Connectionist NLP software available

[Risto Miikkulainen]

The code and data for the DISCERN story processing model and the SPEC
sentence understanding model are now available from our ftp/WWW site.
These software packages are not general-purpose neural network
simulators, but cleaned-up code for specific connectionist NLP models.
I am making them available because they contain implementations of general
ideas for debugging complex neural network systems through X11 graphics
interface, for analyzing the performance of the models, and running
experiments with such models.  I've tried to pay special attention on
making the code portable across platforms (it is based on ANSI/K&R C and
X11R5 with Athena Widgets), and making the software easy to modify and
built on. I hope the software can serve as a starting point for other
experiments in connectionist NLP --- where building simulation programs
from scratch turned out to be a heck of a lot of work :-)

To get a quick feel of what these programs are like (without having to
port them), take a look at the DISCERN demo under WWW at
http://www.cs.utexas.edu/~nn/discern.html or by "telnet
cascais.cs.utexas.edu 30000". The demo runs on remotely on
cascais.cs.utexas.edu, with a display on your X11 screen.

-- Risto Miikkulainen


Here's a short discription of the software:

DISCERN
-------
DISCERN is a large modular system for processing script-based stories.
It includes component models for lexical processing, episodic memory,
and parsing, paraphrasing and question answering.  The main reference is
Miikkulainen (1993): "Subsymbolic Natural Language Processing: An
integrated Model of Scripts, Lexicon and Memory", Cambridge, MA: MIT
Press (a precis of this book was recently posted in the connectionists
list).

The DISCERN software consists of four components: (1) the full DISCERN
performance system (i.e. the "demo" program), (2) training the simple
recurrent and feedforward backprop networks for parsing, generating, and
question answering, (3) training the lexicon feature maps and Hebbian
associative connections, and (4) training the hierarchical feature maps
of the episodic memory. All these are available by anonymous ftp from
cs.utexas.edu:pub/neural-nets/discern, or in WWW, from
http://www.cs.utexas.edu/~nn.


SPEC
----
SPEC is a model of parsing sentences with embedded relative clauses. It
consists of the parser (a simple recurrent network), the stack (a RAAM
network) and the segmenter (feedforward) networks that are trained
together and generalize to novel sentence structures. For a quick
description of the model, see our paper in AAAI-94, or a longer
tech. report version from our ftp/www site.  The SPEC software and
papers are available by anonymous ftp from cs.utexas.edu:pub/neural-nets/spec
or in WWW, from http://www.cs.utexas.edu/~nn.