Conclusions & Future Work


This being a descriptive model, there is no one metric to evaluate the results. Therefore we conducted a series of experiments - both automated and manual to understand where the model can be improved. Some of the insights are as follows -

  • Extracting aspects turned out to be a major bottleneck for the whole process. Dependency grammar rules can be improved by studying more reviews.

  • Contextual knowledge should be used while calculating polarity scores. This will help give a better quantitative understanding of the reviews.

  • Our innovation of using word vectors and clustering algorithm did prove to be useful when 100 odd products were manually examined. This step made sure that similar aspect words were clubbed under one umbrella.

  • Any model is only useful when it solves a problem. Creating an interactive UI does that for us. Upon various demonstrations to a bunch of user, we realised that the UI really brings the model into life.

Future Work

  • Improving dependency rules

  • Contextual knowledge should be used while calculating polarity scores.

  • Developing a web-plugin for the UI - this will make sure that the user do not have to visit a new environment for accessing our model.

  • Currently, the model training time is around 7-8 hrs for 1M reviews. We can further work on optimizing this to reduce the model run time.