A social-event based approach to sentiment analysis of identities and behaviors in text

ABSTRACT

We describe a new methodology to infer sentiments held toward identities and behaviors from social events that we extract from a large corpus of newspaper text. Our approach draws on affect control theory, a mathematical model of how sentiment is encoded in social events and culturally shared views toward identities and behaviors. While most sentiment analysis approaches evaluate concepts on a single, evaluative dimension, our work extracts a three-dimensional sentiment “profile” for each concept. We can also infer when multiple sentiment profiles for a concept are likely to exist. We provide a case study of a large newspaper corpus on the Arab Spring, which helps to validate our approach.

KEYWORDS:

• Affect control theory
• Arab Spring
• Bayesian inference
• machine learning
• natural language processing
• sentiment analysis

Acknowledgments

We would like to thank Deena Abulfottouh, Michael Martin, and the other CASOS graduate students for fruitful discussions on this work. We would also especially like to thank Jonathan Morgan for all of his time spent in discussion with us and Jesse Hoey for clarifying discrepancies in the change equations across ACT publications. Support was provided, in part, by the Office of Naval Research (ONR) through a MURI N00014081186 on adversarial reasoning and the ONR through a MINERVA N000141310835 on state stability under the auspices of the ONR. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Department of Defense, the ONR, or the U.S. government.

Notes

¹ These values were computing using the INTERACT Java program (Heise, 2010a) with the Indiana 2002–2004 dictionary (Francis & Heise, 2006).

² http://www.indiana.edu/~socpsy/ACT/data.html

³ Values from the Indiana 2002–2004 sentiment dictionary (Francis & Heise, 2006).

⁴ The methodology involved in these surveys has evolved over time, and a thorough discussion can be found in Heise (2010b).

⁵ Note that while one could provide a regression model in which either the constant c₀ or c₁ is zero, in practice such an occurrence is unlikely and in any case it can be shown via simple algebra that the equations used here do not fit this case.

⁶ http://www.lexisnexis.com/hottopics/lnacademic/?verb = sf&sfi = AC00NBGenSrch&csi = 237924

⁷ Note that a news article may be indexed by LexisNexis as being relevant to more than one country; hence the values in Table 1 sum to a value greater than 600,000.

⁸ Lemmatization is a process by which words are normalized in a deterministic fashion to facilitate analysis. Lemmatization includes stemming (e.g., changing “walking” to “walk”) as well as other steps, like synonym replacement (e.g., replacing “better” with “good”). It is standard practice in the NLP literature to perform lemmatization before analysis.

⁹ For an introduction to plate notation, and to Bayesian modeling more generally, we refer the reader to the general texts from Gelman et al. (2013). We will here, out of necessity, assume some familiarity with such models.

¹⁰ For a formal proof, see Bromiley (2013).

¹¹ Results for all identities can be found at https://www.dropbox.com/s/oas8rvlzgw4o6dj/all_identities.png