In this paper we present a collection of semantically-encoded biographies of authors who were born in former colony countries from 1945. The data set relies on an ontology that represents the life of an author through the two key concepts of migration from birth place and legal status in a country, both modeled on two Ontology Design Patterns: Time Indexed Person Status and Basic Execution Plan. Together with the resource, we describe a pipeline to convert the textual biographies of the authors gathered from Wikipedia into the roles experienced by them in migrations. The pipeline includes modules for linguistic preprocessing and named entity recognition, and an entity linking step relying on Wikipedia and Wikidata APIs to link places and organizations to their respective countries. A set of lexico-semantic patterns based on verb classes from the Uni ed Verb Index has been developed in order to extract migration-related knowledge from unseen text biographies.
Under-representation of non-Western people is an open issue with a long
tradition [
Our work addresses this topic by providing structured knowledge about
writers who su er a lack of representation on Wikipedia due to their ethnic
origin [
Commons License Attribution 4.0 International (CC BY 4.0).
biographical events from Wikipedia through the adoption of Lexico-Semantic
Patterns [
The paper is structured as follows. In Section 2 we discuss the Linked Data projects that inspired our work, and review the state-of-the-art approaches to event extraction and encoding. In Section 3 we present the Ontology of UnderRepresented Writers, describing how we encoded their biographies through recurrent semantic patterns, and how we modeled the interplay between the authors and their places of birth. In Section 4, we present the pipeline for the automatic extraction of biographical events through Lexico-Semantic patterns. Finally, in Section 5, we analyze, and evaluate results. A discussion about open issues and future work concludes the paper. 2
In recent years, thanks to the availability of sources in a digital form, a new
interest in the study of biographies has arisen is literary, cultural and historical
studies. In particular, three existing Knowledge Graphs share many similarities
with ours: the Orlando project1, Enslaved2, and WeChangEd3. The Orlando
Project is a collection of biographies of 1; 300 British women writers; Enslaved is
a data set of 509; 783 people of the historical slave trade developed from 8
preexisting archives; WeChangeEd is a collection of 1; 800 female editors born between
1710 and 1920, aligned with Wikidata. All these data sets rely on Semantic Web
technologies [
The URW project has a similar perspective to these projects in terms of aiming to represent a group of persons sharing a speci c condition. However, the concept of \being under-represented" is challenging to model, because it has blurred boundaries and it can be very subjective. Our project intentionally does not rely on a taxonomy of ethnicities, choosing instead to fully describe the interplay between a person and the places where they lived their life, in order to avoid a Western representation of non-Western writers biographies.
Several approaches aimed at encoding and annotating events have been
proposed in the last years. Despite the common representational goal, these
approaches vary signi cantly, since events can be formalized at di erent levels of
granularity. The Biography Ontology [
Other works analyze events at a word level. The ACE/ERE projects [
Biographical event extraction from raw text is the subject of works relying
on Wikipedia as a source of knowledge. The Pantheon 1.0 data set [
Other projects have attempted to extract time and geographical information
from biographical texts. Russo et al. [
Our proposal aims at extracting geographical knowledge and life events jointly,
to provide a semantic model for representing biographies. Unlike existing
approaches, which are focused on detecting the lexical entries triggering an event [
A Semantic Model for Under-Represented Writers The semantic model is designed with the purpose of providing a formal and objective description of authors who are potentially under-represented due to the context where they were born. In particular, it encodes biographical events and situations in which a dul:Person is: (i) a writer and (i) has experienced the condition of being under-represented. In this way, a correlation between biographical events and literary production of under-represented authors can be drawn, and employed to gain insight on the motivations and themes re ected in their narratives. The main components of this semantic model are: the condition of being under-represented and the identi cation of objective criteria to classify countries which correlate with this condition.
