Exploiting Emotions via Composite Pretrained Embedding and Ensemble Language Model

Abstract

Decisions in the modern era are based on more than just the available data; they also incorporate feedback from online sources. Processing reviews known as Sentiment analysis (SA) or Emotion analysis. Understanding the user's perspective and routines is crucial now-a-days for multiple reasons. It is used by both businesses and governments to make strategic decisions. Various architectural and vector embedding strategies have been developed for SA processing. Accurate representation of text is crucial for automatic SA. Due to the large number of languages spoken and written, polysemy and syntactic or semantic issues were common. To get around these problems, we developed effective composite embedding (ECE), a method that combines the advantages of vector embedding techniques that are either context-independent (like glove & fasttext) or context-aware (like XLNet) to effectively represent the features needed for processing. To improve the performace towards emotion or sentiment we proposed stacked ensemble model of deep lanugae models.ECE with Ensembled model is evaluated on balanced dataset to prove that it is a reliable embedding technique and a generalised model for SA.In order to evaluate ECE, cutting-edge ML and Deep net language models are deployed and comapared. The model is evaluated using benchmark datset such as MR, Kindle along with realtime tweet dataset of user complaints . LIME is used to verify the model's predictions and to provide statistical results for sentence.The model with ECE embedding provides state-of-art results with real time dataset as well.