the opposite of seasonal affective disorder

[alexr_rwx]

My superpowers seem to be powered by sunlight.

I don't have to kick my research group in the pants: that's not my job. Will Byrd said it best: if it's a net loss of emotional energy to try to drag people into collaborating with me, then it's not worth it! There are other smart people out there; I'll find people to bother with ideas.

I spoke with Mike about how I was feeling, wrt the group, and he's been really supportive; I mentioned that I wanted to work more closely with him as well (he often writes papers where he's the only author, and I think by default his students work solo too), and since then, he's been cc'ing me on emails and talking more in terms of what we're doing instead of what he's doing, and what I'm doing.

So, good! Let's hack some hacks.

(I'm doing a pretty good job of pushing what I'm working on and thinking about onto the wiki: I think it's good for me. http://hackmode.org/wiki if you're interested. It's all public, but editable just by me.)

Comments

[winterkoninkje.livejournal.com]

So Mike's in email range? Hmm, I haven't heard back from an email I sent a while ago...

[winterkoninkje.livejournal.com]

Re HackMode:

ThingsToLearn > You should just learn Joshua instead of Moses. I can give some pointers on Moses, but it's really a mess to work with; Joshua was born out of the desire to make Moses more usable... The only reason to learn Moses specifically is if you want to use the factored translation models (which, while great for hacks, unfortunately aren't a big win in terms of quality output).

ThingsToLearn > GibbsSampling > An important thing here is to compare Gibbs to other sampling methods like Metropolis--Hastings and MCMC. It's good to know when to sample, but it's also good to know when various samplers are better.

[oniugnip.livejournal.com]

Thanks for taking a look!

Is Joshua a complete replacement for Moses, would you say? So nobody needs to use Moses at all?

Maybe I need comment sections on my wiki pages; if knowledgeable people are going to be dropping knowledge in my general direction, then I want to make it easy...

[winterkoninkje.livejournal.com]

Yes and no. There are differences between how they do what they do ---e.g., Moses is "phrase-based MT" aka IBM models (with factored LMs) whereas Joshua is "hierarchical phrase-based MT" aka synchronous PCFG--- so if you're specifically interested in doing things a certain way, then that could be a reason to go with Moses. Or if you need to avoid the JVM for some reason. Or if you're stuck in a group with too much inertia that's unwilling to upgrade.

But those details aside, yes Joshua is a complete replacement for Moses. The only thing I'm aware of that Moses does which Joshua can't is the factored LM thing; but we intentionally decided against reimplementing that since it didn't work out that well and it'd just add maintenance burden. On the other side, there are lots of things that Joshua does that Moses can't. Perhaps most salient is the ability to handle hierarchical grammars with integrated LM scoring. Also, Joshua aims to be a complete SMT pipeline, so in addition to the decoder it also has stuff for grammar extraction, MERT, a few different kinds of language models, a few different kinds of grammars, etc. Joshua is designed as an extensible research platform so it's easy to add your own feature functions, parsing algorithms, etc. And Joshua has better construction quality when it comes to things like giving error messages, not failing silently or corruptingly on malformed inputs, etc.

Joshua is state of the art and includes all the major bells and whistles (http://cs.jhu.edu/~ccb/publications/joshua-open-source-toolkit-for-statistical-machine-translation.pdf). Whereas Moses is very long in the tooth. You'd definitely be better served to learn about Joshua first and then go look at Moses if and when it has something specific to offer that Joshua lacks.