This paper presents MTab, an automatic tool for tabular data annotation with knowledge graphs. MTab tool could provide helpful information for tabular data such as structural annotations (e.g., table headers, subject column) or semantic annotations with knowledge graph concepts from Wikidata, DBpedia, and Wikipedia (e.g., cells with entities, columns with types, and column pairs with properties). The tool supports multilingual tables and could process many table formats such as Excel, CSV, TSV, markdown tables, or a pasted table content. MTab achieves impressive empirical performance on many datasets: 1st on HardTable CEA, CTA, CPA tasks, BioTable CTA, CPA tasks, and HardTablesR3 CPA task. Additionally, the system also got the 1st on usability track with advanced features: easy-to-use, generic solution, welldesigned user interface. MTab's graphical interface, public APIs, documents are available at https://github.com/phucty/mtab_tool.
The Open Data movement has made many valuable tabular resources available on the Internet and Open Data Portals. However, due to insu cient data descriptions, various data formats, and terminology issues, the use of tabular data in applications is constrained. Many tabular data lack a description, or the description is not adequately described the data. Table structure and layout are also lacking in many tabular resources. Furthermore, many tables do not employ conventional vocabularies, such as multilingual expressions, abbreviations, ambiguous or many misspellings, and encoding issues. To improve tabular data usability, it is necessary to have a tabular data annotation system capable of providing explicit information about table content.
This paper introduces MTab, an automatic tool that generates structural and semantic annotations for tabular data. MTab tool, as illustrated in Figure 1, Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEA Cell CTA Column CPA Relation
Entity Q1490 (Tokyo)
Type Q5119 (capital) Property P1082 (population)
could provide helpful information for tabular data such as structural annotations (e.g., table headers, subject column) or semantic annotations with knowledge graph concepts from Wikidata, DBpedia, and Wikipedia, e.g., a cell with entity annotation (CEA task), a column with type (or class) annotation (CTA task), and a column pair with property annotation (CPA task). The tool supports multilingual tables and could process many table formats such as Excel, CSV, TSV, markdown tables, or a pasted table content.
MTab archives impressive performance on many datasets: 1st on HardTable CEA, CTA, CPA tasks, BioTable CTA, CPA tasks, and HardTablesR3 CPA task. Additionally, the system also got the 1st on usability track with advanced features: easy-to-use, generic solution, well-designed user interface. The user could access MTab's graphical interface, APIs, documents at https://github.com/ phucty/mtab_tool. 2
an aggregation of multiple cross-lingual lookup services and probabilistic
graphical models [
In SemTab 2020, the matching target knowledge graph is Wikidata
including new set of di culties such as larger-scale of data, graph shifting, rich and
complex data schema in Wikidata. Beside the generated tabular data from
Wikidata, there was a new manually curated dataset (tough tables [
However, most solutions or systems are not available to use or require
extensive con guration, setup, high computing power, or high time complexity [
This section describes MTab tool, started with the system assumptions in Section 3.1, then the overall framework is described in Section 3.2. 3.1
Assumptions Assumption 1 MTab tool is built on a closed-world assumption. It means that the tool could return incorrect answers if table elements are not available in the knowledge graph.
Assumption 2 We assume that the input tables are horizontal relational types. A horizontal relational table contains semantic knowledge graph triples in [subject, predicate, object]. The table also has a subject column containing entity names and the relation between the subject column and other columns representing the predicate relation between the entities (subject) and attribute values (object).
Assumption 3 We assume that all the cell values of the same column have the same data type, and the entities related to cell values are of the same type. Assumption 4 MTab tool treats input tables independently. -Text (string) -File (CSV, TSV, EXCEL) -Table Object
Preprocessing In this paper, we focus on the usability factor of the annotation system. So, we implement the MTab tool to support multilingual tables and could process various table formats. The system e ciency also is an important concern of the implementation so that we optimize the annotations run time by about 1.52 sec/table on average (tested on SemTab 2020 dataset). Moreover, we also provide graphical interfaces to visualize the annotation results as in Section 4.
