Abstract
In this paper we address the question of assigning semantic roles to sentences in Chinese. We show that good semantic parsing results for Chinese can be achieved with a small 1100-sentence training set. In order to extract features from Chinese, we describe porting the Collins parser to Chinese, resulting in the best performance currently reported on Chinese syntactic parsing; we include our headrules in the appendix. Finally, we compare English and Chinese semantic-parsing performance. While slight differences in argument labeling make a perfect comparison impossible, our results nonetheless suggest significantly better performance for Chinese. We show that much of this difference is due to grammatical differences between English and Chinese, such as the prevalence of passive in English, and the strict word order constraints on adjuncts in Chinese.
Introduction
Semantic Role Labeling (SRL)—the task of automatically determining "who did what to whom"—is a cornerstone of natural language understanding. While algorithms for English were well-developed by the early 2000s, it was unclear how these methods would translate to languages with different grammatical structures, especially when large-scale hand-labeled datasets were unavailable. This study investigates whether high-performing Chinese semantic parsers can be built with minimal training data and how language-specific idiosyncrasies affect the task's difficulty compared to English.
Methods and Data
The authors employ a supervised learning approach using a Support Vector Machine (SVM) classifier, treating SRL as a classification problem for nodes in a syntactic parse tree. Key technical contributions include:
- Feature Engineering: Standard features like syntactic paths and head words were adapted for Chinese, including a "governing category" to handle the unique "DE" (的) construction.
- Parser Porting: The Collins (1999) parser was ported to Chinese, involving new head rules to handle rigid right-headed noun phrases and specific punctuation.
- Datasets: A hand-labeled dataset of 1,138 sentences containing 10 specific Chinese verbs was used. For comparison, a matching English dataset was created from the Penn English PropBank.
Linguistic Advantages of Chinese
A striking finding is that Chinese semantic parsing is significantly more accurate than English parsing. This is largely attributed to two factors:
- The Adjunct Advantage: In English, adjuncts (time, location, etc.) can appear in various positions. In Chinese, they have a strict preference for being preverbal.
- The Passive Problem: English frequently uses the passive voice, flipping Agent and Theme positions. Marked passives are rare in Chinese, leading to a more consistent mapping from syntax to semantics.
As shown in the table above, the vast majority of adjunct roles (argM-*) occur before the verb, providing a strong positional cue for the classifier that is absent in more flexible languages like English.
Results
On seen verbs with gold-standard parses, the Chinese system achieved an F-score of 83.1, significantly outperforming the English system (71.5 F-score) under the same conditions. The effectiveness of the feature set is demonstrated through incremental improvements in precision and recall as more context is added.
The "head word" feature proved particularly powerful, as specific words like "today" or "at" are highly correlated with specific roles. When using the ported Collins parser instead of gold trees, the F-score remained strong at 76.7, a milestone for Chinese syntactic and semantic parsing at the time.
Conclusion and Limitations
While the system excels on seen verbs, performance drops on unseen verbs (average F-score 58.3). Idiosyncratic verbs like chūxiàn (emerge) remain challenging due to their unique argument structures. However, the study confirms that the structural regularities of Chinese—specifically adjunct placement and the rarity of passives—make it inherently more amenable to semantic parsing than English when using standard feature sets.
Summary updated