Representing and Parsing Korean Constituency Structure at Different Levels of Granularity
Authors: Jungyeul Park, KyungTae Lim, Zihao Huang, Eunkyul Leah Jo, Yige Chen, Chulwoo Park
Organizations: Korea Advanced Institute of Science & Technology, South Korea · The University of British Columbia, Canada · The Chinese University of Hong Kong, Hong Kong · Anyang University, South Korea
Korean constituency parsing raises a representational challenge because the terminal units of a phrase-structure tree do not straightforwardly correspond to simple surface words. Korean eojeols are morphologically complex spacing units, and existing constituency resources differ in how they represent eojeol-internal morphology and non-overt elements. This paper compares three constituency parsing representations derived from the Penn Korean Treebank: Morpheme+XPOS, Eojeol+XPOS, and Eojeol+UPOS. We construct these representations by removing null elements, aligning Penn Korean phrase structure with overt eojeol tokens, preserving Penn Korean phrase labels where possible, and varying the terminal and preterminal layers. We then evaluate canonical non-binary transition-based constituency parsers in top-down, in-order, and bottom-up orders under a shared modeling and evaluation setup. All experiments use gold terminal segmentation and gold preterminal labels and therefore evaluate constituency parsing conditioned on gold morphosyntactic annotation. Eojeol terminals yield shorter transition sequences, but Eojeol+UPOS parsing substantially underperforms the morphologically richer conditions. Eojeol+XPOS narrows this gap, while Morpheme+XPOS gives the strongest results even after its predictions are projected to the eojeol terminal domain. Under these gold-annotation conditions, the results show that fine-grained morphological and XPOS representations provide valuable evidence for the evaluated parsers. This empirical finding concerns the information available for parsing and does not by itself determine the linguistically preferable terminal domain. Independently, linguistic and resource-design considerations motivate eojeol as a stable and interpretable surface domain for phrase-structure annotation, with morpheme-level and XPOS information retained as aligned morphosyntactic evidence.
Figures & tables
Representation
Terminal unit
Preterminal label
Morphological information
Morpheme+XPOS
morpheme
XPOS
Projected into terminal yield
Eojeol+XPOS
Eojeol
Combined XPOS sequence
Encoded in preterminal label
Eojeol+UPOS
Eojeol
UPOS
Abstracted to coarse category
Table 1: Three constituency treebank representations compared in this paper.
Figure 1: Three terminal/POS representations derived from the same sentence.
Split
Files
Sentences
Terminals
Train
85
3,877
189,117 / 102,613
Dev
9
428
21,133 / 11,435
Test
18
685
31,631 / 17,354
Table 2: Statistics of the converted Penn Korean Treebank under the fixed file-level split. Files and sentences are shared across the three representations. The terminal counts are reported as morphemes/eojeols: the first value corresponds to the Morpheme+XPOS representation, and the second to the Eojeol+XPOS and Eojeol+UPOS representations. Files ending in 1–8 are used for training, files ending in 9 for development, and files ending in 0 for testing.
Representation
System
EVALB F1
jp-evalb F1
Morpheme+XPOS
Top-down
79.45
77.19
In-order
83.13
80.68
Bottom-up
84.61
82.02
Eojeol+XPOS
Top-down
69.62
67.73
In-order
81.01
78.62
Bottom-up
80.80
78.48
Table 3: Native constituency parsing results on the test set. Each system is evaluated in its own output representation.
Representation
System
EVALB F1
jp-evalb F1
Morpheme+XPOS → Eojeol+UPOS
Top-down
79.58
77.32
In-order
83.26
80.79
Bottom-up
84.76
82.17
Eojeol+XPOS → Eojeol+UPOS
Top-down
69.62
67.73
In-order
81.01
78.62
Bottom-up
80.80
78.48
Table 4: Normalized constituency parsing results in the Eojeol+UPOS representation. Morpheme-based and Eojeol+XPOS predictions are converted to Eojeol+UPOS before evaluation, allowing all systems to be compared against the same gold representation.
System
M+XPOS avg.
M+XPOS total
E+XPOS avg.
E+XPOS total
E+UPOS avg.
E+UPOS total
Top-down
217.88
844713
150.94
585201
150.94
585201
In-order
217.88
844713
150.94
585201
150.94
585201
Bottom-up
84.55
327798
62.24
241294
62.24
241294
Table 5: Transition statistics on the training split. The counts are computed from implemented gold-oracle transition sequences and exclude Finish . M+XPOS, E+UPOS, and E+XPOS denote the Morpheme+XPOS, Eojeol+XPOS, and Eojeol+UPOS representations, respectively.
