Semantic Change and Semantic Stability: Variation is Key

Claire BowernACL 2019

TL;DR

This paper argues that understanding language change, particularly semantic change, requires focusing on variation within language systems, not just differences between them. It proposes an evolutionary framework that models language change through variation, selection, and transmission, offering a more nuanced approach than traditional methods and current word embedding techniques.

What Problem Does This Paper Address?

Studying language change, especially semantic change (how word meanings evolve), is challenging because historical records are often sparse and geographically limited. Traditional linguistic approaches often focus on identifying differences between languages at different points in time, treating language as a system that changes. However, this paper argues that this "what" of change is insufficient without understanding "how" and "why" it happens. Current computational methods, like word embeddings, often treat words as uniform entities and can conflate changes in word frequency with actual meaning shifts, failing to capture the underlying variation that drives change. The paper aims to bridge this gap by proposing a more robust framework for studying semantic change that accounts for variation and its role in evolutionary linguistic processes.

Methods

The paper advocates for an evolutionary framework to model language change, drawing parallels to Darwinian evolution. This framework centers on three key properties:

  • Variation: Language systems are inherently variable, with different ways of speaking (dialects, idiolects) and different social, psychological, and physiological factors influencing language use.
  • Selection: Not all variations are equally likely to spread. Biases in language transmission (acquisition, cognitive, social) act as selection pressures, favoring some variants over others.
  • Transmission: Language is passed down through generations and across individuals, with innovations spreading through social networks.

The paper critiques word embeddings as a primary tool for studying semantic change, highlighting their limitations: they assume uniform knowledge across all speakers, conflate frequency changes with meaning shifts, and struggle to distinguish between relative shifts in subsenses and genuine innovations.

For studying phylogenetic language splits (how languages diverge over long periods), the paper emphasizes the need for semantically stable items. It discusses how evolutionary approaches use "cognate evolution models" (like Brownian Motion) to model language splits based on stable meaning categories. The paper also introduces a method for evaluating the suitability of meaning classes for phylogenetic analysis by considering factors like loan rates, phylogenetic signal, and the amount of missing data.

Datasets

The paper does not introduce new datasets. It refers to existing work and datasets used in computational phylogenetics, specifically mentioning:

  • Pama-Nyungan languages: Used for illustrations and case studies in phylogenetic analysis.
  • Swadesh wordlists: A common resource for constructing cognate matrices in phylogenetic studies, though the paper critiques their reliance on historical convenience rather than principled stability.
  • Corpora used for word embeddings: The paper discusses the general use of large corpora (like Wikipedia) for training word embedding models, noting that these can skew towards encyclopedic information.

Key Insights and Ideas

  • Variation is the engine of change: The paper strongly emphasizes that understanding language change requires looking within a system for variation, not just comparing differences across systems or time points. Variation provides the "seeds" for change.
  • Evolutionary framework offers a unified view: Modeling language change through variation, selection, and transmission provides a coherent way to understand "what," "how," and "why" languages change.
  • Critique of word embeddings: Current word embedding models are too simplistic for nuanced semantic change studies because they treat words as monolithic and can mistake frequency shifts for meaning evolution.
  • Semantic stability is crucial for phylogenetics: To reconstruct language family trees, researchers need to identify meaning categories that have remained relatively stable over long periods, allowing for the tracking of cognate evolution.
  • Meaning categories for phylogenetics need careful selection: The suitability of a meaning class for phylogenetic inference depends on its stability, loan rates, and the amount of data available, not just its presence in traditional lists like Swadesh.

Key Results

The paper is primarily theoretical and methodological, not presenting quantitative experimental results. Its "results" are conceptual:

  • It provides a conceptual framework for understanding semantic change through an evolutionary lens.
  • It offers a critique of existing computational methods (word embeddings) for studying semantic change.
  • It outlines criteria for selecting appropriate meaning categories for phylogenetic analysis, illustrated by PCA plots showing how meaning classes cluster based on stability and data characteristics.

Limitations and Future Work

  • Limitations of word embeddings: The paper extensively details the limitations of word embeddings, particularly their inability to capture variation and distinguish frequency from meaning change.
  • Challenges in phylogenetic analysis: The paper acknowledges that selecting stable meaning classes for phylogenetics is difficult, and that items that are too stable might lead to underestimating splits, while those that change too fast or are heavily loaned can obscure phylogenetic signals.
  • Future Work: The paper implicitly suggests future work in developing more sophisticated computational models that can better incorporate variation and evolutionary principles. It also points to the need for more principled selection of vocabulary for phylogenetic studies, moving beyond convenience lists. The paper's conclusion emphasizes the ongoing need to study both stability and shift across millennia.

Summary updated

BibTeX
@inproceedings{bowern2019semantic,
  author = {Claire Bowern},
  title = {{Semantic Change and Semantic Stability: Variation is Key}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}

Loading related content...

Figure