KR102567896B1 - Apparatus and method for religious sentiment analysis using deep learning - Google Patents

Abstract

An apparatus and method for analyzing religious sentiment using deep learning are disclosed. Religious sentiment analysis method using deep learning includes generating comment data by collecting comments on news articles related to Corona 19 (COVID-19) and religion, and performing sentiment labeling on the generated comment data. , performing data pre-processing on the comment data for which sentiment labeling has been performed, and performing sentiment analysis on the pre-processed comment data using a deep learning model to obtain emotional values for positive and negative emotions for each religion. It includes the step of calculating

Description

Apparatus and method for religious sentiment analysis using deep learning}

The present invention relates to an apparatus and method for analyzing religious sentiment using deep learning.

People's daily lives have been greatly affected by Corona 19 (COVID-19), which spread around the world in January 2020. In addition, on February 18, 2020, a group infection began from a specific religious group, another group infection in June 2020, and a large-scale rally related to Liberation Day in August 2020. Social events like this facilitated the spread of COVID-19 in Korea, and the number of confirmed cases increased rapidly.

Accordingly, with the recognition that religion can be a medium of COVID-19 infection, the media has produced various online information about religion.

So, the present applicant proposed the present invention to approach the impact on religion among the impacts of Corona 19 on our society from a data analysis point of view.

Republic of Korea Patent Registration No. 10-0935828 (2009.12.30)

The present invention collects comments on news articles related to Corona 19 (COVID-19) and religion, analyzes the collected comments using a deep learning model to calculate the emotional level for each religion, and analyzes the number of comments per month for each date. It is to provide a religious sentiment analysis device and method using deep learning that calculates sentiment values.

According to one aspect of the present invention, a method for analyzing religious sentiment using deep learning performed by an apparatus for analyzing religious sentiment is disclosed.

Religious sentiment analysis method using deep learning according to an embodiment of the present invention includes generating comment data by collecting comments on news articles related to Corona 19 (COVID-19) and religion, and sentiment about the generated comment data. Performing labeling, performing data pre-processing on the comment data for which the sentiment labeling has been performed, and performing sentiment analysis on the pre-processed comment data using a deep learning model A step of calculating emotional values for positive and negative emotions for each religion is included.

In the performing of the sentiment labeling, replies showing a positive tendency in the comment data are labeled with positive identifiers, and replies showing negative tendencies are labeled with negative identifiers.

The deep learning model applies a Keras model, one of the Python packages, and is configured to have 10 hidden layers between the input layer and the output layer. .

The 10 hidden layers include a first hidden layer to which TextVectorization, a function that vectorizes strings, is applied, a second hidden layer to which an embedding function is applied that converts positive integers, which are vectorized strings, into dense vectors of a fixed size, and the dense vectors are flattened. With a third hidden layer and 32 units to which the Flatten function that converts into an array is applied, the activation function is set to be ReLU (Rectified Linear Unit), and overfitting in the learning process In order to prevent this, a fourth to tenth hidden layer to which L2 regularization is added is included.

In the step of calculating the sentiment value, sentiment analysis is performed on the comment data using the KoBERT model, which is a Korean BERT model.

The religious sentiment analysis method using the deep learning further includes calculating a sentiment value for each date using the preprocessed comment data.

The emotional value per day is derived from a weighted average per day, and the weighted average (W) per day is calculated using the following equation.

where k is the number of articles per day, i is the article index, count(i) is the total number of comments for article index i, avg(i) is the average sentiment value of the article, and M is the total number of all articles on that day is the sum of the number of comments.

The sentiment average (X) for each date obtained by correcting the weighted average per day is calculated using the following equation.

Here, N is the total number of comments per day, L is the minimum number of comments per day for the sentiment value to gain trust, e is the average sentiment value for the entire period, and W is the weighted average value per day calculated in Equation 1 .

According to another aspect of the present invention, an apparatus for analyzing religious sentiment using deep learning is disclosed.

An apparatus for analyzing religious sentiment using deep learning according to an embodiment of the present invention includes a memory for storing commands and a processor for executing the commands, wherein the commands include news articles related to COVID-19 and religion. Collecting comments of and generating comment data, performing emotion labeling on the generated comment data, and performing data pre-processing on the comment data on which the emotion labeling has been performed. and performing sentiment analysis on the preprocessed comment data using a deep learning model to calculate positive and negative sentiment values for each religion.

