Philgoo Han writeup of Huang, Zweig and Padmanabhan

This is a review of Huang_2001_Information_Extraction_From_Voicemail by user:Ironfoot.

This paper compares the three information extractino methods used over voicemail. Hand-crafted rules as a baseline, maximum entropy as the most fancy technic of the time and probabilistic transducer as a novel approach. The paper also address the novelness of using spoken data. However the core is totally based on text IE. I don't get the role of 'spoken'.

Focusing on the probabilistic transducer method, the results are showing quite low scores and still leaves so many questions. Would the problem lay on merging step of the finite state automata? or the hierarchical induction step? or any other reason? Will building probabilistic rules be a good idea?

And I find out this wasn't the paper I'm suppose to writeup.