Automatic Monitoring Social Dynamics During Big Incidences: A Case Study of COVID-19 in Bangladesh

Topic modeling is a process of discovering hidden topics in a collection of texts Bashar et al. (2020a); Balasubramaniam et al. (2020). It can be considered as a factual show of topics through text mining. One of the most popular topics modeling technique Latent Dirichlet Allocation (LDA) (Blei et al., 2003; Bashar et al., 2020a) discovers topics based on word recurrence in a set of documents. LDA is incredibly valuable for finding a sensibly precise blend of topics inside a given record.

Topic modeling has been well studied for English text mining. For instance, Zhao et al. (2011) used unsupervised topic modeling in their research and compared the content of Twitter with the traditional news media “New York Times”. They used the Twitter-LDA model to find topics from a representative sample of the entire Twitter and then used text mining techniques to compare these Twitter topics with New York Times’ topics, taking into account the topic category and type. Wang and Blei (2011) developed an algorithm to recommend scientific articles to users in online communities. Their method combines the advantages of traditional collaborative filtering and probabilistic topic modeling. They applied collaborative topic modeling for recommending scientific articles. Wayasti et al. (2018) applied the Latent Dirichlet Allocation function in the research and extracted topics based on ride-hailing customers’ posts on Twitter. In their research, they used 40 parameter combinations of LDA to obtain the best combination of topics. According to the perplexity value, the customers discussed 9 topics in the post, including keywords for each topic. Tong and Zhang (2016) recommended two experiments to build topic models on Wikipedia articles and Twitter users’ tweets.

However, topic modeling has not been well studied for Bengali text mining, unlike English text mining. Das and Bandyopadhyay (2010b) used topic wise opinion summarization from Bengali text. They applied K-Means clustering and document-level theme relational graph representation. However, they did not use any topic modeling technique, such as LDA. Rakshit et al. (2015) applied a Multi-class SVM classifier for analyzing Bengali poetry and poet relations. They performed a subject-wise classification of poems into foreordained categories. Hasan et al. (2019) compared the performance of the LDA and LDA2vec topic model in Bengali Newspaper. Al Helal and Mouhoub (2018) used LDA for detecting the primary topics from a Bengali news corpus. However, they did not directly apply LDA in the Bengali text. Instead, they translated the Bengali text into English and then applied LDA to detect the topics. Rahman et al. (2019) used lexical analysis for sentence wise topic modeling. Their topic modeling was based on sentiment analysis. None of the existing works used Bengali text topic modeling for monitoring a pandemic or a major event.

In addition to English and Bengali, topic modeling in various languages is also studied. De Santis et al. (2020) analyzed a system that uses NLP pipelines, a theoretical framework for content aging to determine the qualitative parameters of tweets, and co-occurrence analysis to build topic maps chart splits to identify topics related to posts from Italian Twitter users. Han et al. (2020) extracted topics related to COVID-19 from Sina Weibo(Chinese microblogging website) text dataset through the LDA topic model.

The dynamic topic model is a cumulative model that can be used to analyze changes in document collection over time Bashar et al. (2020a). There are many studies on dynamic topic modeling for the English language. For example, AlSumait et al. (2008) showed that the LDA model could be extended to the online version by gradually updating the current model with new data, and the model has the ability to capture the dynamic changes of the topics. Dieng et al. (2019) researched D-ETM on three data sets and discovered the word probabilities of eight different topics that D-ETM learned over time. Nguyen et al. (2020) discovered latent topics from the financial reports of listed companies in the United States and studied the evolution of the themes discovered through dynamic topic modeling methods. Marjanen et al. (2020) discussed humanistic interpretation’s role in analyzing discourse dynamics through historical newspapers’ topic models. Bashar et al. (2020a) extracted five COVID-19 related topics from the Twitter dataset through LDA topic modeling, and they showed the changes in the extracted topics over time. Nevertheless, for the Bengali language, so far, there is no research on dynamic topic modeling. In this study, we study the evolution of the extracted COVID-19 related topics over time using dynamic topic modeling.

