Hi Hieu (and others who might be interested),
So I created three rule tables:
(1) With CCG-internal brackets and which mentions the source lexical item of
interest:
说了算 [X] ||| is [((S\NP[expl])/(S[to]\NP))/(S[adj]\NP)] ||| 0.000113126
6.94e-05 0.00475133 0.5 2.718 ||| ||| 126 3
说了算 [X] ||| is necessary [(S\NP[expl])/(S[to]\NP)] ||| 0.000309866 6.94e-05
0.00475133 0.00028945 2.718 ||| ||| 46 3
说了算 [X] ||| is necessary to [(S\NP[expl])/(S[b]\NP)] ||| 0.000208847
6.94e-05 0.00475133 1.07891e-05 2.718 ||| ||| 68.25 3
(2) Without CCG-internal brackets and which mentions the source lexical item
of interest:
说了算 [X] ||| is [((S\NP_expl_)/(S_to_\NP))/(S_adj_\NP)] ||| 0.000113126
6.94e-05 0.00475133 0.5 2.718 ||| ||| 126 3
说了算 [X] ||| is necessary [(S\NP_expl_)/(S_to_\NP)] ||| 0.000309866 6.94e-05
0.00475133 0.00028945 2.718 ||| ||| 46 3
说了算 [X] ||| is necessary to [(S\NP_expl_)/(S_b_\NP)] ||| 0.000208847
6.94e-05 0.00475133 1.07891e-05 2.718 ||| ||| 68.25 3
(note that the [...] have been replaced with _..._)
(3) With CCG-internal brackets but *without* the lexical item of interest (I
just deleted some of the characters from the Chinese "word"):
算 [X] ||| is [((S\NP[expl])/(S[to]\NP))/(S[adj]\NP)] ||| 0.000113126
6.94e-05 0.00475133 0.5 2.718 ||| ||| 126 3
了算 [X] ||| is necessary [(S\NP[expl])/(S[to]\NP)] ||| 0.000309866 6.94e-05
0.00475133 0.00028945 2.718 ||| ||| 46 3
算 [X] ||| is necessary to [(S\NP[expl])/(S[b]\NP)] ||| 0.000208847 6.94e-05
0.00475133 1.07891e-05 2.718 ||| ||| 68.25 3
(4) Same as (3), but with no CCG-internal brackets.
算 [X] ||| is [((S\NP_expl_)/(S_to_\NP))/(S_adj_\NP)] ||| 0.000113126
6.94e-05 0.00475133 0.5 2.718 ||| ||| 126 3
了算 [X] ||| is necessary [(S\NP_expl_)/(S_to_\NP)] ||| 0.000309866 6.94e-05
0.00475133 0.00028945 2.718 ||| ||| 46 3
算 [X] ||| is necessary to [(S\NP_expl_)/(S_b_\NP)] ||| 0.000208847 6.94e-05
0.00475133 1.07891e-05 2.718 ||| ||| 68.25 3
Now I tested each of these with an appropriate version of the glue grammar
(the whole thing with 4K+ rules in it) -- where "appropriate" just means
that it has CCG-internal brackets or _..._ escaped CCG brackets.
All for of these settings produced a translation. Here are the results:
----------------------------------------------
Rule table (1) and original glue grammar (w/ 4K+ entries)
Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
(1) [1,2]=X (1) [2,2]=X (1)
Num of hypo = 10 --- cells:
0 1 2
1 3 0
3 0
1
BEST TRANSLATION: 8 Q </s> :0-0 : pC=0.000, c=-1.002 [0..2] 5
[total=-5.893] <<-1.303, 0.000, -6.910, -9.087, -9.576, -5.349, -0.693,
1.000, 1.000>>
is
----------------------------------------------
Rule table (2) and glue grammar with [...] => _..._ transformation (w/ 4K+
entries)
Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
(1) [1,2]=X (1) [2,2]=X (1)
Num of hypo = 10 --- cells:
0 1 2
1 3 0
3 0
1
BEST TRANSLATION: 8 Q </s> :0-0 : pC=0.000, c=-1.002 [0..2] 5
[total=-5.893] <<-1.303, 0.000, -6.910, -9.087, -9.576, -5.349, -0.693,
1.000, 1.000>>
is
----------------------------------------------
Rule table (3) and original glue grammar.
Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
(1) [1,2]=X (1) [2,2]=X (1)
Num of hypo = 3222 --- cells:
0 1 2
1 1587 0
1 0
1
BEST TRANSLATION: 3176 Q </s> :0-0 : pC=0.000, c=-1.002 [0..2] 1589
[total=-22.789] <<-1.303, -1.940, -46.302, 0.000, 0.000, 0.000, 0.000,
0.000, 1.000>>
说了算
----------------------------------------------
Rule table (4) and glue grammar with [...] => _..._ transformation.
Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
(1) [1,2]=X (1) [2,2]=X (1)
Num of hypo = 2007 --- cells:
0 1 2
1 1587 0
1 0
1
BEST TRANSLATION: 1993 Q </s> :0-0 : pC=0.000, c=-1.002 [0..2] 1589
[total=-22.789] <<-1.303, -1.940, -46.302, 0.000, 0.000, 0.000, 0.000,
0.000, 1.000>>
说了算
----------------------------------------------
So, it seems that escaping the CCG bracketings (for features) does nothing,
and I got what I expected in all four cases. moses_chart must be pruning or
dropping some rule-table or glue grammar entries somewhere.
Could it have something to do with my having binarized the original rule
table? (Should've mentioned that earlier, sorry.) I think I set the
ttable-limit to 100 then...but I thought that setting was a
per-source-phrase limit, so it shouldn't matter here, since we have just
three entries for the source "word" 说了算 in the original rule table (the
large one, that is).
[Scratches head.]
--D.N.
2011/6/23 Hieu Hoang <[email protected]>
> **
> hmm, strange. if you can send me the model, i'll look into it.
>
> to get the categories in each cell, uncomment line
> ChartManager.cpp line 104
> feel free to make it into a verbose flag option if you wish
>
>
> On 23/06/2011 09:55, Dennis Mehay wrote:
>
> Hi Hieu,
>
> with ttl's = 100 and 0
> --------------------------------------------------------
> Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
> (1) [1,2]=X (1) [2,2]=X (1)
>
> Num of hypo = 4 --- cells:
>
> 0 1 2
> 1 3 0
> 0 0
> 0
> NO BEST TRANSLATION
> --------------------------------------------------------
>
> and with ttl's 100 and 100000000
> --------------------------------------------------------
> Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
> (1) [1,2]=X (1) [2,2]=X (1)
>
> Num of hypo = 4 --- cells:
> 0 1 2
> 1 3 0
> 0 0
> 0
> NO BEST TRANSLATION
> --------------------------------------------------------
>
> This is from a fresh svn checkout that I compiled just before running. The
> glue rules seem to be failing when trying to combine the chart cells that
> cover "<s> 说了算".
>
> My glue grammar has 4666 entries in it, for what it's worth. I can send it
> to you if you want, but it might be too big to put up here on the forum.
>
> Is there a quick-and-dirty way to see what categories are inserted into
> which cells when (some verbosity setting, perhaps)?
>
> > I corrected this behaviour recently
> >
> http://mosesdecoder.svn.sourceforge.net/viewvc/mosesdecoder/trunk/moses/src/ChartTranslationOptionCollection.cpp?r1=4004&r2=4003&pathrev=4004
>
> Ah, yes. I named the binary ...19-june-2011, but I had copied it from a
> previous svn checkout, sorry. These things are still happening on the
> latest checkout, though.
>
> --D.N.
>
> 2011/6/22 Hieu Hoang <[email protected]>
>
>> hi dennis
>>
>> You're right, it should be working. The entries in the glue rules might be
>> pruned. Can you try to change the [table-limit] in the ini file to
>> [ttable-limit]
>> 100
>> 10000000
>> or
>> [ttable-limit]
>> 100
>> 0
>>
>> Each row correspond to the table pruning limit for each table. If you
>> provide only 1 entry, then it prune every table uniformly.
