Translingual Topic Tracking with PRISE

The University of Maryland participated in the topic track ing task submitting four runs for the required condition four English training stories This paper presents those results along with additional contrastive runs Compar isons are presented between di erent translation selection strategies retention and removal of Mandarin stopwords one pass translation and post translation document expan sion and source dependent and source independent normal ization Source dependent normalization was found to be helpful even for the monolingual English case Two translin gual techniques also yielded substantial improvements post translation document expansion and top two translation se lection Both outperformed the baseline Systran full machine translations demonstrating the potential for developing ef fective and easily implemented word for word techniques for other languages