8 NED with open LLMs and domain ontologies

This chapter introduces a practical, domain-agnostic approach to Named Entity Disambiguation that marries open large language models with rich domain ontologies. It begins by motivating the need to go beyond traditional tools such as ScispaCy, which, while strong in biomedical contexts, can be hard to update, are tied to a single domain, and underuse the relational knowledge embedded in ontologies. The proposed solution runs locally on consumer hardware using an open model (Llama 3.1 8B via Ollama) and leverages SNOMED as a reference ontology, addressing privacy, latency, and extensibility while explicitly exploiting ontology structure during disambiguation.

The implementation loads SNOMED into Neo4j, normalizing many relationship types under a single relationship with a type property and introducing a dedicated hierarchical link to propagate semantic categories from top-level concepts down the taxonomy. These propagated categories feed an ontology-guided NER step in which the LLM is prompted to extract only ontology-derived entity types; character offsets are then computed deterministically in post-processing. Candidate selection does not rely on the LLM: it uses Neo4j full-text search constrained by the NER labels to retrieve plausible SNOMED candidates for each mention (e.g., multiple interpretations for “Zika”), establishing a focused pool for the final decision.

Candidate disambiguation combines graph algorithms and LLM reasoning in three steps: shortest-path detection between candidates within the ontology (with hub filtering and limited hops for relevance), path-to-text translation that turns the discovered graph paths into natural-language statements, and summarization that condenses many sentences into a compact context. The final LLM prompt uses this distilled, ontology-grounded context to pick the best candidate per mention—e.g., preferring “Congenital Zika virus infection” when “microcephaly” co-occurs. The result is an end-to-end NED workflow that is accurate in sparse or ambiguous contexts, portable across domains with suitable ontologies, privacy-friendly when run locally, and extensible with improvements such as vector-based candidate retrieval.

A sample of the hierarchical structure of the SNOMED ontology. Leveraging this hierarchical structure, nodes located on a deeper level, such as “Ecallantide” and “Retinopathy associated with AIDS”, can be categorized using the information from the first-level nodes, such as “Pharmaceutical product” a­nd “Disease”, which represent the archetypal entities of the ontology.

A comprehensive workflow for a Named Entity Disambiguation (NED) system designed to leverage Large Language Models (LLMs) and domain-specific ontologies, such as SNOMED, for biomedical text processing. This workflow is organized into distinct stages, each involving various processes and interactions between the LLM and domain ontology to disambiguate entities within input text accurately.

The first stage of NED is Named Entity Recognition (NER). In our scenario, the collection of named entities is derived directly from the ontology. In SNOMED, the categories are defined by the first-level nodes of the ontology, whose information is propagated to all the other nodes.

The second stage of NED is the Candidate Selection. In this scenario, for each entity mention detected in the previous step, this stage retrieves a collection of potential candidates that can refer to each mention. The current implementation employs a full-text search but can be potentially extended with more advanced techniques.

The third stage of NED is the Candidate Disambiguation. The key goal is to select the best match among all the possible candidates identified in the previous step. To reach this goal, the disambiguation phase exploits an advanced approach combining graph-based algorithms (shortest path detection) and LLMs.

NED Candidate Disambiguation is divided into three main phases: (1) Shortest Path Detection between all the candidates related to the different entity mentions in the sentence. (2) Path-to-text Translation to transform detected paths into natural language sentences. (3) Textual paths summarization to summarize all the natural language sentences into a unique and valuable piece of text useful for the disambiguation

• Named Entity Disambiguation (NED) is essential for accurately identifying and distinguishing entities in complex domains, particularly in biomedical text.
• Traditional Natural Language Processing (NLP) tools such as ScispaCy have some limitations:
• they can not be used in diverse domains, they can not leverage the relationships between entities, and their reference knowledge can not be extended and updated.
• The combination of general-purpose Large Language Models (LLMs) and domain ontologies allows us to address these issues: LLMs can be driven by the continuously updated knowledge incorporated by the ontology and leverage its relational structure.
• To reach this goal, we can deploy a flexible end-to-end process for NED, including multiple phases involving LLMs and domain ontologies, such as Named Entity Recognition (NER), NED Candidate Selection, and NED Candidate Disambiguation.
• To fully leverage the capabilities of LLMs combined with the graph dimension of domain ontologies, the disambiguation phase is divided into three different stages:
1. Shortest-path detection.
2. Path-to-text translation.
3. Textual path summarization.
• Future NED applications can leverage this framework and adapt to different domains, which are characterized by rich ontologies describing the relational nature of their specific entities.

FAQ