CN107918487A - A kind of method that Chinese emotion word is identified based on skin electrical signal - Google Patents

Abstract

The invention discloses a method for recognizing Chinese emotional words based on electrodermal signals. This method uses the advantage of physiological parameters to identify emotions to identify Chinese emotional words. It specifically includes electrodermal collection, preprocessing of the collected data, feature extraction, normalization processing, feature selection, using the improved simulated annealing artificial neural network algorithm to obtain classification results, and finally adding emotional word comparison to the classification results, to identify. As an embodiment, the present invention recognizes 50 emotional words with the highest emotional intensity selected from "Modern Chinese Dictionary", "Modern Chinese Classified Dictionary", and "New Century Chinese New Words Dictionary". Experiments have proved that the present invention can complete the recognition of Chinese emotional words with high accuracy, which fully shows that it is feasible to use physiological parameters to extract text emotional words, and provides new ideas for later text analysis, and the system structure of the present invention is clear, Simple and easy to implement.

Description

A method for recognizing Chinese emotional words based on electrodermal signals

technical field

The invention belongs to the field of processing natural language data of electric digital data processing technology, and in particular relates to a method for recognizing Chinese emotional words based on skin electro-signal emotion recognition technology.

Background technique

Using electrodermal signals to identify emotions has its unique advantages. For example, it is a physiological parameter, so it is more objective. For example, it is easier to collect than other physiological parameters, and it is the most effective and sensitive to neuroemotional changes. For the study of emotions using electrodermal signals, the current technology is relatively mature, so it is timely to use this technology to introduce text analysis. At present, text analysis is hindered by its strong subjectivity, difficulty in grammatical and semantic separation, and imperfect emotional lexicon.

However, with the vigorous development of the Internet, the text information is even more vast, which contains a large amount of useful and practical information, so that people are unwilling to give up the extraction of text information. Objectively, it is necessary to use more objective physiological parameters to identify Chinese emotional words, which will provide a new way of thinking for pure text analysis.

Contents of the invention

The purpose of the present invention is to propose the use of electrodermal signals to identify Chinese emotional words, to provide a new idea for simple text emotional analysis, and to make text emotional analysis more accurate.

In order to achieve the above object, the technical solution adopted in the present invention is a method for identifying Chinese emotional words based on electrodermal signals, which specifically includes the following steps:

S1: Electrodermal collection;

S2: Preprocessing the collected data;

S3: feature extraction;

S4: Normalization processing;

S5: feature selection;

S6: Using the improved simulated annealing artificial neural network algorithm to obtain the classification result;

S7: Add emotion word comparison to the classification result for recognition.

Preferably, the preprocessing in step 2 above adopts wavelet transform for denoising processing.

Further, the feature extraction in step 3 is to extract the statistical value that can represent the change of the electrodermal signal in the time domain and frequency domain of the signal as the original feature of the emotion recognition research.

Further, the above-mentioned time-domain original features include the mean value, median value, maximum value, minimum value, standard deviation, minimum value ratio, maximum value ratio, maximum and minimum difference of the electrodermal signal, and the above-mentioned signal features are first-order difference, The 24 time-domain features generated after the second-order difference calculation and the extraction of the above statistical features.

Further, before extracting the frequency domain features, discrete Fourier transform is performed on the electrodermal signal, and then the frequency mean, median, standard deviation, maximum value, minimum value, and maximum and minimum difference are calculated to obtain 6 frequency domain features .

Further, the normalization process in the above-mentioned step 4 makes the value range of each feature value limited between 0 and 1, and the method for removing individual differences is as follows:

where X _G is the original signal, is the average value of each subject at rest, after normalization, we get:

X=(X _G -X _mean )/(X _max -X _min ) (2).

Further, in order to recognize emotion with the minimum number of features and the highest recognition rate, some groups are randomly selected in the standardized data during feature selection in step 5, and are divided into three parts: the first part is the classifier training set , the second part of the test set tests the classification effect, and the last part of the data is used to verify the effectiveness of the feature set in emotion recognition.

