Children’s Sentiment Analysis From Texts by Using Weight Updated Tuned With Random Forest Classification

Abstract

The sentiment analysis of children is fundamental to understand the emotional responses in literary narratives, where significant implications are made in terms of educational and therapeutic use. Traditional text classification models, which include Bi-LSTM, Decision Tree, and ensemble models, are often characterized by reduced accuracy that can be explained by low data quality, overfitting, and inadequate optimization of hyperparameters. This paper uses a sentiment-annotated 4,000 short stories dataset. The preprocessing pipeline included the cleansing of the text, the process of breaking down into sentences and the removal of unnecessary symbols to standardize the inputs. An eight-layer LSTM network was used to perform feature extraction followed by dimensionality reduction using PCA and Truncated SVD. Classical machine learning models were trained simultaneously with a transformer-based ELECTRA model using weighted cross-entropy loss to counter class imbalance, including Random Forest and Decision Tree. The accuracy, precision, recall, and F1-score were used to evaluate the performance of the model, with ELECTRA achieving the best scores in each of the measures (93.4%). Flask-based web interface was created to enable interactive user input, prediction and presentation of sentiment results. The explainable AI approaches, LIME and SHAP were used to clarify predictions at the feature level, providing transparency and actionable insights. This approach has a significant improvement in anticipated precision and interpretation as opposed to traditional methods.