论文

详细信息

发表期刊/会议

Research at the University of Copenhagen (University of Copenhagen)

发表日期

2017-01-01

发表年份

2017

摘要

NLP tasks are often limited by scarcity of<br/>manually annotated data. In social media<br/>sentiment analysis and related tasks,<br/>researchers have therefore used binarized<br/>emoticons and specific hashtags as forms<br/>of distant supervision. Our paper shows<br/>that by extending the distant supervision<br/>to a more diverse set of noisy labels, the<br/>models can learn richer representations.<br/>Through emoji prediction on a dataset of<br/>1246 million tweets containing one of 64<br/>common emojis we obtain state-of-theart<br/>performance on 8 benchmark datasets<br/>within emotion, sentiment and sarcasm detection<br/>using a single pretrained model.<br/>Our analyses confirm that the diversity of<br/>our emotional labels yield a performance<br/>improvement over previous distant supervision<br/>approaches.