Eclipse Community Forums: Code Recommenders » Help Wanted: Ranking Strategy for Subwords Completion Engine

Help

Home

Home » Eclipse Projects » Code Recommenders » Help Wanted: Ranking Strategy for Subwords Completion Engine

Show: Today's Messages :: Show Polls :: Message Navigator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #690420 is a reply to message #689209]

Wed, 29 June 2011 09:49

Johannes Lerch
Messages: 18
Registered: February 2011

Junior Member

I just integrated your latest contribution. You might have to search a little to find it, as i refactored a lot of stuff. I also added some tests and found some bugs. For example i think qgrams is symmetric and your implementation was not. I fixed those issues, but i would appreciate that you will take a look on that to verify that i didn't broke it.

The main classes of interest are now SubwordsRelevanceCalculator, QGramSimilarity and the corresponding test classes SubwordsRelevanceCalculatorTest, QGramSimilarityTest and ExpectedScoringsTest. As you can see your current approach matches my expectations. Nice job!

But can you give me more details on the weighting function? What was the intent of the factor 20 and why is the difference of the length of both strings relevant?

For all who want to test the current state: Get it from the head channel. I turned off the default java proposals and will check what it feels like to only have the subwords completion.

[Updated on: Thu, 30 June 2011 03:10]

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #690604 is a reply to message #690420]

Wed, 29 June 2011 16:03

Marcel Bruch
Messages: 188
Registered: July 2009

Senior Member

Hi,

I just installed the subwords completion. This engine is great!

During my surface tests I found a situation I'm a bit surprised of the ranking results. Please have a look at the screenshot:
index.php/fa/3182/0/

Is there any special reason why "getDisplay" is ranked before "getLayoutData" ?

Attachment: Screen shot 2011-06-29 at 21.57.45.png
(Size: 86.90KB, Downloaded 246 times)

tw: @MarcelBruch
tw: @Recommenders

[Updated on: Thu, 30 June 2011 00:59]

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #690729 is a reply to message #690604]

Thu, 30 June 2011 01:46

Johannes Lerch
Messages: 18
Registered: February 2011

Junior Member

There is a reason. The weighting function gives a higher rank if the difference between token and completion string length is low.
In your case lay is shorter than both completions, so the shorter completion will be preferred. I asked in my previous post what the idea behind this is, as i am also surprised by the result.

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #690730 is a reply to message #690729]

Thu, 30 June 2011 01:49

Marcel Bruch
Messages: 188
Registered: July 2009

Senior Member

Ah, ok. Given the example I think it's a bit unexpected behavior. I think the earlier a token could be found the higher the proposal should be, right?

tw: @MarcelBruch
tw: @Recommenders

[Updated on: Thu, 30 June 2011 01:49]

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #690748 is a reply to message #690730]

Thu, 30 June 2011 02:56

Johannes Lerch
Messages: 18
Registered: February 2011

Junior Member

I'm able to follow that argumentation, but i think it's more of theoretic nature. I would prefer to see any completions that have the same matchings (ignoring the position of that match if it's not a prefix) in alphabetic order, because that is what i usually will see.

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #690756 is a reply to message #690748]

Thu, 30 June 2011 03:20

Marcel Bruch
Messages: 188
Registered: July 2009

Senior Member

Agreed. Let's drop the string length stuff for the moment, ignore my idea about matching bigram locations, and stick to a plain alphabetical order + prefix + bigrams scoring (+ JDT base score).

Thanks,
Marcel

tw: @MarcelBruch
tw: @Recommenders

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #691307 is a reply to message #690756]

Fri, 01 July 2011 03:51

Marcel Bruch
Messages: 188
Registered: July 2009

Senior Member

Just to keep you updated:

We updated the ranking strategies and also added a highlighting to subwords:
index.php/fa/3205/0/

You can install the latest version directly from the head update site (subwords feature). Your feedback is appreciated and always welcome!

Attachment: subwords-highlighting.png
(Size: 53.12KB, Downloaded 254 times)

tw: @MarcelBruch
tw: @Recommenders

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #696040 is a reply to message #673616]

Wed, 13 July 2011 02:57

Paul-Emmanuel Faidherbe is currently offline

Paul-Emmanuel Faidherbe
Messages: 30
Registered: April 2011
Location: france

Member

Hi,
Sorry for my absence but I just had some hard days at work... I will update your modifications and try to improve my code regarding your latest remarks if still needed. As you infered it, my weighting strategy was to prior shortest proposal matches (as Levenshtein proposed). But your choices seem interesting. I will give you my feedback as soon as I tested it.
Thanks for your investment!

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #696053 is a reply to message #696040]

Wed, 13 July 2011 03:43

Marcel Bruch
Messages: 188
Registered: July 2009

Senior Member

Hi Pef,

we adapted some parts of your code. You find the latest version in the org.eclipse.recommmenders git repository. We came across an issue with relevance scoring since JDT does not support updating the relevance of a proposal. This is annoying and confusing when typing.

I'm not sure whether we can (yet) deliver the bigrams solution with that issue. I reported the issue here: https://bugs.eclipse.org/bugs/show_bug.cgi?id=350991 JDT is yet not very responsive because of the upcoming Java7 release. But I hope we'll get a solution rather soon.

In the meanwhile it would be cool if you could install the latest version from HEAD update site and give it a try?

Best,
Marcel

tw: @MarcelBruch
tw: @Recommenders

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #698749 is a reply to message #696053]

Wed, 20 July 2011 03:24

Paul-Emmanuel Faidherbe
Messages: 30
Registered: April 2011
Location: france

Member

Hi,

Just to inform you, I still had no time to take a look at the new code... So sorry. And I am now leaving for my two weeks summer holidays, so I will not be able to help you more for this time.

I hope you will do great things, but I know you all rock! Smile

Regards.

Report message to a moderator

Re: Help Wanted: Ranking Strategy for Subwords Completion Engine [message #698761 is a reply to message #698749]

Wed, 20 July 2011 03:43

Marcel Bruch
Messages: 188
Registered: July 2009

Senior Member

Have a good time. When you are back, subwords will be part of dev - unfortunately without relevance ranking until https://bugs.eclipse.org/bugs/show_bug.cgi?id=350991 is fixed.

Best,
Marcel

tw: @MarcelBruch
tw: @Recommenders

Report message to a moderator

Pages (2): [ « ‹ 1 2]

Previous Topic:	Early Access: Extended Javadoc View Prototype
Next Topic:	.recommenders folder is not removed along with its nature

Goto Forum:

-=] Back to Top [=-

[ Syndicate this forum (XML) ] [

]

Current Time: Wed May 29 00:33:14 EDT 2013