Disambiguation of Named Entities in Cultural Heritage Texts Using Linked Data Sets

Abstract

International audienceThis paper proposes a graph-based algorithm baptized REDEN for the disambiguation of authors’ names in French literary criticism texts and scientific essays from the 19th century. It leverages knowledge from different Linked Data sources in order to select candidates for each author mention, then performs fusion of DBpedia and BnF individuals into a single graph, and finally decides the best referent using the notion of graph centrality. Some experiments are conducted in order to identify the best size of disambiguation context and to assess the influence on centrality of specific relations represented as edges. This work will help scholars to trace the impact of authors’ ideas across different works and time periods

REDEN: Named Entity Linking in Digital Literary Editions Using Linked Data Sets

“…It is also possible to a posteriori assign weights to these edges based on user preferences, in particular, the higher weight is attributed to an edge, the more priority is given to this edge during centrality calculation. In [8], we describe an attempt to evaluate the impact of relations by setting a higher weights on them.…”

“…In this paper we shall first present previous approaches to NE disambiguation and linking, then introduce our graph based disambiguation algorithm, named REDEN, which includes strategies to consistently handle multiple LD sets, showing how this and other design characteristics make it better suited to work with texts from the humanities and literary domain. REDEN has been introduced in previous publications 5 [9], [8], [10]. Here, in contrast to our previous work, we shall extensively describe the algorithm in its two phases, namely, candidate retrieval and candidate selection, along with new improvements such as the exploitation of any KB linked to an entity using equivalence links (e.g., sameAs).…”

“…For the connections between linguistics, NLP and linked data see[6] 4. http://obvil.paris-sorbonne.fr/corpus/critique/5 The present paper is an extended version of[8], also a peer-reviewed version.…”

“…Domain-Specific Graph-Based NEL ApproachThe algorithm we are going to present was developed in the context of Labex OBVIL8 , an interdisciplinary French laboratory where computational methods are developed and applied to the research in the literary domain. OBVIL has developed a large corpus of TEI digital editions, both of primary and secondary literary sources and essays more generally.…”

“…The analysis of available algorithms and tools for linking soon revealed their limitations, in terms of supported input -no TEI support, difficult to work with pre-detected mentions -but also in terms of supported document bases; in particular manual annotators observed that, in essays from the 19th and early 20th centuries, minor authors and members of the cultural life in general that did not have a Wikipedia (and thus DBpedia entry) could be identified by searching on data.bnf.fr, the linked data catalog of the Bibliothèque Nationale de France (BnF), that is publicly available. Moreover, they observed that often the right referent was found by looking at the context of other mentions, using the relations between individuals, and combining information from different8 http://obvil.paris-sorbonne.fr 9 http://obvil.paris-sorbonne.fr/corpus/critique 10 IDREF -http://www.idref.fr -is the French reference base for individuals, also used in library catalogs;IDREF is an almost complete reference base for the domain in question, but contains very little contextual information on people and is not provided with a SPARQL endpoint; at the same time, many other LD sets are connected to it.…”