Adpositional Supersenses for Mandarin Chinese

Yilun Zhu, Yang Liu, Siyao Peng, Austin Blodgett, Yushi Zhao, Nathan Schneider

Abstract

This study adapts Semantic Network of Adposition and Case Supersenses (SNACS) annotation to Mandarin Chinese and demonstrates that the same supersense categories are appropriate for Chinese adposition semantics. We annotated 15 chapters of The Little Prince, with high interannotator agreement. The parallel corpus gives insight into differences in construal between the two languages' adpositions, namely a number of construals that are frequent in Chinese but rare or unattested in the English corpus. The annotated corpus can further support automatic disambiguation of adpositions in Chinese, and the common inventory of supersenses between the two languages can potentially serve cross-linguistic tasks such as machine translation.

Introduction

Adpositions—words like in, on, at, or through—are notoriously difficult for natural language processing (NLP) because they are highly polysemous, with meanings shifting drastically based on context. While the SNACS framework was developed to categorize these "supersenses" in English, its applicability to Mandarin Chinese remained an open question. Mandarin presents unique structural challenges: it lacks a dedicated class of prepositions, relying instead on coverbs (verbs that function like prepositions) and localizers (words that follow a noun to refine spatial relations).

This study demonstrates that the SNACS framework can effectively bridge the gap between these distinct linguistic structures, providing a foundation for improved cross-lingual NLP tasks like machine translation.

Methodology: Construal Analysis

The researchers employed construal analysis, a key feature of the SNACS framework. This approach assigns two labels to every adpositional instance:

  1. Scene Role: The underlying semantic relation in the specific context (e.g., a physical location vs. a topic of conversation).
  2. Function: The "prototypical" or lexical meaning of the word itself.

By distinguishing between these two, the framework captures metaphorical uses. For example, in "I care about you," about has a SCENE of STIMULUS but a FUNCTION of TOPIC. The authors developed specific criteria for Mandarin, annotating coverbs when they occur pre-verbally and annotating all localizers due to their high semantic significance.

The SNACS hierarchy organizes 50 supersenses into three main subhierarchies: Circumstance, Participant, and Configuration.

Data and Annotation

The study utilized a parallel corpus of 20 chapters from The Little Prince. The Chinese corpus consists of over 13,000 tokens, with 602 manually annotated adpositional instances. The English comparison used existing SNACS-annotated data for the same chapters. The study achieved high reliability, with Cohen’s kappa scores ranging from .88 to .93, indicating that the adapted guidelines are clear and reproducible.

Key Insights and Results

The analysis revealed several critical findings regarding the semantic behavior of Chinese adpositions:

  • Universal Applicability: The 50 supersenses developed for English were remarkably robust for Chinese. The authors identified 40 types of construals in their Chinese data, covering nearly half of the available hierarchy.
  • Functional "Unproductivity": Chinese adpositions have more limited lexical "functions" than English ones. For example, while the English in can function as TIME, MANNER, or TOPIC, the Chinese localizer shàng is almost always lexically a LOCUS (location), even when the context implies a TOPIC.
  • Higher Literalness: Chinese adpositions are more "literal" than English ones. In Chinese, the SCENE and FUNCTION matched in 86% of cases, compared to only 70% in English.
Table showing that Chinese adpositions have a higher rate of matching Scene and Function labels compared to English.

Cross-lingual alignment further highlighted these differences. When comparing aligned adpositions, 73% shared the same Scene Role, but only 51% shared the same Function. This confirms that while the two languages often describe the same "scene," they use words with different core lexical meanings to do so.

Conclusion

This research successfully adapts the SNACS framework to Mandarin Chinese, proving that a common inventory of supersenses can capture adpositional semantics across typologically different languages. Future work will involve expanding the annotation to larger, more diverse corpora and developing automatic disambiguation systems to support downstream NLP tasks.

Summary updated

Abstract

This study adapts Semantic Network of Adposition and Case Supersenses (SNACS) annotation to Mandarin Chinese and demonstrates that the same supersense categories are appropriate for Chinese adposition semantics. We annotated 15 chapters of The Little Prince, with high interannotator agreement. The parallel corpus gives insight into differences in construal between the two languages' adpositions, namely a number of construals that are frequent in Chinese but rare or unattested in the English corpus. The annotated corpus can further support automatic disambiguation of adpositions in Chinese, and the common inventory of supersenses between the two languages can potentially serve cross-linguistic tasks such as machine translation.
BibTeX
@misc{zhu2019adpositional,
  author = {Yilun Zhu and Yang Liu and Siyao Peng and Austin Blodgett and Yushi Zhao and Nathan Schneider},
  title = {{Adpositional Supersenses for Mandarin Chinese}},
  year = {2019},
  eprint = {1812.02317},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.02317},
}
Figure