Hybrid Deep Neural Network G-LSTM for Sentiment Analysis on Twitter: A Novel Approach to Disaster Management

Social media sites, which may provide both real-time updates and user-generated content, have been increasingly important in recent years as information resources during crisis occurrences. Twitter, with its massive user base and simple interface, has attracted the most attention among these channels. The ability to track public opinion, feelings, and reactions in the wake of disasters has made social media sentiment analysis a crucial tool for emergency managers. This information can help emergency response teams, government agencies, and humanitarian organisations prioritise needs and allocate resources more efficiently [1].

However, it is extremely difficult to manually analyse and extract useful information from the large amount of user-generated content on social media due to the data's volume and velocity. When applied to the text data collected from social media platforms, traditional sentiment analysis approaches frequently fall short. This highlights the expanding want for cutting-edge computational methods that can automate sentiment analysis and supply real-time insights during crisis scenarios.For Twitter sentiment analysis in the field of disaster management, this research proposes the use of a hybrid deep neural network model called G-LSTM (Gated Long Short-Term Memory). Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTM) networks are both potent variations of recurrent neural networks (RNNs) noted for their capacity to capture sequential dependencies and handle long-term dependencies, respectively; both abilities are combined in the G-LSTM design. The G-LSTM model integrates the strengths of GRUs and LSTMs to enhance the precision and speed with which disaster-related Twitter data can be analysed for sentiment.

Objectives:

• One goal was to create a G-LSTM model, a hybrid deep neural network, to analyse Twitter data for sentiment in the context of disaster management.

• Second, want to see how well the G-LSTM model can predict emotion labels.

• In order to (3) show how the G-LSTM model may be used in practise for real-time sentiment analysis during catastrophic events; and (4) get people talking about it.

• This paper's remaining sections are structured as follows: The second section includes a survey of related research on the topic of catastrophe managementusing social media sentiment analysis.

• The approach is laid out, which covers topics including data collecting, preprocessing methods, and the G-LSTM framework. Results and metrics for assessing the experiment's success. The results and potential uses of the proposed G-LSTM model.The report finishes with a review of the findings and recommendations for future study in Section 6.

In order to help emergency response and aid organisations make better decisions and allocate resources more efficiently during disasters, this study aims to contribute to the expanding field of social media analytics by developing an advanced sentiment analysis model tailored to Twitter data in the context of disaster management[2].

It's possible to send timely alerts and cautions, as well as helpful information and instructions, using social media. The increasing usage of social media could increase the number of people who hear warnings and notifications sent by emergency services [3]. This includes tornadoes and other forms of extreme weather. SM is the medium for the broader category of social interactions known as "extra-communicative." Everything from desktop computers to smartphones and tablets falls under this category of technology. With its help, users may create, share, and discuss content online, facilitating two-way communication between them and the wider world. The term "social media" (SM) is used to describe a group of online resources that build upon the ideas and architecture of Web 2.0 and enable users to create and share content collaboratively. Users of SM technologies should take an active role in content production, curation, and sharing. Word documents, audio files, video files, still photographs, electronic document files, PowerPoint presentations, and even GPS coordinates all count as forms of content. Sharing via social media makes it easier to talk to people on different platforms. It opens the door for people of both sexes to have a voice in shaping the future of social media. This technique allows for the rapid expansion of involvement through the production of real-time online events, the extension of online interactions offline, and the enhancement of live events via the Internet.

Cooperation between many organisations and sectors is essential for effective disaster risk management, regardless of the scale, severity, or scope of the disasters that must be mitigated [4, 5]. Some things you can do before a disaster strikes are more evident than others (like securing your home and storing up on non-perishable commodities like water and food), but you may also take some precautions online. Create your own 'group' on Facebook or any of the other many social media platforms out there. In case of an urgent situation, please share this link with your loved ones. Create a website where people can record their whereabouts and discuss their escape plans. Cities and towns provide PEP accounts for citizens to follow so that they are aware of emergency plans in their area. The locals should tune in to the reliable police, fire, and media outlets for updates [6]. There are a variety of apps for smartphones that can come in handy in a need. Having news and weather applications on your phone may help you keep informed and ready for any emergency circumstance, therefore they are a must-have. Disarray follows any major disaster. Having a plan in place beforehand might help you relax in tense situations. Social media allows for instantaneous communication between individuals, communities, and even governments [7]. Online users are recording and sharing images and videos of a disaster as it happens. Social media posts typically reach a wider audience than those made by mainstream media channels. In the case of a crisis, authorities will use social media to keep the public apprised of the whereabouts of aid and the anticipated arrival time of relief supplies. The magnitude of the damage can be documented and survivors can be located with the help of photos and videos published on social media. These images notify emergency personnel of the severity of the issue before they arrive. Connecting with the relevant groups and people ahead of time is advantageous because social media provides instant updates on road closures, evacuation routes, designated help spots, shelter places, and more in the event of a disaster [8]. You might utilise your social media accounts to inform those close to you of your current health situation. Update your Facebook, Twitter, and other social media profiles to let your friends and family know where you are and that you are safe. During a disaster, social media may be a great resource for finding out the latest information about road closures, evacuation routes, designated help zones, shelter places, and more. When disaster strikes, individuals use social media to coordinate efforts to find loved ones, animals, and misplaced items. There will be a great deal of people in need of medical care, food, and other basics, and this will assist get the word out about relief efforts and fundraising campaigns. Updates can be shared with clients, consumers, and vendors through a company's social media sites. The company's internet presence can be used to update clients on the status of the rebuilding, relocating, and reopening processes [9].

