Active Learning Group discussion on AL for NLP

From the Active Learning Group about AL for NLP(23.02.2011): 

Below are part of the email: 

[Problem]

- A project to explore the suitability of active learning for a number of specialized annotation tasks

- difficult to reliably reproduce some fairly "common results" in the literature. 

- the problem is specifically on sequence tagging tasks such as NER using common datasets such as CoNLL-2003 and MUC6.

- The "common results" means learning curves (f-measure vs tokens of training data) where the active learning selection strategy outperforms a random strategy at all points along the curve.

[Responses]

- using DL-Learner and OWL.

- used Active Learning over pos tags to learn passive in the Tiger Corpus Navigator:

http://dspace.utlib.ee/dspace/bitstream/10062/15937/1/tlt9_submission_33.pdf

- There is also a web demo: http://hanne.aksw.org

- The method was not compared to approaches *not* using active learning, because the algorithm only  works feasible with few examples.

--

♥ ¸¸.•*¨*•♫♪♪♫•*¨*•.¸¸♥