Using large corpora to explore the framing of concepts

Psychologists and Social Scientists have long observed that the way in which a question or problem is presented to people can impact their attitudes and decisions. Framing is a widely discussed instance of this phenomenon: The choice of words and metaphors in talking about a given issue can affect hearers’ interpretations and biases, making some actions or strategies appear more plausible than others (1). The framing of issues is thus of great practical importance to those with an interest in steering the course of decision processes or justifying actions to a wider audience. For instance, different ways of dealing with drug abuse can become salient depending on whether the problem is presented as one of social policy or of law enforcement (2).
Consequently, the framing of issues is of interest to decision makers and is a prominent topic of research in Political Science, Sociology, Economics, Psychology and related fields. Researchers interested in these questions tend to rely on controlled experiments and manually annotated document collections. The recent explosion in the amount of textual data available electronically provides an opportunity for investigating framing on a large scale and tracking how it influences, and is influenced by, decisions made by governments and businesses. However, analyzing this massive amount of data requires tools to facilitate a fast and efficient analysis of the available data sets. To meet this demand, researchers have turned to machine-learning methods from computational linguistics for applications in topic analysis and opinion classification.
Most of these methods rely in some way on word co-occurrence patterns. A prominent example is Latent Semantic Analysis (LSA) (3), which has been applied to a wide range of tasks, including word sense discrimination, text summarization, and identifying semantic change (4-6). Here we present an LSA-based approach designed to observe and quantify variation in the framing of concepts across time or speaker/author populations. We illustrate this methodology using two examples of framing in political debates in the US senate: the rise and time-course of the framing of terror as a military struggle following the events of September 11th, 2001, and the different framings of abortion by democrats and republicans.

References
1. Tversky, A. & Kahneman, D. (1981) The Framing of decisions and the psychology of choice. Science 211 (4481): 453–458.
2. Mark, E. (2003) War on Drugs: Legislation in the 108th Congress and Related Development, Congressional Research Service Report IB10113 ( http://opencrs.com/document/IB10113/ ).
3. Landauer, T. K., & Dumais, S. T. (1997). A solution to Plato's problem: The Latent Semantic Analysis theory of the acquisition, induction, and representation of knowledge. Psychological Review, 104, 211-240.
4. Schütze, H. (1998) Automatic word sense discrimination. Computational Linguistics 24(1):97-124.
5. Marcu, D (2003) Automatic Abstracting, In Darke, M. A. (ed.) Encyclopedia of Library and Information, 245-256.
6. Sagi, E., Kaufmann, S., & Clark, B. (2009). Semantic Density Analysis: Comparing Word Meaning across Time and Phonetic Space. In Basili R., and Pennacchiotti M. (Eds.), Proceedings of the EACL 2009 Workshop on GEMS: Geometrical Models of Natural Language Semantics. Athens, Greece.