Daejeon, July 27–31
The study of semantic change lies at the heart of intellectual and conceptual history. Since the foundational work of Reinhart Koselleck and Quentin Skinner in the second half of the twentieth century, scholars have emphasised that words and concepts are historically situated instruments whose meanings evolve alongside social, political, and cultural transformations (De Bolla 2023; Palti 2024). Recent advances in distributional semantics – particularly through vector-based word embeddings – have made it possible to approach these dynamics quantitatively, by modelling meaning as patterns of contextual use and tracing their evolution over time (Hamilton et al. 2016; Feng et al. 2022; Fedchenko et al. 2025).
This article builds on previous work to address a more focused question: can individual authors be semantically ahead of their time relative to their contemporaries, or are shifts in meaning primarily collective phenomena constrained by the norms of a given period? We approach this question by concentrating on concept-words – lexical units that are both structurally central within semantic space and subject to marked diachronic instability. Once identified, such words provide especially revealing observables of semantic change, as suggested by both computational approaches (Dubossarsky et al. 2017; Tang 2018) and interpretive traditions in lexical semantics (Rastier 2009; Koch 2016).
Using a large corpus of eighteenth-century French texts (1643–1800), divided into four distinct historical periods, we first identify a set of highly central and semantically unstable concepts through a composite concept score. We then analyse how these words evolve across periods at the level of the language, within specific textual genres, and finally within individual works. This multi-scale approach allows us to assess both large-scale semantic ruptures – particularly around the Revolutionary period – and more localised, potentially anticipatory usages associated with particular authors or texts.
All experiments are based on the ModERN corpus, which brings together a large collection of French works published between 1643 and 1800 ModERN (ERC Consolidator Grant 101043369) - Modelling Enlightenment: Reassembling Networks of Modernity through data-driven research. See https://modern.huma-num.fr/.
P1: 1643-1720 (84 million tokens)
P2: 1721-1750 (97 million tokens)
P3: 1751-1789 (277 million tokens)
P4: 1790-1800 (68 million tokens)
For each work, the corpus provides detailed bibliographic metadata, including author information, title, year of publication, and genre.
To identify words that are both structurally important and semantically dynamic, we implemented a multi-step selection protocol. This protocol deliberately adopts a sampling-based approach to keep computation times reasonable; in future work, both the scope of selection and the thresholds used will be expanded.
To begin, we restrict the vocabulary to words occurring at least 500 times in each of the four periods, excluding numbers and other function words. We then estimate distributional centrality by constructing a lexical co-occurrence network per period (Steyvers & Tenenbaum 2005). Two words are linked if they co-occur at least five times in the same sentence; only words with a degree ≥ 10 are retained, ensuring a minimal level of semantic connectivity.
Next, for each remaining word we compute semantic drift using contextualized embeddings. Up to 500 sentence-level occurrences per period are embedded and averaged to form a diachronic semantic prototype. Cosine distances between prototypes of successive periods provide measures of both mean and maximal semantic drift.
Finally, structural information (centrality) and dynamic information (semantic drift) are combined into a single concept score:
concept_score(w) = 0.4 × centrality_norm + 0.4 × drift_max_norm + 0.2 × drift_mean_norm
This score combines three normalised indicators: distributional centrality, maximum semantic variation between periods, and mean semantic change, weighted to prioritise words that are both structurally well connected and subject to strong diachronic rupture, while still accounting for more gradual forms of semantic evolution (See Figure 1).
Figure 1: Concept profile for ‘hommes’
After excluding grammatically non-relevant forms, we retained the 500 words with the highest concept scores as our core analytical set.
A key challenge in linking lexical-level semantic change to individual works of cultural production is avoiding anachronistic or genre-blind comparisons. Many concept-words are unevenly distributed across genres (e.g., amour in literary texts versus political treatises). To address this, we define for each concept-word a relevant analytical scope. Thus, for each of the 500 selected concept-words, we compute its distribution across genres and retain the four genres that together account for the largest share of occurrences, provided each represents at least 10% of the total. These dominant genres define the interpretive field in which the concept-word can be meaningfully analysed.
Within this scope, we generate contextualised embeddings of each concept-word at multiple analytical levels: as period-level prototypes capturing dominant historical meanings, as genre-specific representations tracing semantic variation across periods, and as work-level aggregates reflecting usage within individual texts To capture the linguistic specificities of each historical period in our embeddings, we fine tuned two transformer based models, CamemBERT https://huggingface.co/docs/transformers/en/model_doc/camembert and BERT https://huggingface.co/docs/transformers/en/model_doc/bert on each of the four sub corpora, yielding eight period specific models. An initial validation was conducted on sampled data (200,000 lines from each sub-corpus), with full corpus fine tuning planned to further refine semantic sensitivity.
