© 2010 AT&T Intellectual Property. All rights reserved. AT&T and the AT&T logo are trademarks of AT&T Intellectual Property. Multimodal Interaction in Speak4it Patrick Ehlen AT&T Rethink Possible
Page 2 This talk will discuss…. Multimodal interaction approaches – mode choice – mode integration Grounding (It’s context!) Grounding in multimodal local search
Page 3 What is multimodal interaction? The most common implementation of “multimodal interaction” – mode choice Let people use more than one mode of input or output – Input: Graphical UI or voice (ASR) – Output: Visual (graphics) or voice (TTS) – Interact using one mode at a time
Page 4 Another approach…. mode integration – Use more than one mode at the same time – Provide simultaneous information using different channels – Combine information from different modes into one interpretation “Italian restaurants near here”
Page 5 Advantages…. It’s natural (underspecification is the norm) Adapt to environment Speech can be shorter and more simple and/or communicate more complex information Complete tasks more quickly “Italian restaurants near here”
Page 6 Advantages…. Some content is better communicated by modes other than speech (e.g., gesturing to communicate spatial information) Information from different modes can complement one other and resolve ambiguities (“mutual compensation”) “Italian restaurants near here”
History of research prototypes – MATCH (Johnston et al 2002) – AdApt (Gustafson et al 2000) – SmartKom mobile (Wahlster 2006) – Multimodal Interactive Maps (Oviatt 1997) Page 7
The Next Big Thing? New technologies (touch screens, GPS, accelerometer data, video-based recognition) will spur an evolution in multimodal interface design – Beyond mode choice to mode integration Speak4it sm – only commercially available product we know of that performs multimodal integration at semantic level – Available for free on iPhone, iPad, Touch Page 8
Multimodal interaction in Speak4It Speak4it gesture inputs – point, line, area (drawn with finger) – when user hits ‘Speak/Draw’ button map display becomes a drawing canvas Page 9
Multimodal integration provides more headaches for designers Problems: – More ‘dimensions’ of context – Demands more focus on “common ground” and aspects of knowledge that have already been grounded with users (Clark 1996)
What is grounding? Mutual knowledge: Things that all parties in a conversation know, and know that other parties in the conversation also know – shared physically, linguistically, or via community When people introduce references, either verbally or by other means, they are grounding those references In dialogue, grounding helps to determine what people say, and what they don’t say – What we do or don’t say reveals a lot about aspects of context we believe are already shared
Grounding in telephony queries – Search queries are very basic dialogue Single exchange of query & response – Telcos have dealt with these queries for a long time…. Cable Car Pizza What listing please? Here’s that number….
Grounding in telephony queries – 411 systems assumed an implicit grounded location because phones had a fixed location (tied to area code) To refer to another location, you called a different area code The area code provided a source of mutual knowledge about the grounded location in a query Cable Car Pizza in San Francisco What listing please? Please call
Then phones lost their tethers (and their implicit grounding mechanisms)…. – With mobile phones, not as much shared knowledge about location – Location became “part of the conversation” again – Spoken query dialogue systems: Google-411, Bing-411, 800-Yellowpages Phone apps etc San Francisco, California What City and state? What listing? Cable Car Pizza
Evidence of grounding problems found in Speak4it Logs Frequency of specific locations in queries: 18% “police department in jessup maryland” “office depot linden boulevard” Most are unlocated: “gas station” “saigon restaurant” Location grounding breaking down: “Serendipity” … followed shortly by “Serendipity Dallas Texas” Page 15 Corrections: “Starbucks Cape Girardeau” … followed six minutes later by “Lowes”.. then right away “Lowes Cape Girardeau”
Location grounding sources in multimodal mobile search Page 16 “italian restaurants” PHYSICAL User’s current location (GPS) GUI Location shown on map display GESTURE Where user touched VERBAL “Sorry I could not find french restaurants in madison” Place spoken in prior query
Example Page 17 “new york, new york”“pizza restaurants”<scroll>
Collecting grounding data in the wild Gathered ground truth from users when they are “in the wild” Present users with a “grounded location disambiguation” screen to collect user-reported intentions Display to ~20% of unlocated queries Use these data to train a context model and to judge model comparisons Page 18
Page % 38.04% 13.59% 69.29% Selected grounded locations (relative to presentation)
Page 20 ncast ncast
Speak4it multimodal architecture Page 21 Gesture Recognition Gesture Recognition ASR NLU Location Grounding Interaction Manager Interaction Manager Multimodal Search Platform Multimodal http data stream (speech, text, ink) Results/ Requests Ink trace Gestures Speech Platform Geo-coder Search Audio Parsed string Location string Lat/Lo n Query Results Listings index NL model Geo index Features Salient Location ASR Gesture Recognition Gesture Recognition SLM
Page 22 Conclusions Multimodal UIs will soon move from mode choice to mode interaction We’ll need richer context models to predict grounding of locations and other references across modes, to align system actions with user expectations Mobile voice searchers don’t always consider their “GPS” location as the grounded one; location shown on the map is considered grounded 37% of the time User groundings from touch are highly salient
Page 23 Acknowledgments Thanks to Jay Lieske, Clarke Retzer, Brant Vasilieff, Diamantino Caseiro, Junlan Feng, Srinivas Bangalore, Claude Noshpitz, Barbara Hollister, Remi Zajac, Mazin Gilbert, Barbara Hollister, and Linda Roberts for their contributions to Speak4it.