TL;DR
This paper reviews the intersection of computational and human approaches to understanding sound change in language. It highlights how computational models excel at analyzing historical sound changes from etymological data, while human perception studies focus on ongoing changes using recorded speech. The authors argue for a unified perspective that combines these methods to gain a more comprehensive understanding of linguistic evolution.
What Problem Does This Paper Address?
The study of sound change in language has traditionally been approached from two distinct perspectives: computational modeling and human perception. Computational methods typically analyze historical sound changes using etymological datasets, relying on phonetic transcriptions. In contrast, human perception studies often examine ongoing sound changes by analyzing recorded speech corpora and observing how listeners perceive these shifts. This paper identifies a gap in research that bridges these two approaches, arguing that a combined perspective is crucial for a more complete understanding of how sounds evolve over time. The authors aim to fill this gap by reviewing the methods and tasks of both approaches and proposing a unified view.
Methods
The paper categorizes computational approaches into several key areas:
- Sound Change Detection: This involves using diachronic distributional embeddings to track historical sound changes. The core idea is that phonemes appearing in similar phonetic contexts are likely related, and changes can be observed by tracking the distance between phoneme embeddings over time.
- Phonetic Alignment Between Cognate Words: This frames the translation of phoneme sequences between cognate words (words with a common origin) as a machine translation task. Both statistical and neural machine translation models are employed.
- Markedness of Phonemes: This approach models sound change by estimating phoneme frequencies over time, aiming to identify patterns related to phonological markedness (common vs. less common sounds).
- Sound Convergence: This focuses on ongoing sound changes where speakers adjust their speech to align acoustically and phonetically with others. This can occur in both interactive and non-interactive settings.
Human perception approaches are primarily discussed through:
- Perceptual Similarity Task: This listener-driven model assesses how phonetic changes are perceived. Listeners are asked to determine which of two recordings of a speaker sounds closer to a target speech, often after being exposed to a model speaker's speech.
- Sound Convergence (Human Perception): This uses the perceptual similarity task to study how native speakers converge their speech towards non-native speakers, exploring factors like perceived accent.
The paper also proposes a Unified Perspective that combines computational and human methods, suggesting that computational tools can refine human perception by revealing subtle acoustic and phonetic changes, while human insights can improve computational models.
Datasets
The paper discusses the use of various datasets:
- Etymological Datasets: These datasets contain phonetic transcriptions reflecting historical sound changes and are used for computational models. Examples mentioned include those derived from Wiktionary (EtymDB) and WordNet (EtymWordNet, CogNet).
- Recording Corpora: These are databases of recorded speech used for human perception studies to observe ongoing sound changes.
- Artificial Setups: One approach to markedness was evaluated on an artificial setup.
The paper notes that while etymological datasets are crucial for computational models, their quality can be unclear due to automatic generation processes and loose definitions of cognacy.
Key Insights and Ideas
- Complementary Strengths: Computational and human approaches to sound change have distinct strengths. Computational methods are powerful for analyzing historical data and identifying patterns across vast datasets, while human perception is essential for understanding how real-time changes are experienced and adopted by speakers.
- The Value of a Unified View: Combining computational analysis with human perception can lead to a more nuanced understanding of sound change. Computational tools can reveal subtle phonetic shifts that listeners might miss, and listener feedback can inform and refine computational models.
- Interconnectedness of Linguistic Change: Sound change is not an isolated phenomenon. It is deeply intertwined with syntactic and semantic changes, as seen in processes like grammaticalization. A comprehensive understanding requires modeling these multi-faceted changes.
- Challenges in Etymological Data: The quality and consistency of etymological datasets are critical for computational linguistics tasks like phylogenetic inference. Issues like borrowing and the inherent difficulty of phonetic alignment pose challenges.
Key Results
The paper is primarily a review and a call for a unified approach, rather than a presentation of new experimental results. However, it highlights findings from reviewed studies:
- Computational Detection of Sound Change: Studies have successfully used phoneme embeddings to detect historical sound changes, such as the shift from voiceless to voiced plosives in Danish.
- Phonetic Alignment Accuracy: Machine translation models have shown promise in aligning phonemes between cognate words, with varying performance depending on the direction of translation and dataset size.
- Markedness Recognition: Probabilistic models have demonstrated the ability to recognize phonological markedness by analyzing phoneme frequencies over time.
- Human Sound Convergence: Research indicates that native speakers can exhibit sound convergence towards non-native speakers, even in non-interactive settings, and this convergence can be influenced by the perceived foreignness of the speech.
- Phylogenetic Tree Reconstruction: Sound-based phylogenetic trees have been shown to sometimes underperform cognate-based trees, potentially due to the influence of language borrowing.
Limitations and Future Work
- Limitations of Computational Approaches: Etymological datasets often lack acoustic features, limiting the depth of analysis. Computational models may also struggle with one-to-many phoneme alignments.
- Limitations of Human Perception: Listeners can misperceive acoustic and phonetic features, leading to inaccurate judgments of sound changes.
- Future Work: The authors strongly advocate for comparative studies that investigate the influence of historical sound changes on ongoing changes by analyzing both etymological datasets and recording corpora. They also emphasize the need for computational and human approaches to model multi-faceted changes that involve phonetics, syntax, and semantics simultaneously. The quality assessment of etymological datasets is also identified as an important area for future research.
Summary updated