Matt and I are living in Bellevue this summer. He’s doing Microsoft internship and I’m working on research. This week has largely been set-up, grocery shopping, traffic, and (successfully) compiling XTAG. I’ve been rewriting my resume and setting up this website. The plan is to post weekly updates and keep pages describing my projects. Also, I’m making a grad school list and studying for the GRE.
The goal this summer is to make a system for planning-based natural language interaction. I’m adopting this method for plan based natural language generation, and applying plan recognition to do natural language parsing. I have three code goals:
Create an automatic system for doing the above. Given a game’s domain and some sample text of words, this system holds the user’s hand through pairing words with their in-game meaning, and hopefully indicates what game actions aren’t yet describable.
Integrate these language-capable PDDL domains with the GME game engine. This will use plan recognition to interpret player commands.
- Alternatively, I may opt to just start running more scientific evaluations on the PDDL, and modify GME during the school year.
I’ll be writing my honors thesis as I go. Besides that, I’ve been compiling a list of grad schools to apply to, studying for the GRE, polishing up my online presence for easier job searching, and enjoying a productive summer without schoolwork stress. It’s nice for work to end at 5:00.
I’ve managed to finally get XTAG compiled and working, at least enough to poke around their tree database easily. For context, XTAG is a database for ‘lexicalized tree adjoined grammars’ (LTAGs). That is, for any english word or phrase, they have all the grammatical forms it might take, each in tree form. The work I’m basing mine off of relies on LTAGs to generate language. XTAG was built in the 90s, so I had a devil of a time getting it to even compile during the semester. I knew it would just take a week or two of concerted effort, but I couldn’t find that much uninterrupted time during the semester. (Hence why I’m taking the summer to do this.)
After going through such hassles for XTAG, I’ve realized I can’t rely on it, especially if I want others to be able to make language-capable PDDL. Eventually I will need to write my own tool just for exploring the LTAGs—one that makes it easy to add game-specific meaning to each. Fortunately, XTAG uses a GNU GPL license, so I can just use their database without inheriting the old code.
That said, my first job is to make a sample PDDL domain. Now that I have XTAG displaying trees, I can focus on writing a planning domain that encodes these trees. The CRISP system has code for this, so my next task is to master that. (Fortunately, it’s rather friendly compared to XTAG!)
In Other News:
Everyone here has adorable dogs and adorable babies.