TL;DR
This paper challenges the long-held notion that English orthography is "near optimal" by developing computational metrics for phonemic transparency and morphological consistency. Their novel algorithm generates a more optimal spelling system, demonstrating that traditional English spelling is significantly further from theoretical optimality than previously assumed.
What Problem Does This Paper Address?
English spelling is notoriously irregular, making it difficult to learn and process. While linguists like Chomsky and Halle (1968) argued that English orthography is "near optimal" because it reflects abstract underlying forms with general rules, this claim has faced considerable skepticism. The authors aim to computationally investigate this "optimality claim" by quantifying two key aspects of an ideal orthography: phonemic transparency (how well spelling reflects pronunciation) and morphological consistency (how consistently morphemes are spelled). They seek to determine if a system can be designed that is demonstrably more optimal than traditional English spelling.
Methods
The core of the paper is a novel algorithm designed to generate an "optimal" English orthography. The process involves:
- Morpheme Representation: Identifying morphemes and their various phonemic forms, then creating a single, consistent phonemic representation for each morpheme. This is achieved by aligning all allomorphs of a morpheme and selecting a representative phoneme for each position, often based on a hierarchy reflecting articulatory effort.
- Word Representation: Combining morpheme representations to form word spellings. A key innovation here is the use of a "predictor" (a discriminative string transducer called DIRECTL+) that estimates the surface pronunciation of a concatenated morpheme sequence. If the predicted pronunciation matches the actual surface phoneme, the underlying representation is kept; otherwise, it's substituted with the surface phoneme. This balances consistency with phonetic accuracy.
- Metrics for Optimality:
- Phonemic Transparency: Measured by average perplexity between grapheme substrings and phonemes, quantifying how many different graphemes correspond to a phoneme and vice-versa.
- Morphemic Consistency: Measured by the average edit distance between morpheme representations across different word forms, normalized by the number of phonemes. Lower edit distance indicates higher consistency.
Datasets
The authors utilized two primary resources for their lexicon:
- CELEX Lexical Database: Provided morphological analysis of words.
- Combilex Speech Lexicon: Provided high-quality phonemic transcriptions.
After intersecting and filtering these resources, they obtained a lexicon of approximately 51,000 word-forms annotated with both morphological and phonemic information.
Key Insights and Ideas
- Chomsky and Halle's claim is questionable: The paper provides a computational counter-argument to the "near optimal" claim for English orthography.
- Optimality is a trade-off: Phonemic transparency and morphological consistency are inherently at odds. A purely phonemic system would sacrifice morphological consistency, and vice-versa. The goal is to find the best compromise.
- Morpheme representation is crucial: The algorithm's success hinges on creating consistent representations for variable morphemes (allomorphs).
- Predicting pronunciation is key to bridging the gap: Using a predictor (DIRECTL+) allows the system to adjust spellings when the underlying morpheme representation would lead to a significantly different pronunciation, thus improving phonemic transparency without completely abandoning morphological consistency.
- Traditional orthography has significant flaws: Even after accounting for common spelling rules, traditional English orthography scores poorly on both phonemic transparency and morphemic consistency compared to their generated system.
Key Results
- The authors' generated system (ALG) significantly outperforms traditional orthography (T.O.) on both phonemic transparency and morphemic consistency.
- Traditional orthography scores poorly across all evaluated measures, including orthographic and phonemic transparency, and morphemic consistency.
- A phonemic transcription (IPA) achieves perfect phonemic transparency (score of 1) but very low morphemic consistency.
- A simple morpheme concatenation system (M-CAT) achieves perfect morphemic consistency (100%) but very low phonemic transparency.
- The ALG system, designed to balance both, achieves a much better compromise, demonstrating a substantial improvement over traditional orthography. For example, traditional orthography's morphemic consistency is 96.11, while ALG achieves 98.90 (after applying spelling rules).
- The ALG(L) system, which respells IPA symbols using the standard alphabet, also shows improved orthographic perplexity compared to T.O.
Limitations and Future Work
- Predictor's limitations: The DIRECTL+ predictor, while effective, tends to make mistakes with unstressed vowel reduction in English. The authors refrain from replacing underlying phonemes with [ə] or [1] in such cases to mitigate this.
- Spelling rules' reliability: While common spelling rules can improve morphemic consistency, they are not always reliable for readers to infer morpheme spellings from surface forms. The paper notes that the f-voicing rule, in particular, has low applicability.
- Future work: The paper suggests further investigation into the impact of spelling rules and exploring how to incorporate them more effectively. The authors also mention the potential for further research into the trade-offs between phonemic and morphemic optimality.
Summary updated