@TomMorris has been very clear with his answer and I agree that FreeBase (or similar approaches) can only be used indirectly because a global taxonomy might not have a direct mapping to your problem.
My advise, and what I would do if no topic information can be provided is the following:
- Apply clustering techniques to your data.
- Try to decide (automatically or not) the meaning of each cluster.
- Assume that all the document in the cluster belong to that "class".
- Use that information to feed a classifier.
Main problems:
1. I have no idea about the size of your data but it could be a problem for the clustering and/or the manual labelling of clusters.
2. Quality might be way lower than using manual judgments.
I hope this gives you some hints at least.