each sentence user using an attention mechanism; finally, weighting and summing the hidden states of the users of each sentence to obtain the user representation of the document;
(3.3) concatenating the semantic representation of the document and the user representation as a final representation of the document;
and (3.4) respectively mapping the final document representation to two numerical values, namely emotion scoring offset calculation and emotion scoring fluctuation by using two multilayer perceptrons.
5. The sparse data-oriented personalized emotion analysis method of claim 4, wherein the user representation of the sentence is:
Figure FDA0002387303170000021
wherein ,
Figure FDA0002387303170000022
Figure FDA0002387303170000023
Figure FDA0002387303170000024
Figure FDA0002387303170000025
Figure FDA0002387303170000026
Figure FDA0002387303170000027
ekis the global vector for the k-th group,
Figure FDA0002387303170000028
is the word wijThe corresponding hidden state, u is the user vector corresponding to the document,
Figure FDA0002387303170000029
and
Figure FDA00023873031700000210
is a model parameter, softmax (·) is a normalized logistic regression function, and tanh (·) is a hyperbolic tangent activation function.
6. The sparse data-oriented personalized emotion analysis method of claim 5, wherein the user representation of the document is:
Figure FDA00023873031700000211
wherein ,
Figure FDA0002387303170000031
Figure FDA0002387303170000032
Figure FDA0002387303170000033
Figure FDA0002387303170000034
Figure FDA0002387303170000035
Figure FDA0002387303170000036
Figure FDA0002387303170000037
Figure FDA0002387303170000038
is the sentence siThe corresponding hidden state is set to be in a hidden state,
Figure FDA0002387303170000039
and
Figure FDA00023873031700000310
are the model parameters.
7. The sparse data-oriented personalized emotion analysis method of claim 1, wherein joint loss is used for optimizing the network, and comprises: using mean square error loss for a basic emotion analysis model based on a deep neural network; the group-based personalized emotion analysis model is subjected to Gaussian penalty loss, so that emotion fluctuation is learned from a loss function, and the influence of samples with overlarge fluctuation on a network is reduced; adding a group vector-based penalty term to enable the group vector obtained by learning to have discriminability; l joining network parameters2The regularization term avoids overfitting.
8. The sparse data-oriented personalized emotion analysis method of claim 7, wherein a loss function of the network is defined as:
Figure FDA00023873031700000311
wherein ,
Figure FDA00023873031700000312
Figure FDA00023873031700000313
Figure FDA00023873031700000314
λ‖Θ‖2is a network parameter L2The regularization term, T is the number of samples,
Figure FDA00023873031700000315
is the true sentiment score for the tth document,
Figure FDA00023873031700000316
is the output result of the basic emotion analysis model based on the deep neural network of the t document, ytThe group-based personalized emotion analysis model for the tth document outputs a final emotion score,
Figure FDA00023873031700000317
the emotion scoring fluctuation of the output of the t document based on the personalized emotion analysis model of the group, wherein I is an identity matrix, and E is { E ═ E }1,…,eKIs a matrix of group vectors, | - |FIs the Frobenius norm of the matrix.
9. A sparse data oriented personalized emotion analysis apparatus comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the computer program when loaded into the processor implements the sparse data oriented personalized emotion analysis method according to any of claims 1-8.
CN202010102417.4A 2020-02-19 2020-02-19 Sparse data-oriented personalized emotion analysis method and device Active CN111325027B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010102417.4A CN111325027B (en) 2020-02-19 2020-02-19 Sparse data-oriented personalized emotion analysis method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010102417.4A CN111325027B (en) 2020-02-19 2020-02-19 Sparse data-oriented personalized emotion analysis method and device

Publications (2)

Publication Number Publication Date
CN111325027A true CN111325027A (en) 2020-06-23
CN111325027B CN111325027B (en) 2023-04-28

Family

ID=71171083

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010102417.4A Active CN111325027B (en) 2020-02-19 2020-02-19 Sparse data-oriented personalized emotion analysis method and device

Country Status (1)

Country Link
CN (1) CN111325027B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112990430A (en) * 2021-02-08 2021-06-18 辽宁工业大学 Group division method and system based on long-time and short-time memory network

