Abstract
Chinese is a "pro-drop" language where subjects and objects are frequently omitted, creating "zero pronouns" (ZPs) that complicate coreference resolution. This paper addresses this challenge by explicitly detecting and modeling these invisible mentions. By integrating ZPs into the syntactic parsing and coreference pipeline, the authors achieve a significant performance boost on the CoNLL-2012 benchmark, demonstrating that accounting for hidden discourse elements is crucial for robust Chinese natural language processing.
Introduction
Coreference resolution—the task of identifying all expressions that refer to the same entity—is a fundamental challenge in natural language processing. In Chinese, this task is uniquely difficult due to the prevalence of zero pronouns (ZPs). Unlike English, where pronouns like "he" or "it" are usually explicit, Chinese allows for the omission of these words when they can be inferred from context.
These "invisible" words are not rare; they constitute over 10% of mentions in Chinese coreference chains. Standard systems that ignore ZPs often produce fragmented discourse analyses, losing the "thread" of an entity's actions across a text. This paper demonstrates that making these hidden mentions visible to the machine significantly boosts the performance of coreference resolution systems.
Methodology
The researchers proposed a two-stage framework to integrate ZPs into the coreference pipeline:
1. Zero Pronoun Detection
The system first identifies potential locations for ZPs using a simplified Semantic Role Labeling (SRL) framework. An SVM classifier predicts whether a subject or object is missing before a predicate phrase (usually a verb). This detection relies on features such as clause type, the path to the root node in the parse tree, and neighboring syntactic siblings.
2. Exploiting ZPs for Coreference
Once detected, ZPs are leveraged in two ways:
- Syntactic Refinement: The Berkeley parser is retrained to explicitly include ZPs in its output. This helps the parser better understand the grammatical structure of sentences by filling "missing" slots.
- Learning Example Generation: ZPs are treated as special noun phrases during model training. By including ZPs in training pairs, the model learns the "salience" of entities more accurately, using the ZP as a bridge to connect explicit mentions that are far apart.
Experimental Setup
The study utilized the CoNLL-2012 shared task dataset, based on OntoNotes v5.0. This massive corpus includes newswire, broadcast news, and web data, serving as the standard benchmark for evaluating coreference resolution in multiple languages.
Key Results and Insights
Detection Accuracy
The ZP detector achieved an F-measure of 88.29% using gold-standard parse trees and 75.95% using automatic parse trees. This indicates that ZPs can be detected with high reliability even in noisy, automatically-parsed text.
Coreference Performance Boost
Incorporating automatically detected ZPs improved the average F-measure of the coreference system from 60.71% to 63.07%. This jump is significant in the field of coreference resolution. The most striking insight is that ZPs act as "structural glue"; adding them back into the parse tree improved the parser's own F-measure by 1.4%–1.8%.
Bridging the Gap
Coreference chains in Chinese often span many sentences. A ZP in one sentence might refer back to a name several sentences prior. By identifying that ZP, the system gains a "stepping stone" to maintain the entity's salience, making it easier to link subsequent mentions back to the original entity.
Conclusion
This work highlights the importance of addressing language-specific phenomena like pro-drop in NLP. Even though final evaluation metrics often remove ZPs before scoring, their internal use during the resolution process provides the contextual "lift" necessary to improve the resolution of visible words. Future work may involve more sophisticated "Zero Pronoun Resolution"—specifically identifying which entity a ZP refers to—to yield even higher gains.
Summary updated