Re: [Moses-support] MERT's Powell Search

[Adam Lopez]

>
> On line 6 does the "score" in "compute line l: parameter value → score"
> refer to (i) the MT evaluation metric score  (e.g. BLEU) between the
> translation and the reference sentence or (ii) nbest list weighted overall
> score as we see in the last column of a moses generated nbest list (e.g.
> http://www.statmt.org/moses/?n=Advanced.Search)?
>

Neither. It is the model score of that sentence w.r.t. the parameter you're
optimizing. Once you have the model score for each sentence as a function
of λj, you can then construct a function representing BLEU as a function of
λj by finding the convex hull representing the result of the argmax
operation. That is what is happening in slides 31-36. See below.


> At line 8 of the pseudo code, when it asks to "find line l with steepest
> descent", is it looking for each sentence find the (i) line with the
> highest λj or (i) the line with the highest g(ei|f).
>

In this context, "steepest descesnt" means "steepest slope"; i.e. choose
the sentence i with the greatest value ai.


> Then at line 15 of the pseudo code, it says "compute score for value
> before first threshold point". Is this "score" different from the "score"
> at line 6? At line 6, it's a sentence-level score (which I hope it means
> BLEU and not the weighted overall score), and at line 15, it seems to be
> computing the corpus-level score given the initial parameter values.
>
> If at line 15, it is computing the corpus level score, is it only taking
> the best score of the n translations for each reference? And if this is
> BLEU, it's doing not a simple case of averaging sentence-level BLEU which
> might be kept from line 6, is that right? If it is BLEU, then this score
> could be pre-computed before the powell search too, right?
>

Remember what we're trying to do: choose λj to maximize BLEU. The algorithm
here does that exactly w.r.t. the N-best list. That is, over a corpus of M
sentences for which we have N-best translations, we want to find:

1) argmaxλj BLEU(λj)

Let's unroll this computation. Let êm(ℓ) be the translation that the
decoder chooses for the m-th training example when λj=ℓ, and bm(ê) be a
function returning the vector of sentence-level statistics used in the
computation of BLEU when ê is the translation of the m-th training example
(i.e. n-gram matches and reference counts). BLEU is a function of the
aggregate results of calls to b, so (1) becomes:

2) argmaxλj BLEU(∑m ∈ 1,...,M b(êm(λj)))

But êm(λj) is just argmaxn∈ 1,...,N g(em,n,fm,λj), where em,n is the n-th
element of the N-best list for the m-th training example and fm is the
source sentence of the m-th training example, and g is the model score we
compute from this pair as a function of λj (holding the remaining elements
of λ constant, remember). So this becomes:

3) argmaxλj BLEU(∑m ∈ 1,...,M b(argmaxn∈ 1,...,N g(em,n,fm,λj)))

And since we have g(em,n,fm,λj) = ∑k∈ 1,...,|λ| λkhk(em,n,fm) = λjhj(em,n,fm)
+ ∑k∈ 1,...,j-1,j+1,...,|λ| λkhk(em,n,fm), we get:

4) argmaxλj BLEU(∑m ∈ 1,...,M b(argmaxn∈ 1,...,N λjhj(em,n,fm) + ∑k∈
1,...,j-1,j+1,...,|λ| λkhk(em,n,fm)))

Since both h and and the remaining elements of λ are fixed, this becomes
(using a variant of the notation in slide 31, where a and b are functions
of these constants):

5) argmaxλj BLEU(∑m ∈ 1,...,M b(argmaxn∈ 1,...,N λja(em,n,fm) + b(em,n,fm)))

The function inside the outer argmax in (4) is exactly the function that's
being constructed piece-by-piece in slides 31-35, and illustrated in slide
36. Here's how that happens:

- On slide 31, we construct the model score the n-th element of the N-best
list for the m-th training example em,n as a linear function of λj, as
we've discussed. This is the bit inside the inner argmax.

- On slide 32, we repeat the construction of 31 for *every* element of an
N-best list for the m-th training example.

- Slide 33 shows the max of the function inside the inner argmax. Each
point on the convex hull is a point where the argmax changes, and the
argmax of any interval over the x-values of these points is just the
element of the n-best list giving rise to the line whose value is maximal
in that interval.

