TL;DR
This paper proposes a refined method for mapping Korean Part-of-Speech (POS) tags to the Universal Dependencies (UD) tagset, specifically focusing on predicates. By leveraging Korean-specific XPOS tags and considering the unique morphological structure of Korean predicates, the authors aim to improve the accuracy and consistency of dependency parsing for Korean.
What Problem Does This Paper Address?
The Universal Dependencies (UD) project aims to create a unified framework for cross-lingual natural language processing by standardizing morphological and syntactic annotations. However, the standard CoNLL U-Format, designed with resource-rich languages in mind, doesn't fully capture the nuances of Korean. A key challenge is the Korean "eojeol" (a unit separated by whitespace), which is often composed of a stem and various endings (josa or eomi). These endings significantly alter the function of the eojeol, especially when it acts as a predicate. Previous UD approaches for Korean often assign a single UPOS tag to an entire eojeol without adequately considering the role of these endings, leading to potential inaccuracies in dependency annotation. This paper specifically addresses the problem of accurately tagging Korean predicates (verbs, adjectives, copulas, and auxiliaries) within the UD framework.
Methods
The core of this paper's contribution lies in its proposed methodology for mapping Korean POS tags to UPOS, with a strong emphasis on predicates. The authors advocate for a more nuanced approach than simply assigning a VERB or ADJ tag to an entire eojeol. Instead, they propose:
- Leveraging XPOS: Utilizing the language-specific XPOS tags (which provide more detailed morphological information) to inform the UPOS assignment.
- Function-based Tagging: Assigning UPOS tags based on the actual function of the eojeol within the sentence, rather than solely on the base stem. For instance, a verb stem combined with a final or connective ending is tagged as VERB, while the same stem combined with an adnominal ending (functioning as a modifier) is tagged as ADJ.
- Predicate-Specific Rules: Developing specific rules for verbs, adjectives, copulas, and auxiliary verbs, considering how their stems combine with different endings to form predicates or modifiers. This includes differentiating between main and auxiliary verbs.
- Dependency Annotation Refinements: Proposing adjustments to dependency annotation, particularly for verb sequences and compound-complex sentences, to better reflect the head-final nature of Korean and the relationships between predicates.
Datasets
The paper discusses and analyzes existing Korean UD treebanks:
- Google Korean Universal Dependency Treebank (GKT)
- Parallel Universal Korean Dependency Treebank (PUD)
- KAIST Korean Universal Dependency Treebank (KTB)
These corpora are used to examine current POS and dependency annotation practices for Korean predicates. The authors also reference the Korean National Corpus (KNC) POS tagset for its morpheme-unit analysis, which informs their proposed method. The application results are presented on the GSD (General Stanford Dependencies) dataset.
Key Insights and Ideas
- Eojeol Complexity: The Korean "eojeol" is not a simple unit; its function, especially as a predicate, is heavily determined by the attached endings. A one-size-fits-all UPOS tag for an eojeol is insufficient.
- Function Over Form: The proposed method prioritizes the functional role of an eojeol in a sentence (predicate, modifier) over its base lexical category when assigning UPOS tags.
- Adjectives as Verbs: Korean adjectives can function as predicates similarly to verbs, suggesting that tagging them as VERB in such cases is not irrational and can improve accuracy.
- Copula Nuance: The Korean copula "-이-" plays a crucial role in forming predicates and requires careful annotation, with suggestions to tag it as VERB when acting as a predicate and AUX when segmented.
- Auxiliary Verb Distinction: Clearly distinguishing between main verbs and auxiliary verbs is crucial for accurate dependency parsing, especially in verb combinations. The paper proposes using AUX for auxiliary verbs and VERB for main verbs.
- Head-Final Principle: Korean's head-final structure, where complements precede heads, is important for dependency parsing. The paper suggests consistently applying the head-final principle, particularly in compound-complex sentences, by identifying the rightmost predicate as the root.
Key Results
The paper presents an application result on the GSD dataset (Table 1) showing the performance of their proposed UPOS tagging. While specific overall accuracy figures are not highlighted as the primary result, the table demonstrates the impact of their method on different XPOS categories:
- VV+EC: The proposed VERB tag achieved a high correction rate (94%) with a low revision rate (4%).
- VV+EF: Similar high performance for VERB tagging.
- VV+ETM: The proposed ADJ tag achieved a 100% correction rate.
- VX+*: The proposed AUX tag showed a high correction rate (97%) with a low revision rate.
Qualitatively, the paper argues that their method leads to more accurate UPOS annotation and, consequently, more reliable dependency relations, especially for predicates.
Limitations and Future Work
The authors acknowledge that their study is limited to POS annotation for predicates. They state that eojeols with other functions not considered in this paper will be examined in future studies. This suggests that a comprehensive UD annotation for Korean would require extending this methodology to other parts of speech and sentence structures.
Summary updated