Apoorv Saxena
commited on
Commit
•
d4223b2
1
Parent(s):
16de72d
Update README.md
Browse files
README.md
CHANGED
|
@@ -84,7 +84,7 @@ out
|
|
| 84 |
|
| 85 |
You can further load the list of entity aliases, then filter only those predictions which are valid entities then create a reverse mapping from alias -> integer id to get final predictions in required format.
|
| 86 |
|
| 87 |
-
However, loading these aliases in memory as a dictionary requires a lot of RAM + you need to download the aliases file (made available here)
|
| 88 |
|
| 89 |
The submitted validation/test results for were obtained by sampling 300 times for each input, then applying above procedure, followed by filtering known entities. The final MRR can vary slightly due to this sampling nature (we found that although beam search gives deterministic output, the results are inferior to sampling large number of times).
|
| 90 |
|
|
|
|
| 84 |
|
| 85 |
You can further load the list of entity aliases, then filter only those predictions which are valid entities then create a reverse mapping from alias -> integer id to get final predictions in required format.
|
| 86 |
|
| 87 |
+
However, loading these aliases in memory as a dictionary requires a lot of RAM + you need to download the aliases file (made available here https://storage.googleapis.com/kgt5-wikikg90mv2/ent_alias_list.pickle) (relation file: https://storage.googleapis.com/kgt5-wikikg90mv2/rel_alias_list.pickle)
|
| 88 |
|
| 89 |
The submitted validation/test results for were obtained by sampling 300 times for each input, then applying above procedure, followed by filtering known entities. The final MRR can vary slightly due to this sampling nature (we found that although beam search gives deterministic output, the results are inferior to sampling large number of times).
|
| 90 |
|