Deep Learning Framework for Sentiment Prediction using Residual Connections in Bidirectional – Gated Recurrent Unit

Sentiment analysis plays an essential role in Natural Language Processing (NLP) for differentiating emotions and opinions expressed in various pieces of text. However, existing algorithms face challenges in handling complex language patterns and capturing long-term dependencies, thereby increasing overall computational cost. This research aims to design an improved sentiment analysis model that enhances accuracy and efficiency while addressing gradient-related limitations in deep networks. This research proposes a Residual Bidirectional Gated Recurrent Unit (RBi-GRU) algorithm for effective sentiment analysis, leveraging residual connections to improve accuracy and efficiency. Residual connections are incorporated into the Bi-GRU network to facilitate gradient flow across layers and mitigate the vanishing gradient problem during training. It also enables deeper networks by protecting data from earlier layers, which further enhances feature representation. Additionally, tokenization, stemming, and global vector-based word representations (GloVe) are employed during preprocessing to capture the semantic relationships and meanings of words, thereby improving contextual understanding in sentiment analysis. The developed RBi-GRU algorithm achieves 98.74% accuracy, 98.99% precision, 98.32% sensitivity, and 98.64% F1-score on the Sentiment140 dataset, compared with the Rectified Linear Unit-based Gated Recurrent Unit (ReLU-GRU). ABSTRAK: Analisis sentimen memainkan peranan penting dalam Pemprosesan Bahasa Semula Jadi (NLP) bagi membezakan emosi dan pendapat yang dizahirkan dalam teks; namun, algoritma sedia ada menghadapi cabaran pengendalian corak bahasa yang kompleks serta kebergantungan jangka panjang, sekaligus meningkatkan masa pemprosesan. Kajian ini bertujuan mereka bentuk model analisis sentimen berketepatan tinggi dan cekap sambil menangani kekangan berkaitan kecerunan rangkaian mendalam. Sebuah algoritma Unit Kawalan Berulang Baki Dua Arah (RBi-GRU) dicadangkan dengan gabungan baki ke dalam rangkaian Bi-GRU bagi memudahkan aliran kecerunan antara lapisan dan mengurangkan masalah lenyap kecerunan semasa latihan, di samping membolehkan pembinaan rangkaian lebih mendalam dan meningkatkan perwakilan ciri. Selain itu, teknik prapemprosesan seperti penandaan token, pengakaran (stemming), serta penggunaan Representasi Kata Vektor Global (GloVe) diaplikasi bagi menangkap hubungan semantik dan makna kontekstual perkataan dengan lebih berkesan. Dapatan kajian eksperimen menunjukkan bahawa algoritma RBi-GRU mencapai ketepatan 98.74%, kejituan 98.99%, kepekaan 98.32%, dan skor F1 sebanyak 98.64% pada set data Sentimen140, sekaligus mengatasi prestasi model Unit Kawalan Berulang berasaskan Unit Pembetulan Linear (ReLU-GRU).