An increasing amount of data is shared on the Web through heterogeneous spreadsheets and CSV les. In order to homogenize and query these data, the scienti c community has developed Extract, Transform and Load (ETL) tools and services that help making these les machine readable in Knowledge Graphs (KGs). However, tabular data may be complex; and the level of expertise required by existing ETL tools makes it di cult for users to describe their own data. In this paper we propose a simple annotation schema to guide users when transforming complex tables into KGs. We have implemented our approach by extending T2WML, a table annotation tool designed to help users annotate their data and upload the results to a public KG. We have evaluated our e ort with six non-expert users, obtaining promising preliminary results.
An increasing amount of data is shared on the Web by multiple organizations using Excel and CSV formats. Content creators usually prefer to use tabular data because it is simple to generate, manipulate and visualize by humans; and there is a signi cant number of tools to help explore and edit the contents of spreadsheets. These data need to be properly understood by others, and hence documentation (e.g., variables captured, provenance, usage notes, etc.) is usually included in auxiliary les or the spreadsheets themselves. As a result, many of these spreadsheets have comments, clari cations, notes and references to other les explaining how to interpret the information contained in them.
In order to convert tabular data to a machine readable format, the Semantic
Web community has created Extract, Transform and Load (ETL) tools (e.g.,
[
In this paper we describe an approach to help non-experts transform their
data into a structured representation through dataset annotations. Our
contributions include 1) a dataset annotation schema that helps generating templates
for translating datasets into KGs; 2) an extension of the T2WML dataset
annotation tool [
In order to assess our approach, we conducted a preliminary evaluation with 6 users unfamiliar with Knowledge Representation or Semantic Web technologies, who were able to describe and integrate their annotated datasets as a KG. 2
We focus on those datasets that are not straightforward to map into a structured representation. Consider for example Table 1, which depicts the food prices in di erent regions of Ethiopia at di erent points in time. The table has a time series for the price value of di erent items at di erent dates, a repeated column with the item being described (ignore), the item category and di erent information about the region where that item was produced. The dataset has also some missing values and labels marked as "unknown", which we may want to skip. This dataset is representative of many open datasets with statistical/time series information, and presents some interesting challenges: { The main subject of the annotation is not clear: The table describes the price of an item in a location at a particular time. One possibility would be to assert that the subject of the triple is the item (e.g., Sorghum), having the price column as the object; and the rest of the columns as quali ers. Alternatively, we could use the country (or the administrative name) as main subject, as it is relevant to create aggregates. Finally, we could also generate a blank node or URI to link together the contents of all columns. { Repeated columns and incomplete cell values: Spreadsheets contain empty values, cell values (or columns) that need to be ignored and comments (specially at the beginning and end) that complicate processing the data. { Distinguishing variables from quali ers: In some cases, it may be difcult to distinguish whether a column is the object associated to a subject or whether it is qualifying other values. For example, if Table 1 contained a \quality" column, it could be interpreted as a new variable, or as a quali er indicating the quality of the information source.
Other problems that frequently occur include complex headers that sometimes join the meaning of two columns (e.g., values and units, location and country, etc.); comments in certain parts of the le; or critical missing information, which is externally provided to the le. For example, there are cases where the year in which the le was produced is part of the title of the CSV instead of a column with a constant name.
All these challenges make the automated annotation of datasets a challenging problem. We need an approach for incorporating user feedback from content creators or domain experts that are familiar with these datasets, but do not necessarily know Semantic Web technologies or mapping languages. 3
Our approach has three main elements: an annotation schema, which we use to create mapping templates (Section 3.1); an extension of the T2WML tool to use the proposed vocabulary when converting datasets into KGs (Section 3.2); and an approach to integrate the mapped results with a reference KG (Section 3.3). 3.1
We have created a simple annotation schema1 by adding a set of headers to the start of spreadsheet as shown in Table 2. The schema was designed to capture basic metadata and to be easy to understand by content creators unfamiliar with Semantic Web technologies. Therefore we capture 1) the dataset identi er to be used when referring to the dataset; 2) the role of each column, i.e., whether it is a variable, a unit or a quali er (location, time or other); 3) the type of each column, i.e., whether the column should be the main subject, the format used to represent dates, whether the variables to annotate are a number or a string, etc.; the 4) column description in case users need to clarify any of the columns to the persons reusing the data; 5) the variable name represented in a column, as in some cases the headers used are di cult to understand; 7) the variable unit; and 8) the header where the original dataset headers start.
An example of our schema is represented in Table 2 by annotating Table 1. As shown in the example, it is not necessary to complete all headers, in case the information is not known or missing. 3.2
We have implemented our approach by extending the Table to Wikidata
Mapping Language Tool (T2WML) [
T2WML is designed for the Wikidata data model [
Figure 1 shows how the T2WML extension would process a dataset similar as
the one shown in Table 2. T2WML recognizes the di erent headers annotated in
the spreadsheet to generate a template YAML following the T2WML mapping
language [
Once users nish annotating a dataset, they can export their results in a structured format like RDF. However, creating a KG with this information still needs signi cant expertise. Therefore, we have created the USC Datamart, a catalog which includes 1) key dataset metadata (i.e., creator, variables included, etc.) of the datasets uploaded by users; and 2) the contents of those annotated datasets (with variables and their quali ers like location, date, units, etc.). We have extended T2WML to allow uploading the structured results into the USC Datamart through a dedicated API2, enabling users to share their results online (see the Upload to Datamart button in Figure 1). Each dataset has its own id, which can be updated with new data. This way if a time series consists on a set of spreadsheets with the same structure for di erent regions, they can all be uploaded using a similar mapping template and the same dataset id.
With the USC Datamart, users may retrieve dataset metadata (e.g., to nd out which variables does a dataset include, or the time period they cover) and once they nd the desired information they can download it as a table for their own analysis. A usage example of the Datamart API can be seen online.3 4
In order to assess our approach, we performed a preliminary evaluation with six users. None of these users were familiar with Semantic Web technologies or mapping languages, but three of them had expertise in data science and scripting languages like Python or R. All users received a training in T2WML (one hour) to understand the main capabilities of the tool and the annotation schema.
The goal of the evaluation was to assess if users could understand the proposed schema and use it in T2WML to annotate and upload datasets similar to the one described in Table 1 (with their corresponding challenges). The evaluation included three datasets with di erent indicators (economic, demographic, production, etc.) in African countries. Each dataset was assigned to two di erent users. As a result, all users were able to upload their datasets successfully to the USC Datamart, with on the y corrections for one of the datasets where the temporal information was part of the title of the le, instead of in its contents.
When asked for feedback, users reported that the proposed annotation approach was preferable to creating their own scripts for data cleaning, but they claimed that it can be di cult to 1) align their own terminology to Wikidata and 2) understand the di erence between a variable and their corresponding quali ers. This means that while our approach successfully tackled the rst two challenges described in Section 2 (annotating the main subject and incomplete columns), additional work is required to guide users in the annotation process. We are improving tutorials and documentation to address these issues.
Garijo et al.
A signi cant number of tools (e.g., [
Other work has focused on automated table understanding to label the
structure of tables without having users to annotate datasets themselves (e.g., [
In this paper we have described our approach for allowing content creators to describe their own datasets to transform them into structured KGs. Our preliminary results show that users are able to understand and use our schema for annotating their datasets easily, enabling them to create and populate an existing KG. Our next step will focus on incorporating table understanding approaches which will make the process easier for users describing their own data.