Eojeol-Based Constituency Representation

• The paper presents the eojeol-based framework that treats each eojeol as a unique terminal, effectively preserving phrase-level syntactic structure.
• It employs layered annotation separating morphology from syntax, facilitating cross-treebank normalization and direct conversion to dependency representations.
• Empirical evaluations indicate competitive parsing performance, with normalization yielding attachment scores in the low 90% range and constituency F1 in the mid-80% range.

An eojeol-based constituency representation is a structural framework for Korean treebanks in which eojeol tokens—orthographic spacing units composed of lexical stems with any attached functional morphemes—are the unique terminal elements of constituency trees. Morphological segmentation and fine-grained part-of-speech information are segregated into a parallel, non-constituent layer. This approach addresses mismatches that arise when morpheme-level terminalization conflates internal morphology with phrase-level syntactic structure and facilitates direct interoperability with eojeol-based dependency resources (Park et al., 27 Dec 2025).

1. Formal Definitions and Schematic Grammar Notation

Let $W = \langle w_1, ..., w_n \rangle$ denote the linear sequence of eojeol tokens within a Korean sentence. Each eojeol is a surface spacing unit, typically consisting of a stem with following grammatical affixes.

An eojeol-based constituency tree is specified as:

$$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$

where:

• $V_t = \{t_1, ..., t_n\}$ is in bijection with $W$; each $t_i$ is terminal, corresponding uniquely to $w_i$.
• $V_{nt}$ is the set of nonterminal phrase nodes.
• $E$ is the parent-child relation such that $(V_{nt} \cup V_t, E)$ forms a rooted, ordered tree.
• $\mathit{root} \in V_{nt}$ is the unique top node.
• $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$0 maps nonterminals to category labels (S, NP, VP, NP-SBJ, NP-AJT, NML, etc.)
• $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$1 maps terminals to surface forms.

A crucial design constraint is that each terminal $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$2 dominates only the eojeol $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$3; morphological composition is not further projected in the syntactic tree. Directly above each $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$4 is a preterminal $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$5 with a universal POS (UPOS) tag: $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$6 {NOUN, VERB, ADJ, PROPN, ADV, PUNCT, ...}. Morphological segmentation is isolated as a parallel mapping:

• $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$7 list of $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$8 pairs, e.g., $$T = (V_{nt} \cup V_t, E, \mathit{root}, \ell_n, \ell_t)$$9.

The main grammar schemata in major Korean treebanks can be summarized as:

Treebank	Grammar Shape
Sejong	Strict binary ($V_t = \{t_1, ..., t_n\}$0, $V_t = \{t_1, ..., t_n\}$1)
Penn-Korean	Mixed arity ($V_t = \{t_1, ..., t_n\}$2) or ($V_t = \{t_1, ..., t_n\}$3)
KAIST	Morpheme-projecting ($V_t = \{t_1, ..., t_n\}$4)

2. Cross-Treebank Normalization and Representational Equivalence

Normalized representations $V_t = \{t_1, ..., t_n\}$5 are derived for each treebank to achieve eojeol-level representational equivalence:

1. Debinarization (Sejong): Collapse all chains of artificial binary nodes (where $V_t = \{t_1, ..., t_n\}$6, $V_t = \{t_1, ..., t_n\}$7, etc., with $V_t = \{t_1, ..., t_n\}$8 recursively expanded) into flat, n-ary branching.
2. Null-element Removal (Penn): Delete all null element leaves (e.g., pro, trace), reconnecting siblings directly to the parent.
3. Eojeol Alignment: Ensure both treebanks yield the same sequence $V_t = \{t_1, ..., t_n\}$9 at their terminals, adopting uniform tokenization at the eojeol level.

Let $W$0 Sejong, $W$1 Penn, and define:

$W$2

$W$3

The resulting trees are isomorphic in unlabeled structure and terminal sequence, modulo minor relabeling conventions like NP-SBJ versus NP. KAIST morpho-syntactic trees are similarly normalized by SPMRL-style conversion, which reattaches functional morphemes to their stems at the eojeol level. This scheme enables direct structural comparability and supports conversion between treebanks (Park et al., 27 Dec 2025).

