Hi everyone, I think now is as good time a time as any to mention my feelings about Joshua. You may have noticed that I haven't done much active development over the past year; you likely also know that the reason is that the research community has shifted entirely from work on statistical models to work on neural machine translation. On the research side, neural models now consistently outperform phrase-based systems on BLEU score on language pairs where there is enough data (roughly, around 15 million words of training), and work there has injected a lot of new life into a field that many had felt was starting to stagnate. From a production standpoint, neural systems are also a big win: the models do best with a GPU and take some time to train, but the architecture and pipeline are simpler, and the resulting models are constant-sized and on the order of a few gigabytes at most, instead of scaling with training data into the tens of gigabytes, as statistical systems do. Test-time inference can also be run fairly efficiently on CPUs where throughput demands are low enough. All commercial systems are now neural or are quickly moving in that direction, including relatively surprising places like Systran, which until recently was known as the world's best-known rule-based system. As GPUs become more ubiquitous and cheap, this situation is only going to get better, even for the end user. There is little doubt that neural MT has supplanted statistical approaches to machine translation, across both academic research and industry. And it is still in its relative infancy, with lots of interesting research problems and engineering issues to investigate and resolve.
It's somewhat sad for me because I've been working on or with Joshua for almost seven years, but I also find my feelings here interesting in contrast to a previous time I've felt tugged away from Joshua. As many of you know, Philipp Koehn joined JHU a few years ago, which brought some tension to JHU with respect to collaborating on research. There was pressure for me to switch. Moses had a much bigger development community and was much more feature rich, but despite this, I was reluctant to let go of Joshua, for a number of reasons. Java is nicer to work with than C++ (and not really that much slower); our code is better written, IMO; jar files are easier to distribute than C++ in compiled or source form; and, of course, I had much more familiarity with the codebase, not to mention something of a personal stake in Joshua. But with neural MT, I have none of these reservations. It's nice for one to have the Moses/Joshua tension resolved (sometimes, ignoring a problem does make it go away!), but for all the reasons I listed in the opening paragraph, NMT is now the clear way to go. And the bottom line for me is that I can no longer justify spending time on Joshua during my working hours, and with a young family and other interests that I want to pursue, I don't have time for it outside of work. I am happy to still linger on the project, but am unlikely to be much of an active participant unless I'm explicitly asked for something. As I've written before here, I think there may still some role for statistical systems, and therefore, for Joshua. In low-resource situations, StatMT may still be the right approach overall, or even simply the best way to quickly build up a working system. There is some promise I think in deploying models easily on older hardware that people have, and perhaps getting people to hep contribute translations and translation memories that could be used to build and improve systems. There are surely more good ideas in this space in the vein of providing a good tool to users. It's been a great experience for me working with the Apache community on Joshua. I am grateful to Chris for convincing us to make Joshua an Apache incubator project, which put a lot of new life into the project. Lewis has been a lot of help throughout helping smooth over the transition; Tommaso has repeatedly helped with tasks large and small; and that is just three of you. It's too bad therefore that the timing just didn't work out, but neural MT ascended very rapidly. I know there are other members here who are also thinking along these lines. At the same time, I hope my departure from active development doesn’t mean the end of the project for those of you who wish to keep working on it. Sincerely, matt > Le 25 sept. 2017 à 23:10, Tommaso Teofili <tommaso.teof...@gmail.com> a écrit > : > > I would also think we're ready for graduation. > My only concern relates to how many of the current committers are willing > to keep contributing to the project, basically if we have a PMC which is > big enough for the graduation. > > Regards, > Tommaso > > > Il giorno sab 23 set 2017 alle ore 01:21 Chris Mattmann <mattm...@apache.org> > ha scritto: > >> Tom, glad you raised this issue, IMO, Joshua is ready for TLP. >> >> We’ve: >> >> 1. Added new PPMC/committers >> 2. Made a release >> 3. Been friendly and cordial and welcoming on the lists >> 4. Vetted the software >> 5. Have some decent, emerging docs >> >> Graduation time…Thoughts? >> >> Cheers, >> Chris >> >> P.S. Subject line change to officially turn this into a [DISCUSS] and >> hopefully >> a [VOTE] >> >> >> >> On 9/22/17, 4:19 PM, "Tom Barber" <t...@spicule.co.uk> wrote: >> >> So I've not checked against the checklist on the podling page yet, but >> what >> do people feel is missing from Joshua prior to graduation? >> >> I'd like to see some non mentors ship a release so we know we've got >> the >> docs right, but of course it doesn't have to be a major release. >> Similarly >> was all the licensing stuff resolved etc? >> >> I'm curious as its not a very fast paced project and it feels like ones >> like Joshua could sit in the incubator for years without causing much >> trouble but also not graduating. I'm not in any great rush, but what do >> people feel about it? >> >> Tom >> >> >> >>