Biographical patterns. According to our formalization, a writer who is
underrepresented is a person who published one or more literary works, and may have
experienced the process of migrating, intended as the movement from a country
to another, and the condition of living in a given country after leaving one's
place of birth. Within the latter situation, the author's legal or professional
status may be expressed. Our solution to encode these situations draws from
the ODP framework, which provides foundationally sound, re-usable building
blocks for representing common patterns across ontologies. More speci cally,
the urw:Migration pattern refers to the BasicPlanExecution ODP [
In our formalization, an under-represented country must be a former colony, it must have a medium or lower HDI (below 0:8), and it must fall within the second half of the ranking of countries by mobility score. The Named Authority List of countries maintained by the European Union6, an authoritative, comprehensive, and multilingual reference for country names, has been used to standardize and index all these sources of geographical knowledge.
Concerning the linguistic resources, we rely on the Ontolex-Lemon model,
which [
Finally, the PROV-O Ontology [
6 https://op.europa.eu/en/web/eu-vocabularies/dataset/-/resource?uri=http: //publications.europa.eu/nresource/dataset/country the LSPs as prov:SoftwareAgent, and the textual Wikipedia biographies as the source of knowledge from which biographical patterns have been derived. 4
From Ontology Patterns to Lexico-Semantic Patterns Before collecting the biographies from Wikipedia, under-represented writers have been identi ed through the occupation Wikidata property (WDT:P106). Each person who worked as a writer, novelist, or poet has been collected and classi ed by retrieving the country of origin associated to her/his birthplace (WDT:P19). For each author, the biography in English language, if present, has been retrieved from Wikipedia. The total amount of collected person entities is 114; 675. Writers who were born from 1945 on, in any Asian or African under-represented country (see Section 3) have been chosen to highlight only on biographies of people who experienced or born after the Decolonization process.
Starting from this initial corpus, a pipeline to convert raw texts biographies
in TIPS, and Migration classes based on Lexico-Semantic Patterns (LSP) [
The rule $subject : Concept COMP RB? IN? $object : Concept matches the phrase Administrative territory of Prague is divided into localities retrieving a mereological relation between Prague, and localities to be stored in an ontology.
Our pipeline is based on three steps: text parsing, LSP development, Information Extraction. Text parsing. Using the SpaCy library7, each biography has been split in sentences, and only the ones containing at least one entity of the type Organization (ORG), or Geopolitical Entity (GPE) have been stored in JSON format, together with the name of the author, and her/his country and year of birth. Below, there is an example of an item, in JSON format, referring to the Nigerian writer, and radio presenter Dotun Adebayo: author: Dotun Adebayo, birthPlace: Nigeria, birthYear: 1960, places: [(Stationers' Company's Comprehensive School, ORG),(Stockholm University)], sentence: He then went on to Stationers' Company's Comprehensive School in Hornsey, North London, followed by
Stockholm University, where he studied Literature.
In parallel, each ORG and GPE has been linked with the respective country. All the strings identi ed as geopolitical entities or organizations by the SpaCy Named Entity Recognition module have been used as an input for search through the Wikipedia API. The rst 10 results of the search have been subsequently analyzed, and, among them, the rst candidate that holds the Wikidata property `country' (WDT:P17) has been selected, if any. Only the 25; 554 sentences containing an ORG or GPE belonging to di erent countries than the birth country of an author have been selected for the next step.
LSP development. After the sentences have been collected, a random subset of them has been analyzed in order to de ne LSP rules for encoding the biographic facts contained in the raw text into the two main patterns of the URW ontology: urw:TimeIndexedPersonCondition (TIPS), and urw:Migration.
Given the structure of these patterns, three key elements have been identied as necessary in an input sentence to make it a candidate trigger: a verb expressing a change of place or a condition (e.g.: eeing a country, obtaining a graduation), a preposition, and an entity of the type Organization (ORG) or Geo Political Entity (GPE) belonging to a di erent country from the place of birth. For instance (see Figure 4), the elements in bold face in the sentence \He [Dotun Adebayo] then went on to Stationers' Company's Comprehensive School in Hornsey, North London." match the pattern (escape-51.11)(to|at|in)(GPE|ORG). So, from this rule, the following RDF triples are extracted: [ a urw:Migration; dul:isSettingFor [ a urw:MigrationRole; dul:isDefinedIn Study_Abroad dul:isRoleOf Dotun_Adebayo.