Religious sentiment analysis apparatus and method using deep learning according to an embodiment of the present invention collects comments on news articles related to Corona 19 (COVID-19) and religion, and analyzes the collected comments using a deep learning model The emotional level for each religion can be calculated, and the emotional value for each date can be calculated by analyzing the number of comments per month.

1 is a flowchart illustrating a religious sentiment analysis method using deep learning performed by an apparatus for analyzing religious sentiment according to an embodiment of the present invention.
2 is a diagram showing an example of the structure of a deep learning model according to an embodiment of the present invention.
3 to 5 are diagrams showing experimental results for a religious sentiment analysis method using deep learning according to an embodiment of the present invention.
6 is a diagram schematically illustrating the configuration of an apparatus for analyzing religious sentiment using deep learning according to an embodiment of the present invention.

Singular expressions used herein include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as "consisting of" or "comprising" should not be construed as necessarily including all of the various components or steps described in the specification, and some of the components or some of the steps It should be construed that it may not be included, or may further include additional components or steps. In addition, terms such as "...unit" and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software. .

Hereinafter, various embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a religious sentiment analysis method using deep learning performed by an apparatus for analyzing religious sentiment according to an embodiment of the present invention, and FIG. 2 is a diagram showing an example of the structure of a deep learning model according to an embodiment of the present invention. am. Hereinafter, a method for analyzing religious sentiment using deep learning according to an embodiment of the present invention will be described with reference to FIG. 1 , but reference will be made to FIG. 2 .

In step S110, the device for analyzing religious sensibility generates comment data by collecting comments on news articles related to COVID-19 and religion.

In order to grasp public awareness, comments on domestic news articles during a preset collection period may be collected. For example, in order to set a data collection space, the ranking of visits of a search engine from August 2019 to October 2020 disclosed in Internet Trend (www.internettrend.co.kr) may be used. According to Internet trends during this collection period, the share of domestic portal sites was ranked first with Naver (58.5%), second with Google (32.8%), and third with Daum (6.63%). Here, Google, the second place, was judged to be an inappropriate web page for collecting comments on news articles, and was excluded from the data collection space. That is, as a data collection space, Naver, which is the first place, and Daum, which is the third place, may be set.

Keywords for searching for news articles related to COVID-19 and religion can be set to five keywords: Christianity, Buddhism, Catholicism, Shincheonji, and religion. Here, Christianity, Buddhism, and Catholicism rank from 1st to 3rd in the proportion of domestic believers in order, accounting for 20%, 15.5% and 8%, respectively (announced by the Ministry of Culture, Sports and Tourism, 2018). And, although Shincheonji has less than 1% of domestic believers, it can be selected as a keyword because it is the religion that caused the mass infection of Corona 19 and caused the first domestic pandemic of Corona 19. Finally, in order to find out the change in perception including all keywords, the keyword of religion may be additionally set.

The data collection period may be set differently for each keyword. That is, for Shincheonji, February 17, 2020, one day before the mass infection occurred, can be set as the reference date for collection, and for other keywords, January 20, 2020, when Corona 19 was introduced, can be set as the reference date for collection. In addition, comments on news articles generated for 5 months before the collection reference date and for 5 months after the collection reference date may be collected.

For example, comments on news articles may be collected using a crawler implemented in the Python language using Selenium and the BeautifulSoup4 library. In order to search for news articles suitable for a date, a date search condition setting function provided by a portal site may be used. Comment data generated by collecting comments on news articles may include URLs of news articles, dates, and all comments and replies, and may be stored and managed in a database (MariaDB). An example of generated comment data can be shown in Table 1 below. In Table 1, before and after the collection reference date are indicated as Before and After.

In step S120, the apparatus for analyzing religious sentiment performs sentiment labeling on the generated comment data.

Comment data generated through the collection of comments on news articles may be sentiment labeled for sentiment analysis using deep learning.

Comments showing a positive trend in the comment data may be labeled with a positive identifier (eg, 1), and comments showing a negative trend may be labeled with a negative identifier (eg, 0). That is, the apparatus for analyzing religious sentiment may receive a positive identifier or a negative identifier for each comment from the user and perform sentiment labeling on each reply of the reply data.

For example, as described above, comment data will be collected in the two data collection spaces of Naver and Daum, with five keywords of Christianity, Buddhism, Catholicism, Shincheonji, and religion, during the period before and after the reference date of collection. can Therefore, the total number of training data sets for deep learning can be 2 × 5 × 2 = 20. Sentiment labeling can be performed by extracting 5 to 10% of comments from each training data set. When a user works on emotion labeling for each comment, bias may exist because each individual has different criteria for determining positive or negative. In order to compensate for this point, eight people conducted emotional labeling together, so that the criteria of a specific individual were not reflected as a result of deep learning. Through this, emotion labeling was performed on a total of 32,474 comments, and emotion-labeled learning data was created.

In step S130, the apparatus for analyzing religious sentiment performs data pre-processing on the comment data for which sentiment labeling has been performed.

Here, data preprocessing is a process of making data into a form suitable for analysis and processing. In other words, the purpose of data preprocessing is to improve the prediction accuracy of artificial intelligence by transforming data into a form that can be processed by a computer and removing unnecessary information.

That is, the apparatus for analyzing religious sentiment may remove unnecessary morphemes by analyzing morphemes using a morpheme analysis library.

For example, since Hangul data can undergo various semantic changes according to research, tokenization is required. To this end, morphological analysis using KoNLPy's Okt (formerly Twitter) library can be performed. Here, various morpheme analysis libraries such as Komoran, Kkma, Hannanum, and Mecab exist in KoNLPy. In an embodiment of the present invention, since morpheme analysis is performed to use a method of removing only some unnecessary morphemes, the Okt library capable of simple and fast morpheme analysis can be used. Data pre-processing may be performed by removing numbers, josa, punctuation, and stop words from the morphologically analyzed comments. An example of data preprocessing can be shown in Table 2 below. In Table 2, Raw data is the comment data before data preprocessing, and Preprocessed Data shows the result of data preprocessing using the Okt library on the comment data.

In step S140, the apparatus for analyzing religious sentiment calculates sentiment values for positive and negative emotions for each religion by performing sentiment analysis on the comment data using a deep learning model.

For example, as shown in FIG. 2, the deep learning model according to an embodiment of the present invention applies a Keras model, one of the Python packages, and includes an input layer and an output layer. It can be configured to have 10 hidden layers between (Output Layers).

For the first hidden layer, TextVectorization, a function that vectorizes strings among the built-in functions provided by Keras, can be applied. At this time, as a factor of TextVectorization, max_tokens, which means the size of the entire word dictionary, can be set to 100,000, output_mode, which determines the output mode, can be set to an integer (int) type, and output_sequence_length, which indicates the size of the output vector, can be set to 100,000. Can be set to 64. After that, learning data can be applied to the TextVectorization function.

An embedding function that converts a positive integer (index), which is a vectorized character string, into a dense vector of a fixed size may be applied to the second hidden layer. Among the parameters of the Embedding function, input_dim, which is the size of the vocabulary list, that is, the maximum integer index + 1, can be set to a value obtained by adding 1 to the max_tokens value, one of the TextVectorization factors. output_dim, which determines the dimension of the dense vector of the output value, may be set to 200.

In the third hidden layer, a flatten function that flattens an input dense vector and converts it into an array may be applied.

The fourth to tenth hidden layers may have 32 units, and an activation function may be set to be a Rectified Linear Unit (ReLU). At this time, in order to prevent overfitting in the learning process, L2 regularization with a value of 0.001 may be added.

The output layer has one node, and the activation function can be set to be sigmoid. Since there must be only one output value per one comment data, the output node must be one and must have a decimal output value between 0 and 1, so the sigmoid function can be used. In this case, the closer the output value is to 1, the more positive it is, and the closer it is to 0, the more negative it is.

On the other hand, the indicator that enables the neural network to learn is the loss function. In an embodiment of the present invention, the purpose is to select one of two binary values, and since labels of data are independent, binary cross entropy can be set as a loss function. Neural network training must find optimal parameters that minimize the loss function value based on these loss function indicators. This process is called optimization, and the optimal parameter is one of the variables required to compile a Keras model. In an embodiment of the present invention, Adam, which is a part of gradient descent, may be set as an optimization function while learning the neural network. Here, Adam is an algorithm that adjusts the renewal strength of learning by reducing the amount of learning and calculating the speed.

