An Improved Bees Algorithm for Training Deep Recurrent Networks for Sentiment Classification

Recurrent neural networks (RNNs) are powerful tools for learning information from temporal sequences. Designing an optimum deep RNN is difficult due to configuration and training issues, such as vanishing exploding gradients. In this paper, a novel metaheuristic optimisation approach proposed RNNs the sentiment classification task. The employs enhanced Ternary Bees Algorithm (BA-3+), which operates large dataset problems by considering only three individual solutions in each iteration. BA-3+ combines collaborative search of bees find optimal set trainable parameters recurrent architecture. Local with exploitative utilises greedy selection strategy. Stochastic gradient descent (SGD) singular value decomposition (SVD) aims handle gradients decision stabilisation strategy SVD. Global explorative achieves faster convergence without getting trapped at local optima has been tested on task classify symmetric asymmetric distribution datasets different domains, including Twitter, product reviews, movie reviews. Comparative results have obtained advanced language models Differential Evolution (DE) Particle Swarm Optimization (PSO) algorithms. converged global minimum than DE PSO algorithms, it outperformed SGD, DE, algorithms Turkish English datasets. accuracy F1 measure improved least 30–40% improvement standard SGD algorithm all Accuracy rates model trained ranged 80% 90%, while was able achieve between 50% 60% most performance good Tree-LSTMs Recursive Neural Tensor Networks (RNTNs) models, achieved up 90% some show that efficient, stable complex task, can problem RNNs.