Nlao writeup of Jansche and Abney

This is a review of Jansche_2002_information_extraction_from_voicemail_transcripts by user:Nlao.

This paper essentially says that if you want to extract from voicemail transcrpts, positions and lengths are very important features, the actually words are not very helpful due to ASR error.

Beware, however, this is mainly because we are trying to match ASR words with dictionaries here. Given large training copora (which is very hard to get), ASR words can still be potentially very helpful without using any dictionary.

[minor points]

- Lack of comparision between log-linear model and decision tree

- Lack of discription about the hand-crafted grammar