Gareth Jones and Karen Sparck Jones
We were unable at Cambridge, for independent reasons, to do the actual SDR speech test. We therefore carried out only some rather simple, lightweight tests using the baseline SRT data. But these were of the kind drawing on IR experience that the QSDR option was intended to encourage. Thus we treated QSDR as a way of checking out the Okapi-type retrieval methods, already applied to speech data in the Cambridge VMR project (Jones et al, 1997), with (a) new, different, document data (b) known-item searching rather than conventional queries. We set aside questions of whether the Cambridge speech recognition system could do any better than the one used for the baseline (without or with tuning to the application), and also any retrieval strategies tied to speech (eg various fusion ones combining different recognition strategies)....