This paper proposes a graph based methodology for automatically disambiguating authors' mentions in a corpus of French literary criticism. Candidate referents are identi ed and evaluated using a graph based named entity linking algorithm, which exploits a knowledge-base built out of two di erent resources (DBpedia and the BnF linked data). The algorithm expands previous ones applied for word sense disambiguation and entity linking, with good results. Its novelty resides in the fact that it successfully combines a generic knowledge base such as DBpedia with a domain speci c one, thus enabling the e cient annotation of minor authors. This will help specialists to follow mentions of the same author in di erent works of literary criticism, and thus to investigate their literary appreciation over time.
Named Entities (NE) are linguistic expressions that stand like rigid designators for referents; such entities normally include names of persons, geographical places, organizations, but also temporal references such as dates. Enriching mentions with a link to its referent by means of a unique identi er is crucial for the semantic annotation of texts. This is done by pointing to an external resource, such as a Universal Resource Identi er (URI) in the Linked Open Data (LOD) cloud. Segments in text referring to a Named Entity are known as entity mentions.
Named Entity Linking (NEL) [
Besides ensuring disambiguation, linking also performs an important additional task, namely textual enrichment, in that it connects the mention with sources of additional information - such as DBpedia in the previous example that needs not be stored in the text but can be accessed when required. In the case of Edmond de Goncourt, additional information from DBpedia can tell us what books he authored, where he was born, ....
The main issue with NEL in digital humanities is that mentions of persons often refer to individuals that are not listed in general ontologies such as Yago or DBpedia, that constitute the typical knowledge base for linking in other domains. Such individuals are often present in other knowledge bases, notably bibliographical linked data repositories (such as the French National Library BnF linked data repository). On the other hand, linking requires access to ontological knowledge, in that choosing between two individuals having the same name may requires comparing the context of the mention with a priori knowledge. In this respect, knowledge bases such as DBpedia remain an important source of general knowledge of the World. Thus the ideal linking algorithm for literary criticism texts combines general and domain speci c sources. The experiment here described goes in this direction.
The paper will rst present previous approaches to NEL, then the proposed graph based disambiguation algorithm based on the notion of centrality, nally describe the experiment carried out on the corpus and the results. Some conclusions and suggestions for further improvement of the algorithm are nally given. 2
Previous approaches for NEL can be divided in two main families. Those using text similarity and those using graph based methods. Both these methods are unsupervised, and they do not rely on pre-annotated corpora for training.
The best known tool of the rst group is DBpedia Spotlight [
Graph-based approaches rely on formalised knowledge described in graph form that is built from a Knowledge Base (KB) (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 candidates available 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 (such as a paragraph) containing possibly more than one mention, so that the disambiguation of one mention is helped by the other ones.
This kind of approach is similar to the one used in Word Sense
Disambiguation [
Centrality is an abstract concept, and it can be calculated by using di erent
algorithms4. In [
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 rich set of relations to disambiguate mentions. More speci cally in [
to ArticleB or whether ArticleA has CategoryC. Here too \local" centrality is then used to assign the correct link to the ambiguous mention. We have chosen a graph-based approach to NEL that will be described in the next section. 3
Our approach to disambiguate NE mentions is a graph-based one. Vertices are represented by URIs of mention candidates (e.g. dbpedia:Victor Hugo) as well as URIs of concepts (e.g. foaf:Person) or individuals connected to at least two different candidates. Edges are semantic relations de ned explicitly between URIs (e.g. \type"). The graph is undirected and their vertices and edges are a priori unweighted. We take advantage of the notion of centrality in Graph Theory to link a NE mention with the URI of the most probable candidate for that mention. In other words, we want to nd the subset of vertices of di erent candidates having the greatest number of edges among them. The edges and vertices of the graph are built leveraging knowledge from di erent LOD sources whose nature is graph-based.
We illustrate the proposed approach with an example. Let us consider the following phrase of a French text of literary criticism written by Albert Thibaudet (1936) :
Quant au rythme, si Victor Hugo a depasse Lamartine, il n'a pas ete plus loin que Vigny.
In bold there are three mentions automatically recognized by a NER algorithm, that need now be linked to an identi er.
For each mention, the NEL algorithm selects possible candidates by exact string matching of the current mention and dictionary entries (e.g. Hugo, M. Hugo) and retrieves the corresponding URIs of the listed LOD sources. An excerpt of the candidates of the three named-entities from the example is listed below by distinguishable personal information instead of URI for readability sake.
