BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY: Neural Architectures for Sequence Labelling - Marek Rei\, Univers
 ity of Cambridge
DTSTART:20170519T110000Z
DTEND:20170519T120000Z
UID:TALK72538@talks.cam.ac.uk
CONTACT:Anita Verő
DESCRIPTION:Many NLP tasks\, including named entity recognition (NER)\, pa
 rt-of-speech (POS) tagging\, shallow parsing and error detection can be fr
 amed as types of sequence labelling. The development of accurate and effic
 ient sequence labelling models is thereby useful for a wide range of downs
 tream applications. Work in this area has traditionally involved task-spec
 ific feature engineering – for example\, integrating gazetteers for name
 d entity recognition\, or using features from a morphological analyser in 
 POS-tagging. Recent developments in neural architectures and representatio
 n learning have opened the door to models that can discover useful feature
 s automatically from the data. Such sequence labelling systems are applica
 ble to many tasks\, using only the surface text as input\, yet are able to
  achieve competitive results.\n\nIn this talk\, we investigate various met
 hods for further improving neural sequence labelling models. \nWe start wi
 th a sequence labelling model that combines bidirectional LSTMs and CRFs\,
  and then explore two extensions:\n1) character-based representations\, fo
 r capturing sub-word features and character patterns\n2) semi-supervised m
 ultitask objectives\, providing the network with additional training signa
 ls for learning useful general-purpose features.\nWe evaluate the impact o
 f these architectures on datasets that cover four different tasks: NER\, P
 OS-tagging\, chunking and error detection in learner texts.
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR