The questions Q&A can answer are determined by four major factors:
1. Vocabulary in your model. This includes table and column names, additional synonyms you teach the system and any optional relationship wording you choose to add.
2. Core BI concepts recognized by the system. This includes things like aggregates, sorting, relative dates, grouping, visualization type selection, etc. We're continually enhancing this set, based on usage and feedback from customers. We haven't yet published
a comprehensive list of what's currently supported, but that sounds like a good idea, thanks.
3. Basic linguistic understanding. This is the core NLP component and is where NLP benchmarking typically takes place. However, since the NLP system is designed to work as a part of the whole rather than as an isolated component, we haven't spent much time
on traditional NLP benchmarking metrics.
4. Heuristics based on the structure of your model. When the system isn't entirely sure what you meant, it has a set of heuristics to try to make a best guess and get as close as possible to the data in which you were interested.
While I'm afraid I don't have any statistics-laden dry technical papers at which to point you, I can recommend installing the samples and playing with them. That way, you'll quickly get a sense of the current capabilities and
limitations of the system first-hand.
Microsoft is conducting an online survey to understand your opinion of the Technet Web site. If you choose to participate, the online survey will be presented to you when you leave the Technet Web site.