CN112396094B - Multi-task active learning method and system simultaneously used for emotion classification and regression - Google Patents

Multi-task active learning method and system simultaneously used for emotion classification and regression Download PDF

Info

Publication number
CN112396094B
CN112396094B CN202011204366.2A CN202011204366A CN112396094B CN 112396094 B CN112396094 B CN 112396094B CN 202011204366 A CN202011204366 A CN 202011204366A CN 112396094 B CN112396094 B CN 112396094B
Authority
CN
China
Prior art keywords
sample
model
task
value
label
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
CN202011204366.2A
Other languages
Chinese (zh)
Other versions
CN112396094A (en
Inventor
伍冬睿
蒋雪
孟璐斌
黄剑
曾志刚
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.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
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 Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202011204366.2A priority Critical patent/CN112396094B/en
Publication of CN112396094A publication Critical patent/CN112396094A/en
Application granted granted Critical
Publication of CN112396094B publication Critical patent/CN112396094B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques

Abstract

The invention discloses a multi-task active learning method and system simultaneously used for emotion classification and regression, and belongs to the field of emotion calculation. The method combines the value measurement of the active learning classification method on the EC task and the value measurement of the active learning regression method on the ER task on the unlabeled samples to obtain the total value measurement of the active learning on a plurality of tasks on the unlabeled samples, and simultaneously excavates the information of the category emotion and the dimension emotion, so that an EC model with good performance and an ER model on a single dimension or a plurality of dimensions can be trained simultaneously only by selecting as few samples as possible for marking, and the method has better performance compared with the EC model obtained by training of the active learning method of a single task and the ER model of a plurality of dimensions under the same inquiry times, and greatly reduces the marking cost.