For example, a total of 32,474 emotionally labeled data by the user is about 4.9% of the total data, so it may be insufficient as training data. To compensate for this, 200,000 Naver movie review data can be additionally included in the training data. As a result of neural network training, the accuracy of the training data was about 95%, and the loss value was about 0.16. The accuracy of the test data is about 83%, and the loss value is about 0.49.

The deep learning model created in this way was applied to the entire comment data. For a total of 748,020 comments, data pre-processing was performed in the same way as the training data, and then the sentiment value of each comment was predicted through the trained deep learning model. The predicted sentiment value may be output as a real number between 0.0 and 1.0 through sigmoid, which is an activation function of a final output node.

According to another embodiment, a KoBERT model may be applied to the deep learning model.

The BERT model is designed to solve various natural language processing problems through simple fine-tuning by pre-training on large data sets. The BERT model is a model that utilizes the encoder part of a transformer, and has strengths in various tasks of natural language processing. In the present invention, sentiment analysis can be performed on comment data using the KoBERT model, which is a Korean BERT model developed by SKTBrain.

For example, the PyTorch API provided by KoBERT can be used. Training data and test data can be divided into 7:3. The maximum data length max_len can be set to 64, batch size to 64, epochs to 8, and learning_rate to 5e-5. As a result of learning, the accuracy of the training data was about 92%, and the loss value was about 0.17. And, the accuracy of the test data was about 94%.

The KoBERT deep learning model generated in this way can be applied to all comment data in the same way as the Keras deep learning model described above.

In step S150, the apparatus for analyzing religious sentiment calculates sentiment values for each date using the comment data for which sentiment labeling has been performed.

The emotional value per first date may be derived from a weighted average. That is, the weighted average (W) per day can be calculated using the following equation.

where k is the number of articles per day, i is the article index, count(i) is the total number of comments for article index i, avg(i) is the average sentiment value of the article, and M is the total number of all articles on that day is the sum of the number of comments.

In Equation 1, the number of comments per date may be biased. For example, if there is one comment on a certain day and it is positive, the positive weight may be 100%. And, if there are 100 comments on another day and 90 positives, the positive weight can be 90%. That is, the positive weight may appear higher on a day when there is one positive comment than on a day when there are 90 positive comments out of 100 comments.

Therefore, after calculating the weighted average for each date, the weighted average can be corrected by considering the total number of replies for the entire period using the following equation. That is, the sentiment average (X) for each date considering the total number of comments for the corresponding date for the entire period may be calculated using the following equation.

Here, N is the total number of comments per day, L is the minimum number of comments per day for the sentiment value to gain trust, e is the average sentiment value for the entire period, and W is the weighted average value per day calculated in Equation 1 .

3 to 5 are diagrams showing experimental results of a method for analyzing religious sentiment using deep learning according to an embodiment of the present invention.

For the religious sentiment analysis method using deep learning according to an embodiment of the present invention, emotional values were predicted for all comments using a Keras framework-based deep learning model designed and implemented directly.

In the graph of FIG. 3, the green dots are the sentiment values before COVID-19, and the red dots are the sentiment values after COVID-19. The Y-axis is a real number between 0.0 and 1.0. The closer to 1.0, the more positive the trend is, and the closer to 0.0, the more negative. At this time, considering the fact that the data with a sensitivity value of 0.7 or more is less than 1% of the total data, the range of the Y-axis in the graph is set from 0 to 0.7. The X-axis is the date, one scale is the date of the news article, and only the year and month are displayed as numbers. Changes in sentiment by religion according to the period of collection for each keyword are shown in FIG. 3 .

