You're a bit lacking on specifics there. Appropriate optimizations are going to depend upon things like the document format, the average document size, how you are processing them, and what sort of information you are storing in your data structure. Not knowing any of them, some general optimizations are:
Assuming that the pre-processing of a given document is independent of the pre-processing of any other document, and assuming you are running a multi-core CPU, then your workload is a good candidate for multi-threading. Allocate one thread per CPU core, and farm out jobs to your threads. Then you can process multiple documents in parallel.
More generally, do as much in memory as you can. Try to avoid reading from/writing to disk as much as possible. If you must write to disk, try to wait until you have all the data you want to write, and then write it all in a single batch.