Description

Multi-task active learning method and system simultaneously used for emotion classification and regression
Technical Field
The invention belongs to the field of emotion calculation, and particularly relates to a multi-task active learning method and system for emotion classification and regression at the same time.
Background
Emotion calculation enables machines to recognize, understand, express and adapt to human emotions, and is the core and basis of human-computer interaction. Emotion recognition is an important step in emotion calculation, and the emotional state of a person is obtained by analyzing and processing collected physiological signals or other non-physiological signals. There are two ways to represent emotion: 1) the emotion classification type emotion (discrete) can simply and intuitively express the emotion into a plurality of independent emotion classifications, such as six basic emotions (happiness, sadness, surprise, fear, anger and dislike) proposed by Ekman et al [1 ]; 2) dimensional emotions (continuous), which are considered to have basic dimensions, each of which is a measure of an aspect of an emotion, such as the three-dimensional space representation of the emotion in three dimensions as proposed by Mehrabian [2 ]. Categorical emotions are generally a classification problem in emotion computation, while dimensional emotions are generally a regression problem. The training of both the Emotion Classification (EC) model and the Emotion Regression (ER) model requires a large amount of labeled data.
In practice, it is very easy to obtain a large amount of unlabeled emotion data, but it is very difficult to label them. On the one hand, emotions are themselves very subjective, may have some uncertainty, and sometimes may be very subtle and difficult to capture, so that multiple annotators are often required to annotate each sample to obtain a more realistic label. On the other hand, some emotion samples are long, and the annotator needs to keep attention to observe the emotion samples all the time to obtain the annotation result. Therefore, the annotation of emotions is time-consuming and labor-consuming, and requires a high cost. Active Learning (AL) is one of the key technologies in machine learning, and can effectively reduce labeling work. The main flow of AL is: the most valuable or useful samples are first selected by some active learning algorithm and the expert is asked for their labels, and then the selected samples and labels are added to the training set to retrain the model. The process is iterated, and learning is stopped when a certain number of queries is reached or the model performance reaches a preset value. The core idea of the method is to obtain a more accurate model by inquiring the sample label as little as possible.
The category type emotion accords with the intuition and common knowledge of people, and the dimension type emotion can describe the emotion dynamically and finely, so that the EC task and the ER task have important research significance in emotion calculation. Moreover, in one application, two models may be used simultaneously, so it makes sense to train both EC and ER models from a small amount of labeled data.
The multi-task learning can learn a plurality of tasks simultaneously by mining the correlation and difference among the plurality of tasks, and actively learn to train with the labeling cost as little as possible to obtain the emotion recognition model. However, the current research only focuses on active learning on the EC model or the ER model, but the current research only focuses on active learning on the EC model or the ER model, and requires selection and labeling of samples respectively, so that training cost is high.
Disclosure of Invention
In view of the above drawbacks and needs of the prior art, the present invention provides a multi-task active learning method and system for emotion classification and regression, which aims to train a better EC model and an ER model in three dimensions simultaneously with as little labeling cost as possible.
To achieve the above object, according to one aspect of the present invention, there is provided a multitask active learning method for emotion classification and regression simultaneously, comprising:
s1, selecting M from unlabeled sample pools0Labeling the category type and dimension type labels of single or multiple dimensions of each label-free sample to serve as an initial training set, and removing the label-free samples from a label-free sample pool;
s2, training on an initial training set to obtain an initial EC model and an initial ER model with single or multiple dimensions;
s3, on the current EC model, obtaining a value ordering vector r of the current residual label-free sample on the EC task by using any active learning classification method1The higher the value, the larger the ranking value;
s4, obtaining a value sorting vector of the current residual label-free sample by using any one active learning regression method on the ER model of each current dimension; calculating the weighted sum of the value ordering vectors in all dimensions to obtain the value ordering vector r of the current residual label-free sample on the ER task2；
S5, calculating a value ordering vector r on the EC task1And a value ordering vector r on the ER task2Obtaining a total value sorting vector r on the EC task and the ER task by the weighted sum;
s6, selecting a non-label sample corresponding to the maximum value in the total value sorting vector r, labeling the class type and single or multiple dimension type labels, adding the non-label sample into the current training set, and removing the non-label sample from the current residual non-label sample;
s7, training a current EC model and a current ER model with single or multiple dimensions on a current training set;
s8, repeating the steps S3-S7 until the maximum number of samples of the current training set is reached or the model performance reaches a preset value, and obtaining a trained EC model and an ER model with single or multiple dimensions.
Further, step S1 hasSelecting the quantity M from a label-free sample pool at random or adopting an unsupervised active learning method0And (4) obtaining an unlabeled sample.
Further, selecting the number M by adopting an unsupervised active learning method0The label-free sample specifically comprises:
selecting a sample closest to a clustering center in the unlabeled sample pool as a first sample, labeling the first sample, adding the first sample into a training set, and removing the first sample from the unlabeled sample pool;
select the next M in turn0-1 sample: computing the current remaining unlabeled sample xnTo each selected sample xmThe distance of (c):
Figure BDA0002756518160000031
