these are the glue rules. For convenience, the input tree also contains 'X'
labels for all spans so that the simple hiero glue rule can apply.
If you want proper syntactic glue rules, then replace the glue grammar
<s> [X] ||| <s> [S] ||| 1 ||| ||| 0
[X][S] </s> [X] ||| [X][S] </s> [S] ||| 1 ||| 0-0 ||| 0
[X][S] [X][X] [X] ||| [X][S] [X][X] [S] ||| 2.718 ||| 0-0 1-1 ||| 0
with
<s> [???] ||| <s> [S] ||| 1 ||| ||| 0
[NN][S] </s> [NP] ||| [NN][S] < /s> [S] ||| 1 ||| 0-0 ||| 0
[NNS][S] </s> [NP] ||| [NNS][S] </s> [S] ||| 1 ||| 0-0 ||| 0
[NNP][S] </s> [NP] ||| [NNP][S] </s> [S] ||| 1 ||| 0-0 ||| 0
...
[V][S] </s> [VP] ||| [V][S] </s> [S] ||| 1 ||| 0-0 ||| 0
[VINF][S] </s> [VP] ||| [VINF][S] </s> [S] ||| 1 ||| 0-0 ||| 0
...
as you can see, it's much harder to think about a syntax glue grammar, and
it doesn't really add usefulness to the model. Hence the convenience 'X'
labels.
On 14 May 2014 08:58, Jörg Tiedemann <[email protected]> wrote:
>
> To follow up on the tracing problems. I added now the sourcePhrase
> feature function. My output is now like this:
>
> Trans Opt 0 [0..5]: [5..5]=</s> [0..4]=S : S ->S -> S </s> :0-0 :
> c=-1.35101 core=(0,-1,1,0,0,0,0,0,0) sourcePhrase=X </s>
> 3.97252core=(0,-6,10,-3.48177,-5.5193,-2.44483,-2.63781,3.99959,-10.4206)
> Trans Opt 0 [0..4]: [4..4]=X [0..3]=S : S ->S -> S X :0-0 1-1 :
> c=1.1999 core=(0,-0,1,0,0,0,0,0.999896,0) sourcePhrase=X X
> 2.83722core=(0,-5,9,-3.48177,-5.5193,-2.44483,-2.63781,3.99959,-10.2912)
> Trans Opt 0 [0..3]: [3..3]=X [0..2]=S : S ->S -> S X :0-0 1-1 :
> c=1.1999 core=(0,-0,1,0,0,0,0,0.999896,0) sourcePhrase=X X
> 0.909388core=(0,-4,7,-3.09425,-4.70708,-1.94095,-2.16629,2.99969,-10.2172)
>
> which still includes the nonlinguistic labels X instead of the tree
> labels from my input sentence. Here is the input tree once again:
>
> <tree label="ROOT"> <tree label="NOUN"> Resumption </tree> <tree
> label="adpmod"> <tree label="ADP"> of </tree> <tree label="adpobj"> <tree
> label="det"> <tree label="DET"> the </tree> </tree> <tree label="NOUN">
> session </tree> </tree> </tree> </tree>
>
> Did I miss something else that I have to set to obtain the source
> language labels used as constraints in decoding or is there something wrong
> with my example?
>
> Thanks again,
> Jörg
>
>
>
>
>
> the trace file normally only shows the target side of the rule since
> the source isn't usually required once the rule has been found.
>
> To see source side in the trace file, add an extra feature function to
> your ini file
> [feature]
> SetSourcePhrase num-features=0
>
> Then each line in the trace file will contain
> sourcePhrase=…
>
>
> Thank you, Hieu. I will try that.
>
>
> About the tree-to-string model in Moses -
> it takes a tree as input, but the resulting derivation doesn't need to
> be a valid tree. It uses the syntax labels only as a constraint on whether
> rules can be applied
>
>
>
> OK. I guess that I have to do some more background reading. I thought in
> string-to-tree you would still have some kind of CKY-style parsing which
> would give you a tree even on the target side. What I understand is that
> labeling can be ambiguous with respect to source side labels. But it should
> still be a tree after decoding, shouldn't it?
>
> Otherwise, there is also a note (4) about obtaining consistent labels at
> http://www.statmt.org/moses/?n=Moses.SyntaxTutorial#ntoc25
>
> Is there already a script in Moses that I can run over an extracted
> rule-table to make these changes? Otherwise, are there any pitfalls that I
> should be aware of when doing the projection or is it as straightforward as
> it sounds?
>
> Thanks!
