Nathan Sanders : Research

Specifically, the general theme of all my work: determining that two corpora are different and explaining why in a way that humans can understand. This uses various techniques, including machine learning, which is an important component of Old AI. So I'm really doing Artificial Intelligence here, it's just not popular to call it that any more.

Specifically, what is dialectometry? How is it different from sociolinguistics? What have I done with dialectometry?

Specifically, why should dialectometry work for cochlear implant speech? Does it work? How is it different from dialectometry data? What are the practical applications?

Specifically, what is Optimality Theory? How does it explain the choices we face when pronouncing a word? (Like 'months' or 'clothes'. But also 'Caribbean'.) How is the theory structured? What are its learnability and computability characteristics? What have I done with OT algorithms?

• Comparison of Phonological and Syntactic Distance Measures (Qualifying paper, 2008)
• Sanders, Nathan C. 2007. Measuring syntactic difference in British English. In Proceedings of the ACL 2007 Student Research Workshop, 1–6, Prague, Czech Republic, June. Association for Computational Linguistics.
• ACL 2007 Student Research Workshop poster
• Syntax distance presentation for CL Lunch

• syntax-distance.zip. I compiled this file manually, which means that the code is rather more for reading than for running. If you want to run the code, let me know when you have problems. Start with build.py.
• Still to come: instructions for IUCL students to check out from svn.

Dialectometry (Phonology)

• IULC Student Conference 2008 presentation : cluster analysis
• CI-2008 presentation : Cluster analysis of phonological distance measures of cochlear implant users [PPT]
• Sanders, Nathan C. & Chin, Steven B. 2009. Phonological distance measures, Journal of Quantitative Linguistics, 43:96-114. [DOC, preprint]
• Related code: align.py

Optimality Theory

• Generation of Factorial Typologies in OT : Four Algorithms (Qualifying paper, 2008)
• Code for the four algorithms. It is quite dense and a little messy. All of the relevant code is in the paper, so you may want to read that first.
• Relentless OT : Machine Verification of Linguistic Theory. A term paper in which I explain Andrea Heiberg's implementation of OT and extend it to include Output-Output Faithfulness constraints.
• Code for feature correspondence constraints. I ported most of Heiberg's system to Scala in the process of writing "Relentless OT", but I don't know that she publicly released her code, so I'll just give a sample of the code I wrote myself.
• Repairing OT: Implement OT with a finite GEN. An handout for a talk that I ended up not giving. Similar to Heiberg's method but less developed.
• A nicely packaged version of RCD, BCD and other OT learning algorithms is coming someday. Watch this space.
• In the meantime, you might want to download OTableau, which implements RCD in addition to generating LaTeX formatted tableaux. I wrote the RCD implementation and helped with maintenance of the rest of the application.

Error Correction in Parsers

My undergraduate senior project was the development of an Earley parser and an LR parser that tried to give useful feedback. As research, I would not call it too impressive. But there was a lot of code involved, some of which might be interesting. On the other hand, it was written while I was learning functional programming, so it is hairy in the places where I interspersed recursion and state changes. Oh, well. 5-year-old code. I've improved since then.

• sanders-srproj.zip

Utility functions to bring other languages up to Haskell's level.

Interactively build OT tableaux using an ugly Tk interface. Also some OT learning algorithms.

These are short programs I wrote for the blog. Or just for fun.

• compile.scm : compiles an imperative language to assembler for a virtual machine that supports (1) garbage collection or (2) an unlimited memory space.
• kmeans.pyw : Use k-means to cluster users by movie preferences. As I recall, it treats users as vectors of scores for each movie.
• User.java : The same k-means program, but in Java. It's a lot lower level, so it should be faster. I haven't tested it.
• sudoku.scm : Solve sudoku puzzles that don't require backtracking.
• typed.pyw : Decorators and various other type checking things for Python, in case you want to write down your types or ask the runtime what the complete type of a list is. This is not, I believe, the newest version.
• cons.pyw : An OO implementation of cons, with all the lispy names like car, cdr, reverse, append, etc. I don't use it, but it's cute. Also comes with a decent number of tests.
• subst_cipher.pyw : An unfinished attempt to solve substitution ciphers. I may have switched over to Haskell, or maybe I just dropped it.

An ancient hex editor that displays and edits data with your own custom character encoding.