>> StaticData.cpp (line 894)
>> For a grammar with lots of non-terminals like yours, the table limit may
>> be cutting off the some of the entries in the glue rule table
>>
>> Also, the decoder shouldn't be processing <s> and </s> as unknown words,
>> they should only be translated by the glue rules. This is the reason you get
>> 1587 translations of </s>.
>>
>> I corrected this behaviour recently
>>
>> http://mosesdecoder.svn.sourceforge.net/viewvc/mosesdecoder/trunk/moses/src/ChartTranslationOptionCollection.cpp?r1=4004&r2=4003&pathrev=4004
>>
>>
>> On 23/06/2011 06:14, Dennis Mehay wrote:
>>
>> Just in case it confuses anyone, both commands (below) were run in the
>> same way, I just simplified it for expository purposes to " moses_chart -f
>> moses.ini -cube-pruning-pop-limit 2000" in the first case, but not in the
>> second.
>>
>> --D.N.
>>
>> 2011/6/22 Dennis Mehay <[email protected]>
>>
>>> Hi Philipp,
>>>
>>> Thanks for the reply. I tracked some of the cases down to a *known* word
>>> (or whitespace-tokenized thingie, anyway -- I don't know much of what
>>> constitutes a word in written Chinese) by doing the following:
>>>
>>> ----------------------------------------------------------------------
>>> $ echo "说了算" | moses_chart -f moses.ini -cube-pruning-pop-limit 2000
>>>
>>> Translating: <s> 说了算 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1)
>>> [1,1]=X (1) [1,2]=X (1) [2,2]=X (1)
>>>
>>> Num of hypo = 1591 --- cells:
>>> 0 1 2
>>> 1 3 1587
>>> 0 0
>>> 0
>>> NO BEST TRANSLATION
>>> ----------------------------------------------------------------------
>>>
>>> (An aside: 1587 is the number of categories in the unknown word list.
>>> Why does the last token, viz., "</s>", get that many cells? )
>>>
>>> Anyhow, sure enough, there are three entries for the middle token "说了算"
>>>
>>> ----------------------------------------------------------------------
>>> $ zless rule-table.gz
>>> ...
>>> 说了算 [X] ||| is [((S\NP[expl])/(S[to]\NP))/(S[adj]\NP)] ||| 0.000113126
>>> 6.94e-05 0.00475133 0.5 2.718 ||| ||| 126 3
>>> 说了算 [X] ||| is necessary [(S\NP[expl])/(S[to]\NP)] ||| 0.000309866
>>> 6.94e-05 0.00475133 0.00028945 2.718 ||| ||| 46 3
>>> 说了算 [X] ||| is necessary to [(S\NP[expl])/(S[b]\NP)] ||| 0.000208847
>>> 6.94e-05 0.00475133 1.07891e-05 2.718 ||| ||| 68.25 3
>>> ...
>>> ----------------------------------------------------------------------
>>>
>>> There are entries in the glue table for these three categories --
>>> ((S\NP[expl])/(S[to]\NP))/(S[adj]\NP), (S\NP[expl])/(S[to]\NP) and
>>> (S\NP[expl])/(S[b]\NP) --- so we should be able to hack together a
>>> translation using any of them.
>>>
>>> ----------------------------------------------------------------------
>>> <s> [X] ||| <s> [Q] ||| 1 |||
>>> ...
>>> [X][Q] [X][((S\NP[expl])/(S[to]\NP))/(S[adj]\NP)] [X] ||| [X][Q]
>>> [X][((S\NP[expl])/(S[to]\NP))/(S[adj]\NP)] [Q] ||| 2.718 ||| 0-0 1-1
>>> ...
