Authors:

• Vidyullatha Sukhavasi, Department of CSE, BVRIT HYDERABAD College of Engineering for Women, Bachupally, Telangana, 500090, India
• Venkatesulu Dondeti, Department of Advanced CSE, Vignan’s Foundation for Science, Technology & Research, Guntur, Andhra Pradesh, 522213, India

1. Introduction

Sarcasm is a complex form of expression where the intended meaning is often opposite to the literal meaning, making it difficult for traditional natural language processing (NLP) models to detect. In the context of social media, where both text and emojis are used to convey messages, sarcasm detection becomes even more challenging. Accurate detection of sarcasm is essential for improving sentiment analysis, content moderation, and other NLP tasks on social platforms.

This case study introduces an advanced transformer model designed to detect sarcasm in multilingual data, specifically in Hindi and English, by analyzing both text and emojis.

2. Problem Statement

Traditional transformer models have shown limitations in accurately detecting sarcasm due to the complexity of sarcastic expressions and the influence of emojis. Existing models that focus solely on text analysis are often insufficient in capturing the nuanced meaning conveyed through sarcasm, especially in multilingual contexts. Therefore, there is a need for a more effective and automated approach that can handle the intricacies of sarcasm in both text and emoji forms across different languages.

3. Objectives

The primary objective of this study is to develop a novel attention-based transformer model that enhances sarcasm detection by integrating textual and emoji-based features in multilingual datasets. The key objectives include:

Developing a robust model that can accurately detect sarcasm in both Hindi and English social media posts.

Improving the performance of sarcasm detection models by addressing the limitations of traditional transformers.

Achieving higher accuracy, precision, recall, and F-measure compared to existing sarcasm detection methods.

4. Methodology

4.1 Data Collection and Preprocessing

The study utilizes social media data from platforms like Twitter, consisting of posts in Hindi and English. The data preprocessing steps include:

Stop Word Removal: Eliminating common words that do not contribute to the sarcasm detection process.

Tokenization: Splitting text into tokens (words or phrases) to facilitate analysis.

Emoji Processing: Extracting and converting emojis into a machine-readable format.

4.2 Feature Extraction

Text Feature Extraction: The study employs the ATF-IDF (Advanced Term Frequency-Inverse Document Frequency) method to extract relevant textual features from the data. ATF-IDF helps in identifying the significance of words in the context of the entire dataset.

Emoji Feature Extraction: An emoji-to-vector model (E-VM) is developed to convert emojis into feature vectors. This model captures the semantic meaning of emojis and their contribution to the overall sentiment of the post.

4.3 Model Architecture

4.3.1 Gated Temporal Bidirectional Convolution Network (GT-BiCNet)

The text model is built using GT-BiCNet, which processes text data by considering both past and future context. The bidirectional nature of this network enhances the understanding of the sarcastic tone in the text.

4.3.2 Attention LSTM Model with ALABerT

The study introduces an Attention-based Long Short-Term Memory (LSTM) model, enhanced with ALABerT (Attention Layer-based Advanced BERT), to classify the combined text and emoji features. The attention mechanism allows the model to focus on critical parts of the input, improving the detection of sarcasm.

4.3.3 Deep Feature Fusion

The text features from GT-BiCNet and the emoji features from E-VM are combined using a deep feature fusion technique. This approach ensures that the model can analyze the relationship between text and emojis effectively.

4.3.4 Enhanced Pelican Optimization Algorithm (EpoA)

To minimize network losses and enhance the overall model performance, the Enhanced Pelican Optimization Algorithm (EpoA) is employed. EpoA optimizes the hyperparameters and reduces the error rates during training.

4.4 Classification and Evaluation

The final classification layer uses a softmax function to categorize the data into sarcastic or non-sarcastic classes. The model's performance is evaluated using metrics such as accuracy, precision, recall, and F-measure.

5. Results

The proposed model demonstrated superior performance in sarcasm detection compared to existing methods. On English Twitter data, the model achieved an impressive accuracy of 99.1%. Additionally, it delivered high precision, recall, and F-measure scores on Hindi Twitter data, showcasing its effectiveness in handling multilingual content.

The attention-based transformer model successfully captured the nuanced meanings conveyed through both text and emojis, addressing the limitations of traditional models. The use of deep feature fusion and EpoA further contributed to the model's high accuracy and robustness.

6. Conclusion

This study presents a significant advancement in sarcasm detection on social media by integrating text and emoji analysis in a multilingual context. The attention-based transformer model outperforms traditional models, providing a more accurate and efficient solution for understanding sarcastic expressions. The high accuracy and strong performance metrics indicate that this model could be instrumental in improving sentiment analysis and other NLP tasks in diverse linguistic settings.

7. Future Work

Future research could explore the application of this model to other languages and social media platforms, expanding its utility in global contexts. Additionally, incorporating real-time processing capabilities and expanding the dataset to include more diverse forms of sarcastic expression could further enhance the model's effectiveness.