Web tables comprise a rich source of factual information. However, without semantic annotation of the tables' content the information is not usable for automatic integration and search. We propose a methodology to annotate table headers with semantic type information based on the content of column's cells. In our experiments on 50 tables we achieved an F1 value of 0.55, where the accuracy greatly varies depending on the used ontology. Moreover, we found that for 94% of maximal F1 score only 20 cells (37%) need to be considered on average. Results suggest that for table disambiguation the choice of the ontology needs to be considered and the data input size can be reduced.
Tables on web sites or in scienti c papers represent a valuable source of
information for the human reader. For machines the story is di erent: Although tables
are a structural representation of knowledge, the information itself is
meaningless to machines { unless it is enriched with semantic information. The Semantic
Web, and speci cally the Linked Open Data initiative provide means for
representing any kind of knowledge semantically. If tables were enriched semantically
a variety of new applications could evolve, as is the idea of Google Fusion
Tables [
In this paper we move one step towards fully-automatic semantic table
annotation. We propose a simple algorithm to annotate table headers with semantic
types based on the types of all cells in that column. Further, we investigate the
in uence of the number of cells on the accuracy of the header-type inference.
Current approaches in table annotation pursue a collective approach, i.e., the
annotation model encompasses entities, types and relation between types. The
underlying assumption is that columns and the cells in a column have some
relation in common which is modeled in the semantic knowledge base. Limaye et
al. [
The general assumption behind our annotation algorithm is that the cells of a table column belong to one supertype, which we want to infer. We make no assumptions of interrelationships between columns, i.e. all columns are treated separately. Further, we assume that the tables do not have merged cells.
Let li 1 < i n be the labels of non-header cells i and Ei = feikg is the set
of all possible semantic meanings of label li. The set Tik is the set of all type
labels assigned to entity eik. The annotation of table headers is performed in
three steps (compare Figure 1):
Step 1 { Cell entity annotation: For each cell label li we derive a list of k
possible entity candidates Ei using a search-based disambiguation method [
Step 2 { Entity-type resolution: For each entity candidate eik in Ei a set of types is retrieved by following the rdf:type and dcterms:subject relations yielding the set of types Tik.
Step 3 { Type aggregation: The types assigned to the table header are the t types that occur most frequently in the set of all types of all cells Si Sk Tik. We set t = 1 in our experiments, e.g. only use the most frequent type as result. 4
Data Set. In our experiments we used dbpedia as knowledge base with type relations rdf:type and dcterms:subject from the Dublin Core Metadata Ontology1. We evaluated our approach on 50 tables extracted from Wikipedia pages 1 http://dublincore.org/documents/dcmi-terms/
including all tables mentioned in Limaye et al. [
Table Length. In a second experiment we assessed the in uence of the number of cells on the accuracy of table header disambiguation. From all 192 columns we randomly selected k cells for the cell-entity annotation step and assessed the header disambiguation accuracy using the DublinCore vocabulary. We repeated the experiment 10 times with di erent randomly selected cells for each k 2 f1; 2; ::7; 8; 10; 12; 15; 20g. Figure 2 shows precision, recall and F1 measure averaged over all runs. As expected for small numbers of cells the performance increases signi cantly when adding one more cell (e.g. from 3 to 4 cells the F1 measure increases from 0.27 to 0.35 a growth of 26%). For larger numbers of cells there is less information gain by adding one more cell resulting in smaller increases in performance (all below 10%). Using 20 cells results in F1 of 0.514, which is 94% of the F1 achieved with all cells (0.547). 2 The data set is available at https://github.com/quhfus/table-disambiguation 5 . 0 4 ce .0 n raom .03 fr pe .2 0 1 . 0 0 ● . 0 ● ● ● ● 5 ● ● ● ● ● ●
We proposed an algorithm to annotate table headers with semantics based on the types of the column's cells. We achieved similar accuracy as previous work with more complex methods. We expect the reason for the comparable performance to be the knowledge base with more exhaustive and qualitative annotations. From our experiments it seems reasonable to use only a small number of cells for annotating the header (20 cells lead to 94% of the total achievable accuracy) if performance is an issue. We plan to exploit more relational knowledge (e.g. same-as) to further improve the annotations.
Acknowledgements. The presented work was developed within the CODE project funded by the EU Seventh Framework Programme, grant agreement number 296150.