3. Annotation Scheme and Layering

Annotation proceeds in discrete layers:

• Tokenization: Deterministic eojeol-level tokenization (e.g., Park & Tyers 2019); punctuation is given separate eojeol status.
• Constituency Bracketing: Hierarchical structure is bracketed Penn-Treebank style over the eojeol sequence. Nonterminals comprise a standard syntactic category inventory (S, NP, VP, ADJP, ADV, NML, etc.), enforcing that no morphological terminals appear.
• Preterminal (UPOS) Layer: Each eojeol is dominated by a unique UPOS preterminal as per Universal Dependencies tagset.
• Morphological Segmentation (XPOS Layer): Morphology is encoded as a parallel mapping from eojeol to its (morpheme, XPOS) decomposition, with XPOS values from the Sejong POS inventory. For example, “땅집이어서” is represented as “땅집/NNG + 이/JKC + 어서/EC”.
• Nominal-Modifier Layer: Adjectival or nominal modifiers within NP may be grouped under an intermediate NML node for explicit stacking.
• Bracketing and Tagging Conventions: Employ grammatical function tags (e.g., NP-SBJ, NP-OBJ, S-COMP) and avoid unary chains except for linguistically required cases.

A typical six-column representation is summarized below:

ID	Surface	Morph (XPOS)	UPOS	L-brackets	R-brackets
1	프랑스의	프랑스/NNP+의/JKG	PROPN	(S (NP-SBJ (NML (ADJ
2	세계적인	세계/NNG+적/XSN+인/ETM	ADJ	(ADJ	))
3	의상디자이너	의상/NNG+디자이너/NNG	NOUN	(NML (NOUN	(
4	에마뉘엘	에마뉘엘/NNP	PROPN	(NOUN (PROPN	))
5	웅가로가	웅가로/NNP+가/JKS	PROPN	(NP (PROPN	))

This layered format formally separates phrase structure, coarse and fine POS tagging, and morphological segmentation (Park et al., 27 Dec 2025).

4. Conversion to Eojeol-Based Dependency Structure

Eojeol-based constituency trees support direct conversion to dependency representations via head-percolation algorithms. As terminals in both representations correspond to the same eojeol sequence, conversion requires no recombination of morpheme-split tokens. The conversion process adapts the Xia & Palmer (2001) head-percolation:

1. Head-Rule Definition: For each nonterminal, establish ordered daughter-type priorities (e.g., NP chooses the rightmost NOUN/PROPN node, VP the leftmost VERB, ADJP the first ADJ).
2. Head Percolation: Recursively identify the head for each constituent by priority.
3. Arc Extraction and Labeling: For each constituent, construct arcs from the head to the heads of its children, assigning dependency labels such as nsubj, obj, amod, ccomp, derived from function tags (e.g., SBJ, OBJ, AJT, COMP).
4. Postprocessing: Reattach dependents in any non-projective cases, though these are rare in eojeol-based trees.

This direct mapping leverages the one-to-one alignment of terminal eojeol, circumventing the complexities seen when morpheme terminals are involved.

5. Empirical Results and Evaluation

There is no large-scale intrinsic evaluation for the proposed scheme in the primary source. However, two empirical observations are provided:

• The SPMRL 2013 Shared Task (Seddah et al. 2013) demonstrated that KAIST trees, post-normalization to the eojeol terminal level, could be parsed and mapped to dependencies with unlabeled attachment scores in the low 90% range.
• Work by Kim & Park (2022) indicates that parsers trained on debinarized Sejong trees achieved constituency F$W$4 in the mid-80% range post-normalization, aligning with cross-linguistic norms when the morphological layer is abstracted from terminals.

A hypothetical summary table is presented as a point of reference (not as a report of new empirical findings):

Resource	Normalization Step	Const. F$W$5	UAS (\%)
Sejong → eojeol	debinarized	84.7	92.1
Penn → eojeol	remove-null	85.3	92.5
KAIST → eojeol	SPMRL-conv	83.9	91.8

This suggests that representational unification and conversion are empirically sound, with performance in line with state-of-the-art cross-linguistic parsers once morphological segmentation is removed from terminal representation.

6. Significance, Interoperability, and Future Work

The primary conceptual gain of an eojeol-based constituency framework is restoration of true phrase-level interpretability in syntactic treebanks, disentangling syntax from internal morphological structure. This backbone supports trivial cross-treebank mapping among Sejong, Penn, and KAIST resources and enables direct, lossless conversion to dependency structures.

Major future directions include large-scale comparison of normalized trees to quantify structural correspondence (e.g., via branch-level F$W$6 under label-preserving mappings) and comprehensive end-to-end parsing experiments (e.g., BERT+CKY backbones) to assess the utility of separating morphology from syntax. The disentanglement of morpheme segmentation from the core phrase structure opens new avenues for both Korean-specific and cross-linguistically comparative research (Park et al., 27 Dec 2025).