Hi Matt, On Mon, Mar 14, 2016 at 8:26 AM, Matt Post <[email protected]> wrote:
> Whoa! Lewis, can you give some more detail on this talk, what you > proposed, and what you plan to talk about? > http://sched.co/6OJI > > I haven't ever been to ApacheCon, but am interested in going. I don't have > much of a feel for what motivates folks outside the academic research > community, and that would be good to have in laying out projects that might > interest people. > I agree. Would be great to meet you there. We could have a Joshua meetup. > > Regarding those project, I have a number of them. Perhaps it would be > useful to flesh them out with some more detail, and perhaps post them, for > those who are interested. First, with respect to Tommaso's question, the > following: > > - Use cases. I'd really like to push machine translation as a black box, > where people can download and use models, not caring how they work, and > building on top of them. I think this could be transformative. I've just > added to Joshua the ability to add, store, and manage custom phrasal > translation rules, which would let people take a model and add their own > translations on top of it, perhaps correcting mistakes as they encounter > them. There's a JSON API for it (undocumented). > > Building this up would also require pulling together lots of different > test sets, evaluating changes, and so on. > > - Neural nets. This is a huge research area. I think the advantages are > that it could enable releasing models that are much smaller. However, on > the down side, it's not clear what the best way to integrate these models > into Joshua is. Fully neural attention models would require re-architecting > Joshua, as they are essentially a new paradigm. Adding neural components as > feature functions that interact with the existing decoding algorithm would > be an intermediate step. > OK. This sounds like bang on for a meet up topic. Regardless of who is there, we could have a Webex or something similar for the incubating community, > > For other projects, I'd love: > > - Better documentation, developer and end-user (probably I need to write a > lot of this; if nothing else, it would be hugely useful to me in terms of > prioritizing to know that people want it) > - Rewriting certain components. The tuning modules, in particular, are a > real mess, and should be synthesized and improved. > > - Replacing Moses components. Joshua can call out to Moses to build phrase > tables; it would be nice to get rid of this (and wouldn't be that hard) > with our own Java implementations. It would also be good to add a > lexicalized distortion model to the phrase-based decoder. > > These all sound excellent and would all make very reasonable GSoC projects, Thanks Lewis
