Tabular data, often found in CSV files, is essential for data analytics workflows. Understanding this data in a semantic context, known as Semantic Table Interpretation (STI), is critical but challenging due to issues like label ambiguity. Consequently, STI has garnered significant attention in recent years. To evaluate STI systems efectively, robust benchmarks are needed. Most existing large-scale benchmarks originate from general domain sources and emphasize ambiguity, whereas domain-specific benchmarks tend to be smaller. This paper presents KG2Tables, a framework designed to create large-scale domainspecific benchmarks from a Knowledge Graph (KG). KG2Tables utilizes the internal hierarchy of relevant KG concepts and their properties. As a proof of concept, we have developed extensive datasets in the food, biodiversity, and biomedical domains. One of these datasets was used in the ISWC 2023 SemTab challenge, and the rest have been integrated into SemTab 2024.
CEUR ceur-ws.org
Semantic Table Interpretation (STI) has recently witnessed increasing attention from the community [1]. The goal of this process is to map individual table components, e.g., columns and cells, to entities and classes from a target Knowledge Graph (KG) such as Wikidata [2], or CEUR Workshop Proceedings (a) Horizontal Relational Table. challenge. We believe this is due to domain-specific challenges that general-purpose systems are ill-equipped to handle or require extensive tuning or training data.
State-of-the-art STI tasks propose ways to annotate tabular data semantically and, thus, facilitate a potential transformation into a KG. We summarize them as follows: 1) Cell Entity Annotation (CEA) links a table cell value to a knowledge graph (KG) entity. 2) Column Type Annotation (CTA) assigns a semantic type to an entire column. 3) Column Property Annotation (CPA) connects a column pair (subject-object) to a semantic property from the KG. 4) Row Annotation (RA) maps an entire row to a KG entity, difering from CEA as the subject column might be missing. 5) Topic Detection (TD) classifies the table into a topic, such as a semantic class. Figure 1 gives an overview of the five most common STI tasks in two table types Horizontal Relation Tables in Figure 1 (a), and Entity Tables in Figure 1 (b). The former includes a set of entities row-wise. The latter represents a single entity with a list of its properties. The solution indicates that Wikidata is the target KG.
In this paper, we introduce KG2Tables2, an STI benchmark generator that constructs both horizontal relational tables and entity tables, given a list of domain-specific concepts from Wikidata. This Zenodo dump3 shows the code we refer to in this paper and lists six benchmarks we created using KG2Tables from three domains: food, biodiversity, and biomedical.
KG2Tables accepts a list of related domain concepts in a CSV file, and constructs a tree structure for these concepts. In Wikidata, domain concepts form a graph structure since the KG allows that. However, KG2Tables process each relevant concept only once. We construct the respective 2https://github.com/fusion-jena/KG2Tables 3https://zenodo.org/records/10285835 1 3 4
tree structure using the internal hierarchy of the input concepts. e.g., in Wikidata, we have included all instances and subclasses via wdt:P31, instance of and wdt:P279, subclass of. We use the term “Children” to generalize related instances or subclasses. This tree structure will be diferent in DBpedia. In that case, the internal hierarchy is determined via rdf:type only. We applied a deduplication step since the overall instances and subclasses may overlap. Such overlap may also occur across diferent levels of the tree.
Figure 2 depicts the approach we developed to construct domain-specific benchmarks. It
starts with the children of domain concepts, i.e., based on the current level of the recursion, and
it consists of four steps: (
In this paper, we presented KG2Tables, a code generator that creates domain-specific tabular data benchmarks for Semantic Table Interpretation (STI) tasks. It uses the internal hierarchy of related concepts in a target Knowledge Graph (KG) to generate two types of tables: horizontal and entity tables. KG2Tables addresses five common STI tasks and was tested in three domains: Food, Biodiversity, and Biomedical. While our examples use Wikidata, KG2Tables is adaptable to any KG through SPARQL query modifications. KG2Tables accepts and parses a list of given domain concepts, thus validating and ensuring the domain specificity of the resulting dataset using either a data-driven approach or a reuse of existing domain-specific classes.