lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jim Ferenczi (JIRA)" <>
Subject [jira] [Commented] (LUCENE-8631) How Nori Tokenizer can deal with Longest-Matching
Date Tue, 08 Jan 2019 19:14:00 GMT


Jim Ferenczi commented on LUCENE-8631:

Thanks for reporting [~gritmind]. Since we give the same cost to all user words I think the
easiest way to solve this issue would be to implement a longest-only match in the user dictionary.
We don't check the main dictionary when we have matches in the user dictionary so this should
only ensure that the longest rule that matches wins. This should also speed up the tokenization
since we'd add a single path in the lattice (instead of all user words that match). I'll work
on a patch.

> How Nori Tokenizer can deal with Longest-Matching
> -------------------------------------------------
>                 Key: LUCENE-8631
>                 URL:
>             Project: Lucene - Core
>          Issue Type: Improvement
>          Components: modules/analysis
>            Reporter: Yeongsu Kim
>            Priority: Major
> I think... Nori tokenizer has one issue.
> I don’t understand why “Longest-Matching” is NOT working to Nori tokenizer via
config mode (config mode: []
> Here is an example for explaining what is longest-matching.
> Let assume we have `userdict_ko.txt` including only three Korean single-words such
as ‘골드’, ‘브라운’, ‘골드브라운’, and save it to Nori tokenizer. After
update, we can see that it outputs two tokens such as ‘골드’ and ‘브라운’ when
the input is ‘골드브라운’. (In English: ‘골드’ means ‘gold’, ‘브라운’
means ‘brown’, and ‘골드브라운’ means ‘goldbrown’)
> With this result, we recognize that “Longest-Matching” is NOT working.
> If “Longest-Matching” is working, the output must be ‘골드브라운’, which
is the longest matching word in the user dictionary.
> Curiously enough, when we add user dictionary via custom mode (custom mode: [|]),
we found the result is ‘골드브라운’, where ‘Longest-Matching’ is applied. We
think that the reason is because learned Mecab engine automatically generates word costs by
its own criteria. We hope this mechanism is also applied to config mode.
> Would you tell me the way to “Longest-Matching” via config mode (not custom) or give
me some hints (e.g. where to modify source codes) to solve this problem?
> P.S
> Recently, I've mailed to Jim Ferenczi, who is a developer of Nori, and received his suggestions:
> * Add a way to set a score to each new rule (this way you could set up a negative cost
for the compound word that is less than the sum of the two single words.
> * Same as above but the cost is computed from the statistics of the training (like the
custom dictionary does when you recompile entirely).
> * Implement longest-match first in the dictionary.
> Thanks for your support.

This message was sent by Atlassian JIRA

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message