Representation
System
k=1
k=2
k=3
k=4
k≥5
Morpheme+XPOS → Eojeol+UPOS
Top-down
82.90
70.45
66.89
54.03
59.04
Eojeol+XPOS → Eojeol+UPOS
70.67
59.57
59.56
48.68
57.46
Eojeol+UPOS native
68.80
48.85
54.27
39.74
57.14
Morpheme+XPOS → Eojeol+UPOS
In-order
86.49
75.14
72.70
58.46
67.22
Eojeol+XPOS → Eojeol+UPOS
84.58
70.77
71.17
59.83
67.34
Eojeol+UPOS native
74.07
57.21
64.12
46.90
59.55
Table 6: Arity-specific constituent results on the test split in the unified Eojeol+UPOS representation. Native Eojeol+UPOS predictions are evaluated directly, while Eojeol+XPOS and morpheme-based predictions are converted to Eojeol+UPOS before evaluation.
Representation
System
h=1
h=2
h=3
h=4
h=5
h=6
h=7
h≥8
Morpheme+XPOS → Eojeol+UPOS
Top-down
100.00
91.09
77.37
67.98
58.95
49.66
42.47
22.94
Eojeol+XPOS → Eojeol+UPOS
100.00
86.36
66.45
55.27
47.96
39.33
31.83
13.43
Eojeol+UPOS native
100.00
78.69
50.34
42.38
35.67
28.26
19.17
11.60
Morpheme+XPOS → Eojeol+UPOS
In-order
100.00
92.19
81.03
72.57
65.60
56.93
50.25
33.49
Eojeol+XPOS → Eojeol+UPOS
100.00
91.33
73.78
65.67
59.17
52.84
45.19
28.85
Eojeol+UPOS native
100.00
81.17
57.59
50.01
44.18
35.33
26.50
19.12
Table 7: Results by tree height. Terminals have height 0, and preterminals have height 1. All systems are evaluated after normalization to Eojeol+UPOS trees.
The design of Korean constituency treebanks raises a central representational question concerning the choice of terminal units. Although Korean words are morphologically complex, treating morphemes as constituency terminals can obscure the distinction between word-internal morphology and phrase-level syntactic structure, and can create mismatches with eojeol-based dependency resources. This paper argues for an eojeol-based constituency representation, with morphological segmentation and fine-grained POS information encoded in a separate, non-constituent layer. A comparative analysis shows that, under explicit normalization assumptions, the Sejong, Penn Korean, and KAIST treebanks can be compared over a shared eojeol-based constituency backbone. Building on this result, we outline an eojeol-based annotation scheme that preserves interpretable constituency, supports cross-treebank comparison and constituency-dependency alignment, and provides a surface-form terminal layer for future end-to-end Korean constituency parsing.
Jungyeul Park, Chulwoo Park
KAIST, South Korea. · Anyang University, South Korea.
We revisit punctuation-aware tree binarization for constituency parsing and ask whether dependency-induced headedness improves binary parser supervision. Although learned heads substantially outperform rule-based heads in intrinsic head prediction, they do not yield consistent parsing gains after debinarization. In particular, punctuation-conditioned evaluation shows that learned headedness underperforms rule-based binarization in macro-average punctuation-sensitive F1, despite a small overall gain on CTB. Similar instability appears under cross-treebank transfer. These results suggest that \ycc{linguistically grounded} headedness is not necessarily parser-optimal when used as a binarization control signal. The paper presents a negative result: better head prediction does not imply better punctuation-sensitive constituency parsing.
Zeyao Qi, Yige Chen, Eitan Klinger +2
The Chinese University of Hong Kong, Ma Liu Shui, Hong Kong · The University of British Columbia, Vancouver, Canada · Korea Advanced Institute of Science & Technology, Daejeon, South Korea
We propose a simplified human-in-the-loop workflow for second language (L2) Korean morphosyntactic annotation by leveraging agreement between two domain-adapted parsers. We first evaluate whether parser agreement can serve as a proxy for annotation correctness by comparing it with independent human judgments. The results show strong correspondence between parser and human judgments, supporting the feasibility of semi-automatic L2-Korean UD annotation. Further analysis demonstrates that parser disagreements cluster in linguistically predictable domains such as grammatical-relation distinctions and clause-boundary ambiguity. While many disagreement cases are tractable for iterative model refinement, others reflect deeper representational challenges inherent in parsing and tagging L2-Korean corpora.
Hakyung Sung, Gyu-Ho Shin
Psychology, Rochester Institute of Technology · Linguistics, University of Illinois Chicago