Table 6.

Comparison with the state-of-the-art PRF retrieval models in certain metrics

	PRoc2	PRoc3	TF-PRF	HRoc1
MAP	0.0600	0.0593	0.0580	0.0651 ∗+# (8.50%,9.78%,12.24%)
P@5	0.2867	0.2667	0.2600	0.2933 ∗+# (2.30%,9.97%,12.81%)
P@10	0.2533	0.2300	0.2467	0.2733 ∗+# (7.90%,18.83%,10.78%)
P@20	0.2417	0.2167	0.2317	0.2350 +# (-2.77%,8.44%,1.42%)
F1	0.1031	0.1020	0.1012	0.1112 ∗+# (7.86%,9.02%,9.88%)

The best result obtained is shown in bold, and the superscripts “*”, “+” and “#” denote statistically significant improvements over PRoc2, PRoc3 and TF-PRF, respectively (Wilcoxon signed-rank test with p<0.05). The values in parentheses are the improvements over RM3, BM25+Rocchio, PRoc2 and TF-PRF, respectively