Text classification, moreover known as text labeling or text categorization, is categorizing content into organized bunches Bashar et al. (2020b); Bashar and Nayak (2020); Bashar et al. (2018); Bashar and Nayak (2020). By utilizing NLP, classifiers can naturally label text and, after that, relegate a set of predefined labels or categories based on its substance.

Many researchers worked on text classification in English. For example, Patil and Pawar (2012) used the Naive Bayes algorithm to classify website content. They divided the website content into ten categories, and the average accuracy of the ten categories was almost 80%. Bijalwan et al. (2014) used K-Nearest Neighbors, Naive Bayes, and Term-gram to classify text. They showed that in their research, K-Nearest Neighbors’ accuracy was better than Naive Bayes and Term-gram. Tam et al. (2002) showed that K-Nearest Neighbors was superior to NNet and Naive Bayes for English documents. Pawar and Gawande (2012) showed that Support Vector Machines’ performance is far superior to Decision Trees, Naive Bayes, K-Nearest Neighbors, Rocchio’s algorithms, and Backpropagation networks. Liu et al. (2010) showed that Support Vector Machines is better than K-Nearest Neighbors and Naive Bayes.

In addition to English text classification, some researchers have also classified Bengali text. For example, Mandal and Sen (2014) applied four supervised learning methods: (Naive Bayes, k nearest neighbor, Decision Tree classifier, and Support Vector Machine) for labeled web documents. They classified the documents into five categories: (Business, Sports, Health, Technology, Education). Chy et al. (2014) applied a Naive Bayes classifier to categorized Bengali news. Pal et al. (2015) described Naive Bayes classifier for Bengali sentence classification. They used over 1747 sentences in their experiment and got an accuracy of 84%. Kabir et al. (2015) used Stochastic Gradient Descent (SGD) classifier to categorize Bengali documents. Eshan and Hasan (2017) created an application that identifies abusive texts in Bengali. They applied Naive Bayes, Random Forest, Support Vector Machine (SVM) with Radial Basis Function (RBF), Linear, Polynomial, and Sigmoid kernel to classify the texts and compare the results among them. Islam et al. (2017) applied SVM, Naive Bayes, and Stochastic Gradient Descent(SGD) to classify Bengali documents and compare results of those classifiers. However, non of the existing works used Bengali text classification for monitoring a pandemic or a major event.

Sentiment Analysis refers to computationally recognizing and categorizing opinions communicated in a chunk of text. It is successfully used in commerce where they use it to track online discussions to identify social estimation of their brand, item, or benefit.

A lot of research work has been done in sentiment analysis for the English language. For example, Cui et al. (2006) have reviewed about 100,000 product reviews from various websites. They divided reviews into two main categories: positive and negative. Jagtap and Dhotre (2014) applied the Support Vector Machine and Hidden Markov Model, and the Hybrid classification model is well suited for extracting teacher feedback and evaluating sentiments. Alm et al. (2005) divided the seven emotional words into three polarity categories: positive emotion, negative emotion, and neutral, and the Winnow parameter adjustment method used can reach 63% accuracy. For extracting the Twitter sentiment, Agarwal et al. (2011) applied unigram, tree model, and feature-based model. Bashar et al. (2020a) used Convolutional Neural Network to extract sentiments related to COVID-19 from the Twitter dataset.

Some research used sentiment analysis in Bengali texts. For instance, Das and Bandyopadhyay (2010a) classified emotions into six categories: Happy, Sad, Anger, Disgust, Fear, and Surprise. Chowdhury and Chowdhury (2014) used sentiment analysis in Bangla Microblog Posts. They applied a semi-supervised bootstrapping method utilizing SVM and Maximum Entropy. Hasan et al. (2014) proposed a strategy to identify sentiments in Bengali texts by Contextual Valency Analysis. They employed the methodology of POS Tagger in their approach. Hassan et al. (2016) used recurrent neural networks to Romanize Bengali texts and analyze sentiments. In their experiments, they used Bangla and Romanized Bangla Text (BRBT) dataset. For Sentiment Analysis of Bangla Microblogs, Asimuzzaman et al. (2017) used Adaptive Neuro-Fuzzy Deduction Framework to anticipate extremity and utilized fluffy rules of speech in semantic rules. Mahtab et al. (2018) designed a model for sentiment analysis on Bangladesh Cricket news. They applied TF-IDF and SVM (Support Vector Machine) in their model and found 64.596% accuracy. Tripto and Ali (2018) used sentiment analysis on Youtube comments. Their research built a model based on deep learning that classifies a Bengali text into three classes and five sentiment classes. Tabassum and Khan (2019) used the Random Forest Algorithm to classify the sentiments in Bengali texts. Tuhin et al. (2019) applied Naive Bayes and a topic modeling approach to design an Automated System of Sentiment Analysis in Bengali Text. Their system classifies emotions into six categories: happy, sad, tender, excited, angry, and scared. However, non of the existing works used Bengali text sentiment analysis for monitoring a pandemic or a major event.

