lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Aaron McCurry (JIRA)" <j...@apache.org>
Subject [jira] Commented: (LUCENE-2089) explore using automaton for fuzzyquery
Date Fri, 12 Feb 2010 01:04:28 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-2089?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12832773#action_12832773
] 

Aaron McCurry commented on LUCENE-2089:
---------------------------------------

I have written a levenstein generator today that seems to operate similarly to what is being
discussed here.  It generates all the possible matches to levenstein algorithm given a term
and a character set, it then creates a booleanquery from it.  For a given term with edit distance
of 1 or 2 it is extremely fast.  I tested it on my dev data that has about 8 billion documents
with 20 shards, each shard has about 170,000,000 terms in the field that I'm testing.  The
normal fuzzy query with a term length of 8 and and edit distance of 2 took about 110 seconds
to complete, and the auto generated query took around a 1.5 seconds complete.  However this
method will probably only work with edits distances in the 1 and 2 range, because once I hit
3 it spiked the memory usage and slowed way down (to be expected).  Not sure if you all want
to take a look at my code or not, but if you want me to post it I will.

> explore using automaton for fuzzyquery
> --------------------------------------
>
>                 Key: LUCENE-2089
>                 URL: https://issues.apache.org/jira/browse/LUCENE-2089
>             Project: Lucene - Java
>          Issue Type: Wish
>          Components: Search
>            Reporter: Robert Muir
>            Assignee: Mark Miller
>            Priority: Minor
>         Attachments: LUCENE-2089.patch, Moman-0.2.1.tar.gz, TestFuzzy.java
>
>
> Mark brought this up on LUCENE-1606 (i will assign this to him, I know he is itching
to write that nasty algorithm)
> we can optimize fuzzyquery by using AutomatonTermsEnum, here is my idea
> * up front, calculate the maximum required K edits needed to match the users supplied
float threshold.
> * for at least small common E up to some max K (1,2,3, etc) we should create a DFA for
each E. 
> if the required E is above our supported max, we use "dumb mode" at first (no seeking,
no DFA, just brute force like now).
> As the pq fills, we swap progressively lower DFAs into the enum, based upon the lowest
score in the pq.
> This should work well on avg, at high E, you will typically fill the pq very quickly
since you will match many terms. 
> This not only provides a mechanism to switch to more efficient DFAs during enumeration,
but also to switch from "dumb mode" to "smart mode".
> i modified my wildcard benchmark to generate random fuzzy queries.
> * Pattern: 7N stands for NNNNNNN, etc.
> * AvgMS_DFA: this is the time spent creating the automaton (constructor)
> ||Pattern||Iter||AvgHits||AvgMS(old)||AvgMS (new,total)||AvgMS_DFA||
> |7N|10|64.0|4155.9|38.6|20.3|
> |14N|10|0.0|2511.6|46.0|37.9|	
> |28N|10|0.0|2506.3|93.0|86.6|
> |56N|10|0.0|2524.5|304.4|298.5|
> as you can see, this prototype is no good yet, because it creates the DFA in a slow way.
right now it creates an NFA, and all this wasted time is in NFA->DFA conversion.
> So, for a very long string, it just gets worse and worse. This has nothing to do with
lucene, and here you can see, the TermEnum is fast (AvgMS - AvgMS_DFA), there is no problem
there.
> instead we should just build a DFA to begin with, maybe with this paper: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.16.652
> we can precompute the tables with that algorithm up to some reasonable K, and then I
think we are ok.
> the paper references using http://portal.acm.org/citation.cfm?id=135907 for linear minimization,
if someone wants to implement this they should not worry about minimization.
> in fact, we need to at some point determine if AutomatonQuery should even minimize FSM's
at all, or if it is simply enough for them to be deterministic with no transitions to dead
states. (The only code that actually assumes minimal DFA is the "Dumb" vs "Smart" heuristic
and this can be rewritten as a summation easily). we need to benchmark really complex DFAs
(i.e. write a regex benchmark) to figure out if minimization is even helping right now.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


---------------------------------------------------------------------
To unsubscribe, e-mail: java-dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-dev-help@lucene.apache.org


Mime
View raw message