Natural Language Processing

My objective is to derive product descriptions from online reviews.

Product descriptions are typically more objective than reviews, since reviews are the buyer’s impression of a product.

However, many products have too many reviews that make it hard to interpret them all.

Therefore, shoppers usually only read a few of them and may miss relevant information.

To address this issue, we have come up with the idea to collate accurate review sentences into item descriptions or features.

Some of the reviews are already descriptive enough to be used as features.

To generate the features automatically, I employed semantic similarity to extract sentences from user reviews and remove sentences which do not provide much information or have similar content.

My goal is to select sentences from each review which accurately depict the product without any adjustments. These are the resulting features which form the final product review.

I have calculated the overall rating by extracting the total number of reviews and the average rating from the various websites reviewing the Item.

I have used a weighted average based on the number of reviews. Weighted average takes account of the number of reviews which the rating is based on. For example an item which has a five star rating based on one review from Store A and a 4.5 star rating based on 100 reviews from store B. Would have a weighted average rating of 4.5. Which is a more accurate reflection of its true rating. Rather than calculating the rating by just summing the ratings and dividing the ratings by 2 ,to get in this case 4.75.

I  have excluded review sites. This is because they often already use an accumulated score from several websites. Often for more than one product.

The scoring is from 1 to 5. Some websites score differently e.g. 1 to 10. In these cases I have converted the score of 1 to 5.