> Jörg
>
>
>
> On 7 May 2014 10:29, joerg <[email protected]> wrote:
>
>>
>> I started to run some experiments with tree-to-string models and I
>> would like to look closer at the derivations of the best hypothesis. Could
>> someone explain me how to understand the trace-information and also how to
>> get output that looks more like the ones shown at
>> http://www.statmt.org/moses/?n=Moses.SyntaxTutorial#ntoc5
>>
>> I have input like this:
>>
>> <tree label="ROOT"> <tree label="NOUN"> Resumption </tree> <tree
>> label="adpmod"> <tree label="ADP"> of </tree> <tree label="adpobj"> <tree
>> label="det"> <tree label="DET"> the </tree> </tree> <tree label="NOUN">
>> session </tree> </tree> </tree> </tree>
>>
>> and rule-tables that contain rules like the following:
>> Resumption [NOUN] ||| reprise [X] ||| 0.103716 0.25 0.656868 1 ||| 0-0
>> ||| 19 3 3 |||
>> Resumption [adpmod][X] [ROOT] ||| reprise [adpmod][X] [X] ||| 0.110561
>> 0.25 0.110561 1 ||| 0-0 1-1 ||| 0.999999 0.999999 0.999999 |||
>> Resumption of [adpobj][X] [ROOT] ||| reprise de [adpobj][X] [X] |||
>> 0.110561 0.0699863 0.110561 0.391762 ||| 0-0 1-1 2-2 ||| 0.999999 0.999999
>> 0.999999 |||
>> Resumption of the session [ROOT] ||| reprise de la session [X] |||
>> 0.110561 0.00885433 0.110561 0.101433 ||| 0-0 1-1 2-2 3-3 ||| 0.999999
>> 0.999999 0.999999 |||
>>
>>
>> My trace is then as follows:
>>
>> Trans Opt 0 [0..5]: [5..5]=</s> [0..4]=S : S ->S -> S </s> :0-0 :
>> c=-1.35101 core=(0,-1,1,0,0,0,0,0,0)
>> 4.10236core=(0,-6,10,-3.538,-4.72685,-2.88132,-2.28836,3.99959,-10.4206)
>> Trans Opt 0 [0..4]: [4..4]=X [0..3]=S : S ->S -> S X :0-0 1-1 :
>> c=1.1999 core=(0,-0,1,0,0,0,0,0.999896,0)
>> 2.96706core=(0,-5,9,-3.538,-4.72685,-2.88132,-2.28836,3.99959,-10.2912)
>> Trans Opt 0 [0..3]: [3..3]=X [0..2]=S : S ->S -> S X :0-0 1-1 :
>> c=1.1999 core=(0,-0,1,0,0,0,0,0.999896,0)
>> 1.05327core=(0,-4,7,-3.15851,-3.34055,-2.40175,-2.28836,2.99969,-10.2172)
>> Trans Opt 0 [0..2]: [2..2]=X [0..1]=S : S ->S -> S X :0-0 1-1 :
>> c=1.1999 core=(0,-0,1,0,0,0,0,0.999896,0)
>> -0.608594core=(0,-3,5,-2.93781,-2.65946,-1.07876,-0.937101,1.99979,-10.1715)
>> Trans Opt 0 [0..1]: [1..1]=X [0..0]=S : S ->S -> S X :0-0 1-1 :
>> c=1.1999 core=(0,-0,1,0,0,0,0,0.999896,0)
>> -2.0778core=(0,-2,3,-2.2661,-1.38629,-0.420272,0,0.999896,-9.72632)
>> Trans Opt 0 [0..0]: [0..0]=<s> : S ->S -> <s> :: c=1.2
>> core=(0,-1,1,0,0,0,0,0,0) 1.2core=(0,-1,1,0,0,0,0,0,0)
>> Trans Opt 0 [1..1]: [1..1]=Resumption : X ->X -> reprise :: c=-4.73173
>> core=(0,-1,1,-2.2661,-1.38629,-0.420272,0,0,0)
>> -4.73173core=(0,-1,1,-2.2661,-1.38629,-0.420272,0,0,-10.2344)
>> Trans Opt 0 [2..2]: [2..2]=of : X ->X -> de :: c=-1.83086
>> core=(0,-1,1,-0.671711,-1.27316,-0.658493,-0.937101,0,0)
>> -1.83086core=(0,-1,1,-0.671711,-1.27316,-0.658493,-0.937101,0,-4.64553)
>> Trans Opt 0 [3..3]: [3..3]=the : X ->X -> la :: c=-2.38389
>> core=(0,-1,1,-0.2207,-0.681098,-1.32298,-1.35126,0,0)
>> -2.38389core=(0,-1,1,-0.2207,-0.681098,-1.32298,-1.35126,0,-5.73737)
>> Trans Opt 0 [4..4]: [4..4]=session : X ->X -> session :: c=-4.76921
>> core=(0,-1,1,-0.379489,-1.38629,-0.479572,0,0,0)
>> -4.76921core=(0,-1,1,-0.379489,-1.38629,-0.479572,0,0,-11.0403)
>>
>>
>> I expected to see rule applications that include NTs like ROOT, NOUN,
>> adpmod etc. but I don't see any. Do I have to add other flags to get this
>> kind of information?
>>
>> Thanks in advance!
>> Best,
>> Jörg
>>
>>
>> **********************************************************************************
>> Jörg Tiedemann
>> http://stp.lingfil.uu.se/~joerg/
>>
>>
>>
>>
>> _______________________________________________
>> 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
>
>
>
>
> --
> 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
>
>
> _______________________________________________
> 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
