BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Constructing datasets for multi-hop reading comprehension across d
 ocuments - Johannes Welbl\, University College London
DTSTART:20180223T120000Z
DTEND:20180223T130000Z
UID:TALK100621@talks.cam.ac.uk
CONTACT:Andrew Caines
DESCRIPTION:Contemporary Reading Comprehension (RC) datasets — SQuAD\, T
 riviaQA\, etc. — are dominated by queries that can be answered with a si
 ngle paragraph or document. However\, enabling models to combine pieces of
  textual information from different sources would drastically extend the s
 cope of RC. In this talk\, I will introduce a novel Multi-hop RC task\, wh
 ere a model has to learn how to find and combine disjoint pieces of textua
 l evidence\, effectively performing multi-step (alias multi-hop) inference
 . \nI present two datasets\, WikiHop and MedHop\, from different domains 
 — both constructed using a unified methodology. I will then discuss the 
 behaviour of several baseline models\, including two established end-to-en
 d RC models\, BiDAF and FastQA. For example\, one model is in fact capable
  of integrating information across documents\, but both models struggle to
  select relevant information. \nOverall the end-to-end models outperform m
 ultiple baselines\, but their best accuracy is still far behind human perf
 ormance\, leaving ample room for model improvement. It is our hope that th
 ese new datasets will drive future RC model development\, leading to new a
 nd improved applications in areas such as Search\, Question Answering\, an
 d Fact Checking.\nPaper: https://arxiv.org/abs/1710.06481
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR