Emotion Analysis Using Latent Affective Folding and Embedding

Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 1–9, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Emotion Analysis Using Latent Affective Folding and Embedding Jerome R. Bellegarda Speech & Language Technologies Apple Inc. Cupertino, California 95014, USA jerome @ apple.com Abstract Though data-driven in nature, emotion analy- sis based on latent semantic analysis still relies on some measure of expert knowledge in or- der to isolate the emotional keywords or key- sets necessary to the construction of affective categories. This makes it vulnerable to any discrepancy between the ensuing taxonomy of affective states and the underlying domain of discourse. This paper proposes a more gen- eral strategy which leverages two distincts se- mantic levels, one that encapsulates the foun- dations of the domain considered, and one that specifically accounts for the overall affective fabric of the language. Exposing the emergent relationship between these two levels advan- tageously informs the emotion classification process. Empirical evidence suggests that this is a promising solution for automatic emotion detection in text. 1 Introduction The automatic detection of emotions in text is a necessary pre-processing step in many differ- ent fields touching on affective computing (Picard, 1997), such as natural language interfaces (Cosatto et al., 2003), e-learning environments (Ryan et al., 2000), educational or entertainment games (Pivec and Kearney, 2007), opinion mining and sentiment analysis (Pang and Lee, 2008), humor recognition (Mihalcea and Strapparava, 2006), and security in- formatics (Abbasi, 2007). In the latter case, for ex- ample, it can be used for monitoring levels of hate- ful or violent rhetoric (perhaps in multilingual set- tings). More generally, emotion detection is of great interest in human-computer interaction: if a system determines that a user is upset or annoyed, for in- stance, it could switch to a different mode of inter- action (Liscombe et al., 2005). And of course, it plays a critical role in the generation of expressive synthetic speech (Schr¨oder, 2006). Emphasis has traditionally been placed on the set of six “universal” emotions (Ekman, 1993): ANGER, DISGUST, FEAR, JOY, SADNESS, and SURPRISE (Alm et al., 2005; Liu et al., 2003; Subasic and Huet- tner, 2001). Emotion analysis is typically carried out using a simplified description of emotional states in a low-dimensional space, which normally comprises dimensions such as valence (positive/negative eva- lution), activation (stimulation of activity), and/or control (dominant/submissive power) (Mehrabian, 1995; Russell, 1980; Strapparava and Mihalcea, 2008). Classification proceeds based on an underly- ing emotional knowledge base, which strives to pro- vide adequate distinctions between different emo- tions. This affective information can either be built entirely upon manually selected vocabulary as in (Whissell, 1989), or derived automatically from data based on expert knowledge of the most relevant fea- tures that can be extracted from the input text (Alm et al., 2005). In both cases, the resulting system tends to rely, for the most part, on a few thousand annotated “emotional keywords,” the presence of which triggers the associated emotional …

Content truncated to 500 words