wherein, m is 1,20,n＝m0+1，...，N，m0The number of the samples which are selected and added into the training set is N, and N is the total number of the samples in the unlabeled sample pool; obtaining the nearest distance from the current residual label-free sample to the selected sample:
Figure BDA0002756518160000032
selecting
Figure BDA0002756518160000033
The largest sample is labeled and added into the training set, and is removed from the unlabeled sample pool.
Further, M in step S10Sentiment sample feature dimension + 1.
Further, step S3 specifically includes:
selecting an entropy-based uncertainty sampling method to apply to the EC model:
calculating the information entropy of each sample:
Figure BDA0002756518160000041
where p (y | x)i) Representing that EC model will not have sample x labelediThe probability of prediction as category y.The information entropies of all samples in the label-free sample pool are sorted in an ascending order to obtain a value sorting vector r on the EC task1。
Further, step S4 specifically includes:
the greedy sampling method in input and output space is selected to be applied to the ER model in three dimensions:
computing unlabeled samples xnTo each selected sample xmIs a distance of
Figure BDA0002756518160000042
Figure BDA0002756518160000043
Calculating a predicted value f (x) of the ER model of the first dimension to the unlabeled samplesn) Label y to selected samplemIs a distance of
Figure BDA0002756518160000044
Figure BDA0002756518160000045
Computing
Figure BDA0002756518160000046
And
Figure BDA0002756518160000047
minimum value of product
Figure BDA0002756518160000048
Figure BDA0002756518160000049
For all samples in the unlabeled sample pool in the first dimension
Figure BDA00027565181600000410
Performing ascending sorting to obtain a value sorting vector on a first dimension
Figure BDA00027565181600000411
Obtaining value sorting vectors in the second dimension and the third dimension by the same method
Figure BDA00027565181600000412
And
Figure BDA00027565181600000413
weighting the value sorting vectors of three dimensions to obtain a value sorting vector r on the ER task2：
Figure BDA00027565181600000414
Wherein, beta1+β2+β3＝1。
Further, step S5 is specifically executed by r ═ α1r1+α2r2Calculating to obtain a total value sorting vector r on the EC task and the ER task; wherein alpha is1+α2＝1。
According to another aspect of the present invention, there is provided a multitask active learning system for emotion classification and regression simultaneously, comprising: a computer-readable storage medium and a processor;
the computer-readable storage medium is used for storing executable instructions;
the processor is used for reading executable instructions stored in the computer readable storage medium and executing the multi-task active learning method for emotion classification and regression at the same time.
In general, the above technical solutions contemplated by the present invention can achieve the following advantageous effects compared to the prior art.
(1) The method combines the value measurement of the active learning classification method on the EC task and the value measurement of the active learning regression method on the ER task on the unlabeled samples to obtain the total value measurement of the active learning on a plurality of tasks on the unlabeled samples, and simultaneously excavates the information of the category emotion and the dimension emotion, so that an EC model with good performance and an ER model on a single dimension or a plurality of dimensions can be trained simultaneously only by selecting as few samples as possible for marking, and the method has better performance compared with the EC model obtained by training of the active learning method of a single task and the ER model of a plurality of dimensions under the same inquiry times, and greatly reduces the marking cost.
(2) The invention is applicable to a variety of active learning methods for classification and regression.
Drawings
FIG. 1 is a block diagram of the multi-task active learning method for emotion classification and regression simultaneously proposed by the present invention;
fig. 2 is a result demonstration provided by the embodiment of the present invention, which can show that the multi-task active learning method provided by the present invention can simultaneously achieve better performance on the EC model and the ER model than the single-task active learning method.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Referring to fig. 1, the invention provides a multi-task active learning method for emotion classification and regression simultaneously, which includes:
s1, selecting M from unlabeled sample pools0Labeling the category type and dimension type labels of single or multiple dimensions of each label-free sample to serve as an initial training set, and removing the label-free samples from a label-free sample pool;
s2, training on an initial training set to obtain an initial EC model and an initial ER model with single or multiple dimensions;
s3, on the current EC model, obtaining a value ordering vector r of the current residual label-free sample on the EC task by using any active learning classification method1The higher the value, the larger the ranking value;
s4, obtaining a value sorting vector of the current residual label-free sample by using any one active learning regression method on the ER model of each current dimension; calculating the weighted sum of the value ordering vectors in all dimensions to obtain the value ordering vector r of the current residual label-free sample on the ER task2；
S5, calculating a value ordering vector r on the EC task1And a value ordering vector r on the ER task2Obtaining a total value sorting vector r on the EC task and the ER task by the weighted sum;
s6, selecting a non-label sample corresponding to the maximum value in the total value sorting vector r, labeling the class type and single or multiple dimension type labels, adding the non-label sample into the current training set, and removing the non-label sample from the current residual non-label sample;
s7, training a current EC model and a current ER model with single or multiple dimensions on a current training set;
s8, repeating the steps S3-S7 until the maximum number of samples of the current training set is reached or the model performance reaches a preset value, and obtaining a trained EC model and an ER model with single or multiple dimensions.
In step S1, the embodiment of the invention selects M as the number0Unlabeled exemplars (emotion exemplar feature dimension +1) in order to obtain more stable initial EC and ER models. Selecting the M if an unsupervised active learning method (e.g., Greedy Sampling of input space, GSx) is used0And (3) selecting GSx samples closest to the clustering center in the unlabeled sample pool as first samples to label, adding the first samples into a training set, and removing the samples from the unlabeled sample pool, wherein N samples exist in the unlabeled sample pool. Then sequentially selecting the next M01 sample, for no loss of generality, assume that there is already m0Samples have been selected and added to the training set for the remaining N-m in the unlabeled pool0Sample, GSx first calculates its distance to each of the selected samples:
Figure BDA0002756518160000071
then the closest distance to the selected sample is obtained:
Figure BDA0002756518160000072
selecting the maximum
Figure BDA0002756518160000073
The samples are labeled and added to the training set.
Step S3 specifically includes:
in the embodiment of the present invention, an Entropy-based Uncertainty Sampling by entry, USE (uncertain Sampling by entry, USE) method is selected as an Active Learning Classification method to be applied to an EC model (Single Task Active Learning for indication Classification, STAL-EC), specifically, a class prediction probability p (y | x) of each sample in a current unlabeled sample pool by the EC model is obtained, and then an information Entropy of each sample is calculated:
Entropy(xi)＝-∑yp(y|xi)log(p|xi) (3)
and finally, performing ascending sorting on the information entropies of all samples in the unlabeled sample pool to obtain a value sorting vector r1I.e. the sample with the smallest entropy is at r1The corresponding ordinal value in (1). Other active learning classification methods besides entropy-based uncertainty sampling may also be employed, as the present invention is not limited in this respect.
Step S4 specifically includes:
the embodiment of the invention selects a method of Greedy Sampling on volume Inputs and Outputs (iGS) in input and output space as an Active Learning Regression method to be applied to an ER (Single Task Active Learning for evolution Regression, STAL-ER) model, considers ER tasks of three dimensions (value, aroma and dominance), and applies the method to the ER model in each dimensionThe iGS method is used in one dimension, such as in the value dimension: iGS first calculates the samples (x) in the current unlabeled pool of samplesn) To the selected sample (x)m) Is a distance of
Figure BDA0002756518160000074
(same as formula 1); then, the predicted value (f (x) of the regression model of the value dimension to the unlabeled sample is obtainedn) Calculate its label y to the selected samplemDistance (c):
Figure BDA0002756518160000075
then, calculating:
Figure BDA0002756518160000081
then combine the minimum of the input space and output space distances on the value dimension task for all samples in the unlabeled sample pool
Figure BDA0002756518160000082
Performing ascending sorting to obtain value sorting vector
Figure BDA0002756518160000083
By the same token can obtain
Figure BDA0002756518160000084
And
Figure BDA0002756518160000085
then weighting the value sorting vectors of the three dimensions to obtain a total value sorting vector of the dimension type task:
Figure BDA0002756518160000086
wherein beta is1+β2+β3The weight value may be selected according to the importance of different dimensions, where β is taken in the following embodiment1＝β2＝β3Three dimensions are considered equally important. Other active learning regression methods besides iGS may be selected, and the invention is not limited in this regard.
Step S5 specifically includes:
weighting and then summing the value sorting vectors respectively calculated in the step S3 and the step S4 to obtain a total value sorting vector r of the unlabeled samples, namely, r is alpha1r1+α2r2In which α is1+α21, the weight α here1And alpha2The value can be taken according to the importance of the class type and the dimension type, and alpha is taken in the embodiment of the invention1＝α2Categorical and dimensional tasks are considered equally important.
The loop stop condition in step S8 may be set such that the size of the final training set reaches a preset value or the performance of the EC and ER models reaches a preset value. This is set according to the data set and the performance that can be achieved by the model that are actually used, and in the embodiment of the present invention, the condition is set to that the algorithm is stopped when the training set size is 300, because the performance of the EC model and the ER model is sufficiently improved and substantially stabilized when the training set size is 300 (the performance of the EC model is measured by the classification accuracy, the larger the better, and the performance of the ER model is measured by the root mean square error, the smaller the better).
To verify the effectiveness of the method of the present invention, the following examples employ an IEMOCAP speech emotion data set to perform experiments, and 1555 pieces of speech data in spontaneous experimental data are used, wherein the classification type task includes a "happy" classification and a "sad" classification, the dimension type task includes three dimensions of validity, arousal and dominance, and the values of labels in different dimensions are all between [1,5 ]. The used classification model is Logistic regression, the regression model is Ridge regression, and the performance of the two models is measured by respectively using the classification accuracy and the root mean square error.
FIG. 2 shows the effect of the multi-task active learning method on category-type emotion tasks (EC) and dimension-type emotion tasks (ER-value, ER-aroma, ER-Dominance). Where BL-All indicates that All unlabeled samples were selected for training as the upper bound. BL-RS represents random selection, ST-MMC/ST-USE represents two single-task active learning methods, MT-MMC + EMCM/GSx/GSy/iGS and MT-USE + EMCM/GSx/GSy/iGS represent multi-task active learning methods. The abscissa represents the number of selected samples (training set size), and it can be seen that the performance of all active learning methods is gradually improved as the number of selected samples increases, the classification Accuracy (ACC) is gradually increased for EC tasks, and the Root Mean Square Error (RMSE) is gradually reduced for ER tasks, because the trained EC and ER models are more reliable as the number of samples in the training set increases. It can also be seen that the active learning method performed better than the random selection, which illustrates the effectiveness of the active learning method on the emotion classification and regression tasks. More importantly, it can be seen that the multitask active learning method performs better on the EC model and the ER model than the single-task active learning method, which shows the superiority of the method of the present invention.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (3)

