Unsupervised Learning of Semantic Relations between Concepts of a Molecular Biology Ontology

Massimiliano Ciaramita, Aldo Gangemi, Esther Ratsch, Jasmin Saric, Isabel RojasIJCAI 2005 · OpenReview

Abstract

In this paper we present an unsupervised model for learning arbitrary relations between concepts of a molecular biology ontology for the purpose of supporting text mining and manual ontology building. Relations between named-entities are learned from the GENIA corpus by means of several standard natural language processing techniques. An in-depth analysis of the output of the system shows that the model is accurate and has good potentials for text mining and ontology building applications.

Introduction

The rapid growth of biomedical literature has created a significant need for automated tools that can extract structured knowledge from unstructured text. While Named Entity Recognition (NER) has matured, identifying the complex semantic relations between these entities—such as "Protein A inhibits Protein B"—remains a challenge. Traditional rule-based systems are often limited by their reliance on manually crafted patterns, which are difficult to scale across the vast diversity of biological interactions. This research explores an unsupervised approach to discover arbitrary, biologically relevant relations directly from the GENIA corpus.

Methodology

The proposed system utilizes a two-stage pipeline designed to treat relation extraction as a statistical collocation problem.

Dependency Parsing and Path Extraction

The first step involves parsing sentences into dependency trees. To improve parser accuracy and simplify the linguistic structure, complex entity names are substituted with their corresponding ontology tags (e.g., "EPO receptor" becomes Protein_molecule). The system then identifies the shortest path between two entities in the tree, typically mediated by a verb that acts as the "pivotal element" of the relation.

Figure 2: Example of a dependency parse tree where entities have been replaced by their ontology tags to simplify extraction.

Statistical Filtering and Generalization

Not all extracted paths represent meaningful biological relations. The system employs a log-likelihood chi-square test ($G^2$) to filter out noise, identifying concept pairs and linguistic paths that appear together significantly more often than chance.

Following this, the system attempts to generalize these relations using the Clark and Weir algorithm. By "climbing" the GENIA ontology hierarchy, the model can elevate specific relations (e.g., "Protein_molecule binds DNA_molecule") to more general ones if the data supports it across multiple sub-concepts.

Results and Discussion

The system was evaluated against the GENIA corpus, which consists of 2,000 Medline abstracts. Evaluation by domain experts revealed high precision in the initial discovery phase.

Discovery of Semantic Relations

In the first stage, 76.5% of the extracted relations were judged biologically correct (excluding the broad Other_name category). This represents a substantial improvement over shallow-parsing methods. The system successfully identified 287 unique significant patterns, which were then condensed into 240 templates after generalization.

Figure 3: A semantic network showing discovered relations centered on the 'Virus' concept, illustrating the system's ability to map complex interactions.

Generalization Challenges

The generalization stage achieved a lower accuracy of 53.1%. This was primarily due to the structure of the GENIA ontology, which does not always follow a strict "is-a" hierarchy. For instance, DNA_substructure is categorized under DNA in the taxonomy but does not always share the same functional relations as a full DNA molecule.

Conclusion

This unsupervised model demonstrates that meaningful semantic relations can be discovered from scientific text without the need for extensive manual labeling. By combining dependency parsing with statistical significance tests and ontological generalization, the system provides a scalable framework for supporting text mining and the ongoing development of biological ontologies. Future work will focus on improving the handling of coordinate structures and grouping synonymous verbs to create more cohesive knowledge maps.

Summary updated

Abstract

In this paper we present an unsupervised model for learning arbitrary relations between concepts of a molecular biology ontology for the purpose of supporting text mining and manual ontology building. Relations between named-entities are learned from the GENIA corpus by means of several standard natural language processing techniques. An in-depth analysis of the output of the system shows that the model is accurate and has good potentials for text mining and ontology building applications.
BibTeX
@inproceedings{ciaramita2005unsupervised,
  author = {Massimiliano Ciaramita and Aldo Gangemi and Esther Ratsch and Jasmin Saric and Isabel Rojas},
  title = {{Unsupervised Learning of Semantic Relations between Concepts of a Molecular Biology Ontology}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2005},
  url = {https://openreview.net/forum?id=ryVmK7zubB},
}
Figure