International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 05 | May 2025
p-ISSN: 2395-0072
www.irjet.net
Enhanced Classification of Tweets and Emergency Response using BERT with AdamW Optimizer and NER Ranjith Kumar Thodupunuri1, Keerthi Edla2, Rahul Reddy Thoodi 3, Mayuka Andrasu 4, Abhinav Reddy Kolanu 5, Shiva Rama Krishna Chethi6 1Assistant professor, Computer Science and Engineering, Kakatiya Institute of Technology and Science, Warangal,
India
2,3,4,5,6 Computer Science and Engineering, Kakatiya Institute of Technology and Science Warangal, India
---------------------------------------------------------------------***--------------------------------------------------------------------Abstract - Social media sites, particularly Twitter, allow for rapid communication about disasters, they have turned as acute information sources during emergencies. The vital target of this study is the classification of tweets, which employs data from GitHub. The model gains an accuracy of 83.32% in classifying tweets into disaster and non-disaster categories by using Bidirectional Encoder Representations from Transformers (BERT) with the AdamW optimizer. This method outperforms base research that just addressed the identification and categorization of tweets about transportation disasters, and itattained 82% accuracy. Named Entity Recognition (NER) is used not just to classify tweets but also to extract principal entities fromthem, including disaster types, locations, and other pertinent information. This study points the determined location from the tweet on the map and also gives assistance regarding the safety measures and emergency contact details during emergency situations through a chat bot. This data can be instantly used by the government agencies, emergency personnel and can protect the lives of people.
This study includes cutting-edge Natural Language Processing (NLP) methods to amplify the classification of disaster tweets. Diverse machine learning algorithms were first tested for tweet classification, including Support Vector Classifier (SVC), Decision Trees, Random Forest, XGBoost, and BERT with the AdamW optimizer [6]. The BERT model with the AdamW optimizer was the most successful model for this task, showing the highest classification accuracy of 83.32%. A potent pre-trained language model called Bidirectional Encoder Representations from Transformers (BERT) has illustrated outstanding performance in a range of Natural Language Processing (NLP) tasks, particularly because of its capability to find out the contextual relationships in text [5]. We further improved BERT's performance by finetuning it with the AdamW optimizer, which resulted in high accuracy in classifying tweets concerned with disasters [8]. The accuracy attained in this study surpasses the 82% accuracy in an earlier study that considered tweets about transportation disasters [6].
Key Words: Bidirectional Encoder Representations from Transformers (BERT), Named Entity Recognition (NER), Adaptive Moment Estimation with Weight Decay (AdamW), Overfitting, Support Vector Classifier, Decision Tree, Random Forest, XG Boost
1.1 OBJECTIVES Classification of Tweets into Disaster or Non- Disaster Tweets: To differentiate tweets into disaster- related or non-disaster related, diverse machine learning algorithms were utilized, like SVC, Decision Tree, Random Forest, XGBoost and BERT with AdamW optimizer [5]. Out of them, BERT with AdamW Optimizer had given the maximum accuracy [6].
1. INTRODUCTION Social media sites, peculiarly Twitter, have become crucial for providing instantaneous updates during crises, as they allow individuals to share significant information, report disasters, and request help [1] , [2]. Social media has turned into a vital instrument in disaster management due to the rapid dissemination of information during emergencies, which facilitates quicker discovery and reaction [3], [4]. It is very much necessary to categorize tweets correctly about disasters because this information can be instantly used by government agencies, emergency personnel, and can protect the lives of people [5]
© 2025, IRJET
|
Impact Factor value: 8.315
Information Extraction Using Named Entity Recognition (NER): Named Entity Recognition (NER) is employed to find out the principal elements such as disaster categories, affected localities, and other crucial facts from tweets [3], [7]. To support the process of disaster management, NER is implemented as a dominant instrument for obtaining organized data [5].
|
ISO 9001:2008 Certified Journal
|
Page 86