Abstract
In this paper, we use a machine learning framework for semantic argument parsing and apply it to the task of parsing arguments of eventive nominalizations in the FrameNet database. We create a baseline system using a subset of features introduced by Gildea and Jurafsky (2002), which are directly applicable to nominal predicates. We then investigate new features designed to capture the novelties in nominal argument structure and show a significant performance improvement. We also investigate the parsing performance of nominalizations in Chinese and compare the salience of the features for the two languages.
Introduction
Semantic role labeling (SRL) typically focuses on identifying "who did what to whom" in relation to a verb. However, the same information is often conveyed through nouns, known as "eventive nominalizations" (e.g., "the investigation of the crime"). These nominal predicates are challenging for standard parsers because, unlike verbs with relatively predictable subject-object structures, nouns express arguments through possessives, prepositional phrases, or "light verbs" (e.g., "make a complaint"). This research develops a machine learning framework specifically for parsing the arguments of these nominal predicates in English and Chinese.
Methodology
The authors treat semantic parsing as a multi-class classification problem using Support Vector Machines (SVMs), following a two-step pipeline: argument identification and argument classification. While starting with standard features, the system introduces several novel features tailored for nominalizations:
- Intervening Verb Features: Captures "support verbs" (like make, take, give) that often help nouns realize their arguments.
- Genitives in Constituent: Identifies possessive markers (e.g., " 's ") which frequently flag the "agent" of a noun's action.
- Ordinal Position: Tracks the physical proximity of constituents, as nominal arguments are usually immediate neighbors within the same Noun Phrase.
- NP Expansion Rules: Captures the internal structure of the Noun Phrase encompassing the predicate.
Experimental Setup
- English: The authors utilized the FrameNet database, filtering for eventive nominalizations, resulting in 7,333 annotated sentences across 105 semantic frames.
- Chinese: 22 nominalizations were selected from the Penn Chinese Treebank, with 630 sentences annotated using PropBank-style labels.
Results and Analysis
The inclusion of noun-specific features led to significant performance gains in both languages. A key insight was the importance of the "Support Verb" bridge and the physical proximity of arguments to the nominal predicate.
English Performance
The baseline system achieved an F1 score of 51.4%. By adding the new features, the F1 score rose to 57.3% (without the oracle frame) and 63.9% (with the oracle frame). Knowing the "Frame" (the specific sense of the noun) provides a massive boost, acting as an "oracle" feature that sets an upper bound for performance.
Chinese Performance
The Chinese system improved from a baseline F1 of 46.9% to 57.8% with the additional features. The results were more modest than English due to data sparsity and the difficulty of tree-node alignment; Chinese nominal arguments are often "internal" to the Noun Phrase and may not correspond to a single node in a parse tree.
Conclusion
The primary limitations identified are data sparsity and the misalignment between semantic arguments and syntactic constituents. Future work will likely move toward "chunk-based" or "dependency-based" models to better capture these elusive arguments, particularly in Chinese nominalizations.
Summary updated