In our aim to make the information encapsulated by biodiversity literature more accessible and searchable, we have developed a text mining-based framework for automatically transforming text into a structured knowledge repository. A text mining workflow employing information extraction techniques, i.e., named entity recognition and relation extraction, was implemented in the Argo platform and was subsequently applied on biodiversity literature to extract structured information. The resulting annotations were stored in a repository following the emerging Open Annotation standard, thus promoting interoperability with external applications. Accessible as a SPARQL endpoint, the repository supports knowledge discovery over a huge amount of biodiversity literature by retrieving annotations matching user-specified queries.
1 Introduction Big data—huge data collections—are proliferating in many disciplines at a rate that is much faster than what our analytical abilities can handle. One particular discipline that has amassed big data is biological diversity, more popularly known as biodiversity: the study of variability amongst all life forms. On the one hand, researchers in this domain collect primary data pertaining to the occurrence or distribution of species, and store this information in a structured format (e.g., spreadsheets, database tables). On the other hand, findings or observations resulting from their analysis of primary data are usually reported in literature (e.g., monographs, books, journal articles or reports), often referred to as secondary data. Written in natural language, secondary data lacks the structure that primary data comes with, rendering the knowledge it contains obscured and inaccessible. In order to make information from secondary data available in a structured and thus searchable form, we have developed a repository containing information automatically extracted from biodiversity literature by a customisable text mining workflow. To maximise its interoperability with external tools or services, we have made the knowledge repository available as a Resource Description Framework (RDF) triple store that conforms with the Open Annotation standard1. We then demonstrate how the repository, accessible as a SPARQL endpoint, facilitates query-based search, thus making the information contained in biodiversity literature discoverable.
A handful of other tools for storing biodiversity
information in RDF format exist. Most of them,
however, do not have the capability to
automatically understand text written in natural language.
Tools such as RDF123
Methodology In this section, we present in detail our framework for constructing the knowledge repository. We begin by briefly describing the corpus of biodiversity documents that was utilised, and then outline the various steps in the text mining workflow. We finally proceed to explaining how the Open Annotation specification was adopted in order to store the information extracted from our corpus. 2.1
The Biodiversity Heritage Library (BHL)2 is a database of biodiversity literature maintained by a consortium of natural history and botanical libraries all over the world. A product of the various partners’ digitisation efforts, BHL currently contains almost 110,000 titles, equivalent to almost 50 million pages of text resulting from the application of optical character recognition (OCR) tools on scanned images of legacy materials. For this work, we decided to narrow down the scope of the knowledge repository to the requirements of our ongoing project whose aim is to comprehensively collect both primary and secondary information on biodiversity in the Philippines.
To this end, we retrieved only the subset of English BHL pages which are relevant to the Philippines, i.e., the union of (1) the set of pages which mention either “Philippines” or “Philippine” within their content, and (2) the set of pages contained by books or volumes whose titles mention “Philippines” or “Philippine”. This resulted in a corpus of a total of 155,635 pages (around 12GB in size).
One of the primary interests of our collaborators in the project is the discovery of fundamental species-centric knowledge, particularly information on species’ geographic locations, habitat, anatomical parts as well as authorities (i.e., persons who described them). Guided by user requirements, we cast this work as an information extraction task requiring: (1) named entity recognition (NER) for taxa, locations, habitat, anatomical parts and persons; and (2) binary relation extraction focussing on the following types of associations: taxon-location, taxon-habitat, taxonanatomical part and taxon-person.
To carry out these tasks on our corpus, we integrated various natural language processing (NLP) tools into one workflow using the Argo platform. Argo3 is a web-based, graphical workbench that facilitates the construction and execution of bespoke modular text mining workflows. Underpinning it is a library of diverse elementary NLP components, each of which performs a specific task. Argo’s graphical block diagramming interface for workflow construction provides access to the component library, representing them as configurable blocks that can be interconnected to define processing sequence.
The workflow that we developed, depicted in
Figure 1, combines several components for
preprocessing, synactic and semantic analyses. It
begins with an SFTP Document Reader which loads
the plain-text corpus from a remote server. This is
followed by a Regex Annotator which attempts to
detect paragraph boundaries based on the
occurrence of newline characters. The paragraphs are
then segmented by the LingPipe Sentence
Splitter4 into sentences, each of which is decomposed
into tokens by the GENIA Tagger
The succeeding components in the workflow contribute towards the relation extraction task. Enju Parser performs deep syntactic parsing and extracts syntactic dependencies amongst sentence tokens. Its outputs are used by the next component, the Predicate Argument Structure Extractor, to compute semantic dependencies in the form of predicate-argument structures. The five instances of the Dependency Extractor component then makes use of the predicate-argument structures to detect relationships between names categorised under the specified entity types. The first instance, for example, detects only relationships between taxon and person names, while the last one captures related anatomical parts and qualities. The Type Mapper ensures that all of the named entities and relations extracted conform with the same annotation schema before they are all saved in Open Annotation format by the last component, the Annotation Store Writer. We briefly describe next how our extracted annotations are encoded according to this format.
Adopting the Open Annotation model The Open Annotation (OA) Core Data Model is an emerging W3C-recommended standard for encoding associations between any annotation and resource (i.e., what is being annotated). Built upon the Resource Description Framework (RDF), the OA model represents an annotation as having a body and a target, with the former somehow describing the latter, e.g., by assigning a label or identifier. Following this fundamental idea and other relevant recommendations given in the specification6, we represented the named entity and relation annotations extracted by our text mining workflow in OA format, as depicted in Figure 2. For brevity, prefixes were used in this figure instead of full namespaces, e.g., oa for http://www.w3.org/ns/oa#.
Once the RDF triples had been generated, they were automatically loaded onto a new Apache Jena TDB7 store, which was then exposed as a SPARQL endpoint by Fuseki8. 3
Example use case We present an example of how our repository, now in the form of a SPARQL-enabled triple store, can facilitate knowledge discovery. A user might be interested, for example, in learning which specific geographic locations have been described in the literature as having associations with certain 6http://www.openannotation.org/spec/core 7https://jena.apache.org/documentation/tdb 8https://jena.apache.org/documentation/fuseki2 species, e.g., the bird family of hornbills. Shown in Listing 1 is a query in SPARQL, the query language for RDF, that retrieves a list of all such locations, as well as the number of times that the relationship was mentioned in the source document. Listing 1: An example SPARQL query that will retrieve locations related to hornbills. In this paper, we presented a framework for building a knowledge repository that: (1) applies a customisable text mining workflow to extract information in the form of named entities and relationships between them; (2) stores the automatically extracted knowledge as RDF triples compliant with the Open Annotation specification; and (3) facilitates the discovery of otherwise obscured knowledge by enabling query-based retrieval of annotations from a SPARQL endpoint. We note that the triple store can be exposed via other application programming interfaces, i.e., web services that abstract away from SPARQL to make querying straightforward for non-technical users.
We envision that our knowledge repository will facilitate the enhancement of search applications, e.g., information retrieval systems. It has been made accessible as a SPARQL endpoint9 that accepts POST requests. The body of the request should be set to a valid SPARQL query while the headers should be configured to hold the following name-value pairs: (1) Accept: text/csv and (2) Content-Type: application/sparql-query. Acknowledgments We would like to thank Prof. Marilou Nicolas for her valuable inputs. This work is funded by the British Council [172722806 (COPIOUS)], and is partially supported by the Engineering and Physical Sciences Research Council [EP/1038099/1 (CDT)].