Master Student project at Ruhr West University of Applied Sciences.
|Students||Ivan Cedric Mabou and Franck Wafo|
|Supervision||M. Sc. Stephan Lehmler|
This study focused on the application of Transfer Learning (TL) to Sentiment Analysis (SA) using pre-trained word embeddings, as well as the transfer of different neural layers. We investigated and demonstrated the extent to which the domain and the task influence TL performance in SA. We conducted transfer experiments on three datasets, one of which was chosen as the source dataset and the others as the target datasets. We also considered two tasks: Binary SA and Multi- label SA. First, we performed transfer learning based on a binary SA. Second, we performed transfer learning based on a multi- label SA. Our observations show that even a less semantically similar dataset can provide better text classification than a highly similar dataset. Moreover, we found that the accuracy value after multi-label SA was lower than that of binary SA.