On the Sensitivity of LSTMs to Hyperparameters and Word Embeddings in the Context of Sentiment Analysis

Recurrent neural networks are still providing excellent results in sentiment analysis tasks, variants such as LSTM and Bidirectional LSTM have become a reference for building fast and accurate predictive models. However, such performance is difficult to obtain due to the complexity of the models and the hyperparameters choice. LSTM based models can easily overfit to the studied domain, and tuning the hyperparameters to get the desired model is the keystone of the training process. In this work, we provide a study on the sensitivity of a selection of LSTM based models to various hyperparameters and we highlight important aspects to consider while using similar models in the context of sentiment analysis. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.