1. A multitask active learning method for emotion classification and regression simultaneously, comprising:
s1, selecting M from a non-label sample pool by adopting an unsupervised active learning method0Labeling the category type and dimension type labels of single or multiple dimensions of each label-free sample to serve as an initial training set, and removing the label-free samples from a label-free sample pool; selecting the quantity of M from the unlabeled sample pool by adopting an unsupervised active learning method0The label-free sample specifically comprises:
selecting a sample closest to a clustering center in the unlabeled sample pool as a first sample, labeling the first sample, adding the first sample into a training set, and removing the first sample from the unlabeled sample pool;
select the next M in turn0-1 sample: computing the current remaining unlabeled sample xnTo each selected sample xmThe distance of (c):
Figure FDA0003547170560000011
wherein m is 1,2, …, m0,n＝m0+1,…,N，m0The number of the samples which are selected and added into the training set is N, and N is the total number of the samples in the unlabeled sample pool; obtaining the nearest distance from the current residual label-free sample to the selected sample:
Figure FDA0003547170560000012
selecting
Figure FDA0003547170560000013
Labeling the largest sample, adding the largest sample into a training set, and removing the largest sample from the unlabeled sample pool;
s2, training on an initial training set to obtain an initial EC model and an initial ER model with single or multiple dimensions;
s3, on the current EC model, obtaining a value ordering vector r of the current residual label-free sample on the EC task by using any active learning classification method1The higher the value, the larger the ranking value; step S3 specifically includes:
selecting an entropy-based uncertainty sampling method to apply to the EC model:
calculating the information entropy of each sample:
Figure FDA0003547170560000014
where p (y | x)i) Representing that EC model will not have sample x labelediPredicting the probability of the category y, and performing ascending ordering on the information entropies of all samples in the unlabeled sample pool to obtain a value ordering vector r on the EC task1；
S4, obtaining a value sorting vector of the current residual label-free sample by using any one active learning regression method on the ER model of each current dimension; calculating the weighted sum of the value sorting vectors in all dimensions to obtain the price of the current residual label-free sample on the ER taskValue ordering vector r2(ii) a Step S4 specifically includes:
the greedy sampling method in input and output space is selected to be applied to the ER model in three dimensions:
computing unlabeled samples xnTo each selected sample xmIs a distance of
Figure FDA0003547170560000021
Figure FDA0003547170560000022
Calculating a predicted value f (x) of the ER model of the first dimension to the unlabeled samplesn) Label y to selected samplemIs a distance of
Figure FDA0003547170560000023
Figure FDA0003547170560000024
Calculating out
Figure FDA0003547170560000025
And
Figure FDA0003547170560000026
minimum value of product
Figure FDA0003547170560000027
Figure FDA0003547170560000028
For all samples in the unlabeled sample pool in the first dimension
Figure FDA0003547170560000029
Performing ascending sorting to obtain a value sorting vector on a first dimension
Figure FDA00035471705600000210
Obtaining value sorting vectors in the second dimension and the third dimension by the same method
Figure FDA00035471705600000211
And
Figure FDA00035471705600000212
weighting the value sorting vectors of three dimensions to obtain a value sorting vector r on the ER task2：
Figure FDA00035471705600000213
Wherein, beta1+β2+β3＝1；
S5, calculating a value ordering vector r on the EC task1And a value ordering vector r on the ER task2Obtaining a total value sorting vector r on the EC task and the ER task by the weighted sum; in step S5, r is ═ α1r1+α2r2Calculating to obtain a total value sorting vector r on the EC task and the ER task; wherein alpha is1+α2＝1；
S6, selecting a non-label sample corresponding to the maximum value in the total value sorting vector r, labeling the class type and single or multiple dimension type labels, adding the non-label sample into the current training set, and removing the non-label sample from the current residual non-label sample;
s7, training a current EC model and a current ER model with single or multiple dimensions on a current training set;
s8, repeating the steps S3-S7 until the maximum number of samples of the current training set is reached or the model performance reaches a preset value, and obtaining a trained EC model and an ER model with single or multiple dimensions.
2. The method for simultaneous multi-task active learning and regression as claimed in claim 1, wherein M is in step S10Sentiment sample feature dimension + 1.
3. A multitask active learning system for both emotion classification and regression, comprising: a computer-readable storage medium and a processor;
the computer-readable storage medium is used for storing executable instructions;
the processor is configured to read executable instructions stored in the computer readable storage medium and execute a method of multitask active learning for emotion classification and regression according to claim 1 or 2.
CN202011204366.2A 2020-11-02 2020-11-02 Multi-task active learning method and system simultaneously used for emotion classification and regression Active CN112396094B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011204366.2A CN112396094B (en) 2020-11-02 2020-11-02 Multi-task active learning method and system simultaneously used for emotion classification and regression