Thanks to the URIs, it is possible to retrieve from the Web of Data the associated RDF graph for each candidate and combine them into a single graph. It should contain only those predicates involving at least two candidates of di erent mentions because we only want the predicates that play an important role in the disambiguation process. Calculating the centrality for every candidate will then give us the best candidates for the three mentions. Figure 1 shows an excerpt of the resulting graph where the chosen mention candidates are marked in bold. We can notice that the vertex yago:RomanticPoets is the one that in uences the centrality measure the most because it is shared by the three chosen candidates. Likewise, other vertices connected to the chosen nodes, such as dbpedia:romanticisme and dbpedia:Alexandru Macedonski, are in uential.
Named Entities are disambiguated and referenced within the context of a paragraph, so in principle two (identical) mentions of the same author within one paragraph will always receive the same link, while the same mention in di erent paragraphs might be assigned a di erent referent, depending on the other mentions it occurs with.
The NEL task is commonly de ned in such a way that it does not assume
the existence of the correct referent among the candidates in the knowledge base
[
Our approach relies importantly on a lookup dictionary; this is the subject of the following section. 4
Linked data [
The French DBpedia is constituted of the articles of the French version of Wikipedia. In DBpedia entries are classi ed one or more of the types of the DBpedia ontology. So for instance the author known as Stendhal5 is classi ed as Person, Artist, Writer, and at the top level, as Thing. Moreover, authors are linked to each other by horizontal relations such as In uencedBy, and, indirectly, by being linked to the same concept, such as Romanticism. BnF entries list all authors of books ever published in France; their entries contain information on date of birth and death, gender, alternative names, works authored. For instance the BnF entry for Voltaire6 gives several alternative names such as Franois-Marie Arouet (Voltaire's real name), Wolter, Good Naturd Wellwisher, ...
Most crucially, BnF links its entry to the DBpedia one when existing, thus making it very easy to connect the two resources in one knowledge graph. Moreover, BnF entries also list the author's Idref, which is the o cial identi cation system used by French universities and higher education establishments to identify, track and manage the documents in their possession. The combination of these two sources was considered able to grant a su cient coverage for a corpus of French literary criticism, thus the BnF and the DBpedia SPARQL endpoints were queried for all authors, retrieving their biographic information (name, surname, alternative names, dates of birth and death, title, ...) in structured form.
In order to be able to retrieve all possible mentions of an author, this information was processed into a dictionary of authors, that contains all alternative names of an author, plus a series of alternative forms automatically generated, with the links to BnF and DBpedia entries. Automatically generated alternative names are of the form: { surname only (Rousseau) { initials + surname (J.J. Rousseau, JJ Rousseau, ...) { title + surname (M. Rousseau, M Rousseau) G{ive..n. the domain (French literature) this procedure ensures that the retrieval of at least one candidate URI for most mentions. At the same time, the mass of information present in the BnF repository will generate several homonyms and make most mentions ambiguous; thus good disambiguation becomes crucial. 5
The NEL algorithm processes a le in XML-TEI format7; NE mentions are annotated with NER annotations (e.g. tag <persName>) for every paragraph; the 5 http://fr.dbpedia.org/page/Stendhal 6 http://data.bnf.fr/11928669/voltaire/ 7 http://www.tei-c.org/index.xml algorithm is devised to processes one single class at a time (here Person). It uses a lookup dictionary per class listing super cial forms and their associated URIs from LOD sources, as described in the previous section. The algorithm produces an enriched version of the input le indicating the chosen candidate for each mention. We developed our implementation in Java ; RDF data is processed thanks to the Jena API8; graphs are manipulated by the JgraphT API9 and implementation of centrality measures are available in the Social Network analysis tool, JgraphT-SNA10. In particular, the algorithm performs the following steps for every paragraph of the XML-TEI le:
2. retrieve the RDF graphs of those URIs 3. simplify and combine graphs then compute the selected centrality measure 4. choose URI of candidate with the higher score per mention then write results in TEI le
The algorithm searches for (1) possible candidates of mentions by exact string matching the mentions of the current paragraph and super cial forms in the dictionary; there must be at least one ambiguous mention to continue. It retrieves URIs (BnF, DBpedia) of mention candidates from dictionary entries. Next, the RDF graph is retrieved (2) for every URI and converted to a JgraphT-compatible graph, where RDF objects and subjects are vertices and RDF predicates are edges. Irrelevant edges and vertices are removed from graphs. We keep edges which involve at least two vertices representing URIs candidates. Information coming from di erent sources is combined into a single graph (3); the way we combine graphs is straightforward. The fusion is implicitly done thanks to one of the main LOD principles which consists of reusing vocabularies published in the LOD vocabulary cloud. In other words, edges (predicates) and vertices (URI nodes) should be shared by at least two graphs associated to candidates of di erent mentions. The selected centrality measure (e.g. closeness) is calculated for the resulting graph. Finally, the algorithm chooses (4) the URI of the mention candidate with the higher centrality score and annotates the input XML-TEI le with this information.
