Microsoft Cambridge at TREC 14: Enterprise Track

A major focus of much work of the group (as it has been since the City University Okapi work) is the development and refinement of basic ranking algorithms. The workhorse remains the BM25 algorithm; recently [3, 4] we introduced a field-weighted version of this, allowing differential treatment of different fields in the original documents, such as title, anchor text, body text. We have also recently [2] been working on ways of analysing the possible contributions of static (query-independent) evidence, and of incorporating them into the scoring/ranking algorithm. Finally, we have been working on ways of tuning the resulting ranking functions, since each elaboration tends to introduce one or more new free parameters which have to be set through tuning. We used all these techniques successfully in our contribution to the Web track in TREC 2004 [4]. This year’s relatively modest TREC effort is confined to applying essentially the same techniques to rather different data, in the Enterprise Track’s known item (KI) and discussion search (DS) experiments. The main interest is whether we can identify some fields and features that lead to an improvement over a flat-text baseline, and as a side effect to verify that our ranking model can deliver the benefit.