Introduction
This research introduces a comprehensive annotation guideline and a large-scale corpus of over 1,500 documents designed to identify Chinese Predicate Heads—the structural and semantic centers of sentences. By moving beyond basic part-of-speech tagging, this framework resolves the inherent ambiguity of multi-verb sentences in Chinese, providing a critical foundation for downstream Natural Language Processing (NLP) tasks such as dependency parsing, summarization, and knowledge graph construction.
The Challenge of Chinese Predicate Identification
In English, identifying the main predicate is often simplified by morphological cues (e.g., suffixes like -ed or -ing). However, Chinese is morphologically impoverished and frequently features "serial verb constructions"—strings of successive verbs without clear delimiters.
Existing frameworks like PropBank or FrameNet focus on semantic roles but often fail to designate a single structural center. This lack of a clear anchor leads to "cascading failures" in NLP pipelines. This study addresses this by defining a "Predicate Head" as the unique structural and semantic core of a Chinese sentence, ensuring a stable, tree-like structure for analysis.
Methodological Framework
The authors developed a rigorous annotation guideline through an iterative process involving 21 trained annotators. Key methodological innovations include:
- The "One Predicate Head" Rule: Enforces a single predicate head per labeling unit to maintain structural stability.
- Flattened Annotation Structure: Focuses on top-level dependencies to reduce human error and complexity compared to deep hierarchical parsing.
- Predicate Patterns: Categorizes heads into five patterns: Singleton (S), Reduplicated (R), Coordinated (L), Modified (M), and Specific (V) for idioms.
- Relevant Element Tagging: Defines five core syntactic elements surrounding the predicate (PRE): Subject (SUB), Temporal (TEM), Locational (LOC), Adverbial (ADV), and Complemental (COM).
Dataset and Linguistic Complexity
The study introduces a manually annotated corpus of 1,500+ Chinese adjudication documents. Legal judgments were selected for their neutral, objective style and predefined format, making them ideal for high-performance information extraction.
Statistical analysis of the corpus reveals the significant challenge posed by "multiverbal" sentences. As shown in the data, it is common to find sequences of two to six adjacent verbs, a phenomenon the authors describe as "Inattention to Structure."
Key Insights and Results
- The Erasability Principle: To distinguish adverbials from predicates, the authors employed an "erasability" test: if a word can be removed without altering the sentence's core meaning, it is classified as an adverbial modifier rather than the predicate head.
- Syntactic Priority: The guideline evolved from purely semantic labels to syntactic ones (e.g., "adverbial elements"), which significantly increased annotation consistency among human labelers.
- Handling "Ba" and "Bei" Constructions: Specific rules were established for these unique Chinese structures to ensure "receptor subjects" in passive sentences are consistently tagged as subjects.
- Quality Control: Through rigorous cross-checking, the team achieved a high level of agreement, surpassing the typical 75% consistency rate found in standard Chinese segmentation tasks.
Conclusion and Future Work
This work provides a robust, validated set of rules and a publicly available dataset that addresses long-standing linguistic challenges in Chinese NLP. While currently limited to the legal domain, future work will explore the "innovativeness" of language in social media and literature. Additionally, the authors plan to utilize this gold-standard data to train and evaluate machine learning models, such as Transformers, to automate the identification of predicate heads in complex natural language.
Summary updated