대규모 언어모델（LLM） 기반 중국어 존현문 NP₂한정성 수용도에 대한 연구 : 문맥 효과를 중심으로

다국어 초록 (Multilingual Abstract)

This study examines how preceding discourse context affects the definiteness constraint on NP₂ in Chinese existential constructions. In the NP₁+VP+NP₂ structure, NP₂ is generally regarded as an element introducing new information and is therefore typically realized as an indefinite noun phrase. However, definite NP₂ also occurs under certain discourse conditions. This study investigates this phenomenon quantitatively using large language models and bits per token （bits/t） as an acceptability-related metric.
The experimental materials consisted of 120 minimal-pair items constructed from Chinese existential sentences extracted and modified from the CCL and BCC corpora. The design crossed NP₂ definiteness with the presence or absence of preceding context. Contexts were divided into direct-reference and indirect-reference conditions; the latter was further classified into locative/situational frame, possessive relation, and superordinate–subordinate relation. Four models were evaluated: DeepSeek-LLM-7B-base, Qwen2.5-7B-base, Llama-3-8B-base, and Mistral-7B-base. In the contextual condition, only the target existential sentence was included in the evaluation span.
The results show that, without context, all four models assigned lower bits/t values to sentences with indefinite NP₂ than to those with definite NP₂. This confirms that the general preference for indefinite NP₂ is reflected in LLM-based evaluation. When preceding context was provided, however, the bits/t values of definite NP₂ sentences decreased across all models, indicating that discourse context improves the acceptability of definite NP₂ by activating its referent and providing a referential anchor.
Context type also affected the size of this effect. The direct-reference condition showed a stronger effect than the indirect-reference condition, with average Δctx（+def） values of 2.03 and 1.70, respectively. Within the indirect-reference condition, the effect appeared in the order of locative/situational frame, superordinate–subordinate relation, and possessive relation. This indicates that the acceptability of definite NP₂ is most strongly enhanced when the preceding context establishes a concrete locative or situational frame in which the NP₂ referent can naturally exist or appear. These findings suggest that the definiteness constraint on NP₂ is not an absolute formal rule, but a probabilistic, discourse-dependent constraint modulated by discourse accessibility and contextual framing.

목차 (Table of Contents)

• 1. 서론 1
• 1.1 연구 목적 1
• 1.2 연구 대상 및 방법 4
• 1.3 연구 내용 6
• 2. 선행연구 분석 8
• 2.1 존현문 NP₂및 그 한정성에 관한 연구 8
• 2.2 언어모델(LM) 기반 언어 평가에 관한 연구 10
• 2.2.1 영어 최소대립쌍을 통한 LM의 문법 평가 10
• 2.2.2 중국어 최소대립쌍을 통한 LM의 문법 평가 12
• 2.2.3 한국어 최소대립쌍을 통한 LM의 문법 평가 14
• 3. 존현문 NP₂수용도 평가 실험 설계 16
• 3.1 실험 설계 16
• 3.1.1 존현문 최소대립쌍 설계 16
• 3.1.2 선행 문맥 조건 설정 17
• 3.2 평가 모델 및 분석 방법 21
• 3.2.1 평가 모델 21
• 3.2.2 평가 지표 24
• 3.2.3 분석 방법 26
• 4. 문맥 조건에 따른 NP₂수용도 비교 29
• 4.1. 문맥 유무에 따른 결과 분석 29
• 4.1.1 비문맥 조건에서의 NP₂수용도 비교 29
• 4.1.2 문맥 제시 여부에 따른 한정 NP₂수용도 비교 32
• 4.2 문맥 유형에 따른 결과 분석 36
• 4.2.1 간접 참조 조건과 직접 참조 조건 하의 한정 NP₂수용도 비교 36
• 4.2.2 간접 참조 조건의 하위 분류별 한정 NP₂수용도 비교 41
• 5. 결론 50
• 참고문헌 54
• Abstract 58