Skip to yearly menu bar Skip to main content


Poster
in
Workshop: Table Representation Learning Workshop (TRL)

DynoClass: A Dynamic Table-Class Detection System Without the Need for Predefined Ontologies

Haonan Wang · Eugene Wu · Kechen Liu · Jiaxiang Liu

Keywords: [ Ontology generation ] [ Large Language Models (LLMs) ] [ Table-class detection ] [ Taxonomy ]


Abstract:

Table-class detection plays a crucial role in various data tasks. Traditional approaches typically depend on predefined ontologies such as DBpedia, but these are often insufficient for domain-specific or evolving datasets. In response, we present DynoClass, a novel table-class detection system that leverages the power of large language models (LLMs) and eliminates the reliance on external ontologies. DynoClass uses LLMs to generate table classes and descriptions directly from sample data and existing documentation, dynamically constructing hierarchical ontology classes. This approach matches the performance of traditional methods while eliminating the need for predefined ontologies.

Chat is not available.