We present REDEN, a tool for graph-based Named Entity Linking that allows for the disambiguation of entities using domainspeci c Linked Data sources and di erent con gurations (e.g. context size). It takes TEI-annotated texts as input and outputs them enriched with external references (URIs). The possibility of customizing indexes built from various knowledge sources by de ning temporal and spatial extents makes REDEN particularly suited to handle domain-speci c corpora such as enriched digital editions in the Digital Humanities.
Semantic annotation plays an important role in the construction of enriched digital editions in the domain of Digital Humanities. The TEI annotation standard3 de nes speci c tags to mark-up key words in texts and to link them to background knowledge by using external references. Thus, for instance, it is possible to annotate topics, citations, places, organization and person names. References can point to internal identi ers or to external resources that are available in the form of Linked Data (LD) on the semantic Web. Such annotations can be used to facilitate research on the enriched texts (by allowing the creation of indexes and the performing of advanced queries). Moreover, they can also enhance the utilization of texts by the nal user, since TEI can be easily converted into other formats, notably Electronic Publication (EPUB). Physical supports such as ebook readers could in future make use of the enriched information in the form of external links.
The results of the work described here make up part of an editorial pipeline the aim of which is produce an enriched digital edition of a corpus of French literary criticism and essays from the 19th century4. Such a corpus, once completed, is intended to allow researchers to gain a more comprehensive perspective on
4 http://obvil.paris-sorbonne.fr/corpus/critique/ the evolution of ideas both in literature as well as in the sciences over the years. While the production of high quality digital editions requires manual checking, natural language processing techniques can speed up the process to a great extent. In particular, the detection of person mentions is a very important step for the enrichment of such a corpus, as it allows researchers to trace references to authorities (authors, scientists, critics, etc.) over time and from di erent books.
While existing Named Entity Recognition and Classi cation (NERC)
algorithms can help human beings to detect persons' mentions, previous
experience has shown that this step is not particularly di cult for trained annotators.
Conversely, disambiguation and linking of mentions to an identi er can often
be painstaking when done manually, as it requires searching external sources,
such as mainstream online encyclopedias (Wikipedia) but also domain-speci c
databases, such as library databases. Named Entity Linking (NEL) algorithms[
While current NEL algorithms perform well on the contemporary news texts domain adaptation is, as always in NLP, an issue. Nineteenth century essays for instance often mention persons that were well known at the time but are currently hard to identify even for specialists. Thus although existing NEL tools are very e cient they mostly rely on generic background knowledge such as Wikipedia or DBpedia that may not be su cient to grant ample coverage for all domains.
The solution we present here is a NEL tool, REDEN, that makes use of state of the art graph-based algorithms and at the same time can easily leverage knowledge from di erent sources (both generic and domain speci c). REDEN is applied to already detected mentions, and relies natively on knowledge drawn from linked data sources for disambiguation, requiring only an index of super cial forms and URIs to be used as candidates. Such indexes can be built out of di erent sources using crawlers.
The purpose of this paper is to show that such a solution can be e ciently used for texts in which a domain speci c knowledge base is necessary. We evaluate REDEN on ad-hoc test-sets from our corpus of essays and compare the results with those obtained with a freely available NEL tool. We rst present the current state of the art in NEL, then brie y describe the tool and then focus on the index building facilities; some experiments are presented to compare our proposed solution to the state of the art, and we conclude with some remarks and ideas for future developments. 2
Supervised methods usually perform very well in NLP tasks but they rely on pre-annotated corpora for training which are not available for speci c domains such as French Literature and the Digital Humanities more generally. Therefore we concentrate on existing non-supervised approaches for NEL. These can be broadly divided into two main families. Those using text similarity along with statistical models and those using graph-based methods.
