With the increasing scale of online cultural heritage collections, the e orts of manually adding annotations to their contents become a challenging and costly endeavour. Entity Linking is a process used to automatically apply such annotations to a text based collection, where the quality and coverage of the linking process is highly dependent on the knowledge base that informs it. In this paper, we present our ongoing e orts to annotate a corpus of 17th century Irish witness statements using Entity Linking methods that utilise Semantic Web techniques. We discuss problems faced in this process and attempts to remedy them.
di erent items and DPLA2 hosts almost 21 million resources), it becomes increasingly worthwhile to consider how we might deal with these limitations for collections both great and small.
In this paper we present a discussion on the role of semantic web resources in the task of automatically enriching digitised cultural heritage collections using EL methods. Our discussion is motivated by ongoing e orts to annotate a collection of 17th century Irish witness statements so that they may be integrated as semantic web resources and avail of bene ts such enrichment provides. We present some of the challenges faced and lessons learned in the course of this endeavour. The contribution of this paper is to demonstrate and emphasise the importance of structure in Semantic Web resources. With due consideration it is possible to create new ontologies which may help to facilitate the EL process. 2 2.1
Entity Linking (EL) refers to a speci c challenge in computer science whereby a series of unknown textual mentions of entities (commonly termed \surface forms") are provided as input to a disambiguation service. The service is tasked with mapping each of the surface forms to an unambiguous referent entity. To provide a concrete example, given the input sentence, \I Henry Jones Doctor in Divinity in obedience to his majesties Commission..." and a request to identify the entity \Henry Jones", an EL service might return a reference to the URI http://dbpedia.org/page/Henry_Jones_(bishop), identifying the subject of the reference as the 17th century Anglican Bishop, as opposed to the ctional character played by Sean Connery in the 1989 lm \Indiana Jones and the Last Crusade".
In order to perform this mapping process, an EL system fundamentally requires two components: 1. a knowledge base that stores information about all the entities of which the system is aware, and 2. a referent selection method, which uses evidence extracted from the knowledge base and present in any prevailing information surrounding the surface form to arrive at a set of likely referents for each ambiguous mention.
Given an ambiguous set of mentions, the EL system retrieves from the knowledge base, a set of candidate referents to which an entity mention many be referring. This is usually based on some fuzzy retrieval method. A variety of heuristics are applied and the system eliminates candidate referents which are unlikely to be the subjects of the mentions. Eventually it arrives at a set of mappings from textual mentions to knowledge base URIs which unambiguously identi es the referents.
Numerous di erent methods and approaches to EL may be freely found in the literature [6{8]. Almost universally, these methods use some form of graph based measure as one of the heuristics in the referent selection process. After the candidates have been retrieved from the knowledge base, a graph derived from the relationships between the entities may be constructed. The nature of these relationships varies, but usually it is based on links between corresponding Wikipedia pages. If a strong network exists between a number of candidate referents, then there is a good chance that they are the correct disambiguation choice for the given set of mentions.
It is also common to augment the graph weights using contextual cues
derived from the words surrounding an entity mention [
The structure and content of the knowledge base is crucial, not only for
informing the disambiguation service that an entity exists, but also for providing
information which helps the the disambiguation algorithm to distinguish good
referents from poor referents. Many modern systems make use of DBpedia3 [
EL systems have the potential to be extremely helpful when enriching cultural heritage collections with semantic data. These fully automated systems are capable of deducing suitable annotations for raw, at, textual documents based on information that is fed to them via a knowledge base. It is easy to see how a suitably informed EL system might dramatically ease the process of semantically linking new cultural heritage artifacts as they are digitised.
Further discussion could be had surrounding the precise point in the digitisation process at which EL is applied. Are we linking metadata which has already been normalised by an expert, or is the system capable of dealing with the noisy, original, primary source content from which the digital artifact is derived? In the case of the latter, how does the system manage archaic references, evolving entities and other such anomalies present in the source collection? 2.2
There have been a number of e orts to investigate the e ectiveness of EL methods in the automatic enrichment of cultural heritage collections.
A Europeana led task force produced a series of reports in 2015 which document their experience with evaluating di erent cultural heritage enrichment 3 http://wiki.dbpedia.org/ 4 https://www.mpi-inf.mpg.de/departments/databases-and-information-systems/ research/yago-naga/yago/#c10444 services and sourcing di erent descriptive vocabularies as targets for the annotation process. Their focus was on annotating metadata for digitised artifacts. The content of this metadata ranged from speci c elds comprised of a single entity e.g. dc:creator, dc:publisher to more general, free-form data such as dc:description.
