Hybrid CNN-LSTM Architecture for Bilingual Next-Word Prediction in Punjabi-English Social Media Texts

This paper presents a novel hybrid Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) architecture for next-word prediction in Punjabi-English bilingual social media texts. The proposed model aims to capitalize on the strengths of CNNs for detecting local contextual features and LSTMs for capturing long-range dependencies in a sentence, offering a more comprehensive understanding of the text. Extensive preprocessing tasks, including noise removal, text normalization, and script identification, were performed on a large dataset of Punjabi-English code-mixed social media texts to prepare them for the model. The model displayed considerable promise, achieving a training accuracy of 92.83%. These results demonstrate the effectiveness of this hybrid CNN-LSTM model for bilingual next-word prediction tasks, paving the way for more nuanced language model development for mixed-script and mixed-language scenarios prevalent in social media texts.