I write academic papers, and I've started writing blog posts about them, and I think this post doesn't cover one of the main reasons that academic papers are le

I write academic papers, and I've started writing blog posts about them, and I think this post doesn't cover one of the main reasons that academic papers are less accessible to non-specialists. When you write an academic paper, it's basically a diff on previous work. It's one of the most important considerations when the paper first comes out. The reviewers and the people up-to-the-minute with the literature need to see which bit is specifically new. But to understand your algorithm from scratch, someone needs to go back and read the previous four or five papers --- and probably follow false leads, along the way! It's another reason why academic code is often pretty bad. You really really should write your system to first replicate the previous result, and then write your changes in on top of it, with the a _bare minimum_ branching logic, controlled by a flag, so that the same runtime can provide both results. And you should be able to look at each point where you branch on that flag, and check that your improvements are only exactly what you say they are. When you start from scratch and implement a good bang-for-buck idea, yes, you can get a very simple implementation with very good results. I wrote a blog post explaining a 200-line POS tagger that's about as good as any around.[1] Non-experts would usually not predict that the code could be so simple, from the original paper, Collins (2002).[2] I've got a follow-up blog post coming that describes a pretty good parser that comes in at under 500 lines, and performs about as accurately as the Stanford parser. The paper I wrote this year, which adds 0.2% to its accuracy, barely covers the main algorithm --- that's all background. Neither does the paper before me, released late last year, which adds about 2%. Nor the paper before that, which describes the features...etc. When you put it together and chop out the false-starts, okay, it's simple. But it took a lot of people a lot of years to come up with those 500 lines of Python...And they're almost certainly on the way towards a local maximum! The way forward will probably involve one of the many other methods discussed along the way, which don't help this particular system. [1] http://honnibal.wordpress.com/2013/09/11/a-good-part-of-speechpos-tagger-in-about-200-lines-of-python/ http://honnibal.wordpress.com/2013/09/11/a-good-part-of-spee... [2] http://acl.ldc.upenn.edu/W/W02/W02-1001.pdf‎ http://acl.ldc.upenn.edu/W/W02/W02-1001.pdf‎