As part of the investigation, a comparative evaluation of seven cultural
heritage EL services was conducted [
However, as a variety of previous studies have shown, while the accuracy of
EL methods may be high, quite often only a very small percentage of entities
contained in cultural heritage datasets may actually be linked with a referent. A
recent study we performed on the 1641 depositions (see Section 3.1) showed that
a human annotator could only identify referents for 33% of the people and
locations in the depositions [
One aspect of the problem is simply that cultural heritage collections are so incredibly diverse, complicated and unique that nding a suitable Semantic Web resource with adequate coverage for all purposes is nigh impossible. This presents the question, how should we annotate a cultural heritage collection when an appropriate Semantic Web resource cannot be found? Moreover (and of particular importance to our own research) how should these new Semantic Web resources be structured in order to aid the automatic enrichment process?
Of particular note for this discussion is the work of Brando et al. [
REDEN's candidate selection phase is based on a literal string comparison between the surface form and entities in the knowledge base. All candidates from all source ontologies are retrieved and a resolution step based on owl:sameAs and skos:exactMatch properties resolves duplicate mentions into a single reference. Once the candidates have been appropriately pruned, a degree centrality measure is used to select the referents.
REDEN demonstrated that developing EL methods which can avail of multiple knowledge bases may help with poor coverage, but this requires that it be possible to establish reliable, accurate mappings between ontologies. Indeed, this property also facilitated the evaluation conducted by the Europeana task force. This is an important consideration when developing new vocabularies for cultural heritage collections. 3
Our own research has focused on attempts to automatically annotate a collection of 17th century manuscripts using EL methods. This has been extremely challenging for a variety of reasons. However, we believe our experiences are a reasonably typical example of problems faced in this eld. Below, we document our observations and experience of working with the collection to date. 3.1
The 1641 depositions are a collection of letters and witness statements taken from the people of Ireland during the 1641 Irish rebellion. The physical manuscripts are comprised of approximately 19,000 pages bound in 31 volumes. Ireland in 1641 was a tumultuous place, and while the accuracy of some of the witness statements may be questionable, the depositions provide an unparalleled window into this dark chapter in Irish history.
The depositions have been digitised, transcribed and annotated by a team of historical scholars who extracted references to people and locations, tagged depositions based on the nature of their contents, and preserved as much information about the physical manuscripts as possible including margin notes, original spelling etc. The resulting documents are stored in a combination of TEI annotated les and an SQL database. This data rich digital resource presents many interesting and exciting opportunities for computer scientists to begin experimenting with methods of analysing and extracting new information from this historical collection.
Working with the digital versions of the depositions comes with a number of challenges, not least of which is the inconsistent nature of the spelling and grammar used throughout. English was still a developing language in 1641, which means that a vast array of variant spellings for names and common words exist across the documents. The below extract from The Deposition of Phillip Sergeant 5 provides an example of these anomalies: \And by those faire promisses the said tzpatrick getting possession both of their persons & goodes, they there behoulding daily cruelties & murthers vpon other English and belike suspecting the like to be exercised against themselues, desired ed away secretly o n to to Mountrath" From the historians' work, we nd that the depositions contain references to more than 60,000 people and 7,000 locations. The people in question range 5 http://1641.tcd.ie/deposition.php?depID=815351r406 from individuals of great historical importance such as Sir Oliver Cromwell, Sir Phelim O'Neill, and King Charles I, to individual servants and common folk who were a ected by the rebellion. Locations similarly range from cities such as Dublin which still ourish today, to small plots of land which have been lost either as their names changed or borders shifted.
We know that several of the entities extracted from the depositions are duplicates. However, the huge range in spelling variations and naming conventions makes it extremely di cult to determine which mentions of entities in the depositions might be references to the same person or place. Compounding this problem is the fact that the severity of textual noise means that standard NLP tools can struggle with simpler tasks such as sentence chunking or Named Entity Recognition. Performing reliable analysis based on the language of the depositions is, to say the least, di cult.
We are not the rst to attempt to decipher the contents of the depositions
using computational methods. The CULTURA project [
An earlier study conducted on a manually annotated subset of the depositions
attempted to assess the feasibility of automatically enriching the collection using
standard Entity Linking tools [
While there are many aspects of the depositions that we may choose to model into a potential ontology to generate linked data, we focus on people and places due to their perceived importance in the documents. For historians there are still a number of unanswered research questions about the motivations and in uences behind the rebellion. A linked data solution may assist them when investigating these questions. However, modelling the depositions is extremely challenging for a number of reasons.
First, there is no true, de nitive list of people and places on which to base the ontology. Ideally any ontology that we create would be populated with a set of distinct entities that can be found in the depositions. While there are resources which can help us to determine this set of entities (as discussed below), there is still noise present in these sources. Sometimes people are referred to by lineage rather than their actual name, e.g. \the heirs of Mr. Gale", or even by title e.g. \Bishop of Meath". Given these ambiguities, there is much risk of accidentally omitting or con ating entities when the ontology is being constructed.
