Using reviews for exploratory search
search through reviews
- Select a few random objects with reviews which (partially) match the description of the query.
- Rank (some of) the selected objects.
- Train a ranking engine based on the user rankings of the selected objects and (some of) the available reviews. [The ranking engine answers two questions: a) Which of two objects ranks higher, and b) is an object relevant.]
- Add more - primarily highly ranked by the engine - objects to the selected group of objects.
- Return to step 2 (or publish the results)
Reviews can be grouped in four kinds:
- An object with a published review in a specified group.
- An object with a graded review (scalar value) in a specified group.
- Two objects with published reviews in the same specified group.
- An object with a review in a specified group which compares it with another object.
The first two kinds are absolute reviews of an object, whereas the last two kinds are comparing reviews. For a simple relevance test only the first two kinds of reviews are relevant.
The specification of a group (for the use in reviews as defined above) can be:
- Review source (RSS feed, newsgroup, etc.).
- Data type (MIME type or partial MIME type).
- Term in object (i.e. a string found in the content).
... or a combination of these.