When the Misidentified Adverbial Phrase Functions as a Complement

A Korean dataset that labels complements and adjuncts of verbal predicates, built from the KLUE dependency treebank and the Sejong dictionary

Findings of EMNLP 2024

Yige Chen1, Kyuwon Kim2, KyungTae Lim3,*, Jungyeul Park4,*, Chulwoo Park5

1The Chinese University of Hong Kong, Hong Kong 2Seoul National University, South Korea 3SeoulTech, South Korea 4The University of British Columbia, Canada 5Anyang University, South Korea

*Corresponding authors

Paper (ACL Anthology) PDF Dataset
Three examples of projecting a Sejong subcategorization frame onto a Korean sentence. (a) Perfect match for 제출하다 (jechulhada, to submit): frame X=N0-이, Y=N1-을, Z=N2-에|에게|로 maps to A 씨는 (AGT), 전치 6주 진단서를 (THM) and 경찰에 (GOL). (b) Partial match for 마치다 (machida, to finish): only 고소인 진술을 (THM) appears because the subject is elided. (c) Partial match for 부르다 (buleuda, to call/summon): 경찰은 (AGT) and 김현중을 (THM) match the frame, 조만간 is an adverb (ADV), and 피고소인 신분으로 is an adjunct (AJT).
Figure 1 (paper). Projecting subcategorization frames from the Sejong dictionary onto sentences of the KLUE dependency treebank. (a) In a perfect match every frame argument is found; note that the dative 경찰에 (gyeongchal-e, ‘to the police’), labeled an adjunct (NP_AJT) in the treebank, is a required argument of 제출하다 and is therefore a complement. (b) A partial match caused by subject ellipsis. (c) A partial match where the phrase 피고소인 신분으로 (‘with defendant status’) has no slot in the frame and is labeled an adjunct; the adverb 조만간 (‘soon’) is not an argument candidate.

Abstract

This study investigates the predicate-argument structure in Korean language processing. Despite the importance of distinguishing mandatory arguments and optional modifiers in sentences, research in this area has been limited. We introduce a dataset with token-level annotations which labels mandatory and optional elements as complements and adjuncts, respectively. Particularly, we reclassify certain Korean phrases, previously misidentified as adverbial phrases, as complements, addressing misuses of the term adjunct in existing Korean treebanks. Utilizing a Korean dependency treebank, we develop an automatic labeling technique for complements and adjuncts. Experiments using the proposed dataset yield satisfying results, demonstrating that the dataset is trainable and reliable.

Keywords Korean predicate-argument structure Complements vs. adjuncts Adverbial phrases Subcategorization frames Sejong dictionary KLUE dependency treebank Korean treebanks Categorial Grammar Sequence labeling

Method

Following Categorial Grammar, the paper makes a binary distinction between complements (syntactically obligatory elements that act as arguments) and adjuncts (optional modifiers). Existing Korean treebanks label every phrase with a non-subject, non-object postposition as adverbial and treat it as an adjunct, although many such phrases (e.g. datives) are selected by the predicate and are therefore complements. The dataset is built automatically:

  1. Source data. The 10,000 training sentences of the KLUE dependency treebank (TTA-style dependency relations with canonical forms and XPOS tags).
  2. Predicate detection. Verbal predicates are located by their XPOS tag and kept if their canonical form has an entry in the Sejong verb dictionary.
  3. Frame matching. For each predicate, its syntactic daughters in the dependency tree are compared with the postpositions of the arguments in each Sejong subcategorization frame, and the best-fitting frame is selected.
  4. Labeling. Daughters whose postposition appears in the frame are labeled complements (COMP); daughters whose postposition is not in the frame are adjuncts (AJT). Daughters without postpositions, such as adverbs and clausal constructions, are neither (ADV/O).

The resulting token-level BIO annotation uses eight classes: O, TARGET, B-/I-COMP, B-/I-AJT and B-/I-ADV. It is released under CC BY-SA 4.0.

Main Results

  • Coverage. Of 22,964 predicate instances, 3,577 are perfect matches and 10,128 are partial matches with a Sejong frame (8,610 with all sentence arguments paired, 1,430 with all frame arguments paired, 88 both); 9,259 have no match.
  • Complements dominate. Across 32,223 sentence-frame pairs with a perfect or partial match, the dataset contains 21,155 complements and 6,757 adjuncts.
  • Reclassified adverbial phrases. Phrases such as the dative 경찰에 (‘to the police’) or 점수 차 승리로 (‘with decisive score margins’), labeled adjuncts (NP_AJT) in the treebank, are complements in the new dataset.
  • Trainable and reliable. With 2-fold cross-validation, KLUE-BERT-base reaches the highest exact F1 (0.7762) and XLM-RoBERTa-base the highest partial F1 (0.8144) for complement/adjunct detection.
Table 1. Complement/adjunct sequence labeling on the constructed dataset: cross-validation mean ± standard deviation of exact and partial F1 for four pretrained encoders (Table 2 in the paper).
KoELECTRA-Base-v3KLUE-BERT-baseKR-BERT-char16424XLM-RoBERTa-base
Exact F10.7631 ± 0.00490.7762 ± 0.00920.7427 ± 0.00590.7634 ± 0.0024
Partial F10.8012 ± 0.00570.8125 ± 0.00970.7850 ± 0.00790.8144 ± 0.0034

BibTeX

@inproceedings{chen-etal-2024-misidentified,
    title = "When the Misidentified Adverbial Phrase Functions as a Complement",
    author = "Chen, Yige  and
      Kim, Kyuwon  and
      Lim, KyungTae  and
      Park, Jungyeul  and
      Park, Chulwoo",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.718/",
    doi = "10.18653/v1/2024.findings-emnlp.718",
    pages = "12326--12336"
}
More projects from KAIST MLP Lab →