Multi-Dialectal Representation Learning of Sinitic Phonology

Zhibai JiaACL 2023

TL;DR

This paper introduces a novel graph-based approach to learn representations of Sinitic phonology across multiple dialects. By constructing a knowledge graph of phonetic components and characters, and leveraging the BoxE embedding model, they demonstrate the ability to capture intricate dialectal relationships and infer historical linguistic features. This work offers a promising new tool for historical phonology and dialectology, particularly for languages with complex syllable structures.

What Problem Does This Paper Address?

Investigating the evolution of languages, especially those with dense dialect clusters like the Sinitic family, is crucial for understanding linguistic history. However, manually comparing dialects and reconstructing proto-languages is a laborious and costly endeavor. While machine learning has shown promise in symbolic reasoning, applying it to Sinitic historical phonology faces challenges. These include the monosyllabic nature of Sinitic languages, where a single character often represents a syllable with an initial, final, and tone, leading to numerous homophones that complicate traditional comparative methods. Furthermore, the supersegmental tone feature adds another layer of complexity to historical analysis. This paper aims to accelerate and improve this process by developing a method for learning multi-dialectal representations of Sinitic syllables.

Methods

The core of this paper's methodology lies in constructing a knowledge graph from structured phonological data. This graph represents unique phonetic components (initials, finals, tones) and individual Chinese characters as nodes. Edges connect these nodes, with edge types indicating the dialect and component category (e.g., "Changsha_initial").

To learn representations from this graph, the authors employ the BoxE (Box Embedding) model, a knowledge base completion technique. BoxE embeds nodes using position and translational bump vectors, and defines relationships as hyper-rectangles. This approach is chosen for its expressiveness and ability to handle many-to-one relationships, which are common in phonological data (e.g., multiple characters sharing the same initial). The model is trained to predict missing links (phonetic components) within the graph, implicitly reconstructing a latent "proto-language."

Datasets

The study utilizes pronunciation data from four varieties of Xiang Chinese: Changsha (長沙), Shuangfeng (雙峰), Guanyang Wenshi (灌陽文市), and Quanzhou Xiancheng (全州縣城). This data was sourced from the Comparative Analysis Toolset for Chinese Dialects. Additionally, labels of Middle Chinese readings were obtained from the same source. Middle Chinese, dating back to the 6th century AD, serves as a proxy for a proto-language for the Xiang dialects. The dataset comprises approximately 2805 characters, with a training set of 22300 entries.

Key Insights and Ideas

  • Graph-based representation learning is well-suited for Sinitic phonology: The inherent structure of Sinitic syllables (initial, final, tone) and their connection to characters naturally lend themselves to a graph representation.
  • BoxE captures dialectal nuances: The BoxE model's ability to handle complex relationships and its "translational bump" mechanism effectively capture the subtle phonetic variations and correspondences across different Sinitic dialects.
  • Unsupervised clustering reveals phonological structure: Applying unsupervised clustering to the learned embeddings demonstrates that the representations capture meaningful phonological contrasts, with finals showing a stronger influence on cluster formation than initials or tones.
  • Inference of proto-language features is possible: The learned embeddings can be used to train classifiers that accurately predict historical linguistic features (like grades, voice, tones, etc.) from Middle Chinese, validating the representations' ability to preserve diachronic information.
  • Dialectal divergence can be quantified: The framework provides a quantitative means to estimate divergences between phonological systems, aiding dialectology.

Key Results

  • The BoxE model achieved a hit@1 accuracy of 51.25% on the validation set for link prediction, outperforming other common knowledge graph embedding models like RotatE and ComplEx.
  • Unsupervised clustering experiments showed that the learned embeddings could distinguish phonological features, with finals contributing more to cluster structure than initials or tones.
  • MLP probes trained on the embeddings achieved accuracies significantly higher than a naive baseline for inferring Middle Chinese features, with the best performance for tones (83.0%) and voice (81.1%).
  • Confusion matrices revealed insights into dialectal blending and historical shifts, such as potential confusion between certain "She" categories in Proto-Xiang compared to Middle Chinese.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited transferability: The initial-final-tone decomposition is specific to Sinitic and may not easily transfer to languages with different syllable structures.
  • Underutilization of polyphony: The study dropped extra pronunciations for polyphonic characters, which could have provided richer information.
  • Stochastic train-test split: The random split might have overlooked important data patterns; cross-validation or bootstrapping could improve robustness.

Future work could involve:

  • Integrating multi-modal data: Incorporating acoustic or articulatory evidence could enhance representations.
  • Developing dialect-specific embeddings: Addressing the assumption of shared embeddings for components across dialects.
  • Exploring more sophisticated deep learning models: Further improving representation quality.
  • Integrating with other linguistic resources: Enriching the knowledge base with recordings, time series, and corpora.
  • Handling non-Sinitic languages: Adapting the method for languages with more complex syllable structures.

Summary updated

BibTeX
@inproceedings{jia2023multi,
  author = {Zhibai Jia},
  title = {{Multi-Dialectal Representation Learning of Sinitic Phonology}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}

Loading related content...

Figure