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.