Abstract
In this paper, with the help of a corpus derived from the Harbin Institute of Technology Information Retrieval Laboratory's Chinese dependency-based Treebank and by using the MaltParser as an implementation tool, we propose a Chinese dependency parsing algorithm designed to effectively identify the right boundary of prepositional phrases containing a verb. Empirical results demonstrate that this algorithm significantly reduces errors when analyzing prepositional phrases with internal verbal structures.
Introduction
In Chinese syntax analysis, prepositional phrases (PPs) are ubiquitous, appearing approximately 15 times for every 100 words. While identifying the start of a PP is straightforward—as it begins with the preposition—determining the "right boundary" (the phrase's conclusion) presents a significant challenge for automated systems.
This difficulty is most pronounced when a PP contains an internal verb. Standard dependency parsing algorithms, such as the Arc-eager transition system, often exhibit "short-sightedness." They tend to attach words to the nearest available head, failing to account for the long-distance dependencies common in complex Chinese sentences. For instance, in a phrase like "The story happened in Shandong’s small town," a basic parser might incorrectly link the preposition "in" to "Shandong" rather than the actual object, "town." When verbs are embedded within these phrases, the structural complexity often leads to cascading parsing errors.
Methodology
The Preposition-Shift' Operation
The proposed solution is built upon MaltParser, a data-driven dependency parsing framework. The core innovation is the introduction of a seventh operation to the standard Arc-eager transition system: Preposition-Shift' (PS’).
Unlike standard operations (Shift, Reduce, Left-Arc, Right-Arc), the PS’ operation allows the parser to defer boundary decisions. When a preposition is encountered, if subsequent words do not match specific boundary markers (such as punctuation or the particle 'de'), the algorithm pushes elements into a temporary queue (Queue T). This mechanism prevents the parser from making a "greedy" mistake and closing the phrase prematurely, waiting instead until the true right boundary is identified.
SVM-Based Feature Recognition
To determine when to apply the PS’ operation, the system employs a Support Vector Machine (SVM) classifier. The authors identified 11 specific features to guide the model, ranging from Part-of-Speech (POS) tags to structural counts of verbs and conjunctions between the preposition and the current word.
Experimental Setup
The study utilized the Chinese dependency-based Treebank from the Harbin Institute of Technology (HIT). This dataset is derived from the "People's Daily" corpus (2000) and contains over 26,000 sentences. The researchers focused on identifying the "Verb Trap," where internal verbs act as distractors that cause standard algorithms (like the BPP algorithm) to terminate a PP too early.
Results and Discussion
Comparative Performance
The proposed algorithm was tested against the standard Arc-eager and BPP algorithms using two primary metrics: Unlabeled Attachment (UA) and Labeled Attachment (LA).
As shown in the results, the proposed algorithm achieved a UA of 82.91% and an LA of 80.76%, outperforming both baseline models. This improvement confirms that the deferred decision-making process enabled by the PS’ operation effectively handles the structural ambiguity of PPs containing verbs.
Data Scaling
The researchers also observed a positive correlation between training data size and accuracy. UA scores rose from 78.02% with 1,325 sentences to 81.76% with 7,208 sentences, suggesting that the model continues to benefit from larger, more diverse linguistic datasets.
Conclusion
By introducing the Preposition-Shift' operation and leveraging a targeted SVM feature set, this research provides a robust method for identifying the boundaries of complex Chinese prepositional phrases. While the algorithm significantly improves accuracy, the authors note a trade-off in computational efficiency; as structural analysis becomes more complex, system performance slows. Future work will focus on optimizing the PS’ operation to maintain high accuracy at the scale required for real-world applications.
Summary updated