| Title | Working in Detail: How LSTM Hyperparameter Selection Influences Sentiment Analysis Results |
| Authors | Derra, Nicholas Daniel and Baier, Daniel |
| Year | 2020 |
| Volume | Archives of Data Science, Series A 6(1) / 2020 |
| Abstract | Sentiment analysis of written customer reviews is a powerful way to generate knowledge about customer attitudes for future marketing activities. Meanwhile, Deep Learning as the most powerful machine learning method is of particular importance for sentiment analysis tasks. Due to this current relevance, an LSTM network based on a literature review to solve the challenging classification task of the IMDB LargeMovie Dataset is created. Hyperparameters are varied separately from each other to better understand their single influences on the overall model accuracy. Furthermore, we transformed variants with positive impacts into a final model in order to investigate whether the impacts can be cumulated. While preparing the amount of training data and the number of iteration steps resulted in a higher accuracy, pre-trained word vectors and higher network capacity did not work well separately. Even though implementing the variants with positive influences together raised the model´s performance, the improvement was lower than some single variants. |