When English Meets LEGO: Can Chinese-Style Word Formation Solve the Vocabulary Explosion?

Forum topic · ✨步子哥 · 2025-11-22

Summary

This article explores a Reddit question—why English doesn't build words like 'Pig-meat' instead of 'pork'—as a lens into three interlocking questions: mathematics, linguistics, and AI. Mathematically, Chinese compounds roughly 3,000 characters into millions of word combinations (combinatorial explosion), keeping the marginal cost of new concepts near zero, while English coinage has pushed the Oxford dictionary past 600,000 entries. Linguistically, Chinese compounds show near-100% semantic transparency (zhhu+ru meaning 'pig meat' = pork), supported by three compounding patterns (subordinate, attributive, coordinated), whereas English mixes Germanic compounding with Norman-French borrowings (pig/pork, sheep/mutton), creating opaque vocabulary. Historically, phonological simplification from Old to Modern Chinese caused homophone pressure, driving disyllabification and lexicalization of compounds (per Arcodia 2007 and Feng 1998/2001). From an AI perspective, BPE tokenization in LLMs handles Chinese 40-60% less efficiently than English, wastes ~35% of vocabulary space, and merges semantically wrong spans; morpheme-based tokenization inspired by Chinese compounding could shrink vocabularies dramatically. The article argues English can selectively adopt transparent compounding in technical, educational, and AI-assisted domains for a 'soft landing' from vocabulary explosion.

Key points

A Chinese forum post examines why English coinages like "Pig-meat" (for *pork*) are mocked, while Chinese builds vocabulary by compounding ~3,000 characters. It argues this reflects a fundamental trade-off between vocabulary size and sequence length, relevant to mathematics, linguistics, and AI tokenization.

Mathematical perspective: combinatorial explosion vs. vocabulary explosion

Languages are information-encoding systems balancing entropy and cost (Huffman coding analogy): high-frequency concepts get short, single-morpheme forms ("pork" is cheaper than "pig-meat").
Short-term efficiency becomes a long-term trap: the Oxford English Dictionary exceeds 600,000 entries, adding ~8,500 words yearly, while Chinese generates new words from ~3,000 characters (3,000² = 9 million two-character combinations).
Per Gerlach & Altmann's stochastic model of vocabulary growth, 70–80% of Chinese neologisms are compounds of core characters, suppressing non-core vocabulary growth.
Under Zipf's law, English requires ~20,000 words for 95% text coverage vs. ~5,000 characters in Chinese; compounding flattens the frequency distribution.

Linguistic perspective: semantic transparency

Chinese compounds score ~100% on semantic transparency (knowing 猪 "pig" and 肉 "meat" lets you infer "pork"); "pork" scores 0%.
English's mixed heritage (Germanic base + Norman French borrowings: pig/pork, sheep/beef/mutton) created a class-based semantic gap; Chinese is an isolating language with no inflection and strong morpheme-syllable correspondence (Arcodia 2007).
Three Chinese compounding patterns are described: subordinate (modifier + head), attributive (metaphorical description + noun), and coordinated (parallel morphemes packing category concepts).
Packard (2000): Chinese speakers process compounds via decomposition-recombination, ~15–20 ms faster than holistic retrieval; after learning 3,000 characters, children can comprehend ~20,000 compounds without memorizing each one.

Historical evolution

Old Chinese was monosyllabic and highly analytic (e.g., each character in the Analects quote functions as an independent word).
Phonological simplification (from CCCMVCCC syllables with ~10 codas in Old Chinese to only [n]/[ŋ] in Mandarin) collapsed ~405 syllables (≈1,300 with tones) under 60,000+ dictionary words—homophone pressure.
Per Feng's (1998) prosodic morphology, disyllabification emerged as an evolutionary adaptation: disyllabic feet disambiguate (易 → 容易/交易/易经), create new meanings (电脑 "electric brain" = computer), and satisfy rhythmic units.
Compounds survive via lexicalization—a five-stage chain from phonological change to lexicalized compound—where only high-frequency, semantically drifted combinations persist.

Cross-language comparison

Chinese achieves zero-marked coordination (no "and," no agreement), enabling phrase-to-word lexicalization; Romance languages' gender/number agreement blocks it, limiting compound productivity.
English is a confused hybrid: Germanic compounding exists (toothbrush) but with low frequency (~30% transparent compounds), inconsistent orthography, and semantic opacity (butterfly, pineapple).
Chinese characters act as a catalyst: they are visually anchored meaning units, unlike phoneme strings in alphabetic script.

AI perspective: the tokenization dilemma

Citing Kurzynski (2025, *To Merge or Not to Merge*), general-purpose LLMs tokenize Chinese 40–60% less efficiently than English because BPE merges character pairs by co-occurrence frequency, not semantics (e.g., LLaMA wrongly merging 的+事; DeepSeek merging 他+不).
Consequences: longer sequences (higher compute), wasted context windows (Chinese conveys only 60–70% of English information density in the same window), and ~35% of bert-base-chinese's 21,128-entry vocabulary wasted on near-unused "##-continuation" CJK variants.
Vocabulary contamination: GPT-4o's longest Chinese merges include gambling/pornography terms from training-corpus co-occurrence; Chinese compounds are naturally contamination-resistant since single characters rarely carry negative semantics.
Proposed fix: morpheme-based tokenization—~3,000 character tokens dynamically bound by attention into compound semantic units—could compress a 50,000-token vocabulary to ~10,000, shrink models, and speed training.

Future outlook

Technical English is already drifting toward transparent compounds (smart TV, user-space, kernel-mode), consistent with Arcodia's hypothesis that analyticity correlates with compound productivity.
Over half of English learners are native speakers of analytic languages; learner-generated compounds may feed back into English (as "Long time no see" did).
LLMs accelerate human-machine coining of terms (区块链 for blockchain, 元宇宙 for metaverse).
The post proposes a three-level "soft landing" for English: transparent compounds in technical vocabulary, compound-pattern teaching in ESL education, and AI-assisted transparent coinage of new terms.

Conclusion

Chinese's elegance lies not in being "better" but in its compatibility with the knowledge-explosion era: combinatorial explosion against vocabulary explosion, transparency against memory load, and rule-based regularity against AI processing cost. English need not become Chinese, but it can learn to grow elegantly—eventually, "Pig-meat" may be a badge of linguistic wisdom rather than a Chinglish joke.

References

1. Arcodia, G. F. (2007). *Chinese: A Language of Compound Words?* In Selected Proceedings of the 5th Décembrettes: Morphology in Toulouse (pp. 79–90). 2. Feng, S. (2001). *Prosodic Morphology and Its Implications for Chinese Word Formation*. Journal of Chinese Linguistics, 29(2), 170–215. 3. Shi, Y. (2002). *The Establishment of Modern Chinese Lexicon and Its Evolution*. Chinese Language Press. 4. Kurzynski, M. (2025). *To Merge or Not to Merge: The Pitfalls of Chinese Tokenization in General-Purpose LLMs*. The Digital Orientalist. 5. Packard, J. L. (2000). *The Morphology of Chinese: A Linguistic and Cognitive Approach*. Cambridge University Press.