From qualifiers to quantifiers: semantic shift at the paradigm level

Quentin FeltgenACL 2022

TL;DR

This paper introduces a novel computational method to analyze semantic shifts within "schematic constructions" – linguistic structures with flexible slots. By examining the interactions and frequency dynamics of these slots (fillers), the authors reveal a significant paradigmatic shift in French quantifiers, demonstrating how competition can occur within a construction, not just between distinct constructions.

What Problem Does This Paper Address?

Language change is often viewed as a competition between an established linguistic form and a new, rising one. However, this perspective struggles to explain changes within complex linguistic units like "schematic constructions," which have open slots filled by various words (fillers) that form a "paradigm." The paper addresses the challenge of understanding how these internal dynamics within a paradigm contribute to language change, particularly when the change isn't a simple replacement but a reorganization of how fillers interact and compete for semantic territory. Traditional methods like word embeddings, which rely on simple collocations, are insufficient for capturing these nuanced, system-level shifts.

Methods

The core of the paper's contribution lies in its novel computational approach to analyze "paradigmatic level" semantic shifts. The method involves:

  1. Frequency Profiling: Tracking the frequency of individual fillers within a construction over time, smoothed using a moving average to capture trends.
  2. Correlation Analysis: Building correlation matrices between the frequency time series of fillers over specific time windows (10 decades). This captures how fillers' frequencies rise and fall together (correlation) or in opposition (anti-correlation).
  3. Derivative Correlation: Complementing the above by correlating the derivatives of the frequency time series. This helps distinguish genuine interactions from spurious correlations driven by a common external factor (like the overall construction's frequency).
  4. Network Construction: Filtering the correlation matrices to identify significant interactions (both positive and negative correlations) and visualizing these as networks. Edges represent correlations, with red edges pointing towards declining forms, indicating competition.
  5. Cluster Identification: Analyzing these networks to identify distinct clusters of fillers that interact more strongly with each other than with fillers in other clusters.

This approach allows for a dynamic, system-level view of how fillers within a construction compete and reorganize over time.

Datasets

The study utilizes the Frantext database for the French language. The analysis is restricted to the period 1321-2020, ensuring coverage across centuries. The corpus contains approximately 300 million words, from which 50,000 occurrences of the French quantifier construction ("un Q(N) de N") were identified.

Key Insights and Ideas

  • Competition Within Paradigms: The most significant insight is that language change, specifically semantic shift, can occur within a construction's paradigm. Fillers can compete with each other, leading to a reorganization of semantic space without necessarily replacing the entire construction.
  • Network Dynamics Reveal Shifts: The network analysis effectively visualizes these internal competitions. The emergence of anti-correlations between clusters of fillers is a strong indicator of a paradigmatic shift.
  • Entrenched vs. Innovative Clusters: The study identifies two distinct clusters of French quantifiers: an "entrenched" cluster (e.g., foule, multitude, infinité) and an "innovative" cluster (e.g., dizaine, millier). This division highlights a shift from general quantity descriptors to more specific or nuanced ways of expressing quantity.
  • Coexistence, Not Eviction: The competition doesn't always lead to the complete elimination of one cluster. In this case, the entrenched cluster remains dominant in frequency, suggesting a coexistence rather than a total replacement.
  • Frequency Alone is Insufficient: The paper demonstrates that analyzing frequency profiles alone is not enough to understand constructional change. While an S-curve indicates entrenchment, the subsequent dynamics are driven by individual fillers, and network analysis is crucial for understanding these internal shifts.

Key Results

The study reveals a major paradigmatic shift in the French quantifier construction, particularly evident from the late 18th century onwards.

  • Emergence of Clusters: By the 1791-1800 period, two competing clusters of quantifiers become apparent.
  • Competition Intensifies: Numerous anti-correlations emerge between these clusters in later periods (e.g., 1831-1930, 1851-1950), indicating a decline in the entrenched cluster relative to the innovative one.
  • Qualitative Differences: The entrenched cluster tends to use more generic arguments (e.g., gens, choses), while the innovative cluster shows a split: some quantifiers allow for precise counting (e.g., dizaine), while others emphasize excess or a collection of items (e.g., nuée, profusion, tas).
  • "Un tas de" as a Key Innovator: The quantifier un tas de emerges as the most frequent in the innovative cluster and surpasses une foule de in frequency, becoming the second most frequent overall after un nombre de.

Limitations and Future Work

  • Qualitative Interpretation: While the method is quantitative, the linguistic interpretation of the identified clusters is largely qualitative, based on collexeme analysis (examining frequent collocates). The authors acknowledge that more sophisticated quantitative tools (like collexeme analysis or word embeddings) could be applied for a deeper semantic understanding.
  • Manual Cleaning: The initial data processing involved manual cleaning of quantifier occurrences, which could be further automated.
  • Generalizability: The paper presents a case study on French quantifiers. Future work could explore the applicability of this method to other schematic constructions and languages to test its generalizability.
  • Refinement of Methods: The authors suggest that the analysis could be refined with a wider range of methods, including clustering algorithms for more automated identification of patterns. They also propose further investigation into the semantic logic underlying the different partitioning of nouns by the competing clusters.

Summary updated

BibTeX
@inproceedings{feltgen2022from,
  author = {Quentin Feltgen},
  title = {{From qualifiers to quantifiers: semantic shift at the paradigm level}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure