The Evolution of Neural Language Models: Can we predict their trajectory?
I'm a postdoc at CU Boulder
Hi. Welcome! I Recently joined CU Boulder as postdoc working on conversational systems for educational setting (NLP, speech processing). My interests are around ethical and responsible AI, participatory design, accessibility, textual personalization. You can see here selected papers, repo projects, and the research I've been working on. Hope you enjoy ;-)
M research is driven by the value-based research approach recognizing that developing applied technical systems, should ensure understanding of socio-technical implications, in ways that increase social and ecological justice
Selected Research Papers
S. Dudy, S. Bedrick, B. Webber
Empirical Methods in Natural Language Processing (EMNLP)
The figure shows the percentage of Types (%) predicted stratified by frequency in known SotA language models
S. Dudy, S. Bedrick, M. Asgari, A. Kain
Computer Speech and Langauge (CSL)
S. Dudy, S. Bedrick
Association for Computational Linguistics (ACL)
R. Dong, D. Smith, S. Dudy, S. Bedrick
North American Association for Computational Linguistics (NAACL)
From a few projects I've worked on
A Python library to create Special Fsts as described in OpenFst source code. This specializer library enables sigma, rho, and phi operations over fsts over using the command line operations of OpenFst. To have a small footprint, this functionality was implemented in Cython and included generating C++ files.
DOCKER IMAGES FOR FST TOOLS
Openfst and Pywrapfst library images that provide an openfst with pywrapfst environment on ubuntu 14.4. The most recent images are the pynini_1.4.1 and the fst_1.6.1 and their installation details as well as content is available.
OBFUSCATION DETECTION WITH DEEP LEARNING
The basic code for obfuscation detection provides a trained classifier to analyze and detect the encoding language used to obfuscate parts of code in order to hide there sensitive information or potnetially harmful behavior.
Code for different sequence classifiers as described in the paper “Textual Prediction of Prosodic Prominence in Spontaneous Speech with Sequence Classifiers”. There are viterbi, greedy first and second order and, local search first second and third order.