I should mention that this happened to me for jpg, and PDF content as well as with docx with a slight mapping change.
I turned on TRACE debugging for org.alfresco.repo.action.executer.ContentMetadataExtracter and see that in the addTags method, the rawValue is a collection with a single value that is a comma separated string with my two keywords (foo and bar) in it.
2016-01-05 09:32:53,220 TRACE [action.executer.ContentMetadataExtracter] [http-bio-8080-exec-21] adding string tag 'foo, bar' to workspace://SpacesStore/60458b43-c123-41b5-a7b0-ab2689c1688a