SM's widespread adoption in emergency management can be attributed to its widespread use, effectiveness, and simplicity. Notwithstanding potential disruptions to traditional means of communication, SM services & networking continue as usual. They aid and empower people and groups to work together for the greater good during all phases of emergency management, including prevention, preparedness, response, and cleanup. At times of natural disaster or other emergencies, it serves as the principal source of news. Via the Internet and text messages, it disseminates data on potential catastrophic outcomes. Awareness is raised within the impacted areas, which aids in recruiting emergency service volunteers and financial support. It serves as a vital link for those who have been separated from one another due to relocation. It details available humanitarian organizations and other resource hubs in the impacted region. The four most common uses of SM technologies among citizens identified in reviews of the literature are as follows: communication with family and friends; situation updates; situational/supplemental awareness; support in accessing resources. Social media aids communities before, during, and after disasters by reuniting and strengthening individuals and organizations for more effective response, recovery, and distribution of aid [10]. AI-Social Disaster is a software that collects and analyses social media conversations in real time during a disaster using AI services and algorithms [11].

This section discusses the suggested technique for predicting social media disasters based on Natural Language Processing (NLP) text data. Data from the Twitter API are first gathered, followed by EDA and preprocessing. The data is then split 85:15, with 85% used for training & 15% utilized for testing. Next, deep learning models for evaluation, including LSTM, GRU, as well as 1D Convolutional Neural Networks, will be used.

3.1 Data collection

Natural language processing (NLP) text data https://odsc.medium.com/20-open-datasets-for-natural-language-processing-538fbfaf8e38 from the Twitter API can be used to anticipate social media disasters, and includes both actual and fictitious disasters. use of a database containing 10,000 tweets with manually assigned categories. We also go over the NLP problem-solving basis. ID, keyword, location, text, and victim are only some of the categories found in the disaster dataset. There are more tweets that contain situational information than tweets that do not (a disparity that amounts to 5103 unique ids in the raw data).

In order to anticipate social media disasters using Natural Language Processing (NLP), various opportunities and considerations arise from the availability of a database consisting of 10,000 manually categorised tweets linked to both real and hypothetical disasters. One, it has a large database for training and assessing NLP models (10,000 manually categorised tweets). Due to the abundance of data, numerous facets of social media catastrophe prediction can be examined in depth. The manual labelling of the tweets' categories guarantees the presence of ground truth labels, which may be used as a trustworthy benchmark to measure the NLP models' results. This allows scientists to gauge how well their prediction models are doing.

3.1.1 Limitations

There is the potential for bias to be introduced by human annotators when tweets are manually categorised. Because of this inherent subjectivity, the accuracy and consistency of the labels in the dataset may be compromised.

It is important to assess the dataset's generalizability to other situations or languages, even though it may be a useful tool for analysing tweets about disasters.

It's possible that the dataset doesn't fully represent the intricacies and features of many crisis scenarios, reducing the generalizability of the results.

The time-sensitive nature of social media data, especially in the event of a natural disaster. Real-time social media disaster prediction may be hindered by the fact that the tweets in the database may only cover a certain time period.

Researchers should work on establishing and enhancing natural language processing (NLP) models that are optimised for social media catastrophe prediction. This can involve trying out new architectures, feature engineering methods, and cutting-edge methods like transfer learning and domain adaptability. To evaluate the models' robustness and generalizability, rigorous evaluation processes, such as cross-validation and testing on external datasets, should be implemented.

Due to the urgency of social media data during disasters, it is important that future studies focus on creating real-time prediction algorithms that can continually monitor and analyse incoming tweets. This would make it possible to respond and intervene in urgent circumstances more quickly. Overall, the availability of a database with 10,000 carefully categorised tweets connected to social media disasters gives tremendous prospects for NLP-based prediction models. Researchers, however, need to keep in mind the caveats of bias, generalizability, and temporal sensitivity. To improve the efficacy of social media disaster prediction, more work needs to be done in the future to address these limitations through data augmentation, model development, real-time analysis, and multilingual techniques

3.2 Pre-processing

N-gram analysis, punctuation removal, HTML tag removal, spelling revision, and vectorization for data cleaning are all steps in the pre-processing process. A text document that contains n consecutive objects, including words, numbers, symbols, and punctuation, is said to contain an n-gram. In many text analytics applications where word sequences are important, such as mood analysis, text classification, and text generation, N-gram models are helpful. The Python clean-text library comes with an inbuilt function called remove emoji.