For each concept-word, we construct a semantic prototype by averaging its contextual embeddings. This analysis is then extended at two additional scales: genre-level embeddings computed within each period, and work-level embeddings aggregating occurrences within individual texts For each genre g and pair of periods (p₁, p₂), we compute the cosine distance between the corresponding prototypes w(p₁, g) and w(p₂, g), derived from occurrences of w in works of genre g published in each period. Higher distances indicate stronger intra-genre semantic change. This analysis is restricted to words that already exhibit significant diachronic rupture at the global level, and for each word we retain the triplet (g, p₁, p₂) that maximizes this distance, identifying both the genres that drive semantic change and the periods in which it is most pronounced.
innovation(w, work) = cos(w(work), w(p_future)) − cos(w(work), w(p_now))
Positive values indicate usages that are closer to a future semantic configuration than to that of the word’s own period and are therefore interpreted as innovative or anticipatory.
The full analytic pipeline for tracking semantic change starts by computing concept scores for a sample of 200,000 lines per sub-corpus, capturing both the distributional centrality of each word and the extent of its semantic change across periods. This procedure isolates lexical units (concept-words) that are simultaneously central to the semantic structure of the corpus and subject to significant diachronic variation (see Figure 2).
Figure 2. Sample of the concept-words with the highest concept scores
Next, we scan the full corpus to extract all occurrences of the selected concept-words, computing a contextualized embedding for each instance and recording the associated metadata (work, author, date, and genre). These embeddings are then aggregated into semantic prototypes at the period, genre, and work levels. Figure 3 shows which concepts exhibit the largest semantic shifts between historical periods.
Figure 3. concept-words with semantic shifts between two historical periods
We then identify the most significant intra-genre semantic shifts across periods (Figure 4). Among the most salient cases, roi exhibits a particularly strong shift within historical writing between 1751-1789 and 1790-1800, with a cosine distance approaching 0.3. More generally, the results consistently point to a rupture between the post-1790 period and earlier decades, suggesting a broader reconfiguration of the conceptual lexicon at the end of the 18th century.
Figure 4. Concept-words with the most semantic drift between periods and genres
Finally, at the level of individual works, innovation scores reveal a striking and unexpected distributional pattern: among the 300 highest scores, 277 correspond to usages that are closer to the semantic configuration of 1790-1800, underscoring once again the Revolutionary decade as a major inflection point in the evolution of concepts (see Figure 5).
Figure 5. Sample of concept-words and texts with the highest innovation scores
However, these innovation scores tend to cluster around moderate positive values rather than exhibiting sharp outliers. And, while certain works display locally anticipatory usages of specific concept-words, no single author consistently emerges as semantically avant-garde across multiple words or genres. This finding is perhaps not merely negative: it indicates that semantic innovation is rarely concentrated in isolated texts, but instead unfolds through cumulative, distributed shifts across genres and discursive communities (cf. Underwood & Sellers 2016).
This study confirms the existence of a large-scale, coherent semantic rupture affecting a substantial portion of the French conceptual lexicon around the Revolutionary period. Words central to political, social, and moral discourse undergo marked reconfiguration after 1790, reflecting profound transformations in collective experience, institutional structures, and modes of public argument. From a methodological perspective, the combined use of distributional centrality and diachronic instability proves effective in isolating concept-words that are both structurally salient and historically volatile.
By contrast, the hypothesis of strong individual semantic avant-gardism receives only limited support. Rather than isolated breakthroughs attributable to singular authors, semantic change appears as a collective and progressive process, unfolding across genres and discursive fields. Innovation is distributed, incremental, and socially embedded; a finding that resonates with field-based models of intellectual production and with Koselleck’s emphasis on historically shared horizons of meaning (Bourdieu 1983; Koselleck 2004).
Beyond its substantive results, the article demonstrates the value of multi-scale semantic modelling for intellectual history. By linking period-level change, genre dynamics, and work-level usage within a unified framework, it becomes possible to move beyond anecdotal claims of lexical innovation and to situate individual texts within broader semantic trajectories. Future work will extend this approach by refining preprocessing, consolidating model fine-tuning on the full corpus, and identifying clusters of co-evolving concepts, opening the way to a more systemic mapping of conceptual transformation in the Enlightenment.