TL;DR
This paper explores objective methods for measuring linguistic distance between South African languages, using both text-based (Levenshtein distance) and acoustic (mean pitch) approaches. The findings show that these statistical methods can effectively group languages based on known linguistic families, offering a valuable tool for language resource development and HLT system design.
What Problem Does This Paper Address?
Accurately assessing the similarity or dissimilarity between languages is crucial for various applications, including language learning, resource allocation, and the development of Human Language Technologies (HLT). While subjective human judgments have traditionally been used, they are prone to bias and inconsistency. Objective metrics are needed to quantify linguistic distance, especially for languages with complex variations, dialects, and limited parallel data. This paper specifically addresses the challenge of measuring linguistic distance among South African languages, a region with a rich linguistic landscape but less explored in this context.
Methods
The study employs two primary methods for measuring linguistic distance:
- Levenshtein Distance (Text-Based): This method quantifies the minimum number of single-character edits (insertions, deletions, or substitutions) required to change one word into another. The authors applied this to orthographic transcriptions of parallel word lists. This is a standard technique in string comparison, adapted here for linguistic analysis.
- Mean Pitch Difference (Acoustic-Based): This approach utilizes the fundamental frequency (F0) of speech. Pitch contours were extracted from speech recordings using the Praat software. The mean fundamental frequency for each language was then calculated, and the differences between these means were used to derive a distance matrix. This method is text-independent and leverages prosodic features.
Both distance matrices were then visualized and analyzed using:
- Multidimensional Scaling (MDS): A statistical technique used in data visualization to represent high-dimensional data in a low-dimensional space (typically 2D), allowing for visual assessment of language proximities.
- Hierarchical Clustering (Dendrograms): A method that groups similar objects into a tree-like structure, illustrating the hierarchical relationships and clustering of languages.
Datasets
The study used data from six selected South African languages: English, Afrikaans, Zulu, Xhosa, Northern Sotho (Sepedi), and Tswana.
- Text Data: Parallel word transcriptions were manually collected from multilingual dictionaries and online resources. Initially, 200 common English words were chosen, and 144 words with unique translations across the six languages were selected for the Levenshtein distance calculation.
- Acoustic Data: Pitch contours were extracted from approximately 5000 speech files per language. Each language had around 200 different speakers (25 recordings per speaker), with an equal distribution of males and females aged 18-65. This acoustic dataset appears to be newly compiled for this research.
Key Insights and Ideas
- Complementary Nature of Text and Acoustic Data: Both text-based (Levenshtein) and acoustic (pitch) methods, when analyzed with MDS and clustering, yielded similar groupings of languages, reinforcing the validity of each approach.
- Refined Grouping by Pitch: The pitch-based dendrogram provided a slightly more refined grouping, more accurately separating the Germanic subgroup from the Bantu languages compared to the Levenshtein-based dendrogram. This suggests that acoustic features can capture subtle linguistic distinctions.
- Objective Metrics for Language Relationships: The study demonstrates that statistical methods can provide objective and useful measures of language similarity, aligning well with established linguistic family trees and human perceptual judgments.
- Practical Implications for HLT and Resource Development: The ability to objectively measure linguistic distance is invaluable for organizations needing to prioritize language resources or develop HLT systems for diverse linguistic communities, especially in resource-constrained environments.
Key Results
- Both Levenshtein distance and mean pitch difference analyses, when visualized with MDS and dendrograms, successfully grouped the six South African languages according to their known linguistic families (Germanic, Nguni, Sotho).
- The dendrograms showed clear clusters for the Germanic, Nguni, and Sotho language groups.
- The pitch-based dendrogram offered a more accurate separation of the Germanic subgroup from the Bantu languages compared to the Levenshtein-based dendrogram.
- The study found a high correlation between the Levenshtein distances and perceptual distances in previous work (Gooskens & Heeringa, 2004), suggesting the reliability of this text-based method.
Limitations and Future Work
- Levenshtein Distance Limitations: The paper notes that Levenshtein distance might be less discriminating for languages with different writing styles (e.g., Chinese vs. English). The current study used relatively small word lists (144 words), which might limit its discriminatory power for finer distinctions.
- Acoustic Data Processing: The acoustic approach, while using natural language data, is computationally taxing, especially when dealing with large datasets.
- Future Work Suggestions:
- Investigate Levenshtein distances using larger text collections, such as parallel corpora, rather than just word lists.
- Explore other acoustic parameters beyond mean pitch, potentially in combination with pitch, for more robust language similarity measures.
- Compare these distance measures with other criteria for language similarity, such as historical separation or mutual intelligibility, in a rigorous manner.
- Refine these methods using phonetic transcriptions or other acoustic measurements for finer distinctions, particularly for dialectal variations.
Summary updated