=Paper=
{{Paper
|id=Vol-2980/paper413
|storemode=property
|title=A Framework for
              Automatically Interpreting Tabular Data at Orange
|pdfUrl=https://ceur-ws.org/Vol-2980/paper413.pdf
|volume=Vol-2980
|authors=Yoan Chabot, Frédéric Deuzé, Viet-Phi Huynh, Thomas Labbé, Jixiong Liu, Pierre Monnin, Raphaël Troncy
|dblpUrl=https://dblp.org/rec/conf/semweb/ChabotDHLLMT21
}}
==A Framework for
              Automatically Interpreting Tabular Data at Orange==
<pdf width="1500px">https://ceur-ws.org/Vol-2980/paper413.pdf</pdf>
<pre>
   A Framework for Automatically Interpreting
           Tabular Data at Orange

 Yoan Chabot1 , Pierre Monnin1 , Frédéric Deuzé1 , Viet-Phi Huynh1 , Thomas
                Labbé1 , Jixiong Liu1,2 , and Raphaël Troncy2
                                    1
                                     Orange, France
                               yoan.chabot@orange.com
                                 2
                                   EURECOM, France
                              raphael.troncy@eurecom.fr


On the Importance of Interpreting Tabular Data
Large parts of knowledge of companies are encoded in tabular data. Being able
to interpret such data is key to increase business eﬀiciency and to propose inno-
vative services, and Orange is no exception. With more than 140,000 employees
worldwide and a heterogeneous client portfolio, Orange produces a phenomenal
amount of tabular data every day. These tables are viscerally embedded in inter-
nal services and products (e.g., network logs, multimedia catalogs). Hence, they
are a source to discover new knowledge. Although this encourages the develop-
ment of eﬀicient tools to process them, several issues are negatively impacting
their use. First, the volume curse makes diﬀicult to identify the right dataset for
a given use case. Then, the knowledge gap between data producers/consumers
is exacerbated by our language footprint (seven main languages), the hetero-
geneous tools producing various table formats, and the experience/jobs of em-
ployees leading to similar concepts being expressed by different terms across
tables.


Making Sense of Tables: DAGOBAH
One lever to address these challenges resides in Semantic Table Interpreta-
tion (STI), i.e., making tabular data intelligently processable by matching ele-
ments of tables and constituents of Knowledge Graphs (e.g., Wikidata or spe-
cific domain/enterprise ones). Providing semantic annotations is a key asset for
search/recommendation engines and natural language based services. Moreover,
KGs can be used to drive STI while being themselves enriched by STI’s results.
Indeed, transferring knowledge to KGs allows previously dormant knowledge to
be structured and queried eﬀiciently. In this view, Orange/EURECOM research
teams have developed DAGOBAH, a framework interpreting tables automati-
cally. DAGOBAH is packaged in a RESTful API named TableAnnotation on
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
2        Y. Chabot et al.

the Orange Developer portal3 . DAGOBAH offers four features: i) preprocessing
(data cleaning, primitive typing, table orientation, etc.), ii) Cell-Entity Annota-
tion associating each cell with a KG entity, iii) Column-Type Annotation typing
each column with a KG concept, and iv) Columns-Property Annotation identify-
ing properties between table columns. It comes also in two flavours: DAGOBAH
SL based on a smart lookup of entities which are then disambiguated by com-
paring the context given by the table row under study with the context of the
candidates in the KG; DAGOBAH Embeddings exploiting geometric properties
of vector spaces to cluster candidates. While computationally more expensive,
embeddings provide promising results on highly ambiguous tables. These two
approaches have proven to be competitive during the SemTab challenge [1, 3].


Future Work
Interpreting tables has become a crucial task attracting a lot of attention in re-
cent years and the SemTab challenge [4] provides a relevant forum for comparing
approaches. However, there is still a long way to go to support the richness of
tables stored in company repositories. In particular, two limitations have been
identified in the current gold standards: 1) tables are generally homogeneously
shaped and the noise artificially generated does not represent the real world
complexities spectrum, and 2) target KGs are mostly encyclopaedic or limited
to very specific domains. A challenge will be to propose datasets encoding more
diﬀiculties such as matrix tables or multi-values cells. Another challenge Orange
will face is the application of annotation techniques to tables derived from busi-
ness knowledge. This requires the development of Orange’s own KG that focuses
on its activities, following the accelerating trend of Enterprise KGs. This raises
diﬀicult questions such as how to support our various business areas and how
to bootstrap such a KG from publicly available KGs? It also requires a complex
pipeline as demonstrated by Amazon’s Product Graph [2]. To this aim, Orange’s
future research work will be structured along the harvesting of the company’s
tabular data, the complementarity of text/tables in the knowledge extraction
process, and the reconciliation of knowledge in a high quality KG.


References
1. Chabot, Y., et al.: DAGOBAH: An End-to-End Context-Free Tabular Data Seman-
   tic Annotation System. In: SemTab 2019. pp. 41–48 (2019)
2. Dong, X.L., et al.: AutoKnow: Self-driving knowledge collection for products of
   thousands of types. In: 26th ACM SIGKDD International Conference on Knowledge
   Discovery & Data Mining. pp. 2724–2734 (2020)
3. Huynh, V.P., et al.: DAGOBAH: Enhanced Scoring Algorithms for Scalable Anno-
   tations of Tabular Data. In: SemTab 2020. pp. 27–39 (2020)
4. Jiménez-Ruiz, E., et al.: Results of semtab 2020. In: CEUR Workshop Proceedings.
   vol. 2775, pp. 1–8 (2020)

3
    https://developer.orange.com, API available upon invitation.

</pre>