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011204366.2A CN112396094B (en) 2020-11-02 2020-11-02 Multi-task active learning method and system simultaneously used for emotion classification and regression

Publications (2)

Publication Number Publication Date
CN112396094A CN112396094A (en) 2021-02-23
CN112396094B true CN112396094B (en) 2022-05-20

Family

ID=74599004

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011204366.2A Active CN112396094B (en) 2020-11-02 2020-11-02 Multi-task active learning method and system simultaneously used for emotion classification and regression

Country Status (1)

Country Link
CN (1) CN112396094B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104732203A (en) * 2015-03-05 2015-06-24 中国科学院软件研究所 Emotion recognizing and tracking method based on video information
CN110782876A (en) * 2019-10-21 2020-02-11 华中科技大学 Unsupervised active learning method for voice emotion calculation
CN111079406A (en) * 2019-12-13 2020-04-28 华中科技大学 Natural language processing model training method, task execution method, equipment and system
CN111177374A (en) * 2019-12-13 2020-05-19 航天信息股份有限公司 Active learning-based question and answer corpus emotion classification method and system

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180260860A1 (en) * 2015-09-23 2018-09-13 Giridhari Devanathan A computer-implemented method and system for analyzing and evaluating user reviews

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104732203A (en) * 2015-03-05 2015-06-24 中国科学院软件研究所 Emotion recognizing and tracking method based on video information
CN110782876A (en) * 2019-10-21 2020-02-11 华中科技大学 Unsupervised active learning method for voice emotion calculation
CN111079406A (en) * 2019-12-13 2020-04-28 华中科技大学 Natural language processing model training method, task execution method, equipment and system
CN111177374A (en) * 2019-12-13 2020-05-19 航天信息股份有限公司 Active learning-based question and answer corpus emotion classification method and system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于深度学习的中文评论文本情感分析研究;张苗;《中国优秀硕士学位论文全文数据库 (信息科技辑)》;20200315;全文 *

Also Published As

Publication number Publication date
CN112396094A (en) 2021-02-23

Similar Documents

Publication Publication Date Title
US11875267B2 (en) Systems and methods for unifying statistical models for different data modalities
Chen et al. Efficient ant colony optimization for image feature selection
CN111382565B (en) Emotion-reason pair extraction method and system based on multiple labels
CN112687374B (en) Psychological crisis early warning method based on text and image information joint calculation
CN114743600A (en) Gate-controlled attention mechanism-based deep learning prediction method for target-ligand binding affinity
Ihou et al. Stochastic topic models for large scale and nonstationary data
Rangulov et al. Emotion recognition on large video dataset based on convolutional feature extractor and recurrent neural network
US11948387B2 (en) Optimized policy-based active learning for content detection
CN112396094B (en) Multi-task active learning method and system simultaneously used for emotion classification and regression
Lestari et al. Implementation of k-nearest neighbor (knn) and suport vector machine (svm) for clasification cardiovascular disease
Chaturvedi et al. Analyzing the Performance of Novel Activation Functions on Deep Learning Architectures
Gambo et al. Performance comparison of convolutional and multiclass neural network for learning style detection from facial images
Thakur et al. Offline handwritten mathematical recognition using adversarial learning and transformers
Janko et al. Winning the sussex-huawei locomotion-transportation recognition challenge
CN114048361A (en) Crowdsourcing software developer recommendation method based on deep learning
Liu et al. Suggestion mining from online reviews usingrandom multimodel deep learning
CN110633363B (en) Text entity recommendation method based on NLP and fuzzy multi-criterion decision
Phan et al. Biomedical named entity recognition based on hybrid multistage CNN-RNN learner
Li et al. Variance tolerance factors for interpreting all neural networks
Malik et al. Matrix factorization-based improved classification of gene expression data
SabbaghGol et al. Evolution of the random subset feature selection algorithm for classification problem
Titu et al. Acquiring businesss intelligence through data science: A practical approach
CN116756347B (en) Semantic information retrieval method based on big data
Kianiharchegani Data-Driven Exploration of Coarse-Grained Equations: Harnessing Machine Learning
Maharudra et al. A HIGH-LEVEL ENSEMBLE FEATURE SELECTION ALGORITHM FOR MITIGATING THE DIMENSIONALITY IN STRESS DATA

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