The best known tool from the rst group is DBpedia Spotlight (DBSL) [
Graph-based approaches rely on formalised knowledge in the form of a graph
built from a knowledge base (e.g. the Wikipedia article network, Freebase,
DBpedia, etc.). Reasoning can be performed through graph analysis operations. It
is thereby possible to at least partially reproduce the actual decision process
with which humans disambiguate mentions. A reader may decide that the
mention \James" refers to philosopher \William James" and not to writer \Henry
James" because it occurs in the same context as \Hume" and \Kant". In the
same way such algorithms build a graph out of the available candidates for each
possible referent in a given context and use the relative position of each referent
within the graph to choose the correct referent for each mention. The graph is
built for a context (e.g. paragraph) containing possibly more than one mention,
so that the disambiguation of one mention is helped by that of the others. Hybrid
approaches such as [
The graph-based NEL is similar to graph-based Word Sense Disambiguation
(WSD)[
Centrality is an abstract concept, and it can be calculated by using di erent
algorithms5. In [
Indegree, Betweenness, Closeness, PageRank, as well as with a combination of all these metrics using a voting system. Results showed the advantages of using centrality with respect to other similarity measures.
This graph-based approach has been applied to NEL, where mentions take
the place of words and Wikipedia articles that of WordNet synsets. Here too
centrality measures are performed on the Wikipedia structure in order to use the
large set of relations to disambiguate mentions. More speci cally in [
With REDEN, we propose a graph-based, centrality-based approach that is particularly adapted for digital humanities; in fact it can exploit RDF sources directly (thus allowing users to choose domain adapted knowledge bases) and takes TEI-annotated texts as input. The disambiguation algorithm processes a le where NE mentions are already annotated (e.g. using the tag <persName>). Possible referents from all mentions in a given context and for a given class (e.g. person) are retrieved from an index built on the y from selected LD sources. The fusion of homologous individuals across di erent sources is performed thanks to owl:sameAs or skos:exactMatch predicates; a graph is created where RDF objects and subjects are vertices and RDF predicates represent edges. Irrelevant edges are removed from the graph before calculating centrality: only edges which involve at least two vertices representing candidate URIs are preserved. Thanks to the selected centrality measure, the best connected candidates for each mention are chosen as referents and an enriched version of the input TEI le is produced.
To illustrate how REDEN works, we apply our approach to a paragraph from a French text of literary criticism entitled \Une these sur le symbolisme" (1936) written by Albert Thibaudet (1874-1936)7. Figure 1 shows an excerpt of the resulting graph where the chosen and correct mention candidates out of the nine present in the text are marked in bold. We can observe that the vertices yago:Symbolist Poets, dbpedia:Charles Baudelaire, dbpedia:French Dramatists And Playwrights are the ones in uencing the centrality measure the most. 6 https://github.com/bernhardschaefer/dbpedia-graph but related published work is not available at this time 7 Full text can be found here: http://obvil.paris-sorbonne.fr/corpus/critique/ thibaudet re exions.html
We developed our implementation in Java8. In [
As previously described, REDEN identi es and looks for mention candidates' URIs in indexes built on the y per class. So for instance, the entries of a person in the index correspond to super cial forms of that person's name and the corresponding URIs from di erent Linked Data sets. REDEN then retrieves the RDF data from the corresponding URIs found in the index for each selected source.
Clearly, a source can be more useful for certain domains than another, for instance the French National Library (BnF) linked data repository is the most appropriate source for a domain such as French literary criticism. BnF provides the catalog of authors for all books ever published in France; their entries contain information on dates of birth and death, gender, alternative names, works authored, etc. Most importantly, BnF provides equivalence links (owl:sameAs ) to other sources such as DBpedia or IdRef. Thus, BnF can be chosen as the entry point for building an index of authors. In particular information on gender and full names can be used to generate further alternative names - for instance: surname only (Rousseau), initials + surname (J.J. Rousseau, JJ Rousseau, ...), title + surname (M. Rousseau, M Rousseau), ... Equivalence links may also be stored, so that later on REDEN can build fused graphs from di erent sources.
