Duyu Tang, Bing Qin, Ting Liu, Zhenghua Li HIT-SCIR 4/8/2019

Presentation on theme: "Duyu Tang, Bing Qin, Ting Liu, Zhenghua Li HIT-SCIR 4/8/2019"— Presentation transcript:

1 Duyu Tang, Bing Qin, Ting Liu, Zhenghua Li HIT-SCIR 4/8/2019
Learning Sentence Representation for Emotion Classification on Microblogs Duyu Tang, Bing Qin, Ting Liu, Zhenghua Li HIT-SCIR 4/8/2019

2 Outline Motivation Representation Learning for Emotion Classification
Experiment Results Conclusion and Future Work

3 Emotion Analysis on Microblog
Microblog, such as Weibo, has become a popular platform to share opinions about products or hot events. 微博现已成为大众接受信息和交流情感的主要平台，比如iphone5s上市，大家会从不同角度发表个人的褒贬观点。 当现实生活中发生了某些热门事件，大家也会将自己的情绪反映在一条条微博当中。 比如右侧哈尔滨雾霾的图片，哈尔滨人很愤怒，也可能很恐惧会对身体有损害。 我们需要自动的手段去从大量微博中发现和总结人们的情绪。

4 Task Emotion Classification Prior Work
Classify a text as Happy, Sad, Angry or Surprise Prior Work Supervised [Pang2002, Mishne2006] Distant Supervision [Go2009, Liu2012] Unsupervised [Turney2002, Ding2008] 本篇论文的任务是对一条中文微博进行情绪分类，具体地：将文本分类为高兴、悲伤、愤怒、恐惧；本文并没有考虑中性情感的情况 前人的工作大体可以分为3条思路：有指导、弱指导、无指导 有指导需要标注数据，人工设计特征或者学习特征训练分类器。 弱指导方法则考虑引入更多的带有噪声的数据，用来直接或间接提高分类性能。 无指导方法大多借助词典、搜索引擎、规则等资源判断文本的情感。 本文的工作可以归为弱指导学习，具体地为了解决bag-of-word表示能力不足的问题，我们利用大规模的自动收集的带有情感标签的数据来学习句子的表示，后续用作特征训练分类器。

5 Outline Motivation Representation Learning for Emotion Classification
Experiment Results Conclusion and Future Work

6 Representation Learning
Traditional The bag-of-word representation can not capture the complex linguistic phenomena. Manufacturing feature engines is time-consuming. Learning based method Deep Belief Networks Text with emoticons

7 Deep Belief Networks Architecture
受前人的工作启发，DBN可以有一定的重建和模式互补效果，已被应用于图像压缩和分类任务中

8 Data Collection Emotions from Weibo

9 Input Layer Top frequent unigrams Punctuation Lexicons Emotion word
Onomatopoeia word Function word

10 Outline Motivation Representation Learning for Emotion Classification
Experiment Results Conclusion and Future Work

11 Dataset Emoticon data Labeled data
20,000 weibo (balanced for each category) Labeled data Emotion Happy Sad Angry Surprise Size 548 837 905 567

12 Results

13 Results

14 Outline Motivation Representation Learning for Emotion Classification
Experiment Results Conclusion and Future Work

15 Conclusion Emotion classification on Weibo
Learning sentence representation on emoticon data via deep belief network For future Word representation for sentiment analysis Compositionality for sentiment analysis

16 Questions?

17 Back Up

18 Emoticon Selection

19 Detailed Params Lexicons Architecture