This pandemic situation has changed society and the country by a significant margin. The whole face of the country has changed completely. Some significant sectors of the nation, such as economic, social, political, have been affected massively. The education systems have been hit particularly hard. This research aims to automatically analyze the daily newspapers in Bangladesh to reveal what is going on in society and gain knowledge to comprehend the fundamental topics (or subjects) and sentiment arising and advance in the discussion.

This study will conduct a topic and sentiment analysis on a large collection of COVID-19 related news articles published in Bangladesh both in Bengali and English texts. The study will focus the analysis on both spatial and temporal dimensions. In the topic analysis, we used LDA-based topic modeling and dynamic topic modeling to find the topics, their evolution over time, and their time and space (location). We also analyzed what impact each topic had on particular areas. Then we analyzed the sentiment distribution over time and space to identify social sentiment in space and time. The experimental workflow of this study is shown in Figure 1.

First, we manually gathered a large collection of COVID-19 related news articles from Bangladeshi six most circulated daily newspapers. Along with the news, the collection contains geospatial and temporal information on the news. The dataset was then preprocessed by removing HTML, markers, and other non-relevant information such as adverts.

Then, we manually organized the news articles in a set of classes and sub-classes. Then we extracted the topics and the subtopics from the dataset. We used these classes and sub-classes to perform basic analysis such as comparing similarity and diversity in the news. These classes and sub-classes have also been used to qualitatively evaluate the accuracy of the topics discovered by LDA and labels predicted by classifiers before LDA and classifiers are employed for detailed analysis.

These publicly available News articles related to COVID-19 have been collected from the six most popular newspapers in Bangladesh from 21 January 2020 to 19 May 2020. The six newspapers are The Daily Prothom Alo, Bangladesh Pratidin, Kaler Kantho, The Daily Star, Newage, and The Daily Observer. A total of 15,565 news articles are collected from these six newspapers. From every news article, we extracted the news title, the main body of the news, a summary of the news (i.e., first few lines of the news body), the published date, and the news incident’s location. We used Python’s BeautifulSoup and Newspaper3k tool for extracting the news content. BeautifulSoup is a popular Python package for parsing HTML and XML archives and one of the most popular web scraping tools. Newspaper3k is a user-friendly library for scraping the news articles and other related data from newspaper portals. It is built upon request and used to parse LXML. This module is an improved version of the Newspaper module and is also used for the same purpose. Table 1 summarizes the statistics of the collection. We call this collection Comilla University COVID-19 News Collection (CoU-CNC). We made it available online¹¹1CoU-CNC Dataset: https://cutt.ly/djGILi2 for anyone for further analysis.

Out of these six newspapers, news articles in three newspapers (The Daily Prothom Alo, Bangladesh Pratidin, Kaler Kantho) are composed in the Bengali language, and in the other three newspapers (The Daily Star, The Daily Observer, New Age) articles are composed in the English language. There are 10,913 news articles in the Bengali language, and the remaining 4652 news articles are in the English language.

