Edinburgh Summarization Demo PeiYun Sabrina Hsueh AMIDA Review

Edinburgh Summarization Demo Pei-Yun (Sabrina) Hsueh AMIDA Review Meeting, 7 -8 February 2008, IDIAP, Martigny Name/Last Name 1 www. amiproject. org

System Overview • Our goal: • To help users find material in recorded meetings quickly • indexing to help them find their way around a recording • compressed playback • automatic summarisation. • Let you catch up on what’s happened so far if you’re late for a meeting. 2 Name/Last Name 2 www. amiproject. org

Three components • (1) Topic segmenter • We formulate the problem as detecting the speaker spurts that are near the topic boundaries. E. g. , in the beginning of (2), ’Okay, lets do the look and feel design presentation first. ’ 3 Name/Last Name 3 www. amiproject. org

Three components • (2) Summarizer • Each summary consists of import utterances from the meeting and serves as a navigation tool for indexing the meeting record. • We classify the dialogue acts as extract-worthy or not by training an SVM on the AMI training data set and running the SVM over the test set. 4 Name/Last Name 4 www. amiproject. org

Three components • (3) Compression component • If the user selects a very severe compression ratio, we delete not only dialogue acts but uninformative words as well. • This is determined by the use of the TF · IDF weighting scheme, which weights a term highly if it occurs many times in one document (e. g. a given meeting) and rarely across a set of all documents (e. g. a group of meetings). 5 Name/Last Name 5 www. amiproject. org

Screenshot • Great progress • More progress 6 Name/Last Name 6 www. amiproject. org