TL;DR
This paper adapts the SNACS framework, originally designed for English prepositions, to annotate Korean postpositions. They introduce the first broad-coverage corpus of Korean postposition semantics using The Little Prince and provide a comparative analysis with English annotations, highlighting language-specific challenges and insights.
What Problem Does This Paper Address?
Understanding the semantic roles of words like prepositions (or postpositions in Korean) is crucial for natural language understanding. While frameworks like SNACS exist for English, applying them to typologically distant languages like Korean, which uses agglutinative postpositions (bound morphemes) rather than independent words, presents significant challenges. Existing Korean linguistic resources often focus on syntactic functions or a limited set of semantic categories, failing to capture the nuanced semantic range of postpositions. This paper aims to bridge this gap by creating a comprehensive semantic annotation for Korean postpositions, enabling cross-linguistic comparison and advancing NLP capabilities for Korean.
Methods
The core method involves adapting the existing SNACS (Semantic Network of Adposition and Case Supersenses) framework for Korean. This adaptation required:
- Developing Language-Specific Guidelines: The SNACS hierarchy and guidelines were modified to accommodate Korean-specific phenomena, such as the stacking of postpositions and the unique roles of nominative (NOM) and accusative (ACC) case markers, which can function semantically.
- Annotation Schema Extension: New supersenses were introduced, specifically
CONTEXTas a top-level category andTOPICALandFOCUSas new supersenses within it, to capture pragmatic uses of Korean postpositions not found in English. AQUOTEsupersense was also added underPARTICIPANT. - Construal Analysis: The SNACS framework's "construal analysis" was employed, assigning two supersenses (scene role and function) to reflect different facets of a postposition's usage, especially when they diverge.
- Corpus Creation: The Korean translation of The Little Prince was annotated with these adapted SNACS labels.
- Cross-Linguistic Comparison: A subset of the annotated data was aligned with existing English SNACS annotations for The Little Prince to conduct an "apples-to-apples" comparison.
Datasets
- Korean Little Prince Corpus: A newly created corpus based on the Korean translation of Antoine de Saint-Exupéry's The Little Prince. This corpus consists of 27 chapters, containing 10,939 tokens and 4,166 postpositional targets.
- English Little Prince Corpus: Used for comparison, this corpus is based on the English translation of The Little Prince and has been previously annotated using the SNACS framework.
Key Insights and Ideas
- Adaptability of SNACS: The SNACS framework, despite its English origins, proves remarkably adaptable to a typologically distant and agglutinative language like Korean, demonstrating the potential for cross-linguistic semantic annotation.
- Korean Postpositions are Highly Polysemous: Korean postpositions, particularly those like
의/-uyand으로/-ulo, exhibit a much wider range of semantic roles (scene roles and functions) than traditional grammatical categories suggest, often exceeding the polysemy of their English preposition counterparts. - Pragmatic Roles are Crucial in Korean: Korean utilizes postpositions for pragmatic functions like
TOPICALandFOCUSto a greater extent than English, necessitating the creation of new supersenses to capture these nuances. - Case Marking Carries Semantic Weight: Unlike English, where word order often dictates subject/object roles, Korean NOM and ACC case markers carry significant semantic information, aligning with proto-agent and proto-patient roles, and sometimes even more specific roles like
ORIGINATORorLOCUS. - Translation Divergences: Direct cross-linguistic alignment reveals interesting differences in how languages express similar semantic concepts, often due to preferred lexical choices, idiomatic expressions, or grammatical structures, posing challenges for direct semantic mapping.
Key Results
- Broad-Coverage Corpus: The paper introduces the first broad-coverage corpus of Korean postposition semantics, annotated with 42 unique SNACS labels and 108 unique construals.
- High Inter-Annotator Agreement: The annotation process achieved high inter-annotator agreement rates (e.g., 82.7% for scene role, 90.2% for function label), indicating the robustness of the adapted guidelines. Agreement was particularly high for NOM, ACC, TOP, and FOCUS markers.
- Quantitative Comparison: The Korean corpus has nearly twice the number of postpositional targets compared to the English corpus for the same text, yet uses fewer unique postposition types. This is attributed to the semantic load carried by Korean case markers and the expression of FOCUS.
- Qualitative Differences: The comparative study revealed that while scene roles show good correspondence between Korean and English, function labels and construals diverge more significantly, highlighting language-specific semantic expression.
Limitations and Future Work
- Limitations:
- The study acknowledges that some English supersenses did not appear in the Korean corpus, possibly due to different grammatical expression (e.g., verbal endings, nominal affixes).
- The alignment study for cross-linguistic comparison was limited to specific chapters and a subset of aligned adpositions, suggesting a need for more extensive parallel annotation.
- Translation divergences between the French original and the Korean/English translations can complicate direct semantic mapping.
- Future Work:
- Investigate supersenses that were absent or infrequent in the corpus, potentially exploring multi-word postpositions.
- Further research into nuanced semantics like causality and force dynamics to better bridge translation divergences.
- Expand the SNACS framework to other languages (e.g., Hindi, German) and conduct further cross-linguistic comparisons.
- Explore the application of SNACS to subjects and objects, as has been done in other recent work.
Summary updated