A Dependency Treebank of Classical Chinese Poems

John Lee, Yin Hei KongNAACL 2012

TL;DR

This paper introduces the first large-scale dependency treebank for Classical Chinese poetry, specifically focusing on Tang Dynasty poems. By adapting existing dependency grammar frameworks and introducing new relations, the authors provide a valuable resource for studying the complex syntax and poetic structures of this important literary tradition, enabling new avenues for computational linguistics research and pedagogical applications.

What Problem Does This Paper Address?

While the field of computational linguistics has seen significant progress in creating annotated corpora for modern languages, historical languages and literary texts have lagged behind. Specifically, Classical Chinese, with its unique grammatical structures and flexible word order, presents a challenge for traditional syntactic analysis. Existing resources for Chinese are predominantly for the modern language, leaving a gap in the availability of linguistically annotated data for its classical form. This lack of resources hinders in-depth computational study of Classical Chinese literature, including its poetic devices like parallelism, and its pedagogical use.

Methods

The authors developed a dependency treebank for Classical Chinese poems by adapting the Stanford dependency types. Key methodological aspects include:

  • Dependency Grammar Framework: They adopted a dependency grammar approach, which is well-suited for languages with flexible word order like Classical Chinese, as it focuses on grammatical relations between words rather than constituency.
  • Adaptation of Existing Relations: They mapped existing Stanford dependency relations for English and Modern Chinese to Classical Chinese, making a few adjustments.
  • Introduction of New Relations: To account for features not present or differently expressed in Modern Chinese, they introduced new relations:
  • Locative Modifier (lmod): To mark bare noun phrases that directly modify a verb to indicate location, a common feature in Classical Chinese where prepositions are often omitted.
  • Oblique Objects (obl): To mark nouns that directly modify a verb to express means, instrument, or respect, similar to dative/ablative cases in Indo-European languages.
  • Re-instatement of Relations: They re-instated "noun phrases as adverbial modifiers" (npadvmod) and "indirect objects" (iobj) from the Stanford English dependencies, which were excluded from the Modern Chinese variant but are relevant for Classical Chinese.
  • POS Tagging: They used a revised version of the Penn Chinese Treebank POS tagset, employing a system of nested POS tags for better interoperability.
  • Annotation Process: Two university graduates with Chinese degrees performed the annotation, with a focus on achieving high inter-annotator agreement.

Datasets

  • Primary Dataset: The treebank consists of over 32,000 characters from 521 poems by two prominent Tang Dynasty poets: Wang Wei and Meng Haoran. This collection represents a significant portion of Classical Chinese poetry from the 8th century CE.
  • Basis for Annotation: The annotation framework was based on existing dependency treebanks for Modern Chinese (e.g., Penn Chinese Treebank) and English (Stanford dependencies).

Key Insights and Ideas

  • Classical Chinese Syntax: The study highlights the significant differences in syntactic expression between Classical and Modern Chinese, particularly the frequent omission of prepositions and the direct modification of verbs by nouns for locative and instrumental purposes.
  • Poetic Parallelism: The paper demonstrates the utility of the treebank in studying poetic parallelism, a key feature of Tang Dynasty poetry. They found that parallelism is not always strict at the character level but can exist at the phrase level, suggesting a more nuanced understanding of poetic structure.
  • Interoperability: By basing their work on existing frameworks, the authors ensure that their treebank can be interoperable with Modern Chinese resources, facilitating comparative studies.
  • Pedagogical Value: The explicit grammatical relations provided by the dependency treebank are expected to be highly beneficial for students learning Classical Chinese.

Key Results

  • High Inter-Annotator Agreement: The annotators achieved high agreement rates: 95.1% for POS tagging, 92.3% for head selection, and 91.2% for dependency labeling, indicating the robustness of their annotation scheme.
  • Parallelism Study Findings:
    • In "parallelism expected" categories for Wang Wei's poems, parallelism was high at both character (91%) and phrase (95%) levels.
  • Even in "parallelism not expected" categories, Wang Wei showed relatively high parallelism (57% character, 71% phrase).
  • Meng Haoran's poems also showed higher parallelism when expected (82% character, 81% phrase) compared to when not expected (53% character).
  • The study identified an "AD VV" vs. "JJ NN" combination as a potentially parallel structure in Wang Wei's poetry.
  • Phrase-Level Parallelism: The analysis revealed that phrase-level parallelism is crucial, especially in cases of character-level mismatch, where syntactic units like verb phrases can be considered parallel.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that the definition of parallelism is subjective and that their metric has high precision but lower recall, potentially missing some parallel structures.
  • The dataset is limited to two poets, which might not fully represent the diversity of Tang Dynasty poetry.
  • The paper notes that the "rules" for parallelism have never been codified, making definitive judgments challenging.
  • Future Work:
    • Parsing Classical Chinese poems from other poets to enrich the corpus.
  • Incorporating semantic information to facilitate deeper studies of parallelism, imagery, and metaphorical coherence.
  • Further exploration of the utility of the treebank for NLP tasks and pedagogical applications.

Summary updated

BibTeX
@inproceedings{lee2012dependency,
  author = {John Lee and Yin Hei Kong},
  title = {{A Dependency Treebank of Classical Chinese Poems}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2012},
}
Figure