I wonder if any NLP approaches could yield an improvement on this? I immediately think of doing some NER on the review (looking for the title, authors, chapters, even specific topics) and if none of the above are mentioned, or at least not mentioned "enough", then it can be flagged for review as potential spam. Likewise, if you did sentiment analysis on a 5 star review, but the sentiment was either neutral or negative, it's likely not a very useful review.
I'm sure there's a lot that could be done with this, but some run of the mill NLP seems like it could at least help. I'm not sure the plausibility of this at a large scale, but it seems like an interesting problem nonetheless.
I would say just include a code inside the box that you can type in to validate your review as an actual purchaser. Then Amazon could heavily weight these reviews, at least for products known to be prone to scam reviews.
They already have a system for this: "Amazon Verified Purchase" will appear below the title of a review from someone who bought the book. Of course, they had to have purchased it from Amazon. They could default to only showing verified reviews by default or weighting verified reviews higher than unverified ones. It's too easy to create accounts and reviews to keep letting this just slide.
I'm sure there's a lot that could be done with this, but some run of the mill NLP seems like it could at least help. I'm not sure the plausibility of this at a large scale, but it seems like an interesting problem nonetheless.