Agent-based modelling, sound change, and metaphony in Southern Italian varieties of Italo-Romance.

Lilian von Bressensdorf, Pia Greca, Jonathan HarringtonInterspeech 2025

TL;DR

This paper uses an agent-based computational model to simulate how sound changes spread between dialects. They found that contact between a more "innovative" dialect and a more "conservative" one leads to an asymmetric shift, where the conservative speakers adopt changes from the innovative ones, supporting a cognitive feedback model of language evolution.

What Problem Does This Paper Address?

How do sound changes, which are often subtle phonetic variations, become established linguistic features that alter word meanings (phonologization)? This paper investigates this by focusing on "metaphony," a sound change in Southern Italian dialects where cues to grammatical information (like singular vs. plural) shift from a suffix to the stem vowel. Specifically, they explore the hypothesis that when dialects with different stages of this sound change (one more advanced, one less) interact, the less advanced dialect will adopt the changes from the more advanced one. This addresses the fundamental question of how language change propagates through communities, particularly the role of dialect contact and cognitive mechanisms.

Methods

The core of this research is an agent-based computational model called the Interactive-Phonetic Agent-Based Model of Sound Change (IP model). This model simulates language use and change through a population of "agents" (simulated speakers).

Here's a breakdown of the key methodological components:

  • Agent Initialization: Real speech data from speakers of two Southern Italian dialects (Mormanno - MM, and West coast - West) were used to initialize the agents. These dialects represent different stages of metaphony.
  • Agent Interaction: Agents interact in simulated conversations. During an interaction, one agent (talker) produces a word, and another agent (listener) perceives it. The listener's internal representation of the word's acoustic properties is updated based on the perceived token. This update is probabilistic and depends on the listener's existing memory of acoustic tokens for that word.
  • Morpho-phonological Class Derivation: The model automatically derives "morpho-phonological (MP) classes" for vowels within words. These classes represent how speakers categorize acoustic variations based on their grammatical function. This is done using Gaussian Mixture Modeling (GMM) and Non-negative Matrix Factorization (NMF), allowing the model to learn these categories from the data rather than pre-defining them.
  • Quantification of Sound Change:
    • Diphthongization: The extent of diphthongization in stem vowels was measured using the third moment (skewness) of smoothed F1 (first formant) trajectories. This captures how much the vowel's acoustic profile deviates from a simple monophthong.
  • Phonologization: The degree to which phonetic variation becomes linguistically relevant for distinguishing meaning was quantified using a "phonologization score" (pprop). This score measures how well words with different suffixes are grouped into distinct phonological classes.

Datasets

The study utilized existing speech data from two Southern Italian dialects:

  • Mormanno (MM) dialect: Data from 28 speakers (age 22-88, 15 female).
  • West coast (West) dialect: Data from 26 speakers (age 27-92, 10 female).

The materials consisted of 45 words, including 21 lexical morphemes (verbs, adjectives, nouns) with primary stressed vowels that could be monophthongal or diphthongal. These words were produced in picture-naming tasks, resulting in over 4000 word tokens. This data was used to initialize the agents in the computational model. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Asymmetric Dialect Contact: The most striking insight is the confirmation of an asymmetric shift in sound change. When a more "innovative" dialect (West, with more advanced metaphony) interacts with a "conservative" dialect (MM, with less advanced metaphony), the conservative dialect adopts the innovative features, but not vice-versa. This suggests that the directionality of sound change is influenced by the relative "progressiveness" of interacting linguistic communities.
  • Cognitive Feedback Loop: The model's success supports the idea that sound change arises from a feedback loop between speech production and perception. Agents learn from the speech they hear, and this learning can subtly alter their own speech production over time, leading to systematic shifts.
  • Emergence of Phonological Contrast: The study demonstrates how subtle phonetic variations (like diphthongization) can become phonologically relevant, leading to the emergence of distinct categories that carry grammatical information. This is crucial for understanding how new phonemic contrasts can arise.
  • Individual Variation Matters: The model accounts for individual differences in how speakers categorize acoustic cues, reflecting real-world linguistic variation and its role in driving change.

Key Results

  • Increased Diphthongization in MM: After the agent-based simulation, the MM agents showed a significant increase in diphthongization of stem vowels in high vowel suffix contexts, moving towards the characteristics of the West dialect. The West dialect agents showed no significant change in diphthongization.
  • Increased Phonologization in MM: The phonologization score (pprop) for the MM dialect significantly increased from the baseline to the post-run simulation, indicating that the acoustic variations became more consistently associated with grammatical distinctions. The West dialect already had a higher baseline phonologization score, which did not change significantly.
  • Statistical Significance: Statistical models confirmed significant effects of dialect and state (baseline vs. post-run) on diphthongization and phonologization, with a significant interaction between dialect and state, particularly for MM.

Limitations and Future Work

  • Limited Scope of Acoustic Features: The study focused on F1 trajectories and did not incorporate the acoustic properties of the suffix vowels themselves. This prevents a full simulation of "cue-trading," where listeners shift attention between different acoustic cues.
  • Cognitive Mechanisms of Cue-Reweighting: The precise cognitive mechanisms by which cue re-weighting in perception influences speech production remain not fully understood, posing a challenge for more complex modeling.
  • Inevitability of Sound Change: The authors acknowledge that sound change is not inevitable in their model. Factors like weak or variably distributed coarticulatory cues, or the lack of sufficient acoustic information to distinguish allomorphs, can lead to stability rather than change.
  • Future Directions:
    • Applying the model to standard Italian to test the stability of linguistic features without metaphony.
  • Investigating the connection between morpho-phonological categorization and the broader phonologization process, and whether both need to be involved in interaction.
  • Incorporating more detailed acoustic features and cognitive mechanisms for cue-trading.

Summary updated

BibTeX
@inproceedings{bressensdorf2025agent,
  author = {Lilian von Bressensdorf and Pia Greca and Jonathan Harrington},
  title = {{Agent-based modelling, sound change, and metaphony in Southern Italian varieties of Italo-Romance.}},
  booktitle = {Interspeech},
  year = {2025},
}
Figure