Section: New Results

Named Entity Linking

Participants : Rosa Stern, Benoît Sagot.

The Ph.D. research work started in 2009 lead in 2013 to the development of a joint entity recognition and linking system for the processing of textual data at the Agence France Presse (AFP).

This system, Nomos, allows to use any existing named entity recognition system, as well as combinations of such systems; their results are passed to a linking module in charge of the association between each detected mention and a unique reference within an existing data inventory. The two tasks (recognition and linking) are jointly operated: the recognition module presents a set of possible detections, which are further disambiguated by the linking module concurrently to the search for the best linking solution to each mention. This joint approach is justified by the need to limit the error propagation between two such modules in a pipeline system.

Experiments were achieved in order to evaluate the performance of Nomos over AFP news wires. They showed that the joint approach, relatively to a purely sequential one, improves the system's global precision, i.e. the linking accuracy as well as the named entity recognition task itself. A gain of 3 points (87,6) is observed for the recognition precision with a low recall loss, while a gain of 8 points (92,9) is observed when several recognition systems are combined - although with a more significant loss of recall.

The Nomos system also allows to anchor of the AFP's textual production in the Linked Data network and the Semantic Web paradigm, since the annotations derived from the entity linking associate each entity to an identified resource in repositories such as Wikipedia, DBPedia, Geonames or the New York Times Linked Data.