MLR

1
MLR
DCG PE Bad P _at_ 1
Yahoo CN 6.89 16.9 14.6 4.3
MLR 2.4 1 -13 0
Used Yahoo Simplified-preferred as baseline,
since it has higher relevance than all-language
and cn-preferred. Also both used CNB.
2
MLR vs Yahoo
The only place where MLR does really bad is on
Latin queries. If we fix this the gain could be
much bigger.
3
MLR by Language
Chinese Latin
Baseline 6.6 8.0
JP MLR 4 1
EN MLR 0 3
Google 3 6
Examples 509 91
4
MLR vs Yahoo by Length
Query Length MLR vs Yahoo of Queries
2 -1 141
3 -2 118
4 -3 123
5 7 216
This is because MLR currently always uses bracket.
5
Best MLR

• If query is latin charset use English (3 Gain)
• If query is gt4 characters (or 12 bytes) use
  MLR/Bracket (7 Gain)
• Else use IRF (0 Gain)

6
Yahoo vs Competitor
DCG PE Bad P _at_ 1
Yahoo CN 6.89 16.9 14.6 4.3
Google -2 -8 6 7
Baidu -3 -14 -31 0
MLR 2.4 1 -13 0
7
Yahoo vs Google
Here darker color means longer query.
8
Yahoo vs. Baidu
9
Yahoo vs. Baidu
All of our gain over Baidu is due to Latin
queries.