Individuals, institutions, and innovation in the debates of the French Revolution

Social systems are characterized by heteroglossia: the coexistence, sharing, and competition of different patterns of speech. Heteroglossia makes linguistics and rhetoric (the reception, influence, and propagation of language within a community) (4, 5) core components in the quantitative study of culture. Tracking changes in speech patterns within a social body allows us to examine cultural evolution: the circulation, selection, and differential propagation of speech patterns in the group as a whole (rather than, say, tracking the ideas of a single individual). Patterns of heteroglossia demonstrate existing power relations, create new ones, and are a key method for the definition of both institutions and genres (6–8). Our methods here quantify a key aspect of cultural evolution: the extent to which one agent’s language patterns are used and copied by another (9, 10). To study the flow of rhetorical influence and attention in the NCA over time, we characterize how patterns of language use, uncovered by topic modeling, are propagated from speech to speech. We do so using Kullback–Leibler Divergence (KLD) (11): KLD, or “surprise,” measures the extent to which the expectations of an optimal learner, trained on one pattern, are violated by later patterns. Other work has demonstrated that surprise (in the Kullback–Leibler sense) is a cognitive as well as an information-theoretic quantity. It predicts what a subject will look at in a dynamically evolving visual scene (12) and can be used to map an individual’s higher level activities (detecting, for example, biographically significant transitions in a subject’s intellectual life) (3). Methodologically, this paper extends that work by considering surprise in relation to both past and future.

We use surprise to analyze a corpus of speeches by many different individuals. Surprise here measures both the deviation of one speech from the patterns of prior ones (novelty) and from patterns that appear in the future (transience). High surprise compared with the past indicates the topic mixture is new compared with previous speeches, hence the term “novelty”; high surprise compared with the future indicates that later speeches do not retain that pattern very strongly, hence the term “transience”. We provide a detailed introduction to these methods in SI Appendix.

Novelty and transience track a number of different effects. In addition to capturing intuitive notions of influence—a speaker with high novelty and low transience may have successfully shifted the terms of a discussion—they also track strategic effects (speakers who angle to speak early in a debate may have higher novelty) and external common causes (the first speech after a major event outside the chamber will have higher novelty). We test for a number of these latter effects (see SI Appendix), through day-level fixed effects and debate-level position analysis, to which our results below are robust. Our methods track the replications of patterns of speech and subject matter, rather than (for example) agreement: A speaker may introduce a new subject (high novelty) that is discussed by others (low transience), only to have his position contradicted or rejected.

The turbulent months at the beginning of the French Revolution led to a durable transformation in the very nature of European government (16). Our study of the elite debates at the center of these events focuses on pattern transmission as a proxy for the differential propagation of the ideas these patterns may communicate. This provides an alternate perspective to qualitative analyses, which usually focus on analyzing the logic of particular ideas and arguments over time (17, 18).

Our analysis reveals clear differences between how the left- and right-wing figures created and transmitted patterns of language use. Conservatives may indeed “stand athwart History, yelling Stop” (19); our results show they did so here less by redirecting conversation from one set of patterns to another than by maintaining the already established patterns of the conversations in which they participated. Indeed, the spatial metaphor of left and right is itself misleading: from the point of view of the debates themselves, the right wing appears as an inertial center, holding off conversation drift, while left-wing speakers produce a wide spectrum of innovations on a much larger periphery, only some of which survive. These roles become visible without reference to the ideas that are propagated: Robespierre emerges as a high-novelty, high-resonance speaker even before we consider the content of the speeches that made him an icon of revolutionary politics.

