Not too long ago I wrote about SmartRatings, a product review site that aggregates expert reviews for a wide range of consumer products. To calculate the aggregated product rank SmartRatings uses the fact that many expert reviewers themselves give the product they review a numerical rank, which SmartRatings then brings to a common denominator and uniformly lists in a nice filterable way.
This approach conquers by its simplicity but has a very significant drawback. It doesn’t allow you to differentiate products by qualities since overall rank is often the only number SmartRatings can obtain from the expert reviews and hence the only number it uses for aggregation. The things become even more complicated if you take into account that each reviewer gives its own weight to different product qualities and the resulting overall rank may mean little to you if your priorities are different from those of the reviewer. For example for some long battery life is the single important quality in choosing a laptop while others give the CPU speed higher priority.
One product - one rank?
How do you enhance and give better structure to your product reviews? If you had direct access to the people writing reviews you would try to target each product quality separately, exactly what Buzzillions is trying to do. But what if the thousands of completed reviews was all you had? This is where semantic analysis comes in play. ReviewGist is a small New Delhi, India based startup that aims at building a product discovery and research tool for consumer electronics that will overcome these limitations. In fact they already have a live site that I have found pretty usable.
How does ReviewGist work?
In the heart of ReviewGist is a web scraping algorithm that goes to review sites and parses the pages with product reviews.
ReviewGist gathers review information for different products from almost all trusted online review sites. Our patent pending deep semantic analysis engine then takes over and extracts out the subjective opinion from these collected reviews. Essentially, we figure out the specific opinions expressed by the reviewer about the product in question.
The opinions are then pieced together to give you a concise and quantitative description of strong and weak sides of each product in the ReviewGist database. Here is for example how the Apple iPhone review looks:
The algorithm is not perfect
On some occasions I have noticed facts were misinterpreted and assigned to wrong categories (e.g. “weak flash” was assigned to battery performance in a camera review). Nishant Soni, the CEO, has confessed in an email to me that “a small amount of human intelligence” is involved in decision making to overcome the current limitations of NLP algorithms. In addition the system is learning from human tagging and the precision should improve over time.
ReviewGist uses semantic algorithms in its bottom-up approach to aggregating product reviews. As a developer I personally prefer it over top-down approach used by most other sites since it gives you better flexibility with how you can present the information to the visitor. It may however suffer from a slow adoption rate since it heavily relies on the quality of analysis these these algorithms can produce, something that still has a long way to perfection.