TL;DR
This paper tackles the long-standing linguistic puzzle of how foreign words are adapted into Japanese, using a computational approach to analyze the entire process. By building a finite-state transducer grammar from a corpus of English-Japanese loanwords, they uncover an internal inconsistency in Japanese loanword phonology, suggesting that adaptation is shaped by multiple interacting dimensions beyond just sound-to-sound mapping.
What Problem Does This Paper Address?
Languages constantly borrow words from each other, a process known as loanword adaptation. When foreign words enter a new language, their sounds and structures are modified to fit the phonological rules of the borrowing language. Japanese, in particular, has a complex system for adapting foreign words, which has puzzled linguists for decades. Previous studies often focused on specific phonological phenomena or a limited set of rules, leaving a comprehensive understanding of the entire adaptation process elusive. This paper aims to provide a more complete description of how English words are adapted into Japanese by developing a computational model that accounts for a wide range of sound changes. The core question is: how "regular" or systematic is this adaptation process, and what factors contribute to its complexity?
Methods
The authors employ a computational approach using finite-state transducers (FSTs) to model the phonological rules of Japanese loanword adaptation. This method involves:
- Building a Parallel Corpus: They manually compiled a corpus of 250 English words and their Japanese loanword equivalents, including phonetic transcriptions. This corpus was designed to capture a wide range of phonological interactions observed in previous studies.
- Developing a Rewrite Rule Grammar: Using the
fomatoolkit, they implemented phonological adaptation as a sequence of context-conditioned rewrite rules. These rules are then compiled into FSTs. - Composing FSTs: Individual FSTs representing specific rules are composed together to create a single, monolithic transducer that models the entire adaptation process. This allows for efficient application of multiple rules in sequence.
- Iterative Refinement and Evaluation: The grammar is developed iteratively. Rules are automatically tested against the corpus, and their accuracy is gauged. This process helps identify shortcomings, overgeneralizations, and undergeneralizations, leading to rule modifications and improvements in coverage.
This approach is contrasted with Optimality Theory (OT), which the authors find more difficult to debug. The FST-based rewrite rule system offers a more straightforward way to construct and debug the grammar.
Datasets
The primary dataset used is a manually compiled corpus of 250 English-Japanese loanword pairs. This corpus was derived from a Japanese-Multilingual Dictionary and includes parallel phonetic transcriptions of both English (source) and Japanese (borrowing) entries. The corpus was carefully constructed to include:
- All phonemes present in both English and Japanese.
- A wide variety of phonological interactions.
- "Peculiar instances" and exceptions discussed in previous literature.
- Words of English origin, with attention paid to variations between British and American English transcriptions.
This corpus serves as both the basis for developing the grammar and for evaluating its performance.
Key Insights and Ideas
- Internal Inconsistency: The study reveals an internal inconsistency within the Japanese loanword phonology that may have been underestimated by previous research. This suggests that a single, perfectly predictable system for loanword adaptation might not exist.
- Multiple Dimensions of Adaptation: Loanword adaptation is not solely a phonetic or phonemic process. The study highlights four key dimensions that shape the output form:
- Phonetic: The inherent phonetic characteristics of the source language influence the adaptation.
- Phonemic: The adapted form generally preserves most segments from the source.
- Orthographic: The written form of the source word significantly influences how it is adapted, sometimes overriding purely phonological considerations.
- Historical: Older, established loanword forms, even if they deviate from current phonological rules, persist due to historical reasons.
- Hierarchical Influence: The phonetic and phonemic dimensions appear to be primary, with orthographic and historical factors acting as sub-members that influence adaptation choices, particularly when the phonemic dimension is considered first.
- Computational Power: The study demonstrates the efficacy of computational tools (FSTs) in comprehensively modeling complex linguistic phenomena like loanword adaptation, enabling a more rigorous and efficient analysis than traditional "paper-and-pencil" methods.
Key Results
- The developed finite-state grammar correctly maps 215 out of 250 (86.0%) input English forms to their Japanese counterparts.
- Mistakes made by the system fall into three main categories: "Remainder of historical forms" (28.6%), "Phonotactics relaxed" (22.9%), and "Unexpected segmental lengths" (48.5%).
- Attempts to modify rules to fix specific errors often led to an increase in the overall number of erroneous output forms, indicating that the core set of 34 rules captured most phenomena, and remaining errors are true exceptions.
- The study successfully identified and incorporated the influence of orthography and historical factors, which are crucial for explaining certain adaptations that purely phonological rules cannot account for.
Limitations and Future Work
- Coverage Limit: While achieving 86% coverage is significant, the remaining 14% of errors represent a residue of internal inconsistency that is difficult to capture with current phonological generalizations. This suggests a limit to the predictability of the system.
- Rule Modification Trade-offs: Modifying rules to fix specific errors often negatively impacts overall accuracy, highlighting the delicate balance in linguistic systems and the difficulty of creating a universally perfect grammar.
- Future Work: The authors suggest that future research could explore the hierarchical structure of the dimensions further, potentially leading to a more nuanced understanding of how these factors interact. They also note the invertibility of FSTs could be used for debugging, though this was not explored in depth in this study. The computational approach itself is presented as a valuable tool for future linguistic investigations.
Summary updated