Political scientists have long viewed the construction of narratives as one of the main tools available to political entrepreneurs to alter individual perceptions and behavior. Despite an explosion of interest in automated text analysis methods, however, no agreed-upon approach for principled empirical measurement and comparison of actual narratives exists. Indeed, widely-used unsupervised methods for the representation of text typically begin by discarding information such as event ordering and actor-actor relations that are fundamental to the concept of a narrative. This problem becomes particularly acute when scholars seek to draw parallels between elite and mass speech, such as open-ended survey responses, where differences in style tend to wash out similarities in meaning.

In this paper, I introduce a new method that overcomes these difficulties, providing a general approach for the discovery and representation of narratives that facilitates comparison across multiple textual domains. Beginning from a representation of text as subject-object relations, I estimate a latent network model that classifies entities into latent narrative roles. I demonstrate the value of this approach through three empirical applications, showing that (i) it is able to accurately recover well-recognized divergent patterns of partisan speech in the US context, (ii), it uncovers meaningful shifts in narrative content after an episode of censorship in Russia that are not discoverable via topic modelling, and (iii) it provides a valid means of assessing the closeness of survey respondents to different elite narratives in low-resource Caucasian languages.