On the other hand, the deep learning model learned using the PyTorch API of the KoBERT model was used to predict sentiment values for all comments. In the KoBERT model, the Y-axis range was set from 0 to 0.5, considering that less than 2% of the total data had a sentiment value of 0.5 or higher. Judging from the fact that Christianity had the lowest and highest density of negative emotions even before COVID-19, it appears that negative emotions were strong in the past. In the post-COVID-19 period, strong negative sentiment seems to have continued due to the case of group infection in a church in June and the issue of an assembly related to Liberation Day in August. It is confirmed that Catholics show a narrower range of emotional values in the post-COVID-19 period than in the previous period, and a slight decrease. There was no issue of group infection, but it is assumed that it was influenced by other religions. In Buddhism, the difference in emotional values is minimal in the post-COVID-19 period.

Shincheonji has a wide distribution of emotional values in the pre-COVID-19 period, but maintains a consistently narrow and low number in the post-COVID-19 period. It can be confirmed that the case of Shincheonji group infection had a negative impact. After June, there is a section showing a relatively high sensitivity level. This was due to the issue of the arrest of Shincheonji Chairman Lee Man-hee in June and the issue of donation of blood plasma needed for COVID-19 research by Shincheonji members in July, respectively. However, this case only temporarily showed positive emotions, and it is difficult to say that the overall emotions toward Shincheonji have changed positively. In particular, there is a positive response to the issue of arresting the president of Shincheonji, but due to the nature of the issue of arrest, positive emotions cannot be regarded as positive emotions toward religion. Although there is no significant difference in the sensitivity to the religious keyword between the before and after period, it can be confirmed that a narrower and denser sentiment value appears in the later period.

In order to find out the reliability of the change in sentiment value, the average according to the period for each keyword was calculated. 4 and 5 show averages for each keyword according to the period in the Keras model and the KoBERT model, respectively. The Y-axis of the graph is the average by period, and the X-axis is the keyword. The bar on the left for each keyword is the average sentiment in the pre-COVID-19 period, and the bar on the right is the average sentiment in the post-COVID-19 period.

Referring to FIGS. 4 and 5 , in the sentiment analysis using the Keras deep learning model, the average sentiment value was lowered in all keywords except for Christianity. Christianity rose slightly higher in the later period. And, in the sentiment analysis using the KoBERT model, the average sentiment value was lowered in all keywords except Buddhism.

In the religious sentiment analysis method using deep learning according to an embodiment of the present invention, the significance level was set to 5% based on the reliability of 95%. The null hypothesis is 'there is no difference between the previous period and the later period', and the alternative hypothesis is 'there is a difference between the previous period and the later period'. Before and after emotional values were grouped per keyword and date, and the F-test of each group was conducted to confirm the presence or absence of a difference in variance. Since there is no directionality between the null and alternative hypotheses, a double-tail p-value was used. Using Microsoft Excel's TTEST function, the p-value of the two groups was calculated. The p-value for each keyword is shown in Table 3 below.

Referring to Table 3, as a result of the significance verification of the results derived from the sentiment analysis using the Keras model and the sentiment analysis using the KoBERT model, it is confirmed that significant emotional changes occur in Shincheonji, Catholicism, and religions.

6 is a diagram schematically illustrating the configuration of an apparatus for analyzing religious sentiment using deep learning according to an embodiment of the present invention.

Referring to FIG. 6 , an apparatus for analyzing religious sentiment using deep learning according to an embodiment of the present invention includes a processor 10, a memory 20, a communication unit 30, and an interface unit 40.

The processor 10 may be a CPU or a semiconductor device that executes processing instructions stored in the memory 20 .

Memory 20 may include various types of volatile or non-volatile storage media. For example, memory 20 may include ROM, RAM, and the like.

For example, the memory 20 may store instructions for performing a religious sentiment analysis method using deep learning according to an embodiment of the present invention.

The communication unit 30 is a means for transmitting and receiving data with other devices through a communication network.

The interface unit 40 may include a network interface and a user interface for accessing a network.

On the other hand, the components of the above-described embodiment can be easily grasped from a process point of view. That is, each component can be identified as each process. In addition, the process of the above-described embodiment can be easily grasped from the viewpoint of components of the device.

In addition, the technical contents described above may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program commands recorded on the medium may be specially designed and configured for the embodiments or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. - includes hardware devices specially configured to store and execute program instructions, such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter, as well as machine language codes such as those produced by a compiler. A hardware device may be configured to act as one or more software modules to perform the operations of the embodiments and vice versa.