Further, the above-mentioned improved simulated annealing artificial neural network algorithm includes the following steps:

Step 1: Determine the neural network structure according to the input and output of the sample;

Step 2: Use the simulated annealing algorithm with memory, as follows:

1) Initialize the parameters, so that the initial weight S ₀ is generated. At this time, set the initial temperature T ₀ >0, the number of iterations i=0, and the inspection accuracy ε, let f _out =f(S ₀ ), f ^* =f (S ₀ ),S _p =S ₀ ;

2) Take the network weight S _p as the initial starting point S ₀ , optimize according to the Powell algorithm, and quickly search for a certain local minimum point;

3) Set the memory variables x' and f(x'), which are used to memorize the current optimal solution and the optimal objective function value respectively. At the beginning of the algorithm, x' and f(x') are initialized to be equal to the initial solution x ₀ and its objective function value f(x ₀ ), after the iteration starts, whenever a new search solution is accepted, compare its objective function value f(x _k ) with f(x′), if f(x _k ) is better than f(x′), then replace the original x′ and f(x′) with x _k and f(x _k ) respectively, and finally the global optimal solution is obtained at the end of the algorithm;

4) A new set of network weights S _p is obtained, let S _i =S _p , f _out =f(S _i ), f _* =f(S _i ), and use the network weight S _i as the iteration value x, Assume the current solution S _i =x, set T=T _i , and perform annealing operation to obtain a new set of network weights S _i+1 , anneal according to T _i =T ₀ /(1+ln(i)), i= i+1;

5) After annealing, if the requirements or the number of iterations are met, the algorithm ends, if f(S _i )<f _out , set S _p =S _i+1 , and return to step 4;

Step 3: Neural network training and prediction. The training is to determine the network structure by setting fixed input and output. During the training process, the neural network continuously adjusts the connection weights between each neuron to reduce the training output and the specified output. The error between the prediction is the process of the trained network processing the input data and obtaining the output;

Step 4: Finally, compare the output results with the table information entered by the subjects during the experiment, and complete the comparison of emotion word recognition.

Compared with prior art, the beneficial effect of the present invention:

1. The present invention can complete the recognition of Chinese emotional words with high accuracy, basically achieving the expected results.

2. The present invention fully demonstrates that it is feasible to use physiological parameters to extract text emotion words, and provides a new idea for later text analysis.

3. The system architecture of the present invention is clear, simple and easy to implement.

Description of drawings

Figure 1 shows a schematic flow chart of the entire scheme.

Figure 2 shows part of the emotional word questionnaire.

Figure 3 shows the experimental emotion word recognition table.

Figure 4 shows the comparison chart of the two identifications.

Detailed ways

The present invention will be further described in detail in conjunction with the accompanying drawings and embodiments.

As an embodiment, the present invention first screens out more than 2000 emotional words from "Modern Chinese Dictionary", "Modern Chinese Classified Dictionary", "New Century Chinese New Words Dictionary", and then screens out the most frequently used 100 words from these 2000 indivual. Finally, another screening is carried out among these 100 words, and 50 emotional words with the highest emotional intensity are obtained.

Using laboratory electrodermal collection tools, 20 people in a laboratory were collected. All 20 people in the laboratory are in good health, have no history of heart disease or mental illness, have not taken any psychotropic drugs within a year, and have all age groups from 20 to 50 years old. The experimental material is 50 selected emotional words with strong emotions. The subjects are required to sit in front of the computer screen. At this time, an emotional word appears on the computer screen every 40 seconds. When the emotional word appears, the subjects are asked to associate with the emotional word. Scenes related to emotional words. The first 30 seconds are used for association and the next 10 seconds are used to fill in whether there is feeling and fill in the emotional intensity (0 is very strong, 1 is strong, 2 is average, 3 is weak, and 4 is very weak). Then play in sequence until the 50 emotional words are played.

Preprocess the collected data. Since the electrodermal signal is relatively weak and is easily affected by machine interference, electromyographic interference, and electromagnetic interference, it is necessary to denoise the collected electrodermal signal. The present invention adopts wavelet transform for denoising processing. Wavelet transform has high frequency resolution and low time resolution in the low frequency part, and high time resolution and low frequency resolution in the high frequency part, so it is adaptive to the signal, very It is suitable for the analysis of physiological signals.

Before the experiment started, the procedure and purpose of the experiment were explained to the subjects in detail. First, each subject was asked to sit 80cm away from the computer screen. The experiment begins. The subjects were asked to close their eyes for one minute, and then open their eyes to look at the screen. At this time, an emotional word would appear on the screen every 40 seconds, of which the emotional word would be displayed for 30 seconds, and the emotional intensity scale of the emotional word would be filled in for 10 seconds. When the emotional words appeared, the subjects were stimulated by the emotional words to associate the corresponding scenes, and when the screen was blank, they filled out the questionnaire and let their emotions return to calm. Play the selected 50 emotional words in turn until the end.

After deleting invalid data, 270 sets of valid data were screened out. Referring to the feature extraction method of the University of Augsburg in Germany, the statistical value that best represents the change of the electrodermal signal in the time domain and frequency domain of the signal is extracted as the original feature of the emotion recognition research. In the time domain, 22 time domains such as the maximum value, minimum value, standard deviation, first-order difference standard deviation, first-order difference minimum value ratio, second-order difference standard deviation, and second-order difference minimum value ratio of the electrodermal signal are extracted. feature. In order to extract the frequency domain features of the electrodermal signal, the discrete Fourier transform is first performed on the electrodermal signal, and then the frequency mean, median, standard deviation, maximum value, minimum value, and maximum and minimum difference are calculated to obtain 6 frequency domain features .

