The AMI Meeting Corpus A PreAnnouncement Jean Carletta

The AMI Meeting Corpus: A Pre-Announcement Jean Carletta MLMI-05

AMI Partners

AMI Overview Development of technology to support human interaction in meetings, through meeting browsers and online meeting assistants for multi-modal meeting recordings. Requires: – – – requirements and evaluation specialists organizational psychologists linguists who study groups signal processors computational linguists

Motivation Technological advancement must be driven by data that – – – is plentiful admits effectiveness measurement has been well-recorded on all modalities has been annotated for things that get us from signal to what the users need is realistic

Why elicitation? Limit the domain – – makes it possible to build ontologies understand what's going on Control knowledge and motivation of participants Makes it easier to get zero-history groups Build in outcome measures

AMI Hub Corpus Goal: 100 hours of 4 participant meetings – – 60% arising from a carefully designed elicitation task remaining 40% from a variety of genres and increasingly real find out where our methods and results generalize find out where they don't

Elicitation task: design a remote control in and around four meetings using normal office software roles: project manager, marketing expert, interface designer, industrial designer – – different training at start of day different feedback from organization throughout phases: project kick-off, functional design, conceptual design, detailed design effectiveness and teamworking measurement throughout the day

AMI Meeting Rooms 4 close- and 2 wide-view cameras, 4 head-set and 8 array microphones, presentation screen capture, whiteboard capture, pen devices, plus extra site-dependent devices TNO Edinburgh IDIAP

48 k. Hz 16 bit Audio Microphone arrays Focusrite Octopre MOTU 2408 MK 3 Focusrite Octopre 48 k. Hz Word Clock MTC MOTU Midi Timepiece AV 60 Hz Blackburst Horita BG-50 PAL ASP 2 Aerial Splitter Receivers X 8

280 Kbps Div. X-encoded Video Sony SSC-DC 58 AP (Wide angle cameras) Timecode Inserter Sony DC 700 60 Hz Blackburst Sony DC 700 MOTU Midi Timepiece AV Timecode Inserter Timecode inserter Longitudinal time code Sony DC 700 (Mini DV recorders) Horita BG-50 PAL Sony XC 555 (close up cameras)

Sample camera views (Edinburgh)

Annotations timestamping against signal – emotion coding; head gestures; hand gestures focusing on addressing and interactions with objects; location in room; gaze orthographic transcription with word-level timings derived by forced alignment discourse structure over orthography – dialogue acts w/ addressing, named entities, topic segments, decision phases, extractive summaries linked to participant abstractive summaries