>>> [X][Q] [X][(S\NP[expl])/(S[to]\NP)] [X] ||| [X][Q]
>>> [X][(S\NP[expl])/(S[to]\NP)] [Q] ||| 2.718 ||| 0-0 1-1
>>> ...
>>> [X][Q] [X][(S\NP[expl])/(S[b]\NP)] [X] ||| [X][Q]
>>> [X][(S\NP[expl])/(S[b]\NP)] [Q] ||| 2.718 ||| 0-0 1-1
>>> ...
>>> ----------------------------------------------------------------------
>>>
>>> And just to be sure that it isn't an unknown word problem, let's mangle
>>> the token "说了算" by deleting the last character and see what happens:
>>>
>>> ----------------------------------------------------------------------
>>> $ echo "说了" | ../moses/bin/moses-chart-19-june-2011 -f
>>> dev-test/ZhEn/mert/run1.moses.ini -cube-pruning-pop-limit 2000
>>> Translating: <s> 说了 </s> ||| [0,0]=X (1) [0,1]=X (1) [0,2]=X (1) [1,1]=X
>>> (1) [1,2]=X (1) [2,2]=X (1)
>>>
>>> Num of hypo = 6396 --- cells:
>>> 0 1 2
>>> 1 1587 1587
>>> 1 0
>>> 1
>>> BEST TRANSLATION: 4763 Q </s> :0-0 : pC=0.000, c=-1.002 [0..2] 3176
>>> [total=-22.789] <<-1.303, -1.940, -46.302, 0.000, 0.000, 0.000, 0.000,
>>> 0.000, 1.000>>
>>> 说了
>>> ----------------------------------------------------------------------
>>>
>>> The best "translation" is just a pass-through, as expected (and there are
>>> 1587 nodes for that unknown token -- just as many as there are unknown word
>>> lhs's in the unknown-lhs file).
>>>
>>> Strange. Very strange. Or am I missing the obvious?
>>>
>>> I'm at a loss here. Does anyone have any guesses as to what's going on
>>> here?
>>>
>>> --D.N.
>>>
>>>
>>> 2011/6/22 Philipp Koehn <[email protected]>
>>>
>>>> Hi,
>>>>
>>>> there always should be a rule to combine a span to the left.
>>>>
>>>> Check what labels are chosen for the 13th word, and why there
>>>> are no glue rules for it.
>>>>
>>>> If I would hazard a guess, I would suspect that this is an
>>>> unknown word and a file with the likely labels for unknown words
>>>> is used, but these do not match the glue grammar.
>>>>
>>>> -phi
>>>>
>>>> 2011/6/22 Dennis Mehay <[email protected]>:
>>>> > Hi all,
>>>> >
>>>> > I posted this, but it bounced. My attachments were too big. I'm
>>>> resending
>>>> > without the larger attachment. Apologies for any duplicate posting.
>>>> >
>>>> > I'm running moses_chart to do some syntax-based MT experiments, and,
>>>> during
>>>> > tuning, I'm coming across some instances where the decoder can't
>>>> produce a
>>>> > translation (btw 32 and 38 in a 500 sentence tuning set). This should
>>>> not
>>>> > be happening, so far as I can tell, since I have a glue grammar (where
>>>> all
>>>> > the nonterminals of the training set plus the [Q] nonterminal are
>>>> accounted
>>>> > for), and an 'unknown-lhs' list with the relative frequencies of all
>>>> the
>>>> > categories as they span only a single word in the training set (i.e.,
>>>> the
>>>> > frequency of each category's spanning a single word in the rule table
>>>> / the
>>>> > total number of single-word instances in the rule table).
>>>> >
>>>> > Here is an example of a sentence that there was no translation for:
>>>> >
>>>> > ------------------------------
>>>> > ---------------------------------------------------------
>>>> > Translating: <s> 没有 规划 作 指导 , 就 可能 出现 谁 有 权 谁 说了算 , 谁 官 大 谁 说了算 . </s>
>>>> > ...
