Learning the Ordering of Coordinate Compounds and Elaborate Expressions in Hmong, Lahu, and Chinese

Chenxuan Cui, Katherine J. Zhang, David R. MortensenNAACL 2022 · DOI

Introduction

In many East and Southeast Asian languages, speakers frequently employ Coordinate Compounds (CCs)—such as the Chinese fùmǔ ("father-mother") for "parents"—and Elaborate Expressions (EEs), which are four-word rhythmic phrases. A central question in linguistics is what determines the specific order of these words. While traditional grammar is often thought to be "blind" to sound, previous research suggests these orders follow strict, "unnatural" phonological hierarchies based on tone or vowel quality.

This paper investigates whether these complex linguistic structures can be learned by machines and whether they are encoded solely in phonology or also within the broader lexical distribution of the language.

Methodology

The researchers utilized three primary datasets to explore these patterns:

  • Hmong: 3,253 unique EEs extracted from a 12-million-word corpus.
  • Lahu: 1,400 EEs compiled from linguistic lexicons.
  • Chinese: 254 antonymic CCs using Mandarin and Middle Chinese pronunciations.

Two computational approaches were used: Phonological Classifiers (Decision Trees and SVMs) to predict order based on sound features, and Neural Sequence Labeling (CNN and Bi-LSTM) to test if word embeddings alone could distinguish correctly ordered expressions from swapped ones.

Phonological Hierarchies

A key finding was that simple machine learning models could "rediscover" established linguistic theory. The Decision Trees reconstructed the exact phonological hierarchies proposed by linguists, providing strong empirical evidence for these rules across different languages.

Phonological hierarchies for Hmong, Lahu, and Middle Chinese (MC) tones and rhymes.

In Hmong, for instance, the model prioritized tones as the most efficient predictor, confirming that sound is a primary driver for these ordering rules. Accuracy for these sound-only classifiers reached 96.7% for Hmong and 81.9% for Lahu.

The Two-Route Hypothesis

The study proposes a "two-route" hypothesis for learning word order: one through phonology (sound rules) and another through lexical distribution (usage patterns).

Surprisingly, neural models achieved over 99% accuracy using only word embeddings, without any explicit phonological information. This suggests that words typically appearing in the first position versus the second position are used differently enough in general context that a model can distinguish them based on their "neighborhood" in vector space.

UMAP projection of tagger embeddings showing separation between first-position ($B_1$) and second-position ($B_2$) words.

Results and Insights

  • Redundancy: Adding phonological data to neural models already using word embeddings did not improve performance, suggesting that historical phonological rules have become "fossilized" within the vocabulary's distribution.
  • In-Context Performance: The CNN model reached a staggering 99.55% accuracy in identifying correctly ordered EEs in the Hmong corpus.
  • Feature Importance: While sound is the most efficient way to encode these rules, word embeddings provide a redundant but highly effective secondary signal.

Conclusion

While machines can utilize both phonological and distributional "routes" to learn these patterns, the next step is psycholinguistic testing to determine which route human brains prioritize. This research demonstrates that even "unnatural" linguistic rules are robustly encoded in data, accessible to both simple classifiers and complex neural architectures.

Summary updated

BibTeX
@inproceedings{cui2022learning,
  author = {Chenxuan Cui and Katherine J. Zhang and David R. Mortensen},
  title = {{Learning the Ordering of Coordinate Compounds and Elaborate Expressions in Hmong, Lahu, and Chinese}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.naacl-main.268},
}
Figure