Full disclosure: this post is plugging work I contributed to, but I think it will be of interest to this community.
In collaboration with the University of Trento (Georgiana Dinu), Oxford's Computational Linguistics Group, (namely Phil, Karl and Ed) has put together a tutorial for this year's ACL, which took place in Baltimore.
This tutorial, New Directions in Vector Space Models of Meaning (pun not intended), covers a variety of topics relating to the theme of representation learning for semantics. We discuss neural language models in depth, neural models of composition, deep learning, and the application of convolutional neural networks to natural language processing. Because of the growing popularity of these topics in the NLP community, the tutorial was the most well attended tutorial at this year's conference.
The tutorial does not pre-suppose much prior knowledge of machine learning beyond basic linear algebra and log-linear models, or any complex knowledge of NLP. The slides are meant to be fairly self-contained, and provide all (or most) of the maths required to implement the models we describe. We also discuss some training and architectural tricks and "black magic" that most people use when applying these models.
The slides for the tutorial can be found on our group's resource page (apologies for typos), and the video of the tutorial, lovingly aligned with the slides by yours truly (6+ hours of painful labour and checking), has just been uploaded to youtube (we apologise for the cruddy quality of the image/sound, but we had to film this ourselves, rather than professionally).
We hope you enjoy our tutorial.
[link][7 comments]