# A novel attention-based deep learning model for improving sentiment classification after the case of the 2023 Kahramanmaras/Turkey earthquake on Twitter

## TL;DR

This paper introduces a new deep learning model for analyzing public sentiment on Twitter after the 2023 Turkey earthquake, offering faster and more accurate insights than traditional methods.

## Contribution

The novel MConv-BiLSTM-GAM model with attention mechanism improves sentiment classification accuracy by 3% compared to existing models.

## Key findings

- The proposed model achieves 93.32% accuracy in classifying Twitter sentiment after the Kahramanmaras earthquake.
- The model outperforms traditional deep learning models in capturing semantic dependencies in tweets.
- The study provides actionable insights for policymakers during disaster response.

## Abstract

Twitter has emerged as one of the most widely used platforms for sharing information and updates. As users freely express their thoughts and emotions, a vast amount of data is generated, particularly in the aftermath of disasters, which can be collected quickly and directly from individuals. Traditionally, earthquake impact assessments have been conducted through field studies by non-governmental organizations (NGOs), a process that is often time-consuming and costly. Sentiment analysis (SA) on Twitter presents a valuable research area, enabling the extraction and interpretation of real-time public perceptions. In recent years, attention-based methods in deep learning networks have gained significant attention among researchers. This study proposes a novel sentiment classification model, MConv-BiLSTM-GAM, which leverages an attention mechanism to analyze public sentiment following the 7.8 and 7.5 Mw earthquakes that struck Kahramanmaraş, Turkey. The model employs the FastText word embedding technique to convert tweets into vector representations. These vectorized inputs are then processed by a hybrid model integrating convolutional neural networks (CNNs) and recurrent neural networks (RNNs) with a global attention mechanism. This ensures careful consideration of semantic dependencies in sentiment classification. The proposed model operates in three stages: (i) MConv—Local Contextual Feature Extraction, (ii) bidirectional long short-term memory (BiLSTM)—sequence learning, and (iii) Global Attention Mechanism (GAM)—Attention Mechanism. Experimental results demonstrate that the model achieves an accuracy of 93.32%, surpassing traditional deep learning models in the literature by approximately 3%. This research aims to provide objective insights to policymakers and decision-makers, facilitating adequate support for individuals and communities affected by disasters. Moreover, analyzing public sentiment during earthquakes contributes to understanding societal responses and emotional trends in disaster scenarios.