acceptodds
Under review as a conference paper at ICLR 2027

MHKG: An Ontology-Guided, Evidence-Linked Knowledge Graph for Metal Exposure and Health

Abstract

Preventing metal exposure and related diseases is an important public health issue. However, related research involves various types of information and the relevant evidence is scattered across a large number of documents. The same concepts often appear in abbreviated forms, aliases, or different granular terms, which limits cross-study retrieval and evidence integration. We bulid a Metal-Health Knowledge Graph (MHKG) dataset, which organizes the literature extraction results into a queryable and auditable structured resource. MHKG links exposure factors with biological responses, health outcomes, and intervention measures, and the relationships retain links to the original articles and supporting paragraphs. We construct the dataset using a unified schema to organize heterogeneous entities and relationships. Meanwhile, cross-document entity standardization and evidence verification make the graph more coherent and the content more accessible for auditing. MHKG provides a structured and auditable resource for studying occupational metal exposure, supporting evidence-oriented retrieval, knowledge exploration, as well as future knowledge-enhanced applications.

Primary area:
datasets and benchmarks
Keywords:
Metal Exposure · Knowledge Graph · Ontology-Guided · Evidence-based

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_RfUvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an ontology-constrained knowledge graph for occupational metal exposure and its associated health conditions. The knowledge graph is constructed through a 4 pipeline comprising data mining and document chunking; ontology-guided triplet extraction across document chunks, supported by an expert-validated ontology and an LLM; integration of retrieved knowledge; normalization of concepts and relations; and evidence linking. The resulting knowledge graph has potential applications in occupational health knowledge retrieval, knowledge discovery, and the identification of insights that may inform strategies for improving occupational health outcomes.

Critical strengths
  • The authors propose an ontology-constrained knowledge graph with evidence support to characterize metal exposure and its associated health effects. The key novelty of this work lies in its approach to grounding knowledge across the literature, particularly through the use of an ontology schema validated by domain experts to improve the consistency and reliability of knowledge extraction. In addition, the authors provide a dataset to support the proposed approach and facilitate further research.
Critical weaknesses
  • The main weakness of this paper is that it contains no evaluation of the constructed ontology-constrained knowledge graph. The authors cannot simply describe a schema without any evaluation on specific tasks for a dataset and benchmark type of contribution. A comparison with existing biomedical knowledge graphs (e.g Semnet, Semnet-2, OptimusKG,..) must be evaluated to enforce their claim about the necessity of the proposed MHKG graph. Under the current state, the following contribution falls short for an ICLR conference.

  • Overall the authors should revise the presentation of their framework and the wording, it is easy for a reader to get confused within section 3. Also, a lot of Appendix content is mentioned in section 3 but given that the whole manuscript is 8 pages long, authors could use that space more wisely and bring forward prompts or figures to ease the reading. “LLM” is mentioned multiple times in the pipeline for candidate resolution (stage 3) and triplet extraction (stage 2) yet no precision about the nature of the LLM nor the analysis of results/error analysis are given in the main study. In section, 3.2, some errors are occurring :

    • (l.191-192) “Figure 2 shows the part of the ontology that we define.”

    • (l.279-282) sentence repeated twice.

    The following claim is made in the conclusion “In order to guarantee the quality of the graph, we construct an automated process covering pattern-constrained knowledge extraction, structure and evidence auditing, entity fusion and normalization, as well as relationship reconstruction.” With no concrete evaluation and error analysis, it is not demonstrated yet.

Questions for authors

Can the authors revise figure 2 ? (with the current state it adds little information compared to figure 3 for instance). Is it possible to provide an analysis of the LLM employed in stage 1 and stage 2 ? Any comparative study with multiple LLMs ? Any downstream task to evaluate your knowledge graph ?

Reviewer_uA1EOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents the Metal-Health Knowledge Graph (MHKG), a knowledge graph that captures metal exposure and health. The MHKG graph is assembled from approximately 1000 PubMed full-text articles, and the pipeline used to construct it is clearly described. The MHKG graph is released anonymously and contains 46,384 canonical entities, 60,663 canonical relations, and 76,041 relation-source records.

Critical strengths

The MHKG graph is a valuable resource for understanding metal exposure and its health effects.

The provenance design is sound: graph and evidence information are separated, so deduplication doesn't discard independent evidence, and every edge is traceable to a chunk of text.

Ontology, object properties, and domain-range constraints, and the extraction prompt are provided in the appendix.

Critical weaknesses

There is no evaluation of the knowledge graph constructed in this work. There is no quantitative evaluation in terms of precision/recall against a gold set, no entity-resolution accuracy, no evidence-match rate, no inter-annotator agreement for the human review stage, and no downstream task to illustrate the usefulness of the graph.

The paper has very limited novelty - the pipeline does not introduce any novel steps or components. It is assembled based on components that have been used before in the literature and cited in the paper itself (ontology-guided extraction similar to what is done in SPIRES/OntoGRC, SapBERT-style normalization, LLM-as-judge resolution). The contribution is just the resource/graph, although its quality is not evaluated.

Reproducible pipeline artifacts are not provided. There are lots of details that would be needed for reproducibility of this work or reuse for other works.

Questions for authors

Which LLM(s) were used for extraction, with what settings and cost? Were any outputs manually scored for precision/recall?

How were the 1,086 papers selected, and under what license are evidence spans released?

Reviewer_EA2fOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents MHKG, an ontology-guided knowledge graph for metal exposure and related health effects. The proposed approach integrates large language models, domain ontology constraints, entity normalization, cross-document entity fusion, and evidence provenance to construct a structured knowledge graph from scientific literature. The resulting resource contains 46,384 canonical entities and 60,663 relations, providing support for knowledge retrieval and analysis in occupational health.

Critical strengths

The paper addresses the relationships between occupational metal exposure and health effects, aiming to overcome challenges such as fragmented medical evidence, inconsistent terminology, and difficulties in cross-document knowledge retrieval. The research has potential practical value.

Critical weaknesses
  1. The proposed framework mainly integrates existing techniques, including LLM-based knowledge extraction, ontology constraints, without introducing clear technical innovations. The authors should further clarify the key differences and novel contributions compared with existing knowledge graph construction methods.
  2. Insufficient experimental evaluation. The paper lacks quantitative evaluations of entity extraction, relation extraction, and entity disambiguation accuracy. Furthermore, comparisons with existing methods and ablation studies of key components are missing, making it difficult to adequately validate the effectiveness of the proposed approach.
  3. Inconsistent statistics between the main text and Table 1. On page 7, the numbers of entity mentions and relation-source records are reported as 165,791 and 76,040, respectively. However, Table 1 on page 8 reports the corresponding numbers as 165,801 and 76,041.
Questions for authors

Could the authors provide at least one experimental result demonstrating the practical value of the proposed dataset or method?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.