Hallucinations of large language models (LLMs) commonly occur in domain-specific downstream tasks, with no exception in ontology matching (OM). The prevalence of using LLMs for OM raises the need for benchmarks to better understand LLM hallucinations. The OAEI-LLM dataset is an extended version of the Ontology Alignment Evaluation Initiative (OAEI) datasets that evaluate LLM-specific hallucinations in OM tasks. We outline the methodology used in dataset construction and schema extension, and provide examples of potential use cases.
(J. Jiang) ∗Corresponding author.
CEUR
ceur-ws.org
OAEI-LLM, which serves to measure the degree of hallucination by LLMs in OM tasks. The new dataset compares the original human-labelled results with LLM-generated results, classifies diferent types of hallucinations made by diferent LLMs, and records the information with a new schema extension. We believe this dataset would benefit research in understanding LLM hallucinations in matching tasks and thereby research in improving LLM-driven OM.
For each mapping (1, 2) ∈ , we consider that it is not an LLM hallucination if there exists (1 ′, 2 ′) ∈ such that 1 = 1 ′ and 2 = 2 ′. We categorise LLM mapping errors in three ways, as shown in Table 1.
The matching assessment process identifies the alignment from which a missing mapping is missing. It also determines a categorisation of incorrect mappings as follows: • False-mapping: LLMs map the entity to an irrelevant entity. • Disputed-mapping: LLMs map the entity to a relevant entity, but it does not align with the OAEI Reference. We call this mapping type “disputed” because we cannot guarantee the OAEI Reference is always correct. LLMs may discover a more precise matching entity. • Align-up: LLMs map the entity to its intended superclass. • Align-down: LLMs map the entity into its intended subclass.
In practice, the notions of relevant and intended are implemented by the chosen matching assessment process. We use an LLM-based evaluator for matching assessment, and directly query the LLM for relevance and superclass and subclass relationships. At this time we cannot recommend a particular LLM to be used.
The OAEI tracks use a general alignment format derived from the Expressive and Declarative Ontology
Alignment Language (EDOAL) [
We extend the current EDOAL mapping schema to record the new information related to LLM
hallucinations. The Simple Standard for Sharing Ontological Mappings (SSSOM) [
Code Snippet 1: An example of an OAEI-LLM reference alignment using our proposed extended EDOAL
mapping schema. Text generated by AI tool Agent-OM [
Diferent LLMs can perform diferently on the same OM task. However, running an LLM-based matcher on a traditional OAEI dataset can only observe the diferences in precision, recall, and F1 score. Using the extended version of OAEI-LLM, we can additionally quantify the LLM errors made by counting each of the diferent types of errors we have identified. Therefore, it becomes possible to understand the tendency of LLMs to generate incorrect answers or to provide relevant but not precise answers. This could also be a valuable supplement to enhance an LLM leaderboard for OM tasks. Indeed, this same additional information on matching failures could also be helpful if adopted more widely by non-LLM ontology matchers.
Fine-tuning is a common approach used in LLMs to inject domain knowledge, but a significant amount of data typically needs to be provided. Unlike the traditional OAEI datasets, which only provide information on matched entities, OAEI-LLM provides extensive information identifying when and how LLM hallucinations occur, and even the mismatched entities caused by LLM hallucinations. This could be used to generate high-quality training data to fine-tune LLMs.
In this paper, we only demonstrate the simplest one-to-one equivalent mappings. More generally,
the reference alignment can be incomplete or contain one-to-many mappings, while the matching
type can also go beyond equivalence to include subsumption matching and complex matching. Such
sophisticated cases may cause additional types of LLM hallucinations. Some LLM hallucinations could
be mitigated by providing more ontology structure as context to the LLM, for example, by employing
the locality principle of traditional matchers [
In our work to date we use a moderately simple LLM-based evaluator to perform the matching assessment. The assessment may vary depending on the LLM used. A more sophisticated matching assessment could use a voting process by multiple LLMs or human expert judgement to generate a more reliable OAEI-LLM benchmark dataset.
The OAEI tracks can be divided into three categories: schema matching (i.e. TBox matching), instance matching (i.e. ABox matching), and interactive matching (i.e. matching that involves user interaction). Correspondingly, OAEI-LLM will also have three variants. • OAEI-LLM-T will focus on comprehending LLM hallucinations in the TBox, with a specific interest in the formal use of terminologies in ontology design and implementation. • OAEI-LLM-A will focus on addressing LLM hallucinations in the ABox, in which an ontology contains a schema together with many instance entities (similar to ontology-based knowledge graphs). • OAEI-LLM-I will focus on understanding how user interaction could be used to mitigate LLM hallucinations and estimating the level of user interaction required for LLM-based OM tasks. Acknowledgements The authors thank the reviewers for providing helpful comments. The authors thank the Commonwealth Scientific and Industrial Research Organisation (CSIRO) for supporting this project. According to the OAEI data policy (retrieved December 1, 2024), “OAEI results and datasets, are publicly available, but subject to a use policy similar to the one defined by NIST for TREC . These rules apply to anyone using these data.” Please find more details from the oficial website: https://oaei.ontologymatching.org/doc/oaei-deontology.2.html.