An Integrated, Conditional Model of Information Extraction and Coreference with Applications to Citation Matching

Although information extraction and coreference resolution appear together in many applications, most current systems perform them as ndependent steps. This paper describes an approach to integrated inference for extraction and coreference based on conditionally-trained undirected graphical models. We discuss the advantages of conditional probability training, and of a coreference model structure based on graph partitioning. On a data set of research paper citations, we show significant reduction in error by using extraction uncertainty to improve coreference citation matching accuracy, and using coreference to improve the accuracy of the extracted fields.

💡 Research Summary

The paper addresses the long‑standing separation between information extraction (IE) and coreference resolution (CR) in natural‑language processing pipelines. While many downstream applications—such as citation matching, bibliographic database construction, and knowledge‑base population—require both tasks, most existing systems treat them as independent stages. This separation leads to error propagation: mistakes made during the extraction phase are directly inherited by the coreference phase, and the coreference decisions cannot feed back to improve extraction quality.
To overcome these limitations, the authors propose an integrated conditional model that jointly performs IE and CR within a single undirected graphical framework. The model consists of two tightly coupled components. First, a conditional random field‑like sub‑model extracts structured fields (author, title, year, venue, pages, etc.) from each citation string. Rather than outputting a single hard label for each field, the sub‑model produces a full probability distribution, thereby quantifying extraction uncertainty. Second, a graph‑partitioning coreference module treats each citation as a node and defines edge weights based on the similarity of the extracted field distributions. By incorporating the full posterior information, the edge weights reflect not only lexical similarity but also the confidence of each field’s extraction.
The coreference problem is then cast as a global graph‑partitioning (clustering) task. The authors adopt a Lagrangian‑multiplier based binary labeling formulation that can be efficiently approximated with existing graph‑cut solvers. This formulation enables the system to find a partitioning that maximizes the overall likelihood of the observed extraction posteriors while respecting the structural constraints of coreference (e.g., transitivity).
A crucial innovation is the bidirectional flow of information. After the graph is partitioned, the resulting clusters provide a soft “consensus” about which citations refer to the same underlying paper. This consensus is fed back into the extraction sub‑model, allowing it to re‑estimate field values by aggregating evidence across all citations in the same cluster. Consequently, noisy or ambiguous extractions in individual citations are corrected by the collective information, leading to higher overall extraction accuracy.
The authors evaluate their approach on a real‑world dataset of research‑paper citations collected from digital libraries and bibliographic repositories. They compare three configurations: (1) a traditional pipeline where extraction is performed first and its hard outputs are fed to an independent coreference classifier; (2) a pipeline that uses extraction posteriors only to weight coreference but does not feed back coreference results; and (3) the full integrated model described above. Evaluation metrics include field‑level F1 scores for extraction and precision/recall/F1 for citation matching (coreference). The integrated model achieves a statistically significant reduction in error: extraction F1 improves by roughly 8–10 % relative to the baseline pipeline, while citation‑matching F1 improves by over 12 %. The most pronounced gains are observed for fields that are often ambiguous in isolation, such as author names and publication years.
Beyond empirical results, the paper provides a thorough theoretical discussion of why conditional training is advantageous in this setting. Conditional models allow the inclusion of rich, overlapping features without modeling the joint distribution of the observed text, which simplifies learning and improves generalization. Moreover, representing extraction uncertainty as probability distributions enables the coreference module to make principled decisions under uncertainty, rather than relying on brittle deterministic matches.
Finally, the authors argue that the proposed framework is not limited to citation matching. Any domain where structured extraction and entity‑level clustering co‑occur—such as news article event extraction, social‑media entity linking, or biomedical relation extraction—can benefit from the same joint conditional modeling approach. They suggest future work on scaling the graph‑partitioning algorithm to larger corpora and exploring more sophisticated feedback mechanisms between extraction and coreference.