Research

Rob Speer

[edit]

Current research

I work in natural language processing (NLP), or enabling interaction with computers using human language.

My perspective on the field is that we don't currently have anything close to the computing power that a computer would need to understand language the way a human does. What we can do is use the AI techniques we know to narrow the gulf of understanding between a human and a computer.

An important step towards narrowing the gulf is to give the computer access to common-sense knowledge, so that it can talk about the same world that humans talk about. Common sense won't solve everything, but it puts some grounding in reality behind the words the computer uses.

In my research with the Commonsense Computing Group, I'm working to collect common sense knowledge from the general public over the Web. This has become my thesis proposal.

NLP can also make more progress if its algorithms are accessible and easy to use. I'm a developer for NLTK, the Natural Language ToolKit for the Python programming language.

I also believe that we can apply NLP techniques to understanding music, and by doing so we could create more intelligent music transcription software. My senior project in music was to survey the field of computational music theory, and the result is a paper called Computable Theories of Music Analysis.

[edit]

Published research

  • F0 peaks aligned with nonprominent syllables in American English. In a UROP I worked on from 2003-2004 in the RLE Speech Lab on prosodic labeling, I identified a speech phenomenon that was not covered by the labeling system; this paper, on how we perceive the phenomenon, was the result.
  • Meeting the Computer Halfway: Language Processing in the Artificial Language Lojban. Catherine Havasi and I presented this as a poster at the Student Oxygen Workshop 2004 conference. Working on this project made me realize that common sense was more essential than logic (and also that I wouldn't be likely to find a professor who was interested in this sort of research), and I've also become disillusioned with the Semantic Web idea, so from there I went to work with Push Singh on common sense architectures.
[edit]

Other work

  • My senior project in music, for Prof. Peter Child, was an analysis of how NLP techniques could be applied to music theory: Computable Theories of Music Analysis
  • I developed course software for Prof. Berwick's 6.863 (Natural Language Processing). The list of laboratory assignments includes some assignments that use my code:
    • Laboratory 1 involves PyKimmo, a GUI frontend to the Kimmo two-level morphology parser that can show the transformations step by step, along with the finite state automata that controlled the transformations, in order to effectively teach the concept. I developed this program together with Beracah Yankama.
    • Laboratory 2 involves the Earley natural language parser. The Python Natural Language ToolKit (NLTK) included a simple Earley parser. I augmented it to match Prof. Berwick's course material by using feature structures as non-terminals, with the results visible in an interactive GUI. Lab 3, Semantic Interpretation, also builds on this parser.
Retrieved from "http://torg.media.mit.edu/rob/index.php/Research"