BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning to Detect Stance and Represent Emojis - Isabelle Augenste
 in\, University College London
DTSTART:20161118T120000Z
DTEND:20161118T130000Z
UID:TALK67694@talks.cam.ac.uk
CONTACT:Kris Cao
DESCRIPTION:In this two-part talk\, I will first introduce our work on sta
 nce detection ("EMNLP 2016":https://arxiv.org/abs/1606.05464) and then on 
 learning emoji representations ("SocialNLP@EMNLP 2016\, best paper":https:
 //arxiv.org/abs/1609.08359).\n\nStance detection is the task of classifyin
 g the attitude expressed in a text towards a target\nsuch as Hillary Clint
 on to be 'positive'\, 'negative' or 'neutral'.\nPrevious work has assumed 
 that either the target is mentioned in the text or\nthat training data for
  every target is given. This paper considers the more challenging version\
 nof this task\, where targets are not always mentioned and no training dat
 a is available for the test targets.\nWe experiment with conditional LSTM 
 encoding\, which builds a representation of the\ntweet that is dependent o
 n the target\, and demonstrate that it outperforms encoding\nthe tweet and
  the target independently. Performance is improved further when the condit
 ional\n model is augmented with bidirectional encoding. We evaluate our ap
 proach on the\nSemEval 2016 Task 6 Twitter Stance Detection corpus achievi
 ng performance second\nbest only to a system trained on semi-automatically
  labelled tweets for the test target.\nWhen such weak supervision is added
 \, our approach achieves state-of-the-art results.\n\nMany current natural
  language processing applications for social media\nrely on representation
  learning and utilize pre-trained word embeddings.\nThere currently exist 
 several publicly-available\, pre-trained sets of word\nembeddings\, but th
 ey contain few or no emoji representations even as emoji\nusage in social 
 media has increased. In this paper we release\nemoji2vec\, pre-trained emb
 eddings for all Unicode emojis which are\nlearned from their description i
 n the Unicode emoji standard.\nThe resulting emoji embeddings can be readi
 ly used in downstream social natural language\nprocessing applications alo
 ngside word2vec.\nWe demonstrate\, for the downstream task of sentiment an
 alysis\, that\nemoji embeddings learned from short descriptions outperform
 s a skip-gram\nmodel trained on a large collection of tweets\, while avoid
 ing the need for\ncontexts in which emojis need to appear frequently in or
 der to estimate a representation.
LOCATION:FW26\, Computer Laboratory
END:VEVENT
END:VCALENDAR