3.3 Data splitting

Before developing the classification scheme, the data will be divided 85:15 between training and testing phases. Training data, not validation or testing data, is used when a deep learning tool is being taught to spot patterns or fulfill predefined criteria.

3.4 EDA

In addition to the traditional modeling & hypothesis testing effort, EDA is used to gain a deeper knowledge of the variables inside the data collection & their interrelationships. The appropriateness of the statistical methods you were planning to use for data analysis can also be determined with its aid. An outlier is a data point that deviates notably from both its immediate surroundings and the rest of the dataset by being noticeably higher or lower.

The distribution of tweets about disasters and tweets about no disaster is shown in Figure 1. Class 0 tweets (No catastrophe) outnumber class 1 tweets in terms of frequency. (disaster tweets).

The characters in tweets are depicted in Figure 2. The distribution of both appears to be nearly identical. Red is considered to be a disaster color, while green is considered to be a non-disaster color. The most popular character limit for tweets is between 120 and 140.

Figure 3 shows the words found in tweets. Both look to have nearly identical distributions. Green is regarded as a color that is not associated with disasters, whereas crimson is. The most common Twitter word count is between 15 and 20.

1.png

Figure 1. Class distribution

2.png

Figure 2. Characters in tweets

3.png

Figure 3. Words in tweets

4.png

Figure 4. Average word length in each tweet

5.png

Figure 5. Analysing stop words

Red indicates a disaster and green indicates a situation that is not a disaster, with the average word length of each tweet shown in Figure 4 falling between 2.5 and 12.5 characters.

Following analysis, Figure 5 displays the frequently used stop words, including "the," "a," "to," "and," "of," and others. most common stop word is they have more than 1400 values.

After research, Figure 6 shows the most common punctuation marks, such as "-" "!" ":" "@" "+" and others. The most common puntuation is "-" which has a value of more than 350.

6.png

Figure 6. Analyzing punctuations

3.5 Deep learning models

Disaster response using deep learning models such as 1D Convolutional neural network (CNN), Long Short Term Memory, and Recurrent Neural Networks (LSTM), and Gated Recurrent Unit (GRU) with attention mechanisms are used to identify situational tweets that comprise both situational and non-situational information.

3.5.1 Long Short-Term Memory (LSTM)

Long short-term memory networks, or LSTM, are a key component of deep learning. Specifically in sequence prediction problems, some varieties of neural networks with recurrent neurons (RNNs) can develop long-term dependencies. LSTM contains feedback connections and it can process the entire sequence of data, unlike single data points like pictures. Used in areas such as voice recognition and machine translation. The LSTM variation of RNNs excels in solving a wide variety of problems. Model architecture of LSTM is shown in Figure 7.

7.png

Figure 7. LSTM model

$i_t=\sigma\left(x_t U^I+h_{t-1} W^i\right)$                 (1)

$f_t=\sigma\left(x_t U^f+h_{t-1} W^f\right)$                 (2)

$o_t=\sigma\left(x_t U^o+h_{t-1} W^o\right)$                 (3)

$\tilde{C}_t=\tanh \left(x_t U^g+h_{t-1} W^g\right)$                 (4)

$C_t=\sigma\left({f_t}^* C_{t-1}+{i_t}^* \tilde{C}_t\right)$                 (5)

$h_t=\tanh \left(C_t\right) * o_t$                 (6)

3.5.2 Gated Recurrent Unit (GRU)

Recurrent neural networks can be classified as gated recurrent units or GRUs. It resembles an LSTM but only has two gates: a restart gate and an update gate. Also missing is an output gate. Since GRUs have fewer parameters, they are often easier and faster to train than LSTMs. Each Gated Recurrent Unit (GRU) in a Gated Recurrent Network (RNN) follows the same procedure as an RNN but employs a unique set of operations and gates. To solve the problem that traditional RNNs have, GRU employs two gates, the Update gate or the Reset gate. The Gated recurrent unit Model is depicted in Figure 8.

$z_t=\sigma\left(W_z \cdot\left[h_{t-1}, x_t\right]\right)$             (7)

$r_t=\sigma\left(W_r \cdot\left[h_{t-1}, x_t\right]\right)$              (8)

$\tilde{h}_t=\tanh \left(W \cdot\left[{r_t }^* h_{t-1}, x_t\right]\right)$               (9)

$h_t=\left(1-z_t\right)^* h_{t-1}+{z_t}^* \tilde{h}_t$                (10)

8.png

Figure 8. GRU Model