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 innovative 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 internal services and products (e.g., network logs, multimedia catalogs). Hence, they are a source to discover new knowledge. Although this encourages the development 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 heterogeneous tools producing various table formats, and the experience/jobs of employees leading to similar concepts being expressed by diferent terms across tables. One lever to address these challenges resides in Semantic Table Interpretation (STI), i.e., making tabular data intelligently processable by matching elements of tables and constituents of Knowledge Graphs (e.g., Wikidata or specific 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 automatically. DAGOBAH is packaged in a RESTful API named TableAnnotation on
the Orange Developer portal3. DAGOBAH ofers four features: i) preprocessing
(data cleaning, primitive typing, table orientation, etc.), ii) Cell-Entity
Annotation associating each cell with a KG entity, iii) Column-Type Annotation typing
each column with a KG concept, and iv) Columns-Property Annotation
identifying properties between table columns. It comes also in two flavours: DAGOBAH
SL based on a smart lookup of entities which are then disambiguated by
comparing 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 [