The embodiments of the present invention described above have been disclosed for illustrative purposes, and those skilled in the art having ordinary knowledge of the present invention will be able to make various modifications, changes, and additions within the spirit and scope of the present invention, and such modifications, changes, and additions will be considered to fall within the scope of the following claims.

10: Processor
20: memory
30: Ministry of Communications
40: interface unit

Claims (9)

In the religious sentiment analysis method using deep learning performed by the religious sentiment analysis device,
Collecting comments on news articles related to Corona 19 (COVID-19) and religion to generate comment data;
performing emotion labeling on the generated comment data;
performing data pre-processing on the comment data for which the sentiment labeling has been performed;
Calculating emotional values for positive and negative emotions for each religion by performing sentiment analysis on the preprocessed comment data using a deep learning model; and
Calculating a sentiment value for each date using the preprocessed comment data,
The deep learning model applies a Keras model, one of the Python packages, or a KoBERT model, a Korean BERT model,
The Keras model is configured to have 10 hidden layers between an input layer and an output layer,
The 10 hidden layers,
a first hidden layer to which TextVectorization, a function that vectorizes a character string, is applied;
a second hidden layer to which an embedding function for converting the vectorized character string into a dense vector having a fixed size is applied;
a third hidden layer to which a Flatten function is applied to flatten the dense vector and convert it into an array; and
With 32 units, the activation function is set to be ReLU (Rectified Linear Unit), and in order to prevent overfitting in the learning process, the fourth hidden layer to which L2 regularization is added. Including a 10th hidden layer,
The sentiment value for each day is derived from a weighted average per day,
The weighted average (W) per day is calculated using the following formula,

where k is the number of articles per day, i is the article index, count(i) is the total number of comments for article index i, avg(i) is the average sentiment value of the article, and M is the total number of all articles on that day is the sum of the number of comments,
Religious sentiment analysis method using deep learning, characterized in that the sentiment average (X) for each day corrected for the weighted average per day is calculated using the following formula.

Here, N is the total number of comments per day, L is the minimum number of comments per day for the sentiment value to gain trust, e is the average sentiment value for the entire period, and W is the weighted average value per day calculated in Equation 1
According to claim 1,
The step of performing the emotion labeling,
Religious sentiment analysis method using deep learning, characterized in that in the comment data, comments showing a positive trend are labeled with positive identifiers, and comments showing negative tendencies are labeled with negative identifiers.
delete delete delete delete delete delete In the religious sentiment analysis device using deep learning,
memory for storing instructions; and
Including a processor that executes the instructions,
The command is
Collecting comments on news articles related to Corona 19 (COVID-19) and religion to generate comment data;
performing emotion labeling on the generated comment data;
performing data pre-processing on the comment data for which the sentiment labeling has been performed;
Calculating emotional values for positive and negative emotions for each religion by performing sentiment analysis on the preprocessed comment data using a deep learning model; and
Performing a religious sentiment analysis method using deep learning, which includes calculating a sentiment value for each date using the preprocessed comment data,
The deep learning model applies a Keras model, one of the Python packages, or a KoBERT model, a Korean BERT model,
The Keras model is configured to have 10 hidden layers between an input layer and an output layer,
The 10 hidden layers,
a first hidden layer to which TextVectorization, a function that vectorizes a character string, is applied;
a second hidden layer to which an embedding function for converting the vectorized character string into a dense vector having a fixed size is applied;
a third hidden layer to which a Flatten function is applied to flatten the dense vector and convert it into an array; and
With 32 units, the activation function is set to be ReLU (Rectified Linear Unit), and in order to prevent overfitting in the learning process, the fourth hidden layer to which L2 regularization is added. Including a 10th hidden layer,
The sentiment value for each day is derived from a weighted average per day,
The weighted average (W) per day is calculated using the following formula,

where k is the number of articles per day, i is the article index, count(i) is the total number of comments for article index i, avg(i) is the average sentiment value of the article, and M is the total number of all articles on that day is the sum of the number of comments,
Religious sentiment analysis device using deep learning, characterized in that the emotional average (X) for each day corrected for the weighted average per day is calculated using the following equation.

Here, N is the total number of comments per day, L is the minimum number of comments per day for the sentiment value to gain trust, e is the average sentiment value for the entire period, and W is the weighted average value per day calculated in Equation 1

Images