Software to fit whole-sentence language models using the principle of maximum entropy. For developers of speech recognizers, text prediction interfaces, OCR, machine translation software.
Be the first to post a text review of ftw. Text Modeller. Rate and review a project by clicking thumbs up or thumbs down in the right column.
The most significant change is a new online algorithm for estimating feature expectations. This allows estimation to any desired accuracy without memory limitations. The algorithm is described in a paper to be submitted for the Interspeech 2005 conference. A link will be posted after the review process. I will also make the code available in a version for Matlab. Stay tuned. Other changes include bug fixes, a few new unit tests, a more streamlined code base and a new tutorial at http://textmodeller.sourceforge.net/tutorial.html.
Changes since v0.8: Added support for SciPy 0.32. Fixed various bugs, improved robustness for large problems and/or large feature values. Implemented new complex log routines, obviating the need to store indices of positive and negative feature values, for better memory efficiency. Included support for the limited memory variable metric code of Zhu, Byrd, and Nocedal available as LBFGS-B in SciPy. Fixed a small UTF-8 display bug in the Berger example.
Replaced the use of the Numerical Recipes-derived CG routines (with their restrictive licence) with the equivalents in scipy.optimize. Re-wrote more of the entropy and gradient routines to work in the log domain and added checks for added robustness with very large and small-valued features and parameters. Added rudimentary routines for fitting models on continuous sample spaces. These are not as robust as the routines for discrete sample spaces.
Be the first person to add a text review.
Copyright © 2009 Geeknet, Inc. All rights reserved. Terms of Use
Thanks for your rating!
Would you also like to write a review?