Hi John,
the answer is extract.cpp (I had to look it up myself).
The orientation decision is handled based on alignment points as described
in the tutorial slide:
// orientation to previous E
bool connectedLeftTop = isAligned( sentence, startF-1, startE-1 );
bool connectedRightTop = isAligned( sentence, endF+1, startE-1 );
So what happens when startF == 0, so we are looking up
alignments for alignment points for word "-1"?
Let's look at the isAligned function:
bool isAligned ( SentenceAlignment &sentence, int fi, int ei ) {
if (ei == -1 && fi == -1) return true;
if (ei <= -1 || fi <= -1) return false;
So, if the first target (E) phrase is in the top left corner (translated into
the first source (F) phrase), it is monotone, but not otherwise.
-phi
On Wed, Nov 11, 2009 at 8:52 PM, John DeNero <[email protected]> wrote:
> Hi Philipp & Chris,
>
> Thanks for the help so far. One more question about a special case:
> What is the reordering of the first source phrase with regard to the
> previous phrase? Always mono? (Another reasonable policy might be
> mono if aligned to the first target phrase and discontinuous
> otherwise.) Does the last source phrase with regard to the following
> context have the same policy? If you don't know off the top of your
> head, I'll dig into the data and figure it out.
>
> Thanks,
> John
>
> On Wed, Nov 11, 2009 at 11:59 AM, Philipp Koehn <[email protected]> wrote:
>> Hi,
>>
>> the determination in training, whether a phrase is swap (with regard to
>> previous
>> phrase or next) is based on alignment points around the phrase.
>>
>> Slide 112 in this tutorial defines which alignment points are looked at:
>> http://www.iccs.inf.ed.ac.uk/~pkoehn/publications/tutorial2006.pdf
>>
>> So, yes, swap swap is possible - it happens if a sequence of
>> phrases is in inverse order.
>>
>> -phi
>>
>> On Wed, Nov 11, 2009 at 7:49 PM, John DeNero <[email protected]> wrote:
>>> Thanks, Chris. Just to clarify, am I interpreting the following cases
>>> correctly, where P is the phrase pair in question and X are word
>>> alignments in neighboring corners, and the source goes left to right?
>>>
>>> The "mono swap" case:
>>> $ zcat extract.o.gz | grep "mono swap" | wc -l
>>> 41043
>>>
>>> X X
>>> P
>>>
>>> The "swap swap" case:
>>> $ zcat extract.o.gz | grep "swap swap" | wc -l
>>> 61745
>>>
>>> X
>>> P
>>> X
>>>
>>>
>>> The "swap mono" case:
>>> $ zcat extract.o.gz | grep "swap mono" | wc -l
>>> 50403
>>>
>>> P
>>> X X
>>>
>>> On Wed, Nov 11, 2009 at 11:39 AM, Chris Dyer <[email protected]> wrote:
>>>> Hi John-
>>>> The first label is the orientation of the phrase pair with respect to
>>>> its left context (on the source side), and the second is the
>>>> orientation with respect to its right context. That's why you have to
>>>> have "swap other" or "other swap", since a phrase can only be inverted
>>>> on one side.
>>>> Hope this helps,
>>>> Chris
>>>>
>>>> On Wed, Nov 11, 2009 at 2:34 PM, John DeNero <[email protected]> wrote:
>>>>> Hi all,
>>>>>
>>>>> I'm trying to generate a replacement phrase extraction file to be used
>>>>> in estimating a lexical reordering model. I'm running
>>>>> train-factored-phrase-model.perl with the "-reordering
>>>>> msd-bidirectional-fe" flag, which generates an extract.o.gz file with
>>>>> content like:
>>>>>
>>>>> reanudación ||| resumption ||| mono mono
>>>>> reanudación del ||| resumption of the ||| mono mono
>>>>> ...
>>>>> este ||| this ||| swap other
>>>>> ...
>>>>>
>>>>> I understand that the mono, swap, and other tags correspond to the
>>>>> "(m) monotone order, (s) switch with previous phrase, or (d)
>>>>> discontinuous" types described in the online Moses docs. I don't
>>>>> really understand what the two different tags correspond to, though.
>>>>> What does the first entry vs. the second entry mean in each line?
>>>>> Apologies if this is explained somewhere in the docs or mailing list
>>>>> archives -- I didn't find it.
>>>>>
>>>>> Thanks,
>>>>> John
>>>>>
>>>>> _______________________________________________
>>>>> Moses-support mailing list
>>>>> [email protected]
>>>>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>>>>
>>>>
>>>
>>> _______________________________________________
>>> Moses-support mailing list
>>> [email protected]
>>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>>
>>
>
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
