Experiments in Spoken Document Retrieval at DERA-SRU

A small amount of internal funding allowed DERA-SRU to participate in the TREC-7 SDR evaluations for the first time this year. Since we had almost no experience of entering this or related NIST evaluations (e.g. ARPA HUB-4 LVCSR) there was a rather steep learning curve along with intense development of the experimental infrastructure. The intention was to generate a base for future participation and to build upon this using experience gained from related work on topic spotting. To this end, a straightforward (i.e. non-optimised) speech recogniser was used to generate transcripts and retrieval was performed using the okapi [6,9] search engine. Previous work on topic spotting [7] suggested that term expansion using a semantic network (in this case wordnet [2,3]) might be useful. This hypothesis appeared to be supported by preliminary work on TREC-6 SDR data which yielded text (i.e. R1) results that were comparable with the best achieved elsewhere.