Emotion Intensities in Tweets

Abstract

This paper examines the task of detecting intensity of emotion from text. We create the first datasets of tweets annotated for anger, fear, joy, and sadness intensities. We use a technique called best-worst scaling (BWS) that improves annotation consistency and obtains reliable fine-grained scores. We show that emotion-word hashtags often impact emotion intensity, usually conveying a more intense emotion. Finally, we create a benchmark regression system and conduct experiments to determine: which features are useful for detecting emotion intensity; and, the extent to which two emotions are similar in terms of how they manifest in language.

Lexicon based feature extraction for emotion text classification

“…Whilst GPELs offer useful knowledge about emotion-rich words, they are static and are likely to have poor coverage of the emotion vocabulary used in domains like Twitter. For emotion classification of tweets, Mohammad [17] and [22] demonstrated that DSEL based features offer significant gains over n-grams when compared to those of GPEL based features [32]. However feature extraction using DSELs has not been explored beyond binary and integer counts.…”

“…Different learning methodologies such as Point-wise Mutual Information (PMI) [22], Latent Dirichlet Allocation (LDA) in a semi-supervised setting [35] have been applied to learn DSEls. In addition, supervised LDA (sLDA) [36] offers a more accurate means to model emotion classes as topics and for lexicon generation.…”

“…Note that on the training documents of each data set, DSELs are learnt (see figure 1) using WED [21], PMI [22], sLDA [36] and by the proposed method (refer section 3.1) in order to extract the lexicon based emotion features mentioned in section 3.2.…”

“…• T EC, T EI, MEI, GEI and GEC features extracted using baseline DSELs generated using PMI [22], WED [21] and sLDA [36] (see section 3.2); 7 available for all our experiments. Performance is evaluated using the standard F-score metric in all emotion classification tasks.…”

“…Further, unfair may be associated with anger despite being more dominant in sadness related documents; the crisp binary memberships of words in GPELs do not allow to capture such fuzzy memberships of words to emotion classes, thereby making them limitedly effective for feature extraction. Accordingly, recent efforts in emotion analysis focused on learning domain specific lexicons [20,21] and also utilizing them for emotion feature extraction [17,22]. However the emotion features extracted were limited to simple emotion word counts in a document using the lexicon, which, while being simple, do not exploit the knowledge of the lexicon in its entirety.…”