Applying Co-Training to Reference Resolution

Christoph Müller, Stefan Rapp, Michael StrubeACL 2002 · DOI

Introduction

Reference resolution—the task of identifying which entities in a text are referred to by expressions like "he," "the building," or "it"—is a fundamental challenge in Natural Language Processing (NLP). Building robust resolution systems typically requires vast amounts of human-annotated data, a process that is both time-consuming and expensive. This paper investigates the potential of Co-Training, a weakly supervised learning paradigm, to mitigate this data bottleneck by leveraging small amounts of labeled data alongside a large pool of unlabeled text.

Methodology

The researchers focused on a corpus of German tourism texts (the Heidelberg Corpus), consisting of 250 short documents. Reference resolution was framed as a binary classification task: determining whether a given pair of an (anaphor, potential antecedent) represents the closest correct link.

Feature Engineering and View Splitting

A critical requirement for Co-Training is the existence of "redundant views"—independent sets of features that can each independently classify the data. The authors identified 17 distinct features across document, noun phrase (NP), and coreference levels.

Table 1: Document, NP, and Coreference-level features used for classification.

To optimize the Co-Training process, the authors employed a greedy view-splitting procedure to partition these features into two distinct sets. Instead of the traditional Naive Bayes classifier, they utilized J48 decision trees (a Weka implementation of C4.5), which proved more effective for the linguistic nuances of the German corpus.

Key Insights: The Challenge of "Natural" Splits

The success of Co-Training hinges on the independence of the feature views. However, in reference resolution, features are often deeply intertwined. For example, gender agreement is frequently correlated with semantic class. This lack of a "natural" split makes it difficult for the two classifiers to provide truly independent "opinions" on unlabeled data, often leading to performance degradation if the views are too similar.

Results and Discussion

The effectiveness of Co-Training varied significantly across different types of referring expressions. For pronouns and proper names, which are relatively "easy" for standard supervised models, Co-Training struggled to outperform simple baselines. In fact, for personal and possessive pronouns, as few as 100 labeled instances were sufficient to achieve an 80% F-measure, suggesting that massive annotation for these categories might be unnecessary.

The Definite NP Breakthrough

The most significant gains were observed in the resolution of definite noun phrases (e.g., "the castle"). These are notoriously difficult to resolve because they often require broader context or semantic knowledge.

Figure 1: Co-Training performance for definite noun phrases (def_NP) across different initial training set sizes.

As shown in the results for definite NPs, Co-Training provided a measurable boost. With an initial training set of 1,000 instances, the F-measure improved from 17% to 25%. With 2,000 instances, performance rose from 19% to 28%. This suggests that the unlabeled data helps the system learn the diverse and often subtle ways definite NPs link to their antecedents.

Conclusion

While Co-Training is not a universal solution for reference resolution, it shows clear promise for difficult categories like definite noun phrases. The primary hurdle remains the identification of independent feature views in a domain where linguistic features are highly correlated. Future research may benefit from larger datasets and more complex feature sets to further exploit the "bootstrapping" potential of unlabeled data.

Summary updated

BibTeX
@inproceedings{muller2002applying,
  author = {Christoph Müller and Stefan Rapp and Michael Strube},
  title = {{Applying Co-Training to Reference Resolution}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2002},
  doi = {10.3115/1073083.1073142},
}
Figure