Furthermore, simpli cation of graphs and calculation of centrality measures in the combined graph are crucial parts of the algorithm (3). This step is detailed in the Algorithm 1. It essentially removes edges which are irrelevant to calculate a centrality measure, in other words, it deletes those edges which involve at most one vertex of a non-candidate URI. 6
This section describes the experiments settings used to test our proposal as well as preliminary results which are encouraging. In this experiment, in order to
Algorithm 1 NEL: simplify and combine graphs, compute centrality Require: graphs: graphs of candidates per mention, measure: centrality measure for graph in graphs do initialize vertexToDelete for vertex in graph do if vertex is not a candidate then initialize vertexCheck for edges of vertex do if vertex1 notEqual vertex AND vertex1 is candidate then
vertexCheck.add(vertex1) end if if vertex2 notEqual vertex AND vertex2 is candidate then
vertexCheck.add(vertex2) end if end for if size of vertexCheck < 2 then
vertexToDelete.add(vertex) end if end if end for graph.removeAllVertices(vertexToDelete) chosenURIs = calculateCentrality(measure, graph) end for return chosenURIs, chosen candidate per mention evaluate the performance of the algorithm, the linking is performed on correctly identi ed and classi ed authors. 6.1
Experiment settings The test corpus consists of a French text of literary criticism titled \Une these sur le symbolisme" (A thesis about Symbolism) and it is the rst volume of the work named \Re exions sur la littrature" (Re exions on literature) published by Albert Thibaudet in 1938.
The text is drawn from a larger \Corpus critique"11, published in TEI by the Labex OBVIL and containing a large collection of critical essays by di erent authors.
The chosen text in particular presents a high density of authors' mentions,
so that each paragraph generally contains an average of 2-3 mentions that are
treated at the same time by the algorithm. Mentions concerning authors were
manually annotated by two experts in French literature; URIs assigned to
mentions are those from Idref12. Guidelines to manual annotation were those
proposed by the MUC7 conferences as well as those de ned by the XML/TEI
standard. The resulting test corpus contains 1021 manually annotated mentions of
11 http://obvil.paris-sorbonne.fr/corpus/critique/
12 www.idref.fr
person entities. We measure the precision of the proposed NEL approach in
terms of the attribution of the right URI to a mention with respect to the URI
manually assigned by humans. The authors lookup dictionary was
automatically built in advance thanks to the BnF LOD source which is rich in SameAs
predicates pointing to DBpedia and Idref URIs. The resulting lookup
dictionary is composed of 4,218,798 author names including their alternative names
(e.g. M. Lamartine, Monsieur Lamartine, etc.). We chose 3 centrality measures
commonly used in social network analysis and the word-sens disambiguation
problem, these are: DegreeCentrality [
Results and Analysis
Precision is calculated comparing the number of correctly assigned links over the total of manually annotated entities of authors. The best result is obtained with BrandesBetweennessCentrality, with a precision of 0.74. DegreeCentrality has a comparable performance, FreemanCloseness centrality seems to heavily underperform with respect to the other centrality measures. The last column of table 1 shows the number of empty links over the total.
These rst results are satisfying: though far from the 85% accuracy that is normally achieved by similar algorithms on the news domain, such levels of precision are nevertheless remarkable, considering that in many cases the text discusses minor authors, today unknown, that are not necessarily listed in DBpedia. Moreover, the use of BnF makes the number of candidates (and thus the possibility of error) explode, with sometimes as much as 20 or more possible candidate for a mention.
To quantify authors incompleteness in both the DBpedia and BnF data sets used in this experiment, we count the number of mentions in which the algorithm (using DegreeCentrality measure) does not nd any corresponding URI in the chosen KB. In this manner, there are 160 author mentions, out of 1021 mentions identi ed in the corpus by the algorithm, that have no match in DBpedia, that is around 16%. Remarkably, there are only 23 mentions (i.e. 2%) that have no match in either BnF or DBpedia. Notice that all authors in this test set that are in DBpedia are also in BnF.
The most frequent mistakes considering DegreeCentrality and BrandesBetweennessCentrality measures (the most similar and precise ones) concern the following authors: Viele-Gri n, Francis (1864-1937); Boileau, Nicolas (16361711); Barres, Maurice (1862-1923); Payen, Fernand (1872-1946); Lefranc, Abel (1863-1952); Shakespeare, William (1564-1616); Spencer, Herbert (1820-1903); Goncourt, Edmond de (1822-1896) and brother Goncourt, Jules de (1830-1870); Mentre, Franois (1877-1950). The algorithm makes three types of mistakes.
MISSING CANDIDATES - In 23 cases the algorithm is unable to retrieve any candidate from the lookup dictionary, since the author is not present in any knowledge base. This is the case of author Francis Viele-Gri n. In other cases the correct entity is present but not associated with the required pseudonym. This is the case of William Shakespeare's alleged alter ego William Stanley13. This alias is not listed in the dictionary for Shakespeare, therefore, it is not possible to assign both mentions to the same person (and thus the same URI).
