Learning Emotion Indicators from Tweets: Hashtags, Hashtag Patterns, and Phrases

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1203–1209, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Learning Emotion Indicators from Tweets: Hashtags, Hashtag Patterns, and Phrases Ashequl Qadir School of Computing University of Utah Salt Lake City, UT 84112, USA asheq@cs.utah.edu Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112, USA riloff@cs.utah.edu Abstract We present a weakly supervised approach for learning hashtags, hashtag patterns, and phrases associated with five emotions: AFFEC- TION, ANGER/RAGE, FEAR/ANXIETY, JOY, and SADNESS/DISAPPOINTMENT. Starting with seed hashtags to label an initial set of tweets, we train emotion classifiers and use them to learn new emotion hashtags and hash- tag patterns. This process then repeats in a bootstrapping framework. Emotion phrases are also extracted from the learned hashtags and used to create phrase-based emotion clas- sifiers. We show that the learned set of emo- tion indicators yields a substantial improve- ment in F-scores, ranging from +%5 to +%18 over baseline classifiers. 1 Introduction Identifying emotions in social media text can be benefi- cial for many applications, for example to help compa- nies understand how people feel about their products, to assist governments in recognizing growing anger or fear associated with an event, or to help media outlets understand people’s emotional response toward contro- versial issues or international affairs. On the Twitter micro-blogging platform, people often use hashtags to express an emotional state (e.g., #happyasalways, #an- gryattheworld). While some hashtags consist of a sin- gle word (e.g., #angry), many hashtags include multi- ple words and creative spellings (e.g., #cantwait4tmrw, #Youredabest), which can not be easily recognized us- ing sentiment or emotion lexicons. Our research learns three types of emotion in- dicators for tweets: hashtags, hashtag patterns, and phrases for one of five emotions: AFFEC- TION, ANGER/RAGE, FEAR/ANXIETY, JOY, or SAD- NESS/DISAPPOINTMENT. We present a bootstrapping framework for learning emotion hashtags and extend the framework to also learn more general hashtag pat- terns. We then harvest emotion phrases from the hash- tags and hashtag patterns for contextual emotion clas- sification. First, we make the observation that emotion hashtags often share a common prefix. For example, #angry- attheworld and #angryatlife both have the prefix “an- gry at”, which suggests the emotion ANGER. Conse- quently, we generalize beyond specific hashtags to cre- ate hashtag patterns that will match all hashtags with the same prefix, such as the pattern #angryat* which will match both #angryattheworld and #angryatlife. A key challenge is that a seemingly strong emotion word or phrase can have a different meaning depending upon the following words. For example, #angry* may seem like an obvious pattern to identify ANGER tweets. But #angrybirds is a popular hashtag that refers to a game, not the writer’s emotion. Similarly, “love you” usually expresses AFFECTION when it is followed by a person (e.g., #loveyoumom). But it may express JOY in other contexts (e.g., #loveyoulife). We use probabil- ity estimates to determine which hashtag patterns are reliable indicators for an emotion. Our …

Content truncated to 500 words