I think you've converted the lattice incorrectly. Using the topological ordering of the nodes you specified, this lattice would become:
((('take',0.1,2),('cut',0.2,1),('throw',0.3,3),),(('off',1,4),),(('out',1,3),),(('in',1,1),),(('garbage',1,1),),)
Also, reporting the arithmetic average number of paths isn't a not terribly
useful statistic since they often grow exponentially with the size of the
lattice, and the relevant computational complexity information for decoding
is a function of the log of this, size the geometric mean is arguably
justified here. This would be an easy fix to make.
On Thu, Jun 6, 2013 at 8:02 AM, Hieu Hoang <[email protected]> wrote:
> I don't know much about lattices, but why do you say it's not correct?
>
> it seems there are 2 lattices in the input. What is the exact input you
> gave chekplf?
>
>
> On 4 June 2013 15:41, Wei Qiu <[email protected]> wrote:
>
>> Hi,
>>
>> another question about using word lattices:
>> Is the statistics provided by chekplf reliable?
>>
>> Taking a minimum running example,
>> I want to turn
>> [image: Inline image 1]
>> into PLF:
>>
>>
>> ((('take',0.1,1),('cut',0.2,2),('throw',0.3,3),),(('out',1,4),),(('off',1,3),),(('in',1,1),),(('garbage',1,1),),)
>>
>> checkplf tells:
>> Reading PLF from STDIN...
>> PLF format appears to be correct.
>> STATISTICS:
>> Number of lattices: 2
>> Total number of nodes: 10
>> Total number of edges: 14
>> Average density: 1.4 edges/node
>> Total number of paths: 6
>> Average number of paths: 3
>>
>> which seems to be not correct.
>>
>> Did I misunderstand anything?
>>
>> Many thanks.
>>
>> Best,
>> Wei Qiu
>>
>>
>>
>>
>>
>> On Tue, May 14, 2013 at 10:10 AM, Barry Haddow <
>> [email protected]> wrote:
>>
>>> Hi Wei
>>>
>>> 1. It's in contrib/checkplf. There's no Makefile for it, but it only
>>> depends on PCNTools.cpp and PCNTools.h so is straightforward to build from
>>> the command line.
>>>
>>> 2. Yes, it's possible but a bit fiddly. The first thing to note is that
>>> filtering does not work for lattices, so you need to force EMS to binarise
>>> the whole models. To do this set the variable binarize-all in the GENERAL
>>> section to point to the binarize-model.perl script in scripts/training. You
>>> will also need to pass "-num-lattice-features 1" in training-options to get
>>> the correct moses ini file created. And don't forget to add "-inputttype 2"
>>> to the decoder settings. You need to make sure you use a version of Moses
>>> from immediately before Hieu's refactoring,
>>>
>>> cheers - Barry
>>>
>>>
>>> On 14/05/13 08:50, Wei Qiu wrote:
>>>
>>>> Hi,
>>>>
>>>> I'm following the tutorial about using word lattices
>>>> http://www.statmt.org/moses/?**n=Moses.WordLattices<http://www.statmt.org/moses/?n=Moses.WordLattices>
>>>>
>>>> I've got some questions:
>>>>
>>>> 1. I couldn't find the bin moses-cmd/src/checkplf|. |I installed moses
>>>> in /opt/moses, but there are only 3 folders bin lib scripts. I found the
>>>> source file in the contrib directory in the source code directory, but am I
>>>> supposed to compile it manually?
>>>>
>>>> 2. How can I setup it in EMS? Is it possible to adjustthe weight-i for
>>>> arc probability during the tuning step?
>>>>
>>>>
>>>> Thanks in advance!
>>>>
>>>> Best,
>>>> Wei Qiu
>>>>
>>>>
>>>>
>>>> ______________________________**_________________
>>>> Moses-support mailing list
>>>> [email protected]
>>>> http://mailman.mit.edu/**mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>>>
>>>
>>>
>>> --
>>> The University of Edinburgh is a charitable body, registered in
>>> Scotland, with registration number SC005336.
>>>
>>>
>>
>> _______________________________________________
>> Moses-support mailing list
>> [email protected]
>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>
>>
>
>
> --
> Hieu Hoang
> Research Associate
> University of Edinburgh
> http://www.hoang.co.uk/hieu
>
>
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
>
>
<<Wordlattice.png>>
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
