recall [% Precision-recall curves for term matching, LSI, and PLSI on the 4 test collection have been utilized to evaluate similarities. to achieve this, queries have to be fol ded in, which is done in the PLSA by fixing the P(wjz)parameters and cal culating weights P(ajq)by TEM One advantage of using statistical models vs. SVD echniques is that it allows us to sy stematically com bine different models. While this should optimally be done according to a Bayesi an model combination scheme, we have utilized a much simpler approach in our experiments which has never theless shown excel formance and rob simply combined the cosine s cores of all models with a uniform weight. The resulting method ed to PLSI. Empirically we have found the performance to be very robust w r.t. different (nonuinbination with Figure 7: Perplexity and average precision as a func- i form) weights and also w.r.t. the A-weight used in coml the original ucing benefits of (model )aver aging. Notice that lsa tion of the inverse temperature fi for an aspect model represent ations for different K form a nested sequence with K=48(left which is not true for the st atistical mo dels which are expected to capture a larger variety of reasonable de- tion). The condensed results in terms of average pre ision recall(at the 9 recall levels 10%-90%)are sum- We have utilized the following four medium-sized stan- marized in Table l, while the corresponding precision dard document collection with reley nt: recall cur ves can be found in Figure 6. Here are some D(1033 document abstracts from the National additional details of the experimental setup: PLSA Library of Medicine), (ii) CRAN(1400 do cument ab- models at K=32, 48, 64, 80, 128 have been trained by stracts on aeronautics from the Cranfiel d Institute of T EM for each data set with 10% held-out data. For Technology ),(iii) CACM (3204 abstracts from the PLSI we report the best result obtained by any of these CACM Journal), and (iv) CISI (1460 abstracts in li- models, for LSI we report the best result obtained for brary science from the Institute for Scientific Informa- the optimal dimension(exploring 32-512 dimensions t a step size of 8).The combination weight A with the0 50 100 0 10 20 30 40 50 60 70 80 90 MED recall [%] precision [%] 0 50 100 0 10 20 30 40 50 60 70 CRAN recall [%] 0 50 100 0 10 20 30 40 50 60 CACM recall [%] 0 50 100 0 5 10 15 20 25 30 35 40 45 50 CISI recall [%] cos LSI PLSI* cos LSI PLSI* cos LSI PLSI* cos LSI PLSI* Figure 6: Precision-recall curves for term matching, LSI, and PLSI on the 4 test collections. have been utilized to evaluate similarities. To achieve this, queries have to be folded in, which is done in the PLSA by xing the P (wjz) parameters and calculating weights P (zjq) by TEM. One advantage of using statistical models vs. SVD techniques is that it allows us to systematically com￾bine dierent models. While this should optimally be done according to a Bayesian model combination scheme, we have utilized a much simpler approach in our experiments which has nevertheless shown excel￾lent performance and robustness. Namely, we have simply combined the cosine scores of all models with a uniform weight. The resulting method is referred to as PLSI . Empirically we have found the performance to be very robust w.r.t. dierent (non-uniform) weights and also w.r.t. the -weight used in combination with the original cosine score. This is due to the noise re￾ducing benets of (model) averaging. Notice that LSA representations for dierent K form a nested sequence, which is not true for the statistical models which are expected to capture a larger variety of reasonable de￾compositions. We have utilized the following four medium-sized stan￾dard document collection with relevance assessment: (i) MED (1033 document abstracts from the National Library of Medicine), (ii) CRAN (1400 document ab￾stracts on aeronautics from the Craneld Institute of Technology), (iii) CACM (3204 abstracts from the CACM Journal), and (iv) CISI (1460 abstracts in li￾brary science from the Institute for Scientic Informa- 0.6 0.7 0.8 0.9 1 1200 1400 1600 1800 2000 beta perplexity K=48 0.6 0.7 0.8 0.9 1 30 40 50 60 70 beta average precision K=48 0.6 0.7 0.8 0.9 1 1200 1400 1600 1800 2000 beta perplexity K=128 0.6 0.7 0.8 0.9 1 30 40 50 60 70 beta average precision K=128 Figure 7: Perplexity and average precision as a func￾tion of the inverse temperature  for an aspect model with K = 48 (left) and K = 128 (right). tion). The condensed results in terms of average pre￾cision recall (at the 9 recall levels 10%￾90%) are sum￾marized in Table 1, while the corresponding precision recall curves can be found in Figure 6. Here are some additional details of the experimental setup: PLSA models at K = 32; 48; 64; 80; 128 have been trained by TEM for each data set with 10% held-out data. For PLSI we report the best result obtained by any of these models, for LSI we report the best result obtained for the optimal dimension (exploring 32{512 dimensions at a step size of 8). The combination weight  with the