>>>> > Decoding:
>>>> > Num of hypo = 84813 --- cells:
>>>> > 0 1 2 3 4 5 6 7 8 9
>>>> > 10<1%20%C2%A0%202%20%C2%A0%203%20%C2%A0%204%20%C2%A0%205%20%C2%A0%206%20%C2%A0%207%20%C2%A0%208%20%C2%A0%209%20%C2%A010>
>>>> > 11 12 13 14 15 16 17 18
>>>> > 19 20 21
>>>> > 1 100 77 93 83 99 99 100 100 85 99 43 85 3 99 85 18
>>>> 100
>>>> > 85 3 14 1000
>>>> > 40 960 278 717 916 857 976 276 396 952 958 150 0 0 919 74 402
>>>> 802
>>>> > 0 0 12
>>>> > 200 975 908 849 850 858 968 971 971 862 974 0 0 0 852 865
>>>> 984
>>>> > 0 0 0
>>>> > 200 940 849 889 763 715 990 962 979 905 0 0 0 0 864 984
>>>> 0
>>>> > 0 0
>>>> > 200 868 939 886 863 803 887 861 981 0 0 0 0 0 871
>>>> 0
>>>> > 0 0
>>>> > 200 828 910 801 838 796 722 870 0 0 0 0 0 0 0
>>>> 0
>>>> > 0
>>>> > 200 799 914 832 801 745 926 0 0 0 0 0 0 0
>>>> 0 0
>>>> > 200 756 819 901 693 692 0 0 0 0 0 0 0 0
>>>> 0
>>>> > 200 716 680 665 437 0 0 0 0 0 0 0 0
>>>> 0
>>>> > 200 683 527 929 0 0 0 0 0 0 0 0 0
>>>> > 200 532 588 0 0 0 0 0 0 0 0 0
>>>> > 200 580 0 0 0 0 0 0 0 0 0
>>>> > 200 0 0 0 0 0 0 0 0 0
>>>> > 0 0 0 0 0 0 0 0 0
>>>> > 0 0 0 0 0 0 0 0
>>>> > 0 0 0 0 0 0 0
>>>> > 0 0 0 0 0 0
>>>> > 0 0 0 0 0
>>>> > 0 0 0 0
>>>> > 0 0 0
>>>> > 0 0
>>>> > 0
>>>> > NO BEST TRANSLATION
>>>> >
>>>> > Translation took 4.340 seconds
>>>> >
>>>> ---------------------------------------------------------------------------------------
>>>> >
>>>> > The ASCII-art chart's alignment may be a bit off, but, just
>>>> eye-balling it,
>>>> > it looks as if the 19th word (index 18) has a chart entry count above
>>>> it,
>>>> > but then this entry does not get combined with what's to the left
>>>> using the
>>>> > glue rules.
>>>> >
>>>> > Could this be a pruning or cutoff issue (i.e., stack size,
>>>> > cube-pruning-pop-limit, maximum number of rules per span, etc.)? Or
>>>> maybe
>>>> > it has to do with the fact that my unknown-lhs file has *all*
>>>> categories
>>>> > that spanned a single word in the training set. Maybe I should prune
>>>> it to
>>>> > the top 10 or 20, or so. I'm really at a loss here. I thought the
>>>> glue
>>>> > grammar would make the decoder always return an answer, no matter how
>>>> awful.
>>>> >
>>>> > Any insight?
>>>> >
>>>> > I have attached my moses.ini file in case anyone wants to have a look.
>>>> I
>>>> > can also send the glue rule file later, but, as I said, it seems to
>>>> account
>>>> > for all of the training set's categories (and it was produced
>>>> automatically
>>>> > using the -glue-grammar option).
>>>> >
>>>> > Best,
>>>> > Dennis
>>>> > _______________________________________________
>>>> > Moses-support mailing list
>>>> > [email protected]
>>>> > http://mailman.mit.edu/mailman/listinfo/moses-support
>>>> >
>>>> >
>>>>
>>>
>>>
>>
>> _______________________________________________
>> Moses-support mailing
>> [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
