TL;DR
This paper explores three automated methods for quantifying the prevalence of prefixing versus suffixing across thousands of languages. By comparing these approaches, the authors reveal that while automated methods generally agree on a slight preference for suffixing, they also highlight significant discrepancies that point to the challenges of accurately capturing linguistic features from raw text and grammars.
What Problem Does This Paper Address?
Linguists have long observed that suffixing is more common than prefixing across the world's languages. However, precisely quantifying this tendency and understanding its underlying causes requires comprehensive data on a vast number of languages. Gathering this data is a monumental task, as it traditionally involves manual analysis of grammatical descriptions, which is time-consuming and prone to subjective judgments. This paper addresses the need for more efficient and replicable methods to measure prefixation and suffixation, aiming to provide a broader dataset for linguistic research.
Methods
The authors propose and evaluate three distinct methods for measuring the suffix ratio (SR), a metric representing the proportion of suffixes to prefixes:
- Humans Read Grammars (HRG): This is the classic, albeit labor-intensive, approach. It involves human linguists extracting information from existing grammatical descriptions. The paper uses data from the World Atlas of Language Structures (WALS) chapter on prefixing vs. suffixing, which categorizes languages into broad classes based on an affixing index.
- Machines Read Grammars (MRG): This method leverages computational techniques to automatically scan large collections of digitized grammatical descriptions. It counts occurrences of the terms "prefix" and "suffix" within these texts. While crude, this approach offers significant advantages in speed and replicability.
- Machines Read Raw Text (MRT): This is the most automated method, using raw text data (specifically, New Testament translations) to infer affixation patterns. It employs a technique to measure the "prefixhood" and "suffixhood" of string segments based on their probability of appearing as final or non-final occurrences within words. This method aims to approximate affixation without requiring explicit morphological segmentation.
Datasets
- WALS Chapter 26A (Dryer, 2005): Used for the Humans Read Grammars (HRG) approach, this dataset contains information on 948 languages, categorized by their affixing index.
- Digitized Grammars Collection (Virk et al., 2020): A collection of over 10,000 raw text grammatical descriptions, used for the Machines Read Grammars (MRG) approach. This dataset covers 4,287 languages.
- Bible Corpus (McCarthy et al., 2020): A collection of New Testament translations in over 1,000 languages, used for the Machines Read Raw Text (MRT) approach. For comparability, languages with less than 7,000 verses were excluded, resulting in data for 1,030 languages.
Key Insights and Ideas
- Automated methods offer a scalable alternative: The paper demonstrates that computational approaches (MRG and MRT) can provide large-scale data on prefixation/suffixation, overcoming the bottleneck of manual annotation.
- General agreement on suffix preference: All three methods, despite their differences, generally indicate a slight preference for suffixing across languages, with suffix ratios typically ranging between 0.51 and 0.68.
- Discrepancies highlight methodological challenges: Significant differences between the methods, particularly between MRG/HRG and MRT, underscore the inherent difficulties in accurately measuring linguistic features from different types of data.
- MRG is sensitive to authorial style: The MRG method shows considerable variation even for different sources describing the same language, suggesting that the descriptive aims and styles of individual authors can influence the results.
- MRT's limitations in distinguishing affix types: The MRT method, while fast, cannot differentiate between derivational, inflectional, or fossilized affixation, and its accuracy in reflecting the "ideal token count" remains an open question.
Key Results
- Average Suffix Ratios:
- HRG: 0.67 (or 0.65 if languages with little affixation are excluded)
- MRG: 0.59 (average across all sources) or 0.61 (using the source with the most hits)
- MRT: 0.51 (indicating a minimal suffix preference)
- Correlation: All three measures show a correlation of r > 0.5, indicating a general agreement. The agreement on the polarity of SR (whether it's > 0.5) increases to around 0.7.
- MRG vs. MRT Agreement: The correlation between MRG and MRT is r = 0.50.
- HRG vs. MRT Agreement: The correlation between HRG and MRT is r = 0.54.
- HRG vs. MRG Agreement: The correlation between HRG and MRG is r = 0.67.
- Discrepancy in MRG: For the MRG approach, sources describing the same language can differ significantly, with an average absolute difference in SR of |SRMRG(s1) - SRMRG(s2)| ≈ 0.24.
Limitations and Future Work
- HRG Limitations: Excludes non-inflectional affixes, derivational affixes, and other morphological phenomena. It also relies on discretized labels rather than raw counts.
- MRG Limitations: Highly sensitive to the scope and style of individual grammar descriptions. It cannot discern the type or context of affixes and may be influenced by authorial choices in terminology.
- MRT Limitations: Reads a "noisy reflection" of prefixation/suffixation and cannot distinguish between different types of affixation. The "prolongation" effect (segments being too long) is noted, though hypothesized to affect prefix and suffix extraction uniformly. The paper suggests that more complex measures than simple string counting are needed for MRT.
- Future Work: The authors suggest that more research is needed to understand the sources of discrepancy in MRG and to develop more robust MRT measures. They also highlight the need to investigate the influence of orthography and writing systems on these measures. The resulting database of 4,437 languages is made available for future research.
Summary updated