1st International Workshop on Conversational Approaches to Information Retrieval (CAIR'17)
at
in collaboration with
Follow us: @CAIRWorkshop
Program
- Date: Friday, August 11, 2017
- Time: 09:00 - 17:20
- Venue: Keio Plaza Hotel, 43rd Floor, Moon Light Room
Program at a glance
- Welcome
- Keynote Talk I: Search Failed? Let's talk (Ron Kaplan)
- Hands on with commercial devices (+ poster setup)
- Coffee Break
- Oral Session 1: Datasets
- Panel discussion
- Lunch Break
- Keynote Talk II: End-to-end learning for task-oriented conversational systems (Jason D. Williams)
- Oral Session 2: Models and Interactions
- Coffee Break
- Oral Session 3: User Cases and Evaluation
- Open Discussion + Closing
09:00-:15
Welcome
09:15-10:15
Keynote Talk I: Search Failed? Let's talk
Ron Kaplan (Amazon)
Conversational Search is a complement to traditional search via web search boxes, mobile apps, voice-only speaker appliances, or other interfaces. It provides a friendly fallback mechanism when a first-try search fails to satisfy a user's needs. Search may fail because the engine may be unable to properly decode and match the user’s information intent, especially when it diverges from the big-head patterns that query analysis, relevance ranking, and other standard search technologies are tuned to. This may be because the user's underlying information goal is expressed in an idiosyncratic way, or perhaps because the information goal is too complex for the user to lay out in a single search query. It may also be because the user is still in an exploratory stage of research and doesn't yet know what they are actually looking for. Conversational Search is a way of recovering a retrieval mission when, for whatever reason, an initial search is unsuccessful. The user and the system can collaborate in a multi-turn natural language conversation that helps the user sharpen a possibly fuzzy information intent and helps the system better understand and respond more accurately to the user's goals.
10:15-10:30
Hands on with commercial devices
Poster setup
10:30-10:50
Coffee Break
10:50-11:35
Oral Session 1: Datasets
10:50-11:05
Paul Thomas (Microsoft), Daniel McDuff (Microsoft Research), Mary Czerwinski (Microsoft Research) and Nick Craswell (Microsoft)
11:05-11:20
Johanne R Trippas (RMIT University), Damiano Spina (RMIT University), Lawrence Cavedon (RMIT University) and Mark Sanderson (RMIT University)
11:20-11:35
Johanne R Trippas (RMIT University), Damiano Spina (RMIT University), Lawrence Cavedon (RMIT University) and Mark Sanderson (RMIT University)
11:35-12:30
Panel Discussion
Moderator: Mark Sanderson (RMIT University)
Panelists:
- Nick Belkin (Rutgers University)
- James Allan (University of Massachusetts at Amherst)
- Jason D. Williams (Microsoft Research)
- Emine Yilmaz (University College London)
12:30-14:00
Lunch Break
14:00-15:00
Keynote Talk II: End-to-end learning for task-oriented conversational systems
Jason D. Williams (Conversational Systems Group, Microsoft Research)
Sequence-to-sequence neural models are an attractive approach for modeling the textual back-and-forth of conversation. While early work focused on social (non-task-oriented) conversation, for building practical task-oriented conversational systems, models that operate on text alone are impractical. For example, task-oriented systems interface with structured APIs (not just text), and task-oriented systems often require constraints such as business rules. Moreover, for task-oriented systems, there is generally no in-domain data to train from. In this talk I'll introduce Hybrid Code Networks, which overcome these challenges by combining an RNN with domain-specific knowledge encoded as software and system action templates, and by supporting a mixture of supervised, reinforcement, and interactive learning. In addition to covering this method and results, I'll also speculate on possible touchpoints with information retrieval, and on important open scientific problems for conversational systems more broadly.
15:00-15:30
Oral Session 2: Models and Interactions
15:00-15:15
Tom Kenter (University of Amsterdam) and Maarten de Rijke (University of Amsterdam)
15:15-15:30
Jaime Arguello (University of North Carolina at Chapel Hill), Bogeum Choi (University of North Carolina at Chapel Hill) and Robert Capra (University of North Carolina at Chapel Hill)
15:30-15:50
Coffee Break
15:50-16:35
Oral Session 3: User Cases and Evaluation
15:50-16:05
Zhanyi Liu (Baidu), Zheng-Yu Niu (Baidu), Jian-Yun Nie (Université de Montréal), Hua Wu (Baidu) and Haifeng Wang (Baidu)
16:05-16:20
Julia Kiseleva (University of Amsterdam) and Maarten de Rijke (University of Amsterdam)
16:20-16:35
Rishabh Mehrotra (University College London), Ahmed Hassan Awadallah (Microsoft Research), Ahmed El Kholy (Microsoft) and Imed Zitouni (Microsoft)
16:35-17:20