Quote Erat Detectendum. A Computational Analysis of Reported Speech in News Media

Quotes are central to journalistic practice. Journalists build the “webs of facticity” at the basis of their news stories by weaving into them externally sourced information (Tuchman 1978). Political journalism is no exception. To establish the legitimacy of their narratives, political journalists anchor reporting into relayed facts and discourse coming from pertinent sources within the political landscape. Through repeated quotation, more than mere mentioning, political actors are accredited as relevant and legitimate voices with respect to the context under consideration (Carlson 2009; Fishman 1980; Sigal 1986). Despite this relevance, quotes remain underexplored in large-scale empirical social science research. This is mostly due to challenges in automatic quote extraction, particularly in nonstandard texts. To address this, we introduce Quote Erat Detectendum (QED), an open-source NLP pipeline for detecting quotes in news texts. Building on existing work, we develop a transparent and robust method that identifies direct, indirect, and mixed quotes, before attributing them to identifiable speakers. The results improve on classic techniques, and they clearly outperform more recent, generative LLM-based approaches. We then illustrate QED’s analytical potential. From a corpus of nearly half a million French political news articles, QED extracts over 1.8 million quotes by identifiable individuals. Three applications are presented: (1) comparing quotes to mentions, we show the evolving status ascribed to political parties; (2) examining cues, i.e. reporting verbs, we analyze journalistic framing of political groups; (3) analyzing the content of quoted statements, we address how media ascribe issue ownership to parties. By operationalizing classic social science concepts through large-scale quote analysis, this work opens new avenues for the sociology of media and political communication.