Nora Aguirre-Celis
Working on my PhD dissertation at the UNIVERSITY OF TEXAS
under the supervision of Prof. Risto Miikkulainen
Forming Text Representations with Neural Networks
My research is concerned with the use of a Neural Network model to form
meaningful representations of the content of documents available electronically. The goal is to show that these representations are well-suited for
text management tasks such as categorization, evaluation, summarization, and
retrieval.
Using a large set of texts, we expect the model to find
word ordering, semantic structure and document organization
automatically.
Our basic model uses HAL representations of words to train a RAAM
network to form the representations. RAAM (Recursive Auto-Associative
Memory) is trained to develop representations of sentences,
paragraphs, and the entire document, using 2 hierarchical
RAAMs.
We have been working for the last few months developing a model able
to represent a limited set of children narratives taken from the
CHILDES collection at CMU. Our model seems to encode a good quality
of sentence representations but we are still working on the
whole-narrative representation that embed the central ideas found in
the stories.
Additionally we will be adding DISCOURSE RELATION
tags to the sentence representations, to model the flow of ideas or
events happening in the narratives, in order to represent word ordering
and to compare to a bag-of-words statistical approach in the task of
Categorization.
naguirre@cs.utexas.edu
Links related to my research
Nora's Home Page
naguirre@cs.utexas.edu
Last update: October 25, 2002.