A generative model for unsupervised discovery of relations and argument classes from clinical texts

Bryan Rink, Sanda M. HarabagiuEMNLP 2011

Abstract

Extracting structured information from Electronic Medical Records (EMRs) is critical for clinical decision support, yet much of this data remains locked in unstructured text. We propose the Relation Discovery Model (RDM), an unsupervised generative approach for identifying relationships between medical concepts. By jointly modeling relation-specific "trigger words" and the semantic classes of their arguments, RDM effectively filters linguistic noise and uncovers latent relational structures. Evaluated on the 2010 i2b2/VA Challenge dataset, RDM significantly outperforms standard topic modeling and clustering baselines in both quantitative accuracy and qualitative interpretability.

Introduction

The vast majority of clinical data in EMRs is stored as narrative text. While named entity recognition can identify individual concepts like "aspirin" or "headache," understanding the semantic links between them—such as whether a drug was prescribed to treat a specific symptom—is a complex task. Supervised learning for relation extraction is hindered by the high cost of expert annotation and privacy constraints. Existing unsupervised methods, such as Latent Dirichlet Allocation (LDA), often struggle because they do not distinguish between words that signal a relationship (triggers) and general background vocabulary.

Methods

The authors propose the Relation Discovery Model (RDM), a novel generative Bayesian model. Its key innovation lies in how it decomposes a "pseudo-document" consisting of two medical entities and the words between them. The model's architecture is shown in Figure 1.

Figure 1: Graphical model representation of the Relation Discovery Model (RDM).

Key features of the RDM include:

  1. Joint Argument Modeling: Instead of modeling arguments independently, the RDM samples a joint semantic class for both arguments (e.g., a "Test" and a "Problem"). This captures the intuition that certain types of entities are naturally predisposed to specific relations.
  2. Trigger vs. General Words: The model uses a mixture component to decide if a context word is a "trigger word" (specific to the relation type) or a "general word" (common across the corpus).
  3. Directionality: It includes a binary variable to handle argument order, recognizing that "[MRI] revealed [tumor]" and "[tumor] was revealed by [MRI]" represent the same underlying relationship.
  4. Inference: The model parameters are estimated using collapsed Gibbs sampling, a standard technique for performing inference in complex generative models.

Results

The model was evaluated using the 2010 i2b2/VA Challenge dataset, which contains manual annotations for medical concepts and eight specific relation types. On the DS1 dataset (pairs known to have a relation), RDM achieved a Normalized Mutual Information (NMI) of 18.2, compared to 12.8 for the LDA baseline and 8.25 for K-means. On the noisier and more realistic DS2 dataset, RDM maintained a strong lead with an NMI of 18.1.

Qualitatively, the model successfully identified meaningful trigger clusters. As shown in Figure 2, the model grouped prescription-related tokens (e.g., "p.o.", "p.r.n.") and diagnostic indicators (e.g., "showed", "revealed") into distinct relational categories, effectively filtering out linguistic noise.

Figure 2: Examples of discovered relation trigger word clusters.

Conclusion

The RDM provides a robust framework for unsupervised relation discovery in clinical text. While currently limited by the need to pre-specify cluster counts and a susceptibility to stylistic variance between hospitals, it demonstrates the power of selective context modeling. Future work will explore non-parametric extensions, such as the Chinese Restaurant Process, to automatically determine the optimal number of relations and classes from the data itself.

Summary updated

BibTeX
@inproceedings{rink2011generative,
  author = {Bryan Rink and Sanda M. Harabagiu},
  title = {{A generative model for unsupervised discovery of relations and argument classes from clinical texts}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2011},
}
Figure