Semantic role labeling of nominalized predicates in Chinese

Nianwen XueNAACL 2006

TL;DR

This paper tackles the challenging task of semantic role labeling (SRL) for nominalized predicates (e.g., "development" from "develop") in Chinese. By introducing and utilizing the Chinese Nombank, a new annotated corpus, the authors demonstrate that SRL for nominalized predicates is significantly harder than for verbs, and explore methods to improve performance, including leveraging verb data and reranking techniques.

What Problem Does This Paper Address?

Semantic Role Labeling (SRL) aims to identify the arguments of a predicate and assign them semantic roles (e.g., agent, patient). Traditionally, SRL research has heavily focused on verbs, largely due to the availability of large, annotated datasets like PropBank and FrameNet for English. However, verbs are not the only linguistic elements that function as predicates; nouns derived from verbs (nominalizations) and relational nouns also express actions or states and take arguments. Analyzing these nominalized predicates is crucial for a more complete understanding of sentence meaning. The challenge lies in the scarcity of annotated data for nominalized predicates, especially in languages other than English. This paper addresses this gap by focusing on Chinese nominalized predicates and introducing the Chinese Nombank, a new corpus designed to facilitate SRL research in this area.

Methods

The authors employ a three-stage SRL architecture, a common approach in the field.

  1. Argument Detection: This stage filters constituents in a parse tree to identify potential arguments or adjuncts of a nominalized predicate. They adapt heuristics from previous work on verbal predicates, which exploit syntactic structures to prune unlikely candidates before applying a binary classifier.
  2. Argument Classification: A multi-category classifier is used to assign specific semantic role labels (e.g., ARG0, ARG1, ARGM-TMP) to the detected constituents. They utilize a Maximum Entropy classifier with a tunable Gaussian prior.
  3. Reranking: A preliminary experiment explores reranking techniques to improve SRL accuracy by considering joint probabilities of arguments within a proposition, inspired by work on verbal predicates.

Key features used in their models include the predicate itself, its position relative to the constituent, the syntactic path between them, dominating verbs, head words and their part-of-speech, phrase types, and combination features.

Datasets

The primary dataset used is the Chinese Nombank. This corpus is an extension of the English PropBank and Nombank frameworks, providing semantic role annotations for nominalized predicates in Chinese. It is built upon the Chinese Treebank (CTB), a syntactically annotated corpus. The Chinese Nombank contains annotations for approximately 1,227 nominalized predicate types and 10,497 instances, making it roughly one-quarter to one-third the size of the Chinese Proposition Bank for verbs. For experiments, the data is split into training (688 files) and test sets (72 files). Verb data from the Chinese Proposition Bank is also used in an experiment to investigate its utility for improving nominal SRL.

Key Insights and Ideas

  • Nominalized Predicates are Harder: SRL for nominalized predicates in Chinese is significantly more challenging than for verbs, even when using gold-standard parses. This suggests that the linguistic cues for identifying and classifying arguments are less prominent for nominalizations.
  • Syntactic Structure is Less Informative: Unlike verbs, where arguments often occupy clear syntactic positions like subject and object, arguments of nominalized predicates are less predictable. They can appear inside or outside the noun phrase headed by the nominalization, and their detection relies less on straightforward syntactic cues.
  • Verb Data Doesn't Directly Help: Simply combining verb and nominalized predicate training data without careful feature engineering hurts performance. This is because features like the syntactic path between a constituent and the predicate differ drastically between verbs and nominalizations, creating noise. However, features like head words and POS tags do benefit from verb data, suggesting a need for more nuanced feature adaptation.
  • Reranking Shows Limited Improvement: While joint learning and reranking show promise for verbal SRL, their application to Chinese nominalized predicates yielded only insignificant improvements. This might be due to the less rigid constraint that numbered arguments should not share the same semantic role label in Chinese nominalizations compared to English verbs.

Key Results

  • Gold Standard Parses:
    • When constituents are known (gold standard), the system achieved an F-score of 86.9% for semantic role labeling of nominalized predicates. This is considerably lower than the 93.9% reported for verbal predicates on the same data.
  • When constituents are unknown (requiring argument detection), the F-score dropped to 71.6%, compared to 91.4% for verbal predicates.
  • Automatic Parses: Using an automatic parser resulted in a further performance drop, with an F-score of 51.3%. This highlights the significant impact of parsing errors on SRL accuracy.
  • Verb Data Integration: Adding verb data to the training set without careful feature adaptation hurt overall performance. However, when features like head words and POS tags were considered, there was a slight, statistically insignificant improvement.
  • Reranking: Selecting the top 10 propositions using a reranker did not lead to significant improvement, despite a high oracle score of 97% when constituents were known.

Limitations and Future Work

  • Data Size: The Chinese Nombank, while a valuable resource, is smaller than verb corpora, which likely contributes to the performance degradation.
  • Parsing Accuracy: The significant drop in performance when using automatic parses underscores the dependency on parsing quality. Insufficient training material for Chinese parsers and the inherent difficulties of Chinese segmentation and word identification are identified as major hurdles.
  • Feature Engineering: The authors speculate that the lack of improvement from reranking might be due to the less rigid constraints on argument labels for Chinese nominalizations. They plan to continue experimenting with new features and parameters for reranking.
  • Future Work: The authors intend to explore new features and parameters for reranking. They also suggest that further improvement in SRL accuracy is contingent on advancements in Chinese parsing accuracy, which in turn requires more training material.

Summary updated

BibTeX
@inproceedings{xue2006semantic,
  author = {Nianwen Xue},
  title = {{Semantic role labeling of nominalized predicates in Chinese}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2006},
}
Figure