- Slide 34 shows how we actually get the argmax. We have to find the
intersection points of the upper convex hull, which is why we're sorting
the lines by slope and computing their intersection.

- Finally, slide 36 shows the complete function inside the argmax of (4).
We compute the statistics b for the maximizing sentence in each interval,
and then sum the resulting function over all training examples. This
basically gives us a set of intervals and sufficient statistics for BLEU in
each interval, which we use to compute the complete function.
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

[Moses-support] Compiling Moses with -fPIC

[Miriam Käshammer]

Dear Moses community,

 

My goal is to link Moses (the decoder) as a static library into some other shared library. As far as I understand the compiler/linker output of this other library, I need to compile the Moses library with parameter -fPIC (position independent code). Could you help me in achieving this?

 

I already tried to add "cxxflags=-fPIC" to the bjam command like this:

./bjam -j8 -d2 -a --with-boost="${PREFIX}" --with-xmlrpc-c="${PREFIX}" --with-cmph="${PREFIX}" --with-irstlm="${PREFIX}" --install-scripts="${PREFIX}"/scripts link=static cxxflags=-fPIC

However, the build process just seems to get stuck before it actually starts, see attached log.

 

Any help/comment is appreciated.

Thanks!

Miriam

 

 

 

build.log
Description: Binary data
___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

[Moses-support] 1st CALL FOR PAPERS: Computer Speech and Language Special Issue on Deep Learning for Machine Translation

[Marta Ruiz]

*Computer Speech andLanguage Special Issue on Deep Learning for Machine
Translation
*



Deep Learning has been successfully applied to many areas including Natural
Language Processing, Speech Recognition and Image Processing. Deep learning
techniques have surprised the entire community both academy and industry by
powerfully learning from data.



Recently, deep learning has been introduced to Machine Translation (MT). It
first started as a kind of feature which was integrated in standard phrase
or syntax-based statistical approaches. Deep learning has been shown useful
in translation and language modeling as well as in reordering, tuning and
rescoring. Additionally, deep learning has been applied to MT evaluation
and quality estimation.



But the biggest impact on MT appeared with the new paradigm proposal:
Neural MT, which has just recently (in the Workshop of Machine Translation
2015) outperformed state-of-the-art systems. This new approach uses an
autoencoder architecture to build a neural system that is capable of
translating. With the new approach, the new big MT challenges lie on how to
deal with large vocabularies, document translation and computational power
among others*.*



This hot topic is raising interest from the scientific community and as a
response there have been several related events (i.e. tutorial[1]
<#151a106fdef5395b__ftn1> and winter school[2] <#151a106fdef5395b__ftn2>).
Moreover, the number of publications on this topic in top conferences such
as ACL, NAACL, EMNLP has dramatically increased in the last three years.
This would be the first special issue related to the topic. With this
special issue, we pretend to offer a compilation of works that give the
reader a global vision of how the deep learning techniques are applied to
MT and what new challenges offers.



This Special Issue expects high quality submissions on the following topics
(but not limited):

· Including deep learning knowledge in standard MT approaches (statistical,
rule-based, example-based...)

· Neural MT approaches

· MT hybrid techniques using deep learning

· Deep learning challenges in MT: vocabulary limitation, document
translation, computational power

· MT evaluation with deep learning techniques

· MT quality estimation with deep learning techniques

· Using deep learning in spoken language translation



*IMPORTANT DATES*

Submission deadline: 30th March 2016

Notification of rejection/re-submission: 30th July 2016

Notification of final acceptance: 30th October 2016

Expected publication date: 30th January 2017


*GUEST EDITORS*

Marta R. Costa-jussà, Universitat Politècnica de Catalunya, Spain.
marta.r...@upc.edu

Alexandre Allauzen, Centre National de la Recherche Scientifique, France.
allau...@limsi.fr

Loïc Barrault, Université du Maine, France.
loic.barra...@lium.univ-lemans.fr

Kyunghyun Cho, New York University, USA. kyunghyun@nyu.edu

Holger Schwenk, Facebook, USA. schw...@fb.com



--

[1] <#151a106fdef5395b__ftnref1>
http://naacl.org/naacl-hlt-2015/tutorial-deep-learning.html

[2] <#151a106fdef5395b__ftnref2>http://dl4mt.computing.dcu.ie/
___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

[Moses-support] dictionary based word alignment?