Due to the great individual differences of the electrodermal signals, and the value ranges of the eigenvalues of each statistical feature extracted according to the formula are in different orders of magnitude, in order to facilitate unified comparison, standardize the statistical distribution of data, facilitate subsequent processing, and perform normalization processing For each feature, the value range of each feature value is limited between 0 and 1. Enter the formula below:

Methods for removing individual differences:

where X _G is the original signal, is the mean value for each subject at rest.

After normalization, we get:

X＝(X _G -X _mean )/(X _max -X _min ) (2)

Feature selection is performed on the processed data. Feature selection is to recognize emotion with the least number of features and the highest recognition rate. Randomly select 180 groups from the above 270 groups of standardized data and divide them into three parts: the first 80 groups of data form the classifier training set; the middle 60 groups of data form the test set to test the classification effect; the last 40 groups of data are used for verification Effectiveness of feature sets in emotion recognition.

The neural network algorithm is a distributed storage of information, and its information is stored on the entire network. What is stored in a certain place on the network is not an external information, but part of multiple information. In this way, the advantages of the neural network algorithm are obvious. Because the information is stored on the network, its information storage and processing are combined into one, so that it can process data in a large scale in parallel, and this method makes it have strong fault tolerance. capability and robustness. And its self-learning and adaptability. But the neural network algorithm also has major defects, it is easy to fall into the local minimum, and the convergence process is slow, so it needs to be improved.

Due to the defects of the neural network algorithm, the addition of simulated annealing technology greatly improves the function of the algorithm and largely solves these defects. The present invention first uses the Powell algorithm to quickly converge to the local minimum value. When the local minimum value is found, the simulated annealing search strategy is used to immediately search for whether there is a valley near the local minimum value, and it is carried out multiple times. At this time, memory is added to record the search results. The optimal value and the optimal function value of , so that the global optimal value can be found quickly. Applying this method to the neural network algorithm can solve the defect that the neural network cannot jump out of the local minimum.

The steps of the improved simulated annealing artificial neural network algorithm are summarized as follows:

Step 1: Determine the neural network structure according to the input and output of the sample.

Step 2: Use the simulated annealing algorithm with memory:

1) Initialize parameters. In this way, the initial weight S ₀ is generated. At this time, the initial temperature T ₀ >0 is set, the iteration number i=0, and the inspection accuracy ε. Let f _out =f(S ₀ ), f ^* =f(S ₀ ), S _p =S ₀ .

2) Take the network weight S _p as the initial starting point S ₀ , optimize according to the Powell algorithm, and quickly search for a certain local minimum point.

3) Set memory variables x' and f(x'), which are used to memorize the currently encountered optimal solution and optimal objective function value respectively. At the beginning of the algorithm, x′ and f(x′) are initialized to be equal to the initial solution x ₀ and its objective function value f(x ₀ ), respectively. After the iteration starts, whenever a new search solution is accepted, its objective function value f (x _k ) is compared with f(x′), if f(x _k ) is better than f(x′), replace the original x′ and f(x′) with x _k and f(x _k ) respectively. At the end of the algorithm, the global optimal solution is obtained.

4) To obtain a new set of network weights S _p , let S _i =S _p , f _out =f(S _i ), f _* =f(S _i ). Take the network weight S _i as the iteration value x, set the current solution S _i =x, set T = T _i , and perform annealing operation. Get a new set of network weights S _i+1 . Anneal according to T _i =T ₀ /(1+ln(i)). i=i+1.

5) After annealing, if the requirements or the number of iterations are met, the algorithm ends. If f(S _i )<f _out , set S _p =S _i+1 , turn back to (4).

Step 3: neural network training and prediction. Training is to determine the network structure by setting fixed input and output. During the training process, the neural network continuously adjusts the connection weights between each neuron to reduce the error between the training output and the specified output. Prediction is a process in which a trained network processes input data and obtains an output.

Finally, compare the output results with the form that the subjects filled in during the experiment to complete the comparison of emotion word recognition. Among them, some emotional word questionnaires are shown in Figure 2, the emotional word recognition diagram in the experiment is shown in Figure 3, and the comparison and recognition diagram of the two is shown in Figure 4. The experimental results show that the recognition of Chinese emotional words can be basically completed by using the changes of electrodermal signals.

