|
What are, in your opinion, the most interesting papers from the ACL 2010 conference that just happened in Sweden? Let's make this thread a collection of some of that conference's most interesting papers! If possible, please give only one paper per answer and briefly say why you liked it. This could become a valuable resource for everyone interested in latest NLP papers.
This question is marked "community wiki".
|
|
The Viterbi algorithm used for many sequence tagging tasks runs in quadratic time with the number of labels. This paper shows an exact method (i.e. not an approximation) that makes it orders of magnitude faster. This could even have accuracy implications for some tasks, since sequence taggers often use beam search to make Viterbi more efficient.
This answer is marked "community wiki".
Thanks, will read this immediately!
(Jul 22 '10 at 21:27)
aria42
Does anyone have a sense how it is better than CarpeDiem?
(Jul 31 '10 at 15:35)
Yaroslav Bulatov
2
They are very similar, but this algorithm has less assumptions than carpediem (assumptions in the sense that carpediem is faster when they are true, not that carpediem fails when they are false). Carpediem, as far as I understand, works faster when transition information is less important that current-state information, and this algorithm makes no such assumptions.
(Jul 31 '10 at 16:01)
Alexandre Passos ♦
|
|
I thought this one was fascinating, basically because it's the best use I've seen of that corpus of 19th Century novels. David Elson, Nicholas Dames and Kathleen McKeown. Extracting Social Networks from Literary Fiction.
This answer is marked "community wiki".
|
|
I thought this one was interesting: Mark Johnson: PCFGs, Topic Models, Adaptor Grammars and Learning Topical Collocations and the Structure of Proper Names (PDF). I haven't read it carefully yet but it seems it allows an LDA-type model that finds topics for multi-word phrases, not just single words.
This answer is marked "community wiki".
|