[Julian]

Hello all, would anyone know of a word alignment tool that can take a 
bilingual dictionary as an argument to guide probabilities? Preferably 
with an implementation like fast_align or similar.

Thanks in advance

Julian

---

Julian Myerscough
Quality Assurance Manager - Languages for Business Ltd


___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

[Moses-support] input format of Giza++

[Shaimaa Marzouk]

Dear Support-Team,

I have prepared a small corpus (tokenisation, truecasing are done) and would 
like to convert it into the format of Giza++ described here:
http://www.statmt.org/moses/?n=FactoredTraining.PrepareData

Could you please tell me, how to convert the corpus files (two parallel files 
prepared by the Editor “gedit”) into the right input format of Giza++?

Thanks a lot
Shaimaa

___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support

[Moses-support] First Call for Participation: WMT16 Machine Translated related Shared Tasks

[Barry Haddow]

ACL 2016 FIRST CONFERENCE ON MACHINE TRANSLATION (WMT16)
Shared Tasks on translation, evaluation, automated post-editing and 
document alignment.

August 2016, in conjunction with ACL 2016 in Berlin, Germany

http://www.statmt.org/wmt16

As part of WMT, as in previous years, we will be organising a collection 
of shared tasks related to machine translation.  We hope that both 
beginners and established research groups will participate. This year we 
are pleased to present the following 10 tasks:


- Translation tasks
- News
- IT-domain
- Biomedical
- Multimodal
- Pronoun
- Evaluation tasks
- Metrics
- Quality estimation
- Tuning
- Other tasks
- Automatic post-editing
- Bilingual document alignment

Further information, including task rationale, timetables and data will 
be posted on the WMT16 website, and fully announced in January. Brief 
descriptions of each task are given below. Intending participants are 
encouraged to register with the mailing list for further announcements 
(https://groups.google.com/forum/#!forum/wmt-tasks)


For all tasks,  participants will also be  invited to submit a short 
paper describing their system.


News Translation Task
-
This is the translation task run at most of the past WMT editions. This 
year the language pairs will be English to/from Czech, Finnish, German, 
Romanian, Russian and Turkish. Sponsorship for the task comes from the 
EU H2020 projects QT21 and Cracker, Yandex and the University of Helsinki.


IT Domain Translation Task
-
This guest task will involve translation of queries and their responses, 
on the topic of information technology. It will cover English to/from 
Bulgarian, Czech, German, Spanish, Basque, Dutch and Portugese, and be 
sponsored by the EU FP7 project QTLeap.


Biomedical Translation Task
-
This guest task will focus on the translation of biomedical research 
abstracts from English to and from Spanish, Portuguese and French.


Multimodal Translation Task
-
This task will aim at generating image descriptions in a target 
language, given equivalent descriptions in one or more languages. The 
dataset will consist of 30,000 image--description tuples in three 
languages -- English, German and French.


Pronoun Translation Task
-
This will be similar to the task run last year as part of the DiscoMT 
workshop (https://www.idiap.ch/workshop/DiscoMT/shared-task)


Metrics
--
The idea here is that participants propose evaluation metrics for 
machine translation, which compare the MT output against a reference. 
The metrics will be correlated against the human judgements produced in 
the news translation task. This task is sponsored by QT21.


Quality Estimation
-
This consists of several sub-tasks, all of which are concerned with the 
idea of assessing the quality of MT output without using a reference, at 
different levels of granularity: word, phrase, sentence and document. 
This task is sponsored by QT21.


Tuning
-
Participants in this task are asked to come up with algorithms and 
objectives (i.e. metrics) for tuning the parameters of a given MT system.


Automatic Post-editing
---
In this task participants will aim to create systems that can 
automatically correct machine translation outputs, given a corpus of 
human post-edits. This task is sponsored by QT21.


Bilingual document alignment

The aim is to find translated document pairs from a large collection of 
documents in two languages.


Best wishes
Barry Haddow
(On behalf of the organisers)







The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
___
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support