Table 5.
Comparison with baseline PRF retrieval models in certain metrics
| BM25+Rocchio | RM3 | HRoc1 | HRoc2 | HRoc3 | |
|---|---|---|---|---|---|
| MAP | 0.0490 | 0.0540 | 0.0651 ∗+ (32.8%,20.5%) | 0.0647 ∗+ (32.0%,19.8%) | 0.0642 ∗+ (31.0%,18.9%) |
| P@5 | 0.2733 | 0.2600 | 0.2933 ∗+ (7.32%,12.8%) | 0.2733 + (0.00%,5.11%) | 0.2733 + (0.00%,5.11%) |
| P@10 | 0.2533 | 0.2467 | 0.2733 ∗+ (7.89%,10.8%) | 0.2667 ∗+ (5.29%,8.11%) | 0.2567 ∗+ (1.34%,4.25%) |
| P@20 | 0.2167 | 0.2233 | 0.2350 ∗+ (8.44%,5.24%) | 0.2317 ∗+ (6.92%, 3.76%) | 0.2317 ∗+ (6.92%, 3.76%) |
| F1 | 0.0853 | 0.0932 | 0.1112 ∗+ (30.3%,19.3%) | 0.1108 ∗+ (29.9%,18.9%) | 0.1096 ∗+ (28.5%, 17.6%) |
The values in parentheses represent the improvements over BM25+Rocchio and RM3 respectively. The best result obtained is shown in bold, and the superscripts “*” and “+” denote statistically significant improvements over BM25+Rocchio and RM3, respectively (Wilcoxon signed-rank test with p<0.05)