Introduction
In traditional natural language processing pipelines, Information Extraction (IE) and Coreference Resolution (CR) are typically treated as independent, sequential tasks. For example, a system might first identify fields like "Author" or "Title" in a citation string and then use those extracted fields to determine if two citations refer to the same publication. This serial approach suffers from error propagation—where extraction mistakes inevitably lead to coreference failures—and fails to leverage the synergy between the tasks.
This paper proposes an integrated framework using conditionally-trained undirected graphical models to perform IE and CR simultaneously. By allowing uncertainty and evidence to flow in both directions, the model can use coreference information to correct extraction errors and vice versa.
Integrated Model Architecture
The authors utilize Conditional Random Fields (CRFs) to represent the joint task. Unlike generative models, CRFs allow for the inclusion of arbitrary, overlapping features of the input text without requiring explicit modeling of their dependencies.
The integrated model consists of three primary components: segmentation for IE, pairwise coreference represented as a graph partitioning problem, and entity attribute variables that represent "canonical" values derived from multiple mentions. To handle the complexity of joint inference, the authors employ a structured variant of Iterated Conditional Modes (ICM). A key innovation is the use of N-best integration: instead of passing only the single best segmentation to the coreference step, the model passes the top $N$ hypotheses. This allows the coreference component to resolve extraction ambiguities by choosing the segmentation that best fits the global context of the dataset.
Experimental Evaluation
The model was evaluated on the CiteSeer dataset, which contains approximately 1,500 citations across four topical subsets: Reinforce, Face, Reason, and Constraint. The goal was to segment citations into fields and cluster those referring to the same publication.
Key Results
The integrated approach demonstrated significant improvements over baseline pipeline methods:
- Coreference Accuracy: Using N-best lists (specifically $N=9$) resulted in a 25-35% reduction in coreference error compared to models without segmentation information.
- Extraction Accuracy: Leveraging coreference information to refine field labels led to a 6-14% reduction in extraction error.
Statistical significance was confirmed using McNemar’s test ($p < 0.05$), indicating that the gains from joint inference are robust across different citation domains.
Conclusion and Future Work
The research demonstrates that "closing the loop" between extraction and coreference provides substantial benefits. By maintaining uncertainty through N-best lists and using relational modeling, the system effectively uses the consensus of multiple noisy records to improve overall accuracy. Future work includes scaling these methods to larger datasets using more efficient sampling or variational inference and exploring more sophisticated conditional features.
Summary updated