Claims (8)

1. a method for identifying Chinese emotion words based on electrodermal signal, is characterized in that comprising the following steps: S1: Electrodermal collection; S2: Preprocessing the collected data; S3: feature extraction; S4: Normalization processing; S5: feature selection; S6: Using the improved simulated annealing artificial neural network algorithm to obtain the classification result; S7: Add emotion word comparison to the classification result for recognition. 2. the method for identifying Chinese emotion words based on electrodermal signal according to claim 1, is characterized in that, the preprocessing in step 2 adopts wavelet transform to carry out denoising processing. 3. the method for identifying Chinese emotional words based on electrodermal signals according to claim 1, characterized in that, the feature extraction in step 3 is to extract the statistical value that can represent the electrodermal signal changes in the time domain and frequency domain of the signal As an original feature for emotion recognition research. 4. the method for identifying Chinese emotion words based on electrodermal signal according to claim 3, is characterized in that said time domain original feature comprises mean value, median value, maximum value, minimum value, standard deviation, minimum value of electrodermal signal Ratio, maximum value ratio, maximum and minimum difference, and 24 time-domain features generated by extracting the above statistical features after performing first-order difference and second-order difference calculations on the above signal features. 5. the method for identifying Chinese emotion words based on electrodermal signal according to claim 3, is characterized in that before extracting described frequency domain feature, discrete Fourier transform is carried out to electrodermal signal, then calculates frequency mean value, median value , standard deviation, maximum value, minimum value, maximum and minimum difference, and get 6 frequency domain features. 6. the method for identifying Chinese emotion words based on electrodermal signal according to claim 1, is characterized in that described normalization process in step 4 makes the value range of each characteristic value be limited between 0 to 1, removes The method of individual difference is as follows: <mrow><mover><mi>X</mi><mo>&amp;OverBar;</mo></mover><mo>=</mo><msub><mi>X</mi><mi>G</mi></msub><mo>-</mo><msub><mover><mi>x</mi><mo>~</mo></mover><mrow><mi>c</mi><mi>l</mi><mi>a</mi><mi>m</mi></mrow></msub><mo>-</mo><mo>-</mo><mo>-</mo><mrow><mo>(</mo><mn>1</mn><mo>)</mo></mrow></mrow> where X _G is the original signal, is the average value of each subject at rest, after normalization, we get: X=(X _G -X _mean )/(X _max -X _min ) (2). 7. the method for identifying Chinese emotional words based on electrodermal signal according to claim 1, is characterized in that some groups are randomly selected in the data after standardization during the feature selection described in step 5, and it is divided into three parts: the first part It is the classifier training set, the second part of the test set tests the classification effect, and the last part of the data is used to verify the effectiveness of the feature set in emotion recognition. 8. the method for identifying Chinese emotion words based on electrodermal signal according to claim 1, is characterized in that described improved simulated annealing artificial neural network algorithm comprises the steps: Step 1: Determine the neural network structure according to the input and output of the sample; Step 2: Use the simulated annealing algorithm with memory, as follows: 1) Initialize the parameters, so that the initial weight S ₀ is generated. At this time, set the initial temperature T ₀ >0, the number of iterations i=0, and the inspection accuracy ε, let f _out =f(S ₀ ), f ^* =f (S ₀ ),S _p =S ₀ ; 2) Take the network weight S _p as the initial starting point S ₀ , optimize according to the Powell algorithm, and quickly search for a certain local minimum point; 3) Set the memory variables x' and f(x'), which are used to memorize the current optimal solution and the optimal objective function value respectively. At the beginning of the algorithm, x' and f(x') are initialized to be equal to the initial solution x ₀ and its objective function value f(x ₀ ), after the iteration starts, whenever a new search solution is accepted, compare its objective function value f(x _k ) with f(x′), if f(x _k ) is better than f(x′), then replace the original x′ and f(x′) with x _k and f(x _k ) respectively, and finally the global optimal solution is obtained at the end of the algorithm; 4) A new set of network weights S _p is obtained, let S _i =S _p , f _out =f(S _i ), f _* =f(S _i ), and use the network weight S _i as the iteration value x, Assume the current solution S _i =x, set T=T _i , and perform annealing operation to obtain a new set of network weights S _i+1 , anneal according to T _i =T ₀ /(1+ln(i)), i= i+1; 5) After annealing, if the requirements or the number of iterations are met, the algorithm ends, if f(S _i )<f _out , set S _p =S _i+1 , and return to step 4; Step 3: Neural network training and prediction. The training is to determine the network structure by setting fixed input and output. During the training process, the neural network continuously adjusts the connection weights between each neuron to reduce the training output and the specified output. The error between the prediction is the process of the trained network processing the input data and obtaining the output; Step 4: Finally, compare the output results with the table information entered by the subjects during the experiment, and complete the comparison of emotion word recognition.