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108363753A (en) * 2018-01-30 2018-08-03 南京邮电大学 Comment text sentiment classification model is trained and sensibility classification method, device and equipment
CN108446275A (en) * 2018-03-21 2018-08-24 北京理工大学 Long text emotional orientation analytical method based on attention bilayer LSTM
CN108804417A (en) * 2018-05-21 2018-11-13 山东科技大学 A kind of documentation level sentiment analysis method based on specific area emotion word
CN108984724A (en) * 2018-07-10 2018-12-11 凯尔博特信息科技（昆山）有限公司 It indicates to improve particular community emotional semantic classification accuracy rate method using higher-dimension
CN110334187A (en) * 2019-07-09 2019-10-15 昆明理工大学 Burmese sentiment analysis method and device based on transfer learning

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108363753A (en) * 2018-01-30 2018-08-03 南京邮电大学 Comment text sentiment classification model is trained and sensibility classification method, device and equipment
CN108446275A (en) * 2018-03-21 2018-08-24 北京理工大学 Long text emotional orientation analytical method based on attention bilayer LSTM
CN108804417A (en) * 2018-05-21 2018-11-13 山东科技大学 A kind of documentation level sentiment analysis method based on specific area emotion word
CN108984724A (en) * 2018-07-10 2018-12-11 凯尔博特信息科技（昆山）有限公司 It indicates to improve particular community emotional semantic classification accuracy rate method using higher-dimension
CN110334187A (en) * 2019-07-09 2019-10-15 昆明理工大学 Burmese sentiment analysis method and device based on transfer learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
胡均毅等: "基于情感评分的分层文本表示情感分类方法", 《计算机工程》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112990430A (en) * 2021-02-08 2021-06-18 辽宁工业大学 Group division method and system based on long-time and short-time memory network
CN112990430B (en) * 2021-02-08 2021-12-03 辽宁工业大学 Group division method and system based on long-time and short-time memory network

Also Published As

Publication number Publication date
CN111325027B (en) 2023-04-28

Similar Documents

Publication Publication Date Title
Jin et al. Multi-task learning model based on multi-scale CNN and LSTM for sentiment classification
CN108536679B (en) Named entity recognition method, device, equipment and computer readable storage medium
Qian et al. Hierarchical CVAE for fine-grained hate speech classification
CN110796160B (en) Text classification method, device and storage medium
KR20180125905A (en) Method and apparatus for classifying a class to which a sentence belongs by using deep neural network
CN111061843A (en) Knowledge graph guided false news detection method
CN109189933A (en) A kind of method and server of text information classification
CN112749274B (en) Chinese text classification method based on attention mechanism and interference word deletion
CN112883714B (en) ABSC task syntactic constraint method based on dependency graph convolution and transfer learning
CN112527966A (en) Network text emotion analysis method based on Bi-GRU neural network and self-attention mechanism
CN109214006A (en) The natural language inference method that the hierarchical semantic of image enhancement indicates
CN111538841B (en) Comment emotion analysis method, device and system based on knowledge mutual distillation
Zhang et al. A BERT fine-tuning model for targeted sentiment analysis of Chinese online course reviews
CN114528490B (en) Self-supervision sequence recommendation method based on long-term and short-term interests of user
CN113704460A (en) Text classification method and device, electronic equipment and storage medium
CN111259147B (en) Sentence-level emotion prediction method and system based on self-adaptive attention mechanism
Luo et al. EmotionX-DLC: self-attentive BiLSTM for detecting sequential emotions in dialogue
CN111046157A (en) Universal English man-machine conversation generation method and system based on balanced distribution
CN111325027B (en) Sparse data-oriented personalized emotion analysis method and device
CN113761910A (en) Comment text fine-grained emotion analysis method integrating emotional characteristics
CN116644759A (en) Method and system for extracting aspect category and semantic polarity in sentence
CN111368524A (en) Microblog viewpoint sentence recognition method based on self-attention bidirectional GRU and SVM
CN113779244B (en) Document emotion classification method and device, storage medium and electronic equipment
CN115659990A (en) Tobacco emotion analysis method, device and medium
CN115309894A (en) Text emotion classification method and device based on confrontation training and TF-IDF

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