Likewise, using DBpedia as entry point an index of places can be assembled. For instance, the DBpedia entity Gare de Paris Saint-Lazare has the following name forms (familiar, o cial and old ones) - for instance: Gare Saint-Lazare, Gare Saint-Lazare (Paris), Embarcadere de l'Ouest, ... 8 Code source and useful https://github.com/cvbrandoe/REDEN resources can be found here:
We have therefore developed an index-building module which allows for the con guration and the implementation of a crawler for a chosen LD source. The main purpose of this module is to facilitate the construction and the update of entity indexes which should be suitable for the domain of the exploited corpora. In particular, the temporal dimension and possibly the spatial dimensions can be used as lters. For instance when working with 19th century texts, a time lter can a priori exclude the currently well-known singer Thierry Amiel as a candidate for the mention `Amiel' and thus enable the identi cation of the correct referent, well-known writer Henri-Frederic Amiel.
The proposed module takes into account domain parameters such as the temporal and spatial dimensions. For instance, a time period can be speci ed with respect to the date of birth of the persons described in the corresponding linked data set, or to any other property of the individual (date of death, periods of activity). Concerning the spatial extent of the index, one may be interested in including only persons who were born or have lived in Europe because the corpora probably concerns authors with these features. A bounding box can be de ned by specifying a rectangle consisting of four Latitude and Longitude coordinates9. It is also straightforward to de ne new domain parameters to lter for themes, such as Symbolism or Psychology.
From a technical point-of-view, a crawler performs SPARQL queries using a selected end point. The created indexes can bene t from text search engines which enable the optimal querying of candidates by REDEN and therefore reduce query processing time.
Currently REDEN comes with two already implemented crawlers: one to collect an index of place candidates from DBpedia and one to crawl an index of author candidates from the LD repository of the BnF, containing alternative names for every entity. In the case of authors, as BnF entries point to DBpedia equivalents when available, these two sources can be easily fused using the aforementioned algorithm. In the case of places, there is no need to fuse other LD sources although one may be interested in including a place gazetteer available in linked data such as Pleaides10 when working with Ancient texts. 5
In previous experiments [
http://www.mapdevelopers.com/geocode bounding box.php 10 http://pleiades.stoa.org of information: set of classi ed mentions. Nevertheless, the comparison can help to give an idea of the di erences in coverage that our method o ers.
Since REDEN treats one class at a time, DBSL too is called with the chosen
class as a parameter. Moreover, we use the most relaxed con guration possible,
accepting any matching results with no lter on support or con dence, to
maximize recall. DBSL is installed locally and launched using French DBpedia as
a knowledge base. As for REDEN, the chosen centrality measure is
DegreeCentrality [
The manually linked test corpus consists of a French text of literary criticism entitled \Une these sur le symbolisme" (A thesis about Symbolism) published by Albert Thibaudet in 1936, and a scienti c essay entitled \L'evolution creatrice" (Creative Evolution) written by Henri Bergson and published in 1907. We can consider two measures of accuracy: A1, the proportion of correct links over the mentions for which a link was chosen, and A2, the proportion of correctly assigned links over the number of mentions overall. The second measure considers missing links as mistakes, and can be seen as a measure for coverage11. 5.1
Even though REDEN's strength lies in the possibility of merging resources, we will rst compare the two algorithms on French DBpedia alone12. Since DBSL cannot process more than a paragraph at a time, REDEN too is con gured to use paragraphs as contexts for its graphs. We choose the task of linking location mentions, which, in our texts are rather sparse but not particularly domainspeci c. A priori we expect most referents for locations' mentions to be present in the knowledge base.
As shown by the results, both linking algorithms are comparable. They both achieve excellent correctness rates when they assign links, but they miss a certain number of places. Missing places occur often due to the lack of alternative names, such as \Lacedemone" (another name for Sparta), or from entries that are missclassi ed in DBpedia, such as \Berlin" which is classi ed as a Concept and not 11 These measures may be considered to roughly correspond to precision and recall, though such terms seem less adapted to describe this type of task. 12 REDEN and DBSL use versions of 05/03/15 and of 17/07/2014, respectively. Unfortunately, it is technically not possible for REDEN to access a snapshot of DBpedia RDF resources from that particular date. as a Location (in the French version of DBpedia). REDEN achieves a better coverage even when using the same knowledge base, possibly because it better exploits the available alternate names, or because DBSL does not provide an answer when it is unable to nd a correspondence between the textual context of a mention and the description of a candidate.