The subsequent step in the de nition of the LSPs has been the clustering of
verbs through a mapping to general verb types, aimed at reducing the number
of patterns and increasing their recall. To do so, we employed the Uni ed Verb
Index8, a repository resulting from the mapping of several lexical resources that
provides syntactic and semantic frames of English verbs. In particular, we linked
the verbs in our data to the VerbNet classes in Uni ed Verb Index (UVI) [
As anticipated in Section 3, the mapping between LSPs and VerbNet classes
is expressed in the ontology through the Ontolex-Lemon speci cation [
has a corresponding ontolex:LexicalSense which is the v#2009433 WordNet o set. The latter is one of the possible lexicalization of the escape-51.1-1-1 VerbNet class, which is the ontolex:Concept.
Information Extraction. After the creation and re nement of the LSPs, 53 rules of the form:
VerbNet class $preposition GPE|ORG have been formulated and applied to the annotated sentences.
The following is an example of how the same LSP matches sentences with different verbs and preposition, and encodes them as urw:TimeIndexedPerson Status:
LSP: obtain-13.5.2 fromjforjatjbyjinjas GPEjORG Ajunwa9 received her BA at University of California, Davis in 2003.
He held a master's degree in Theatrical Directing which he obtained from the University of So a 5
From the life events encoding pipeline (Section 4) 12; 147 sentences containing an instance of TIPS, Migration, or both have been obtained.10 Some preliminary statistics can help to assess the relevance of the data stored in the KG. The resulting Knowledge Graph includes 2; 618 di erent authors' biographies, place of birth, and year of birth. 1; 638 of these authors were born in Asia, 980 in Africa. In total, 39; 167 RDF triples have been stored in the data set.
In order to test the precision of the LSP (Lexico-Syntactic Patterns), we manually evaluated a random sample of 2; 555 sentences, which correspond to the 10% of sentences containing at least one GPE or ORG di erent from the author country of birth. Each sentence was labelled as expressing a `TIPS|Migration' (48:5%) or `None' (51:5%), then compared with the patterns.
Table 1 shows that LSPs performed with a precision of 0:68. The manual analysis of prediction errors revealed they had several causes. In some cases, the subject was not the author but another person (e.g., `His father left India
10 The rst version of the data set is publicly available on the GitHub
repository of the Under-Represented Writers project https://w3id.org/
UnderRepresentedWritersOntology/. 10; 569 sentences expressing at least one
TIPS were detected, 3; 549 with Migration patterns. In 1; 971 cases both are present
in the same sentences.
in early 1963 to study at Oxford University'). Another source of error is the
presence of reported speech of the writer (e.g., `Members of her African
audience have asserted that Thiam does not understand why women may support
FGM'). Finally, both the NER and the entity linking pipeline seem to introduce
false positives (e.g., in the phrase `Shatrughan Sinha, has also spoken in Kumar's
favour on Twitter', Twitter is marked as an organization in the United States).
It is important to mention an imbalance in the performance of the two
biographical patterns: Migration situations are retrieved with a precision of 0:805,
in line with recent ndings from the literature [
In this paper we presented a pipeline to extract life events of writers born in an Asian or African Former Colony Countries from 1945 onwards from Wikipedia biographies through Lexico-Semantic Patterns. At the present stage, the data set includes 12; 147 biographical events about 2; 618 authors.
A manual evaluation of a sample of the results showed a good precision of Lexico-Semantic Patterns. However, some rules need to be further specialized in order to extract a taxonomy of TIPS-related conditions. Despite these limitation, it is important to underline that a pipeline based on a small set of rules has produced a relatively large corpus, from which holistic knowledge about life's narratives can be extracted, and generalized to other types of biographies. Future works must take into account the chronological arrangement of Migration and TIPS patterns within a whole biography, and generalize Lexico-Semantic Patterns to other categories of under-represented people { ethnic minorities and second generation migrants, people with other occupations { which can be collected in the URW Knowledge Graph.