We also find an emergent distinction between orators and institutional players. Talented speakers like Robespierre, Pétion, and Maury could achieve power through rhetoric and debate, appealing to the crowds in the galleries as much as their colleagues. Yet not every ambitious delegate could take this road. Quite apart from a delegate’s rhetorical capacities or political stature, the physical venue itself was demanding: loud and (literally) resonant voices were required to be heard. While the NCA provided a place for rhetorical masters to thrive, in other words, it also ended up producing a second class of delegate who concentrated on committee work. Deliberating outside the parliament room, committees contributed specialized knowledge and abilities and became indispensable workhorses for the emerging government (13). The information-processing functions that committees took on distinguish them clearly from other forms of speech, while Fig. 2 shows how these functions emerged over time; once in place, committees were both strong sources of new patterns of speech and new sources of extrarhetorical power. Both individuals and institutions, in other words, mattered, playing distinct roles in the development of the parliament over time.

Many delegates welcomed these extraparliamentary functions: committees, wrote the assembly member Jacques Dinochau, “regulate the order of debate, classify questions, and maintain a continuity of principles, thus preventing an incoherence which might otherwise have menaced our decrees” (13). Yet the private nature of these committees was in dramatic contrast to the public debates in the hall itself—their increasing influence a testimony to the emergent distinction between the spectacle of democracy and the actual ways it functioned in a body and polity too large for direct participation. The dramatic and early appearance of a specialized information-processing role for committees foreshadows their appearance in modern democracies, where they emerge endogenously (20) to serve as essential information-management systems (21).

The power of these committees continued to grow. While early committees were devoted to technical matters such as monetary and fiscal policy, they were also instrumental to key developments such as the dismantling of feudal and religious privileges. When the revolution collapsed into chaos in 1793, it was committees, such as the famous “Committee of Public Safety” with Robespierre at its head, that effectively became the republican government.

The history of human culture is more than just the rise and fall of particular ideas. It is also the emergence of new information-processing mechanisms and media, and roles that individuals and institutions play in creating and propagating these ideas through time. In the language of biological evolution, we must understand not only the characteristics for which an environment selects but the strength of that selection over time and the shifting and heterogeneous nature of the transmission mechanisms.

By quantifying the flow of word patterns between times and speakers, we see not just the traces of a conscious battle of ideas but also—more important and more clearly—the contours of a new rhetorical space. This latter was neither intentionally produced, nor was it the exclusive property of any one political group. New word patterns from the left resonated, but so too did old ones from the right. Political actors did not just take different ideological positions, but played different roles in the propagation of patterns. And together, in both cooperation and competition, they invented new mechanisms for the collective management of information.

The AP is the definitive source for parliamentary transcripts of the Revolution, reconstructed from primary sources including transcripts, minutes, and newspaper reports. The French Revolution Digital Archive (https://frda.stanford.edu) is a digital version, with full-text speeches and text encoding initiative markup from the beginning of the NCA in July 1789 to its end in September 1791. After stop-word removal, speeches are represented as count vectors over a vocabulary of $\mathrm{10,000}$ most common words. The resulting corpus contains $\mathrm{4,765,773}$ words in $\mathrm{44,913}$ speeches. Each speech is matched to a speaker and may be tagged with a “role”: as a speech made by a delegate acting as the assembly president or on behalf of a committee (to introduce a report or comment on it). See SI Appendix for details on corpus preparation. All of the delegates to the Assembly, and all of the speakers in our data, were male.

We use latent Dirichlet allocation (LDA) (2) to quantify the semantic content of speeches. LDA categorizes speeches by identifying co-occurring word patterns (“topics”); any particular speech is a weighted combination of these topics. Many map remarkably well to the semantics of the speeches themselves. In addition to identifying subject matter content (topics in the ordinary sense, such as discussion of the church, finance, or corruption), they are also sensitive to rhetorical moves (e.g., logical arguments vs. appeals to patriotism). The SI Appendix presents interpretations of the LDA topics and a close reading of speeches from a particular debate. LDA allows for a variable number of topics, corresponding to an effective resolution; we use $K=100$ topics.

Having decomposed speeches into topics, we can then track the ways in which these topic combinations deviate from those in previous speeches (novelty) and are discarded (or not) by speeches that follow (transience).

J.H. acknowledges Research Experience for Undergraduates National Science Foundation Grant ACI-1358567 at the Santa Fe Institute.