n this project we aim to 1) Predict if a game review is helpful or not 2) Extract aspects from helpful game reviews. This is helpful for gamers in identifying the most interesting aspects as well as the most disliked aspects of a game before purchasing it.
ULMFit model is trained on Steam reviews to identify if a review is helpful or not. The trained ULMFit model predicts Metacritic reviews if they are helpful or not. Predicted reviews are split by top 3 genres based on number of games - action, sports and fantasy. Predicted Metacritic reviews for each genre are fed into the aspect extraction model. For aspect extraction, we have used an unsupervised neural attention model. Traditional topic models for aspect extraction tend to not have highly coherent aspects and they don't have very high interpretability since they consider all words are generated independently. This model improves coherence since it uses neural word embeddings which consider the distribution of word co-occurrences. Further, the interpretability of the aspects have been improved by splitting the dataset by genre. It was able to predict more granular aspects particular to each game genre as opposed to aspects generated from the entire gaming dataset.