The overall framework of the MTab tool is described in Fig. 2. We build WikiGraph, which is an integrated knowledge graph from Wikidata, DBpedia, and Wikipedia as in Section 3.2.1. The annotation procedure is started with data preprocessing as in Section 3.2.2. Then, the system performs data type prediction, header prediction, and subject column prediction as in the structural annotations section (Section 3.2.3). Finally, MTab performs semantic annotations as in Section 3.2.4. 3.2.1 Knowledge Graph We build a WikiGraph from the dump data of Wikidata, Wikipedia, and DBpedia as the target knowledge graph for the annotation tasks. With the dump data on 1 January 2021, we extracted 91.2 million entities and 249.3 million entity labels in multilingual, including entity labels, aliases, other names, redirect entity labels, and disambiguation entities. We also extracted 3.5 billion triples in WikiGraph. Additionally, WikiGraph will be updated frequently based on the future released dumps of knowledge graphs (Wikidata, Wikipedia, and DBpedia). 3.2.2
Preprocessing
Data Type Prediction The system rstly predicts a table cell's data type into
either non-cell (empty cell), literal, or named-entity (NE). We use the
pretrained SpaCy models [
Next, the system predicts a table column's data type into either a non-match
column (empty column), a literal, or a named-entity column. The column data
type is derived from the majority voting of all cell data types in this column.
Header Prediction We use simple heuristics to predict table headers as follows.
{ Table headers could be located in some of the rst rows of a table.
{ If the list of data types of the header candidate row di ers from most data
types of the remaining rows, the candidate is the table header. For example,
the list of data types of header candidate (row) is [named-entity,
namedentity, named-entity], while the list of the majority data type of remaining
rows is [named-entity, literal, literal].
{ We also found that the length of header text is empirically shorter or longer
than the remaining data rows. If the length of values of the header candidate
row is less than the 0.05 quantile or larger than the 0.95 quantiles of the
length of the value of remaining rows, the candidates are the table header.
Subject Column Prediction We adopt the heuristics proposed by Ritze et al. [
{ A column is a subject column when its data type is a named-entity type. { The average cell value length is from 3.5 to 200. We also add a restriction that only considers non-header cells since the length of table headers could di er from the remaining cells. { The subject column is determined based on the uniqueness score as an increased score for columns with many unique values and reduces the score for columns with many missing values. The subject column is the highest unique score column. If we have many columns that have the same score, the left-most column is chosen. 3.2.4
Semantic Annotations Matching Target Prediction: MTab automatically predicts the matching targets based on data types when the input does not have matching targets. The CEA matching targets are the table cells whose data types are named entity types. The CTA matching targets are columns so that the column data types are named entity types. The CPA matching targets are the relation between the subject column and the remaining table columns.
Entity Search: We perform entity candidate generation for each table cell with
the entity search modules. MTab tool provides the three entity search
modules, i.e., keyword search, fuzzy search, and aggregation search1. We
implement the keyword search using BM25 algorithm with the hyper-parameters as
b = 0:75; k1 = 1:2. The fuzzy search is implemented using Damerau{Levenshtein
edit distance. We perform candidate ltering and hashing with pre-calculating
entity label deletes as the Symmetric Delete algorithm [
Post-Processing: We calculate context similarities with the value-based matching
between statements of entity candidates in the subject column with table row
values. Finally, generate the annotations for entities, properties, and types based
on majority voting of context similarities [
4.1
Entity Search The entity search interface is available at https://mtab.app/mtabes. Fig. 3 depicts an example of entity search with the query of \2MASS J10540655-0031018". MTab tool supports multilingual search so that users could type entity name expressed in any language. 1 Entity Search Documents: https://mtab.app/mtabes/docs The table annotation interface is available at https://mtab.app. Users could submit table les in various table formats, expressed in any language to MTab API, or copy data content and paste it to the interface. Then, users could tap the \Annotate" button to get the annotation results.
Fig. 4 illustrates an annotation example of a SemTab dataset's table. MTab took 0.49 seconds to annotate a pasted table from the text box (left picture). The photo on the right is the annotation results. The table header is in the rst row, and the subject column is in the rst column. Entity annotations are in red and located below the table cell value. The type annotation is in green and located in the \Type" column. Finally, the relations between the subject column and other columns are in blue and located in the property column. This paper presents the MTab tool for table annotation with Wikidata, DBpedia, and Wikipedia knowledge graphs. MTab tool achieves promising performance on many datasets of SemTab 2021. Moreover, the system also got the rst rank of usability track.
In the future work, we will focus on e ciency improvement of the MTab tool by processing only small parts of table content and continues expanding until there is no di erence in the annotation results. Another direction is building downstream applications based on MTab's annotations, such as question answering and data analysis.
The research was supported by the Cross-ministerial Strategic Innovation Promotion Program (SIP) Second Phase, \Big-data and AI-enabled Cyberspace Technologies" by the New Energy and Industrial Technology Development Organization (NEDO). 2 SemTab 2021 Leaderboards: https://www.aicrowd.com/challenges/semtab-2021/ leaderboards