Food Composition Table (FCT) or Food Composition Databases (FCD) are composed of tables that describe foods and its composition. During SemTab 2022, we proposed a Food Composition Table dataset that we called TSOTSATable dataset. In this paper, we present how the annotation of this dataset is being done using Wikidata, FoodOn and Open Research Knowledge Graph (ORKG). The extracted tables are annotated using Wikidata and FoodOn. The scientific papers from which knowledge are extracted are annotated using Open Research Knowledge Graph. The annotation consists of matching the cells of the food tables to the Wikidata Knowledge Graph and FoodOn ontology, the matching of elements of the table describing the scientific papers to ORKG resources, the detection of the type of elements of each columns (CTA) and the matching of the relations between columns to ORKG properties. During the annotation, we found that many tables were not relevant to the food domain. Thus, we added a new annotation task that is Irrelevant Table Detection (ITD). This consists for a domain dataset, to determine the tables that are not relevant to this domain.
GitHub1 and Google Collaboratory2. a video showing how we automatically extract tables from PDFs is also available3.
A study of the TSOTSATable dataset showed that the data extracted from scientific papers were more complete than the ones extracted from Zenodo. On the other hand, Zenodo contains a lot of tables not relevant to the domain of food science and nutrition. Thus, we decided to start the annotation of tables extracted from scientific papers.
In this paper, we describe how this dataset is being curated using Wikidata, FoodOn [
In the rest of this paper, we present the annotation in Section 2, the overview of the dataset in Section 3 and the conclusion in Section 4.
The curation started by the annotation of an excerpt of the TSOTSATable dataset, containing data extracted from papers of the "Journal of Food Composition and Analysis". To this end, two tools were designed: one for the automatic annotation using Wikidata and FoodOn and one for automatic annotation using Open Research Knowledge Graph. In the following sections, a short overview of the ontology and Knowledge Graphs used for the annotation is presented (see Section 2.1). Thereafter, the diferent annotation tasks are presented in Section 2.2 and ifnally, the annotation process is presented 2.3.
To annotate the TSOTSATable dataset, we choose diferent vocabularies corresponding to the diferent domains covered by the dataset. The dataset is composed of two types of tables: the tables describing foods and their composition (food science and nutrition domain) and a table containing the list of scientific papers (digital library domain) from which the tables are extracted.
The first step of the annotation consists of the selection of relevant KGs or ontologies to use. Given that this is a food domain, we compared several excerpts of the dataset with several food 1https://github.com/Neuralearn/pdf-to-excel 2https://colab.research.google.com/drive/1gOPBCVO9VtKcoIewXyr_6nNoxo1Bkqbz 3www.youtube.com/watch?v=HZh31OGiQRQ 4https://github.com/jiofidelus/tsotsa/tree/main/TSOTSATable_dataset/rawData 5https://github.com/jiofidelus/tsotsa/tree/main/TSOTSATable_dataset/annotatedData 6https://github.com/jiofidelus/tsotsa/tree/main/TSOTSATable 7https://github.com/jiofidelus/SemTabTable-Papers/tree/main/sourceCode 8https://creativecommons.org/licenses/by-sa/4.0/ ontologies hosted on Bioportal. We used the ontology recommender of Bioportal to search for the most appropriate ontology in the food domain that can be used to annotate the dataset. Fig. 1 presents an example of the use of ontology recommender to search for the most appropriate ontology.
From the ontology recommender, we found that FoodOn is the most appropriate ontology to
annotate the dataset. FoodOn9 [
On the other hand, we manually searched for a set of terms in the Wikidata KG using its search engine and we found that Wikidata contains a lot of relevant annotations. Wikidata11 is amongst the most popular KGs in the world. It is involved in the SemTab challenge since the challenge was launched in 2019. Once we found that Wikidata contain relevant annotations, we built an automatic tool for the annotation of the dataset using the Wikidata MediaWiki API12. The same disambiguation process used during the annotation of the dataset by FoodOn ontology was used to select amongst the entities the ones that may match to the elements of the table.
Concerning the annotation of scientific papers from which tables are extracted, we rely on ORKG because we have a great experience on the use of this KG for annotating scientific papers. Open Research Knowledge Graph13 (ORKG) is a scholarly KG used to acquire, publish and process structured scholarly knowledge published in the scholarly literature. It is built according to the principle of Open Science, Open Data, and Open Source. We used the automatic annotation feature of ORKG to annotate all the scientific papers from which the tables were extracted.
During the annotation process, we found that many tables were not relevant to the domain of nutrition. On the other hand, ORKG is based on an ontology. This ontology describes a research paper as composed of paper metadata and its semantic description. The semantic description consists of (1) assigning ORKG classes to the diferent key-insights extracted, (2) defining several properties for comparing research contributions, (3) and comparison tables of research contributions dealing with the same research problem. From this ontology, instances are instantiated during the paper annotation. Based on this, and the annotations tasks generally proposed by SemTab challenge, we defined the following annotations tasks: • Column Entity Annotation (CEA): This is to match each cell of the tables to the ontology/KG entity. The entities in the tables extracted were matched to Wikidata and FoodOn. Concerning scientific papers, we used ORKG resources, which can be a class, an instance, or a property. • Column Type Annotation (CTA): this consists of the assignment of classes from the ontology and KGs to columns of the tables. • Column Property Annotation (CPA): This is the assignment of a property to the relationship between two columns in tables. We found it dificult to identify properties amongst columns of the tables. In efect, the majority of these tables contain numbers in the cells and sometimes in the headers, abbreviations of food components (for instance, k=potassium, Fe=Fer, Mn=Manganese, etc.) The fact that the columns are filled with only 11https://www.wikidata.org/ 12https://www.wikidata.org/w/api.php 13https://orkg.org/ numbers make it dificult to build an automatic tool for determining the relations between two columns. Thus, in the current version of the dataset, this annotation task concerns only the scientific papers. • Irrelevant Table Detection (ITD): this task consists of the detection of tables that are not relevant to the domain of Food and Nutrition. It should be noted that this task is currently manual.
The raw data contained the following types of files: • TSOTSATable source: this is a CSV file containing information on scientific papers from which the tables were extracted. It is named 0 − . in the dataset. • TSOTSATable files: these are the CSVs files containing the tables extracted from the scientific papers. Each file is named using a unique identifier. The latter allows linking the file to the corresponding source file in the knowledge source file. The file name of each table is obtained using his ID in the data source plus a number denoting the order of its apparition in the data source. For instance, the 3 table in a scientific paper that has the = 12 is named _3.
Concerning the annotation, we created three folders corresponding to the three vocabularies used to annotate the TSOTSATable dataset. Each folder contains diferent target annotations: • TSOTSATable_CEA: this is the file containing the CEA of the tables. • TSOTSATable_CTA: this is the file containing the CTA of the tables.
• TSOTSATable_CPA: this is the file containing the CPA of the tables.
A subset containing 251 tables were annotated and published on Zenodo repository [
Food Composition tables were annotated using Wikidata and FoodOn and the scientific papers from which data is extracted was annotated using Open Research Knowledge Graph. Table 1 presents the number of entities and types annotated using Wikidata and FoodOn. An expert in Food Science and Nutrition was invited to select these annotations randomly and verify their relevance. Concerning scientific papers, around 500 terms were annotated using ORKG.
In a recent work, we extracted Food Composition data from scientific papers and we built
a tabular dataset with it [
Future work consists of finalizing the annotation and using this dataset to build a TSOTSAGraph, a Food Composition Knowledge Graph.