PPR-SSM: personalized PageRank and semantic similarity measures for entity linking

Andre Lamurias, Pedro Ruas, Francisco M. CoutoBMC Bioinformatics 2019 · DOI · OpenReview

Introduction

In the biomedical domain, entity linking (EL) is a critical yet challenging task. It involves mapping "mentions" in text—such as chemical names, symptoms, or biological processes—to unique, unambiguous identifiers in a Knowledge Base (KB) or ontology. The complexity arises from the high prevalence of synonyms and polysemous terms (acronyms referring to multiple distinct proteins, for instance). While general-purpose EL tools exist, they often fail to capture the technical nuances of scientific literature.

Existing biomedical approaches frequently rely on local similarity (string matching) but ignore "global coherence"—the logical consistency of a set of chosen concepts within a document. Furthermore, many state-of-the-art models require massive amounts of labeled training data, which is scarce in specialized scientific niches. To address these issues, researchers developed PPR-SSM, an unsupervised method that leverages the structural knowledge of ontologies through Personalized PageRank and Semantic Similarity Measures.

Methodology

PPR-SSM is a graph-based approach that operates in three main stages without requiring any training data:

  1. Candidate Generation: For every entity mention in a document, the system uses fuzzy string matching to identify potential matches within an ontology, such as ChEBI (Chemical Entities of Biological Interest) or the Human Phenotype Ontology (HPO).
  2. Graph Construction: A graph is constructed where nodes represent these candidate concepts. Edges are established between candidates if they share a relationship within the ontology (e.g., an "is-a" relationship) within a defined distance threshold.
  3. Personalized PageRank (PPR) + SSM: The algorithm performs random walks on the graph to determine which candidates are most "central" or coherent across the entire document. Crucially, these edges are weighted using Semantic Similarity Measures (SSM) and Information Content (IC). This allows the system to understand not just that two nodes are connected, but how closely they are related based on the ontology's hierarchy.
Example of mapping text mentions to ChEBI ontology IDs within a biomedical snippet.

The Power of Ontology Structure

Instead of treating an ontology as a flat dictionary, PPR-SSM utilizes it as a topographical map. By incorporating Information Content (IC), the model prioritizes specific, rare terms over generic ones. For example, it can distinguish between a broad category like "Neoplasm" and a specific condition like "Bilateral vestibular Schwannoma."

Ontology hierarchies for HPO (A) and ChEBI (B) illustrating the structural relationships used by PPR-SSM to navigate ambiguity.

This structural awareness enables "global coherence." If a document mentions "histamine" and "leukotriene," the system recognizes their shared relationship to "allergic response" within the ontology, which in turn helps disambiguate other nearby terms that might otherwise be ambiguous.

Key Results and Performance

The researchers evaluated PPR-SSM across four gold-standard datasets: ChEBI-patents, HPO-GSC (human phenotypes), CRAFT-BP (biological processes), and CRAFT-CC (cellular components). The method consistently outperformed standard string-matching baselines and basic PageRank implementations:

  • Accuracy Gains: On the ChEBI-patents dataset, the inclusion of semantic similarity improved accuracy by 0.1385 over versions without it, reaching a final accuracy of 0.8039.
  • High Recall: The model demonstrated high recall in disambiguation (up to 0.84), marking a significant improvement over previous benchmarks in the CRAFT corpus.
  • Ontology Sensitivity: The study found that the optimal search radius for relations varies by domain; chemical ontologies (ChEBI) benefit from a shorter search radius (distance 3), while phenotypes (HPO) require a broader one (distance 6).

Limitations and Future Directions

While effective, PPR-SSM faces a "candidate gap": if the initial fuzzy matching fails to include the correct ID, the graph-based ranking cannot recover it. Additionally, very short documents may lack sufficient entity density to build a meaningful graph. Future work may involve integrating word embeddings (like BioBERT) to improve initial candidate selection and exploring methods to extract relations directly from text to supplement the ontology-based graph.

Reference: Lamurias, A., Ruas, P. & Couto, F.M. PPR-SSM: personalized PageRank and semantic similarity measures for entity linking. BMC Bioinformatics 20, 534 (2019).

Summary updated

Abstract

Biomedical literature concerns a wide range of concepts, requiring controlled vocabularies to maintain a consistent terminology across different research groups. However, as new concepts are introduced, biomedical literature is prone to ambiguity, specifically in fields that are advancing more rapidly, for example, drug design and development. Entity linking is a text mining task that aims at linking entities mentioned in the literature to concepts in a knowledge base. For example, entity linking can help finding all documents that mention the same concept and improve relation extraction methods. Existing approaches focus on the local similarity of each entity and the global coherence of all entities in a document, but do not take into account the semantics of the domain. We propose a method, PPR-SSM, to link entities found in documents to concepts from domain-specific ontologies. Our method is based on Personalized PageRank (PPR), using the relations of the ontology to generate a graph of candidate concepts for the mentioned entities. We demonstrate how the knowledge encoded in a domain-specific ontology can be used to calculate the coherence of a set of candidate concepts, improving the accuracy of entity linking. Furthermore, we explore weighting the edges between candidate concepts using semantic similarity measures (SSM). We show how PPR-SSM can be used to effectively link named entities to biomedical ontologies, namely chemical compounds, phenotypes, and gene-product localization and processes. We demonstrated that PPR-SSM outperforms state-of-the-art entity linking methods in four distinct gold standards, by taking advantage of the semantic information contained in ontologies. Moreover, PPR-SSM is a graph-based method that does not require training data. Our method improved the entity linking accuracy of chemical compounds by 0.1385 when compared to a method that does not use SSMs.
BibTeX
@article{lamurias2019ppr,
  author = {Andre Lamurias and Pedro Ruas and Francisco M. Couto},
  title = {{PPR-SSM: personalized PageRank and semantic similarity measures for entity linking}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-3157-y},
}
Figure