Some Bing N-gram notes on tokenization
October 26, 2010
Posted by on
Brendan O’Connor asked me on Twitter what I knew about Bing N-gram tokenization conventions, and I said I would ask someone who knew. The N-gram team plans a future blog post on this, but here’s some things I was told that I could share (and I quote):
Here are a few things we can share right now:
- Segment boundaries use the special symbols <s> and </s>, thus P(“<s> hello there </s>”) != P(“hello there”).
- The GetConditionalProbability uses the last space character as the ‘word’ boundary, even if internally there are multiple tokens represented in that last ‘word’. That is, GCP(“shown as-is”) is P(“as-is”|”shown”), not P(“is”|”shown as”).