Abstract
The public opinion expressed on review or blogging sites and social networking platforms can be the source for the extraction of very critical information related to feelings and emotions of mass towards the subject matter in the field of commerce and governance. Natural Language Processing (NLP) and Artificial Intelligence can be used for sentiment analysis of this textual information. For text processing, NLP applications nowadays rely on pre-trained embeddings derived from large corpora such as news collection and web crawlers. There are many pre-trained word embeddings available. However, no study found which compares the accuracy achieved using these embeddings. In this paper, we worked on different kinds of word embeddings (pre-trained and untrained) and derived a comparison concerning accuracy for sentiment analysis applications using Deep Learning (DL) models. We found that the deep learning models perform better with pre-trained embeddings compared to Keras default (untrained) embedding.
