The goal of the PLOW (Procedure Learning on the Web) is to build a
system with which a user can teach the computer to perform tasks on
the web. PLOW learns from both explicit demonstration of the task
together with natural language instruction. The natural
language play by play
provides key information that allows
rapid and robust learning of complex procedures including conditionals
and iteration in one short session. PLOW demonstrates the power of an
integrated approach to learning, combining deep natural language
understanding, reasoning and machine learning.
A good overview paper describing the system is
PLOW: A
Collaborative Task Learning Agent
, which won an
Outstanding Paper Award at the Twenty-Second AAAI National
Conference on Artificial Intelligence (AAAI-2007) in
Vancouver, BC, Canada. Additional papers are available from our
main Publications page.
The first demonstration of PLOW demonstrated the basic techniques by learning fill in a requisition form for purchasing a book, where the system must learn to look up the process and ISBN number. [QuickTime | MP4]
The second version of PLOW introduced the ability to learn iterations using mixed-initiative interaction where the system attempts to infer and complete iterations for the user as it is being taught. As a result, PLOW can learn iterative procedures without seeing even one complete trace of the task in execution. Traditional machine learning techniques would typically require many complete traces in order to reliably infer an iterative construct. [QuickTime | MP4]
The next version of PLOW (November 2007) includes significant improvements to the graphical interface as well as new capabilities such as the ability to understand information extracted from pages. The graphical display helps the user understand what PLOW is doing and what it has learned, reducing the need for corrections and improving the learned procedures. The ability to understand extracted elements allows PLOW to learn more complex tasks involving online analysis, such as only keeping results whose distance is below some threshold (as illustrated in the clip). [QuickTime]
PLOW has also been ported to work on text-based terminal interfaces, instead of Web Browsers. This video shows the PLOW for Terminals (PLOT) system and its use on a terminal system (CHCS) which the US Military Health System uses for booking appointments. [QuickTime]
Documentation related to PLOW development has been relocated to the
the docs
directory.