50MB is a large Word file. How much of that is text, versus images?
It would be an interesting exercise to try a different implementation. Shouldn't be difficult. Off the top of my head: If you are using Java 5, put the replacement implementation on your classpath (removing JAXB). With Java 6 or later, put it in your endorsed dir.
You will also need to change org.docx4j.jaxb.NamespacePrefixMapperUtils which assumes Sun/Oracle implementation, for example, with MOXy, see
http://stackoverflow.com/questions/4873 ... es-in-jaxbGiven that trunk currently uses a dependency which is yet to be put in Maven Central, and that VERSION_2_7_1 has
http://www.docx4java.org/trac/docx4j/changeset/1707 I'd suggest you consider checking that version out instead for now.
An idea I am toying with for handling very large files is to shred them into eXist XML database, and then use docx4j to do its thing against eXist. Exploring this approach would be a more significant undertaking.
Please let us know what you discover!