[
https://issues.apache.org/jira/browse/JOSHUA-145?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15267521#comment-15267521
]
Matt Post commented on JOSHUA-145:
----------------------------------
Reclassified.
I recently added a related feature to Joshua. If you invoke the decoder with
-lowercase, all the input sentence tokens will be lowercased, and the grammar
lookups will used the lowercase version. It then adds an annotation on each
token of the form
lettercase = {lower, upper, all-upper}
This is available to any feature function, for example. If you also invoke the
decoder with "-project-case", it will use word-level alignments to project
source-language case to the target language, according to the following logic:
- If aligned to the first word, case is only projected if it is "all-upper"
- Otherwise, project the source-language case
This does things like project all caps, and capitalization of names (including
if they were OOVs). It's different from true-casing or re-casing. I haven't
done a thorough comparison, but this was the method that helped put a
relatively simple Joshua system in first place for WMT 2016 en-tr.
> Add truecasing
> --------------
>
> Key: JOSHUA-145
> URL: https://issues.apache.org/jira/browse/JOSHUA-145
> Project: Joshua
> Issue Type: New Feature
> Reporter: Matt Post
> Assignee: Matt Post
> Fix For: 6.1
>
>
> Joshua currently lowercases all data; a better approach is truecasing, where
> the most frequent capitalization pattern is used for each token.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
