Heu, non, désolé, pas d'ours ce matin.

I think I may have seen one PowerPoint presentation that was "really inspiring and enthusiastic."  I suspect I never made one that was either.

Great article from the BBC.

Karsh is a wonderful portrait photographer (one of his many claims to fame is the photograph of an angry Churchill, for which he presumably stole the guy's cigar for added menacing scowl).

The festival takes place in many locations.  For example, the Museum of Science and Technology, in addition to photographs, is displaying large format cameras.   Bring kids to show them what cameras looked like before they became cell phone features!

Not surprisingly, the highest "quality" (at least as measured by BLEU) is Portuguese-to-Spanish.

More surprising: The difference between eg Italian->Spanish and Spanish->Italian.  Or the TER score of anything going to Greek (maybe a tokenization bug here?).

Anyway it's also a great entry point to various mono- and multi-lingual resources.

It should be a good way to learn more about the fancy new stuff that has been developed within and around that project.

I will give a talk on "Improving Statistical Machine Translation by learning the translation direction" in the afternoon.

On se demande juste: pourquoi «en colère»?  En France, quand on est en colère, on fait la grève, (ou au moins une petite manif).  Là, madame Borde fait des articles en plus.  Ils sont fous ces Québecois! (*)

Je vous quitte, je vais chercher du Tamiflu.

(*): avec mes excuses à Obelix.

• Computational Linguistics has gone from symbolic to statistical methods over the 1990s, and with that switch came an extreme reliance on large amounts of data.
• There is widespread (imho correct but poorly supported) belief in Machine Translation that doubling the amount of data produces essentially a constant gain in BLEU score.
• A recent lecturer at our institute was making the point that waiting for the web to grow was arguably a more efficient way to increase system performance than working on better models [but where would the fun be in that, right?]

Of course being interested in statistical learning, I'm not about to argue against the appropriateness of learning from data.

However, I find the over-reliance on data and its natural growth on the web slightly unconvincing, especially when it is used as an excuse for not working on the actual models.

Machine Translation seems particularly interesting because it is regularly evaluated in public bakeoffs or internal project evaluations.  There is a clear trend that increasing data does regularly increase performance.  Note, however, that, before the current state-of-the-art phrase-based MT models, were a set of word-based models collectively referenced as "IBM models".  It is very unlikely that, even with massive amounts of additional data, IBM models would outperform current phrase-based approaches.  (In fact some IBM models are used as preprocessing when learning phrase-based models.)

So: the availability of growing amounts of data, however unreasonably effective it may be, should not, imho, replace sound work on the modelling part.  It should, on the other hand, help us think of what features are important in a model.  Scaling is of crucial interest, as shown by the recent work on fast classifier training.  In that context, I believe it is also important to think about non-parametric models, ie models for which the effective model complexity grows with the amount of training data.

Even in my case, with a fairly distinctive name, it turns out that publication search is not that straightforward:

1. There are a number of other "C. Goutte", including at least one Caroline with fairly well-cited papers in the field of medicine.
2. On one of my well-cited papers I'm credited only as "C. Goutte", making it difficult for search engines like Scholar to tie it to my full name.

Having a unique ID would make search a lot simpler... assuming the ID is truly unique (not the case yet), and that search engines can somehow extract it reliably!

Reminiscent of the exhibition "Imaging a shattering earth", presented last year at the CMCP and originally developed at the Oakland University Art Gallery. (great website with pictures, lots of related material and links to podcasts of a tour of the CMCP exhibition -- et donc disponible en Français aussi).