Output details
11 - Computer Science and Informatics
University of Edinburgh
Output title
Spectral Learning of Latent-Variable PCFGs
Type
E - Conference contribution
DOI
-
Name of conference/published proceedings
Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Volume number
-
Issue number
-
First page of article
223
ISSN of proceedings
-
Year of publication
2012
Number of additional authors
4
Additional information
<22> Originality: First algorithm for learning supervised parsing models with latent-variables in a way which is provably correct (from a learning-theoretic perspective).
Significance: Parsing is at the foundations of NLP. It was followed up by a paper that shows how to use the algorithm in practice -- one of the first practical uses of spectral algorithms. (Cohen is the first author, NAACL 2013.) The algorithm also works much faster than previous approaches.
Rigour: The algorithm is based on spectral analysis and the method of moments, a young learning-theoretic/statistical area of research. Proofs of correctness are given in the paper.
Interdisciplinary
-
Cross-referral requested
-
Research group
D - Institute for Language, Cognition & Computation
Citation count
0
Proposed double-weighted
No
Double-weighted statement
-
Reserve for a double-weighted output
No
Non-English
No
English abstract
-