Second, the inconsistent language of the depositions means that multiple variant spellings for people and places can be found throughout the collection. If a suitable ontology can be constructed to represent each entity, discovering all the possible variant names by which it may be referenced would be a monumental task. Sometimes these variations are minor spelling di erences e.g. \Florence FitzPatrick" being referred to as \F orenc F tz Patrick", but some are more severe, such as the \Barony of Fassadinin" being referred to as the \Barrony of assa and Dyninge". Detecting such di erences is di cult through an automated process and requires an expert to assess its correctness.
Third, if we are to construct this ontology with an eye to automatic enrichment, then the inclusion of links between entities and how to establish them is an important consideration. We could use familial connections, but we are not aware of any reliable sources which document these in a readily adoptable manner. On what basis then are we to establish relationships between our entities? Currently, there is no reliable way to specify that a relation is likely without stating it as a fact in an ontology.
We must also exercise some degree of caution in our attempts to annotate the depositions. If the intention is to assist scholars with their research, then the information conveyed by the proposed solution must be accurate. This can be a subtle problem. For example, if we consider the entity \the Pope", should this be used to describe the role of the head of the Catholic Church, or should it describe an individual who held that role? If we assume the latter, then we must be sure to refer to the correct pope for the source document, which involves additional knowledge that may not be readily available in the knowledge base. Pope Urban VIII held the position until 1644 when he passed away and was replaced by Pope Innocent X. Modelling evolving entities such as these is a common problem in the cultural heritage domain.
In spite of these challenges, resources do exist which can help us to generate lists of distinct entities. Three resources at our disposal are: { The Down Survey: A complete national survey of land in Ireland after the rebellion. The survey was conducted in order to establish which lands should be forfeited as penalty for crimes during the rebellion. { The Statute Staple: A record of transactions between individuals. The staple documents goods bought and sold, and provides information about debts owed between various parties before the rebellion { The Books of Survey and Distribution: A list of properties held by various land owners. These documents were used to determine taxes based on land ownership.
These documents have been the subject of historical research for a number of years and were some of the major contributing sources for the Petty Maps project 6.
Given the three resources above, we have begun the process of constructing
an ontology to model the entities present in the depositions. Using standard
record resolution methods based on the DICE coe cient [
Our goal is to eventually use the information in this new ontology as a knowledge base for an EL system. Using the information derived from the three sources, it may be possible to automatically enrich the depositions with semantics using standard EL methods. Furthermore, if the entities described by the ontology are found to be recorded in other historical sources as they are digitised, it may be possible to use this knowledge to automatically link and integrate them with the depositions and the greater web of knowledge.
In order to facilitate EL methods, we are attempting to capture features that are commonly used by EL algorithms. The variety of surface forms which may link to an entity will be an important consideration. As yet we are contemplating which relationships are likely to be the most helpful. Our initial focus is on attempting to model ownership of land and debts owed between individuals as monetary records are often well documented based on the resources available to us.
Where possible we will attempt to reuse existing knowledge by linking entities in our ontology to corresponding entries in DBpedia, Geonames or other similar ontologies using properties such as owl:sameAs. Existing commonly used vocabularies will be used to describe the entities, though it may be necessary to extend or create additional properties for accurate representations of entities. For example, vCard7 has a number of useful properties for describing people, and while it does capture the concept of honori cs applied to people e.g. \Mr.", \Mrs.", \Dr." etc., it does not quite distinguish between an honori c and a title e.g. \Earl" is a title held that is passed on to other people, rather than an honori c applied to an individual person's name. As yet we are assessing the suitability of vocabularies such as FOAF, vCard and SIOC8 for describing such properties.
7 https://www.w3.org/TR/vcard-rdf/ 8 https://www.w3.org/Submission/sioc-spec/
Performing automatic enrichment of cultural heritage collections is challenging for a variety of reasons. As evidenced by our own experience, and the documented experience of other researchers, nding knowledge bases with adequate coverage for a given cultural heritage resource is extremely di cult. While developing an entirely new ontology that does not reuse existing knowledge is a solution, if not done properly it can lead to inaccurate or incompatible knowledge representations that negate one of the greatest bene ts of linked data i.e. connectivity among disparate collections. It is of far greater bene t to the community if these new vocabularies can be integrated with existing semantic web resources in a seamless fashion.
Due to the issue with well-known knowledge bases not covering a large percentage of the entities in specialised cultural heritage collections, it is likely that curators of such resources will need to develop their own ontologies in order to accurately represent the semantics of their data. While it is good to expand the web of knowledge with this new information, we suggest that due care be given to the structure of these resources and to how this structure may lend itself to informing automatic enrichment processes going forward. Methods such as REDEN may exploit owl:sameAs or similar relationships between a new ontology and more established ones in order to knit together various knowledge bases for the EL process. If automatic enrichment services can make use of the information in new linked data resources, then future annotation processes may be expedited as new collections are digitised and made available.
The ADAPT Centre for Digital Content Technology is funded under the SFI Research Centres Programme (Grant 13/RC/2106) and is co-funded under the European Regional Development Fund.