TL;DR
This paper introduces G-UNIMEM, a novel case-based learning model designed to automatically acquire linguistic regularities for resolving and choosing anaphora in Mandarin Chinese. By leveraging semantic roles as dominant features, G-UNIMEM achieves significantly higher accuracy than previous search-based methods and linguistic criteria, offering a more practical and efficient approach to this complex NLP task.
What Problem Does This Paper Address?
Understanding and processing natural language requires resolving anaphora – how pronouns or other referring expressions link back to their antecedents. This is particularly challenging in Mandarin Chinese, where zero anaphora (omission of a pronoun) is common. Existing approaches to anaphora resolution and choice in Mandarin Chinese fall into two camps: search-based methods that lack linguistic depth, and linguistic theories that are often too abstract and difficult to implement computationally. This paper aims to bridge this gap by developing a system that can learn anaphoric regularities from data, making the process more accurate and computationally feasible.
Methods
The core contribution is the G-UNIMEM model, a modified version of Michael Lebowitz's UNIMEM case-based learning system. G-UNIMEM is designed for incremental learning and can generalize concepts from training instances. Key modifications include:
- Explicit Causal Rule Extraction: G-UNIMEM can extract explicit causal rules from its memory hierarchy, unlike UNIMEM.
- Improved Generalization: It avoids erroneous conclusions by retaining multiple possible causal accounts, rather than committing to a single, potentially incorrect one (e.g., distinguishing between "whiskey and water" causing drunkenness versus just "water").
- Feature Hierarchy: The system organizes training instances into a memory hierarchy based on feature frequencies, creating either a "g-c-hierarchy" (generalized goal-cause) or "c-g-hierarchy" (cause-goal). The g-c-hierarchy is generally preferred for retaining all causal accounts.
- Rule Generation and Filtering: A rule generator extracts causal rules from the hierarchy, and a rule filter removes ill-formed or conflicting rules.
- Feature Selection: A feature selector resolves conflicts by augmenting rules with mutually exclusive contingent cause features.
Crucially, the paper emphasizes the use of semantic roles (case) as dominant features, a departure from previous approaches that relied on more general linguistic criteria.
Datasets
The paper uses a dataset of 120 testing instances for evaluation. These instances were selected from linguists' examples, textbooks, essays, and novels, with half containing zero anaphora and the other half containing pronominal anaphora. The training set size varied from 30 to 120 instances. No new benchmark dataset was introduced.
Key Insights and Ideas
- Learning Anaphoric Regularity is Key: Instead of relying solely on pre-defined linguistic rules or brute-force search, the paper demonstrates the power of machine learning to acquire complex linguistic patterns for anaphora.
- Semantic Roles are Powerful Predictors: Using semantic roles (like agent and theme) as dominant features for anaphora resolution and choice proved highly effective, outperforming general linguistic criteria. This aligns with theories suggesting that thematic roles play a significant role in discourse coherence.
- Case-Based Learning Adapts Well: The case-based learning approach of G-UNIMEM allows for incremental acquisition and adaptation, making it suitable for the dynamic nature of language.
- G-UNIMEM's Generalization is Robust: The model's ability to retain multiple causal accounts prevents premature or incorrect conclusions, a significant improvement over simpler generalization mechanisms.
Key Results
- Significant Accuracy Improvement: G-UNIMEM, particularly with the c-g-hierarchy and semantic roles as dominant features, achieved high accuracy rates: 95.8% for resolving and 90.8% for choosing anaphora with 120 training instances.
- Superior to Previous Methods: These results significantly surpassed those obtained using Tai's criteria (54.2% for resolving, 71.7% for choosing) and were generally better than history list and Hobbs's algorithms.
- Reduced Conflicting Rules: The system effectively reduced the number of conflicting rules after resolution, indicating a more coherent set of learned regularities.
Limitations and Future Work
- Sample Size: The authors acknowledge that the 120 training and testing instances are small compared to a real-world linguistic domain. They suggest that a larger corpus would lead to more objective and reliable results.
- Feature Set Objectivity: While statistical methods were used to validate the choice of dominant vs. contingent features, the initial assignment of these features could be further refined.
- Computational Cost: While G-UNIMEM is described as "easy" in terms of computability for anaphora compared to some other methods, the paper doesn't detail the exact computational complexity or performance on very large datasets.
Future work would likely involve testing G-UNIMEM on larger, more diverse corpora and potentially exploring more sophisticated feature selection mechanisms.
Summary updated