How we produce our summaries
Pluribo's summary engine analyzes millions of user opinions on various products and produces a concise natural language summary for each product. The technical details of this process are complex and patent-pending, but the basic principles are straightforward.
1. Start with many reviews about a product
Pluribo aggregates millions of reviews from Amazon and other sources. The summarization process works best when we have a big enough sample of reviews on a given product (about 30 reviews at minimum).
2. Read each review to determine what product features it mentions
Pluribo automatically scans the text of each review to detect phrases that express an opinion about a product feature, such as "easy to install", or "I was disappointed with the sound quality". The natural language algorithm is smart enough to detect subtleties such as negation (e.g., "not easy to install"), but it isn't perfect -- we're constantly working to improve accuracy.
3. Extract a set of numerical feature scores for each review
Every Amazon review comes with single rating of 1 to 5 stars. This tells us whether the author liked a product, but it doesn't tell us why. On the other hand, Pluribo performs sentiment analysis to extract a distinct numerical score for every feature discussed in a review (potentially dozens of feature scores for just one review). This provides much deeper insight into what the author liked or disliked.
4. Add up the feature scores across all reviews on that product
After scoring the features in each review, Pluribo adds up the feature scores for all reviews on that product. This gives us a net score for each feature, and allows us to filter out features that were not detected in enough reviews to be statistically significant.
5. Give more weight to some opinions than to others
Opinions that are more trustworthy or informative deserve greater consideration. As Pluribo tallies the net feature scores for a product, the algorithm gives more weight to reviews have been voted as more "helpful" by internet users and to reviews that are more recent. The algorithm penalizes reviews that appear to be redundant or poorly formed.
6. Compare the product with its peers to see how it is distinctive
Pluribo compares the extracted feature scores for each product with the average scores for the product category. This tells us which product features stand out -- which features that are distinctively good, or bad, as compared with the category. It is also tells us which are the best products in the category, and which are the ones to be avoided.
7. Summarize the results in textual form
At this point in the analysis, Pluribo could stop and simply give you a statistical summary for each product. Instead, we go a step further and use artificial intelligence to condense the product statistics into a brief, lucid paragraph of text. The details of this last step involve a little bit of magic and are challenging to explain, but we believe that a few lines of text is often more helpful than a page of statistics.
Still have questions? Have a look at frequently asked questions, read about Pluribo technology, or send us an email.