MISSING CONTEXT - In some rare cases only one ambiguous author's mention is present in a single paragraph, thus the algorithm resorts to a fall back strategy, choosing the entity with more links in absolute. Sometimes this strategy causes errors, as in the case of \Vigny", for whom, in isolation, the wrong link to Auriane Vigny is chosen.
INCOMPLETE INFORMATION - In some cases the context of the sentence should be su cient to produce a correct disambiguation but the NEL algorithm makes mistakes due to lack of links in the knowledge base, which prevents the centrality measure to produce the desired result. For instance, \Shakespeare", when mentioned in the context of Shakespearian critic Abel Lefranc, should produce the correct linking to William, but Nicolas is chosen instead. Clearly explicit links between Abel Lefranc and the object of his studies are missing in the knowledge bases. Ancient authors also tend to cause problems due to lack of information, e.g. the Greek author Lysias is mistaken for an homonymous French revolutionary collective.
WRONG, MISLEADING INFORMATION - Sometimes the knowledge bases contain wrong or misleading information. For instance there exist a BnF entry for the \Ronsard family", classi ed as foaf:Person, which is chosen instead the correct assignment, namely one of its members, Pierre de Ronsard. The opposite is also true, so some mentions refer to both Goncourt brothers as a collective noun, but the algorithm chooses one of the two. Finally, wrong or misleading pseudonyms are sometimes listed in BnF for an author, causing wrong candidates to be injected in the graph and sometimes selected. So for instance \Descartes" is listed as a pseudonym for novelist Horace Walpole and thus sometimes Walpole is wrongly chosen as the link for philosopher Descartes.
Error analysis also shows that sometimes relevant information that is present in the knowledge base is not used in the decision process because it cannot be encoded in the graph in the form of links. A typical example is temporal information which is encoded in the form of dates (data-typed literals). In other words the fact that - for a given context - two candidate referents lived in the same period of time cannot be taken into account. 13 Stanley is believed by some to be the real author behind Shakespeare's works.
To evaluate the impact of the temporal dimension, we chose to evaluate against an index from which we removed authors born after the date of publishing of the work. The results show a slight improvement with DegreeCentrality reaching precision 0.78 and BrandesBetweennessCentrality 0.77. A greater improvement may be obtained using a more sophisticated graph building algorithm, that transforms information about dates of birth and death in links that can connect authors in a measurable way. 7
We presented an algorithm to perform NEL on a corpus of 19th century literary criticism, with the speci c goal of disambiguating and referencing author mentions for research purposes. The NEL module is meant to be used in combination with a NER module, and will help researchers in the creation of digital literary editions enriched with information about authors. The main purpose of this work is to help scholars in history of literature to perform complex queries in order to study the literary appreciation of authors over time, and investigate the history of literary criticism in French literature. More speci cally the enrichment of the aforementioned \Corpus critique" is meant to enhance ongoing research in the history of scienti c ideas, and to provide a way to follow the dissemination of theories and concepts de ned by Charles Darwin, Claude Bernard, Henri Bergson in non scienti c texts of their time.
The reported experiment shows how combining di erent sources can be useful to perform linking on a domain speci c corpus with satisfying results. While the precision is not yet state of the art, it is nevertheless remarkable, considering that it is the rst time graph that centrality algorithms have been used for NEL combining DBpedia with a domain speci c source. Tests showed signi cant di erences between one implementation of centrality and the other two. Error analysis suggests possible improvements of the algorithm, including the ad hoc transformation of temporal information - present in the knowledge base in the form of literals - into links of the context graph. Another possible evolution of the algorithm would be to assign di erent weights to the edges so that for instance sharing the same literary circle becomes a more important relation than being born in the same town. Weights would be learned from manually annotated data. Further experiments will be carried out with di erent corpora and on di erent categories of entities, notably places.
Experimenting with the size of the context will also be necessary, in order to nd the best trade-o between e ciency and informativeness. A more ample context (ideally a whole chapter) may produce a better graph of candidates, such that all mentions can disambiguate each other correctly. But at the same time this may introduce noise, and also generate a graph so big that its construction and the calculation of centrality may require too much time.
Another possible evolution of the algorithm could be to improve the graph
fusion procedure. So far, our strategy does not handle the proper fusion of
individuals which are described heterogeneously by the di erent sources (e.g. Victor
Hugo as described by BnF, as described by DBpedia, and so on). In this study
we chose to study the problem from a quantitative point-of-view and thus to
consider existent knowledge as it is without a pre-processing step. In the future,
we foresee to make use of strategies commonly applied in Conceptual Graphs for
information fusion [
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.