As previously stated this evaluation is meant as a comparison of the algorithms, as REDEN works on pre-detected mentions while DBSL has to perform detection too. That said, in this experiment DBSL analyses only paragraphs containing locations and we force the type \location" in input; in most cases when DBSL is missing a link the name of the mention is written in capitalized words and coincides exactly with the title of a DBpedia entry; thus it is possible to conclude that, in such cases, it is not the detection but the linking that fails. 5.2
The second experiment aims to show how our tool can achieve a much better performance on this corpus thanks to the possibility of combining two knowledge bases to target a speci c domain. Here, the more di cult task of identifying authors' mentions is evaluated. DBSL is run using French DBpedia as usual, REDEN is run on an index of authors built from French DBpedia + the BnF linked data. Moreover, we exploit all the ne-tuning options o ered by REDEN: in order to best exploit contextual information we increase the disambiguation context, using the chapter for Thibaudet, which is a text with high mentions' density, and the whole book for Bergson because it is poorer in mentions13. Moreover, we lter the index by preventing authors who weren't born or were too young at the time of publication of each work from becoming candidates.
As shown in Table 2, DBSL achieves a high correctness rate for persons when it makes a choice, assigning almost perfectly each mention found on both corpora. On the other hand, the amount of entities found is much smaller. The reason is obviously that REDEN exploits additional information, granting coverage for authors who aren't present in DBpedia.
At the same time, we notice that DBSL correctly recognizes the mention of
(Alphonse de) \Lamartine" but not those of (Maurice) \Barres" and (Alphred
13 In [
It is important to note here that DBSL is more precise than REDEN in those cases where it actually chooses a link (A1). In particular, it beats our method when a paragraph contains only one ambiguous candidate mention. In this case, the unstructured textual context is a useful source of information to disambiguate, while the graph-based algorithm cannot use background information. Nevertheless, the greater coverage of the graph-based approach makes the overall post-processing work of manual checking less cumbersome in terms of both correction and integration. 6
In this article we presented REDEN, a graph-based tool for Named Entity Linking that can be easily extended to di erent online sources and tailored to suit various di erent needs. Domain adaptation is enabled by entity indexes built on the y using linked data crawlers. New sources for a particular domain can be added by con guring a new crawler with the appropriate SPARQL query. Experiments involving the disambiguation of persons' and locations' names in a corpus of French essays of the 19th century show how the graph-based algorithm gives us state of the art correctness results while allowing for more exibility and coverage.
The experiment with DBSL highlighted some of the practical advantages of our tool. First, DBSL is very dependent on the DBpedia. Con guring it to use new knowledge bases such as BnF does not yet seem possible. Furthermore, REDEN works natively with RDF. The structured information used has thus the potentiality to grow as new links are added to the semantic web. REDEN uses online sources (end points) which allows for the easily downloading of the most up-to-date version of the data from the corresponding repository. These data are cached by REDEN in order to download an RDF resource only once. On the contrary, DBSL uses o ine resources which can be an advantage when bandwidth is limited, however updating the DBSL knowledge base looks to be a very di cult and time consuming process. Another technical limitation is the size of the context that DBSL uses for its decision; the tool runs as a Web service accepting HTTP calls and the URL size is restricted. In REDEN, the size of the context does not need to be xed, but can be set by the user according to prede ned textual partitions in TEI; thus the tool can run using all mentions in a paragraph, a chapter, or a whole text as a disambiguation context. Finally crawlers can be customized in such a way as to apply (time or space) lters on candidates, thus making it domain adapted.
For these reasons we conclude that the proposed methodology seems better adapted to the task of linking named-entities to a knowledge base other than DBpedia. Thus it seems particularly useful for digital editions in the humanities, where texts are already in TEI format, and specialists have a good knowledge of the text and of relevant sources of information.
As prospective work we intend to perform an exhaustive comparison of
REDEN to other graph-based NEL tools such as AGDISTIS[
This work was supported by French state funds managed by the ANR within the Investissements d'Avenir programme under reference ANR-11-IDEX-0004-02 and by an IFER Fernand Braudel Scholarship awarded by FMSH.