Multi-label maximum entropy model for social emotion classification over short text

Abstract

Social media provides an opportunity for many individuals to express their emotions online. Automatically classifying user emotions can help us understand the preferences of the general public, which has a number of useful applications, including sentiment retrieval and opinion summarization. Short text is prevalent on the Web, especially in tweets, questions, and news headlines. Most of the existing social emotion classification models focus on the detection of user emotions conveyed by long documents. In this paper, we introduce a multi-label maximum entropy (MME) model for user emotion classification over short text. MME generates rich features by modeling multiple emotion labels and valence scored by numerous users jointly. To improve the robustness of the method on varied-scale corpora, we further develop a co-training algorithm for MME and use the L-BFGS algorithm for the generalized MME model. Experiments on real-world short text collections validate the effectiveness of these methods on social emotion classification over sparse features. We also demonstrate the application of generated lexicons in identifying entities and behaviors that convey different social emotions.