Aspect-based Sentiment Detection using Deep Neural Networks and Transfer Learning
Date:
A novel approach to jointly detect aspect and the corresponding sentiment in a given sentence using DNN and transfer learning.
In this thesis, we propose a new neural network architecture for aspect-based sentiment analysis (ABSA). Instead of a pipeline approach where an aspect is first detected and then the corresponding sentiment polarities are classified, we use a joint formulation which is similar to the approach used by Schmitt et al. (2018). Furthermore, we also propose an input data transformation technique that provides some interesting advantages to our model. It helps in addressing problems that arise due to the huge number of output classes and it also provides additional capabilities that help our model to apply transfer learning approaches such as Progressive Neural Networks(PNN).
We conducted a thorough study of different ways in which a task such as sentiment analysis or ABSA can be formulated using various sequential modeling techniques such as HMM, CNN, RNN. We performed experiments with different variants of our model on a variety of datasets such as - SemEval-2016 Restaurant, SemEval-2016 Laptops, Organic food, GermEval-2017 Deutsche Bahn. We also experimented with different word embeddings such as GloVe, fastText, ELMo.
We were able to show that transfer learning approaches consistently gives a boost of about 1 or 2 F1 points across all the four datasets. With our experiments especially with GermEval-2017 data, we were able to show that our model performed better than the LSTM variant of the current state of the art approach from Schmitt et al. (2018). However, their CNN variant performed better than our best LSTM based model. To be the best of our knowledge, we are the first ones to present practical results on SemEval-2016 restaurant and laptop dataset for aspect-based sentiment analysis using a joint approach, therefore setting a baseline for future research.