As the dataset has 4,652 articles in English and we wanted all articles in the same language to be better parsed, so we translated the English articles into Bengali via Python’s googletrans module. As a result, after translating these articles, all the articles are in the Bengali language. Then, we applied tokenization to split a string of text into smaller tokens. The news articles are split into sentences, and sentences are tokenized into words. Then, we applied noise removal (e.g., removing HTML tags, extra white spaces, special characters, numbers) to clean up the text. Then, we removed the stopwords from the document. As there is no build-in stopwords module for Bengali nltk, we manually created a stopword list and made it available online²²2Bengali-Stopwords: https://cutt.ly/2jXbDRB. Then, we expanded contraction. We set the minimum letter length to 6. We also removed all the words that were below the minimum letter length. There are no good resources for stemming and lemmatization in the Bengali language. So, we applied stemming and lemmatization to the tokens in our own process. After removing all the stopwords and other noises, there were a total of 80,693 tokens. There are some specific suffixes for the Bengali language. The suffix removal from the word has also been done with the help of Python. We used Bangla_Steamer.Steamer library of Python to improve accuracy. However, it did not show the expected results as the library is effective for a small number of Bengali words. To increase the accuracy of this 80,693 sizes lemmatized dictionary, we manually verified about 30000 most frequent tokens from 80693 words. We lemmatized where we needed to lemmatized manually, and we also corrected the incorrect and misspelled words where it was needed. Many more words are manually lemmatized and corrected through this manually 30,000 words check. We have published verified Bengali words on the Internet and titled “Modified Bengali Words” ³³3Modified Bengali Words: https://cutt.ly/8jE6GIC for further analysis.

To compare the number of news published and the COVID-19 cases of Bangladesh, we collected an open-source dataset⁴⁴4https://www.worldometers.info/coronavirus/country/bangladesh/ of confirmed COVID-19 cases and death cases of Bangladesh from March 8 to May 19.We also collected another open-source dataset⁵⁵5https://data.humdata.org/dataset/district-wise-quarantine-for-covid-19 of confirmed cases based on divisions and districts of Bangladesh from March 8 to May 19.

Time Series or temporal analysis of newspaper articles is utilized to observe the transient expansion during the pandemic. Time series decomposition includes considering a series of components in the time dimension: Level, Trend, Seasonality, and Noise segments. Level refers to the average value in the series, Trend refers to the increasing or decreasing value in the series, Seasonality refers to the repeating short-term cycle in the series, and finally, Noise refers to the random variation in the series. Decomposition gives a powerful supportive model for pondering time series and better arrangement issues during time series analysis and decision making. The additive model (Dagum, 2010) suggests that the segments are added as the following formula:

where $y(x)$ represents the additive model, $l(x)$ represents the observed level, $t(x)$ represents the trend, $s(x)$ represents the seasonality and $n(x)$ represents the noise or residual in the signal $x$. This model is linear. The change over a period of time is reliably affected by the similar sum of the linear trend as a straight line. A linear seasonality has a similar recurrence and abundance. On the other hand, A multiplicative model (Dagum, 2010) recommends that the components are multiplied together as the following formula:

where $y(x)$ represents the multiplicative model, $l(x)$ represents the observed level, $t(x)$ represents the trend, $s(x)$ represents the seasonality and $n(x)$ represents the noise in the signal $x$. A multiplicative model is exponential or quadratic when expanding or diminishing over the long run. A nonlinear pattern is a bent line. In this examination, we disintegrated the time series utilizing the multiplicative model. For Spatial Analysis, we used Tableau Software to compare the number of news published, and the number of COVID-19 confirmed cases geographically.

Analyzing the topics of news articles published during a major incident or a pandemic like COVID-19 can help monitor the situation and understand the public concerns, which is critical for government authorities and charity organizations to disseminate required resources and aids. However, in such a situation, a large number of news articles are published in various newspapers. We observed that as the situation deteriorated during the pandemic, newspapers had to publish much news on various topics. Manually analyzing the topics by reading a large number of articles is time-consuming and expensive. We utilized two unsupervised machine learning techniques: (a) LDA (Blei et al., 2003), a popular topic modeling technique, as static topic modeling to automatically find topics of articles published in newspapers, and (b) dynamic topic modeling in (Blei and Lafferty, 2006) to see how those topics evolve over the long haul.

LDA is a Bayesian probabilistic model that discovers topics and provides topic distribution over documents and word distribution over topics. It has two phases: (a) the first phase models each document as a composition of topics, and (b) the second phase models each topic as a composition of words. LDA utilizes word co-occurrences inside documents for discovering topics in a document assortment. Words occurring in an equivalent document are practically coming from the same topics, and documents containing comparative words will undoubtedly include comparable topics. In this research, the Gensim package in Python was utilized to execute the LDA model. We utilized every news article as a document in the topic modeling. Before applying the LDA topic model, we manually associated documents into general classes and sub-classes to know about the quality of LDA extracted fine-grained topics.

Then, we analyzed each LDA extracted topic’s temporal trends to see when a topic has been discussed more or published more in the newspapers. Finally, we analyzed each topic’s spatial distribution to see what effect each had in a particular place. We used Tableau software to analyze the spatial distribution of each topic.

The static topic modeling treats words as interchangeable and indeed treats documents as interchangeable. However, the presumption of replaceable documents is impractical for some assortments when accumulating along the time. For example, tweets, news articles, and insightful articles as they are advancing substance along time. The subjects in a newspaper article assortment develop, and it is essential to display the elements of the fundamental topics unequivocally.

Dynamic topic modeling extends the static theme, which illustrates the progression of the theme in consolidate. Dynamic topic modeling can catch the development of topics in a successively coordinated assortment of news articles. In this research, the articles are synchronized by week. We used the dynamic topic model to analyze discussion topics and topic changes over time.

Then we built a text classifier to verify their performance and predict the class, sub-class, and topics in the unknown (upcoming in the future) news articles. Such classification is important when we need to monitor a specific category (or class or group) of news. We made Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) models in Python utilizing Keras deep learning library for text classification. RNN is a special kind of neural network where the previous step’s output will be used as the current step’s input. In a traditional neural network, not all inputs and outputs are interdependent. However, interdependence is an important part of text data. In such cases, the model needs to predict the next word given the previous words, so the previous word must be stored. Thus RNN was born, which solved the problem with the help of hidden layers. The primary function of RNN is also essential, namely the hidden state. It can remember some information about the sequence.

RNN is a neural feedback network that operates on the internal memory. Since the RNN has a similar function for each piece of information and the current range’s output is based on the last count, the RNN is essentially recursive. When there is an output, it is copied and sent back to the relay network. The current input and the output of the previous input are taken into account in determining the prediction of the next word. Unlike direct feedback neural networks, RNNs can use their internal state (memory) to manage the input elements’ interdependence. That makes them useful for text data, handwriting recognition, or speech recognition. The architecture of an unrolled recurrent neural network is shown in Figure 5.

In Figure 5, first the model gets $x_{0}$ from the input sequence. Then it produces $h_{0}$, which is used in the next input to the model along with $x_{1}$. That is, both $h_{0}$ and $x_{1}$ become inputs to the next step. Then, $h_{1}$ and $x_{2}$ are input to the next step, and so on. Like this, RNN continues summarizing the unique circumstance in the hidden state while training. Then, it uses the summarized hidden state to classify the sequence (Bashar et al., 2020b).

We proposed a hybrid neural network model based on Convolutional Neural Network (CNN) and LSTM for sentiment analysis in Bengali texts.

Integrated models are used to solve various vision and NLP problems and improve a single model’s performance. The following subsections provide an overview of the LSTM and CNN models offered. In subsection 3.6, we described LSTM. In this research, we used two-layer bi-LSTM, word embedding include words in the news articles and provide sentiments.

Another part of our proposed structure is based on Convolutional Neural Network (CNN). CNN has very successful in various image processing and NLP tasks these last years. They are powerful in exploration, achieving local relevance, and data standards through learning. Generally, to rank text on CNN, different words in sentences (or paragraphs) need to be placed. Stacked to form a two-dimensional matrix, pleated filters (different lengths) are applied to the window. To use CNN for text classification, the different words stacked in a sentence are usually stacked in a two-dimensional matrix, and afterward, a convolution is applied to the word in the window in one word to be created applied a new function declaration. Then, a max-pooling is applied to the new function, and the combined functions of different filters are combined to shape a concealed portrayal. Completely associated layers trail these portrayals for the last estimate. The architecture of our CNN-BiLSTM Hybrid network model is shown in Figure 6.

We created a sequential model that includes an LSTM layer. Then we made our model sequential and adding layers. In the first layer, we applied a conv1D with 200 as a filter for CNN. After that, we applied two Bi-LSTMs on the second and third layer with an error of 0.5. Then we applied a dense network on the remaining levels. We also used Adam as an optimizer with tight hyperparameters and applied L2 adjustments to reduce overfitting as much as possible. We only used five epochs, as using more epochs resulted in overfitting and kept the stack size of 256, as it worked very well.

We built a text classifier to automatically categorize upcoming new news articles into classes, sub-classes, and topics. We implemented an LSTM recurrent neural network model in Python utilizing Keras⁶⁶6https://keras.io/ deep learning library for this classification.

We splited our data into 80% for training, 10% for validation, and 10% for testing. In the data preparation, first, we cleaned the text by removing unnecessary characters and stopwords. After cleaning the text, we tokenized the data using Keras Tokenizer. After that, we built a word index from it. Then we vectorized Bengali text by turning each text into a vector. We limited the dataset to the top 50,000 words and set the max number of words in each article at 1000. After that, we added padding and truncated to our data to make the input sequences uniform and the same length for modeling.

After cleaning the data, we selected pre-trained word embeddings ⁷⁷7Bengali Word Embeddings: https://cutt.ly/KjXmEio. Word embedding maps each word from the vocabulary to a vector of real numbers. We used these pre-trained word embedding in the embedding layer of our LSTM model.

In our classification model, the first layer is the embedded layer that employments 300 length vectors to represent each word. The second layer is an LSTM layer with 100 hidden units. The final layer is a dense layer, also known as the output layer. This final layer has a length of 8, 19, and 9 for the classes, sub-classes, and LDA-discovered topics, respectively. Softmax is used as the activation function for multi-class classification in the final layer. We used categorical cross-entropy as the loss function, Adam as the optimizer, and a batch size of 32. We used only five epochs as it worked quite well.

The experimental results for Precision, Accuracy, F₁ score and Recall are given in Table 4. The Precision, Accuracy, F1 score, and Recall are 47.80%, 44.39%, 45.13%, and 42.80%, respectively, for the eight classes. For the 19 sub-classes, Precision, Accuracy, F1 score, and Recall are 47.33%, 38.51%, 37.20%, and 30.82%, respectively. Furthermore, we also computed the performance of 9 LDA topics. For the 9 LDA topics, Precision, Accuracy, F1 score, and Recall are found 81.37%, 79.55%, 79.67%, and 78.10%, respectively.

We analyzed sentiment in the COVID-19 related news articles to see how positively and negatively the society was affected by the COVID-19 (or any big incident). We also analyzed the effectiveness of a hybrid CNN-BiLSTM model in identifying sentiments in Bengali texts. First, we manually labeled each news article according to positive or negative sentiment in the article. Then we trained CNN-BiLSTM to detect the sentiment of any upcoming new articles. After labeling the articles’ positive/negative sensation, we visualized the results in Figure 22. The figure shows that there were more negative sentiment news articles than positive sentiments.

We prepared our labeled news collection as 80% of for training, 10% for validation, and 10% for testing for sentiment analysis. Using Keras Tokenizer, we tokenized the data after cleaning the dataset. Then we built a word index and vectorize each text. We retrained the dataset to the 60,000 top words and set the max number of words in each article at 200 using feature selection. We added padding and truncated the data to make the input sequences uniform and the same for modeling.

After data preparation, we built our model. The first layer of the model is the embedding layer. We set the embedding dimension to 300 for embedding each word. In the second layer, we started a conv1D with 200 filters for CNN. Then in the third and fourth layers, we applied two Bi-LSTM with a dropout of 0.5. In the final layer, we used a dense network. We used Adam as the optimizer with finely tuned hyperparameters and applied L2 regularizations to reduce overfitting. We kept the batch size of 256 as it worked quite well. We used only five epochs that gave us reasonably good results.

After calculating the Precision, Accuracy, F1 score, and Recall, the performance and the sentiments of the classification presented in Table 5. The Precision, Accuray, F1 score, and Recall are 74.89%, 74.94%, 74.88%, and 74.89%, respectively, for sentiment analysis.

Figure 23 shows the spatial and temporal distribution of the number of positive and negative sentiment news articles during the pandemic. It shows how sentiments are changing over eight divisions for 20 weeks.