A Corpus of Adpositional Supersenses for Mandarin Chinese.

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

TL;DR

This paper introduces the first comprehensive corpus of semantically annotated adpositions in Mandarin Chinese, adapting an existing English-based supersense framework. The authors demonstrate that this framework is robust for Chinese, achieving high inter-annotator agreement and providing valuable insights into cross-linguistic semantic variation of adpositions.

What Problem Does This Paper Address?

Adpositions (like prepositions and postpositions) are crucial for conveying semantic relationships but are notoriously ambiguous and vary significantly across languages. This ambiguity poses a challenge for natural language processing tasks like machine translation and grammatical error correction. A major hurdle in studying this variation and building robust multilingual systems is the lack of annotated corpora for many languages, especially for adposition semantics. While English has seen efforts in this area, Mandarin Chinese, a language with a different syntactic structure and a rich set of adpositional expressions (often referred to as coverbs and localizers), has been largely underserved. This paper aims to fill this gap by creating the first broad-coverage corpus of semantically annotated adpositions in Mandarin Chinese.

Methods

The core of the methodology involves adapting the Semantic Network of Adposition and Case Supersenses (SNACS) framework, originally developed for English, to Mandarin Chinese. SNACS provides a hierarchical inventory of 50 semantic labels (supersenses) designed to be language-independent. The authors developed new guidelines specifically for identifying adpositional targets in Chinese, distinguishing between coverbs (which function like prepositions) and localizers (which follow noun phrases). They carefully filtered out coverbs when they acted as the main predicate of a sentence. The annotation was performed on a Mandarin translation of "The Little Prince." To ensure reliability, the corpus was jointly annotated by three native Mandarin speakers who received advanced training. Inter-annotator agreement was measured using raw agreement and Cohen's kappa, showing high levels of consistency.

Datasets

The primary dataset used is a Mandarin Chinese translation of Antoine de Saint-Exupéry's "The Little Prince." This translation was chosen because it is sentence-aligned with an English translation, facilitating cross-linguistic comparisons. The authors also reference the STREUSLE corpus (an English adposition supersense corpus) for comparative analysis. The final annotated corpus contains 933 manually identified adpositions across 27 chapters, totaling approximately 20,000 tokens.

Key Insights and Ideas

  • Cross-linguistic Robustness of SNACS: The SNACS supersense framework, developed primarily for English, proves to be well-suited for annotating Mandarin Chinese adpositions, despite syntactic differences. This suggests a degree of universality in how adpositions encode semantic relations.
  • Chinese Adpositional Complexity: Mandarin Chinese utilizes both coverbs and localizers to express adpositional meanings, often in combination. The paper highlights the need for careful annotation guidelines to distinguish these from main predicates and to capture their semantic contributions.
  • Construal Analysis is Key: The paper emphasizes the importance of "construal analysis," which distinguishes between the adposition's function and its "scene role" (the underlying relation in the context). This is crucial for understanding how languages adapt semantic categories. For instance, a single English preposition might map to different functions or scene roles in Chinese, and vice-versa.
  • Syntactic vs. Semantic Annotation: The study found that automatic POS taggers (like StanfordNLP's ADP tag) are insufficient for identifying adposition targets in Chinese according to their criteria. This underscores the need for manual annotation and highlights discrepancies between syntactic categories and semantic roles.
  • Unique Chinese Construals: While the SNACS framework is largely sufficient, the adaptation revealed some unique construals in Chinese, such as "BENEFICIARY EXPERIENCER," which are not as prominent or attested in the English corpus.

Key Results

  • The paper presents the first broad-coverage corpus of semantically annotated adpositions in Mandarin Chinese, containing 933 manually identified adpositions.
  • High inter-annotator agreement (Cohen's kappa ranging from 0.88 to 0.93) was achieved, validating the adapted SNACS scheme for Chinese.
  • Comparisons with an English corpus (STREUSLE) reveal differences in the distribution of supersenses. For example, Chinese has a lower proportion of divergent construals (where scene role differs from function) compared to English.
  • The study found that automatic POS tagging for adpositions in Chinese has low precision (41.3% ADP tag for gold adpositions), with many adpositions misclassified as verbs or nouns. A heuristic rule-based approach for adposition identification achieved a higher F1 score (82.4%) than the POS tagger.

Limitations and Future Work

The authors acknowledge that the current corpus is based on a single translated text ("The Little Prince"). They plan to annotate additional genres, including native (non-translated) Chinese texts and learner corpora, to capture a broader range of adpositional usage and semantic behavior. Further work could also explore the application of this annotated corpus to downstream NLP tasks like machine translation and grammatical error correction for Mandarin Chinese.

Summary updated

Abstract

Adpositions are frequent markers of semantic relations, but they are highly ambiguous and vary significantly from language to language. Moreover, there is a dearth of annotated corpora for investigating the cross-linguistic variation of adposition semantics, or for building multilingual disambiguation systems. This paper presents a corpus in which all adpositions have been semantically annotated in Mandarin Chinese; to the best of our knowledge, this is the first Chinese corpus to be broadly annotated with adposition semantics. Our approach adapts a framework that defined a general set of supersenses according to ostensibly language-independent semantic criteria, though its development focused primarily on English prepositions (Schneider et al., 2018). We find that the supersense categories are well-suited to Chinese adpositions despite syntactic differences from English. On a Mandarin translation of The Little Prince, we achieve high inter-annotator agreement and analyze semantic correspondences of adposition tokens in bitext.
BibTeX
@inproceedings{peng2020corpus,
  author = {Siyao Peng and Yang Liu and Yilun Zhu and Austin Blodgett and Yushi Zhao and Nathan Schneider},
  title = {{A Corpus of Adpositional Supersenses for Mandarin Chinese.}},
  booktitle = {Language},
  year = {2020},
  doi = {10.48550/arxiv.2003.08437},
}
Figure