Constructing Sentiment Sentence Embedding Model Using Transfer Learning

TL;DR: A model is constructed by fine-tuning Google's Universal Sentence Encoder using a Deep Neural Network and the embedding layer will be allowed to retrain through backpropagation on the Sentiment dataset.

Abstract: For Sentiment classification a suitable embedding to train a machine learning model is needed. A lot of research has been done on generating embeddings that are capable of storing the context of words, but they are inefficient for sentiment extraction as they cannot differentiate words used in the same situation but contrasting sentiment polarity. To address this issue, a lot of researchers have proposed different embedding models, but most of them are word embeddings where lot of semantic information is lost. We have constructed a model by fine-tuning Google's Universal Sentence Encoder using a Deep Neural Network. The embedding layer will be allowed to retrain through backpropagation on the Sentiment dataset. Once training is completed, the hidden and output layers will be discarded, leaving only the USE layer, the output of this layer will be used as our Sentiment Embedding.