KeisukeKamataki writeup of Jansche and Abney
From Cohen Courses
Jump to navigationJump to searchThis is a review of Jansche_2002_information_extraction_from_voicemail_transcripts by user:KeisukeKamataki.
- They tried to extract the greeting phrase, the name of the caller and returning phone number from voice-mail transcripts. They found that positional information is important to detect the name and greeting phrase since they usually appear the beginning part of the transcript. They relatively achieved better performance with automatic transcriptions comparing with previous work. As for phone number, they basically used the length of numerical representation and combined with their hand-crafted grammar. Their method worked very well with 95 F-measure.
- I like: They did a kind of good job to clarify the empirical probability distribution of the name/phrase occurrence position. They clarified the special difficulties of phone number extraction such spoken rendition and showed their own two-phase approach which worked very well.
- I didn't like: Their principle findings don't sound something new. Their feature selection and classification technique could be stated more clearly. They might want to more discuss about the error analysis rather than performance analysis especially for phrase and name extraction which has a big room for improvement.
