Towards Automatic Sentiment-based Topic Phrase Generation
For obtaining a comprehensive understanding and knowledge of customers’ expectations and demands, analysis of user-generated online product and service reviews is very important. Utilizing crucial insights from customers' feedback for developing business strategies can significantly improve product quality. In this paper, we have proposed a pretrained language model based encoder-decoder framework which generates a topic that describes the concise meaning of a customer's review text based on its corresponding polarity. We performed a comparative performance analysis of three language models, namely, BERT, ALBERT and GPT2 for the topic generation task on a dataset containing 8,124 customer reviewers along with topics and it's associates sentiment. In our experiment, we found that GPT2 model outperformed the other two models by achieving lower perplexity of 1.82.
PDF Abstract
