Methodology

Topics

Organisation Topics

The topics for each organisation are provided using the DBpedia Spotlight Tool

with some subsequent data processing and filtering.

To understand how the Spotlight Tool works, a brief overview of DBpedia and Linked Open Data / Semantic Web is provided in the Appendix below.

DBpedia Spotlight Tool

The Spotlight Tool extracts DBpedia URIs from unstructured data, providing high-quality linked data from text sources. It does this using a four-step approach:

Spotting: the process of identifying strings or substrings (known as surface forms in the literature) that may be entity mentions.
Candidate selection: selecting a set of DBpedia resources that are candidate meanings for the surface forms identified by the Spotting stage.
Disambiguation: deciding on the candidate resource (i.e. the DBPedia URI representing a Wikipedia page) that is most likely represented by the entity mention.
Filtering: adjusting the annotations to user input or task specific requirements. For instance, you can use the DBpedia ontology to narrow the set of possible candidate forms to a subset of the DBpedia taxonomy.

The URIs extracted by DBpedia Spotlight Tool are what we define as Topics for the Organisation whose description was fed as input to the annotation process.

A much more comprehensive explanation of the process can be found in the following papers:

Subsequent Data Processing

The DBpedia Spotlight Tool API provides a number of parameters to tune the annotation process, including confidence, support, types and SPARQL filtering. For this project, we have only used confidence as input for distinguishing annotations. The API accepts text as input, and provides an arbitrary number of DBPedia resources as output depending on the various inputs provided.

The confidence score is a number in the interval [0, 1] which controls the disambiguation step. The higher this number is, the more likely that the annotation for a particular surface form matches its intended candidate. When using the annotation API, you can supply the confidence score to increase the number of annotations provided for the input text, at the cost of allowing potentially erroneous annotations to be included in the final result set. In Machine Learning terms, it controls the balance of Precision vs Recall.

In our implementation, we convert the confidence to an integer number in the range of [0, 100].

For each organisation we annotate its description field. To investigate the range of different annotations and their associated confidence scores, we perform the annotation step 11 times on each input, with a stepwise increase of confidence from 0 to 100 in 10 increments (e.g. 50, 60 etc). We save each resource only once, at the highest possible confidence found for that resource, e.g. if dbr:Brain was found at confidences 10, 20, 30 and 40 but not at 50, then the final output would include dbr:Brain with an associated confidence value of 40.

In this manner, each DBpedia Resource is saved at its highest confidence level, which then allows us to filter out lower quality resources.

Following the recommendation of the paper (a) we filtered out DBpedia resources scoring less than 60.

Missing Organisation Descriptions

A number of organisations were missing their description field, which is necessary to provide entities using the DBpedia Spotlight tool. To automatically extrapolate descriptions for these organisations, we used the free Bing search API

. We provided the organisation’s name as the search query, and used the search results’ top result to populate the missing description field. The script for this process is found at src/bin/ai_map/fillMissingData.mjs, and more thorough documentation is provided at src/bin/ai_map/README.md.

Source code

The source code for the annotation process is hosted on Github at https://github.com/nestauk/dap_dv_backends

The main entry point for the annotation process is the script located at src/bin/es/annotate.mjs. This script takes as input the location of a Spotlight endpoint, an Elasticsearch index and domain which hosts the data to be annotated, and the name of the field to annotate (i.e. description). The script accepts additional parameters, which are fully documented at src/bin/es/README.md.

Internally, the script makes use of various modules we’ve written to allow for the use of the Spotlight API to annotate an Elasticsearch index using the methods we’ve outlined above. The majority of the logic used is contained in the src/node_modules/dbpedia and src/node_modules/es directories.

We host our own private annotation server on AWS, but there is a publicly available endpoint provided by the University of Leipzig at https://api.dbpedia-spotlight.org/en/annotate

. You can also use the Spotlight docker container

to host the API locally and point the script to the localhost port.

Appendix: Semantic Web and DBpedia

The Semantic Web is an extension of the World Wide Web, with the goal of making the web machine-readable. It uses a number of standards to formally describe meta-data about the web and existing data sources. For example, the Resource Description Framework

(RDF) is a standard set by W3C which can be used to describe ontologies, relationships between data points, and graph networks more generally.

W3C gives this definition:

The Semantic Web is a Web of data — of dates and titles and part numbers and chemical properties and any other data one might conceive of. RDF provides the foundation for publishing and linking your data. Various technologies allow you to embed data in documents (RDFa, GRDDL) or expose what you have in SQL databases, or make it available as RDF files.

DBpedia

offers a good example of the type of problem the Semantic Web intends to solve, and it is also the source of data we use in combination with the Spotlight Tool to provide automatic topics for the organisations in the project’s dataset.

The DBpedia project aims to extract structured content from information contained in various Wikimedia projects (primarily Wikipedia). The result of this extraction is a Knowledge Graph, where data are stored in machine readable format to allow for the collecting, organising, sharing, searching and using of the results in an open and transparent manner.

DBpedia data is served as Linked Data, so that it sits as an important part of the Semantic Web and allows for complex, graph-like queries, such as ‘find cities with low criminality, warm weather and open jobs’. DBpedia contains all of the information found on Wikipedia (and other Wikimedia projects), whilst also connecting Wikipedia articles and the information contained within those articles in a machine-readable format, so that the example query above is actually possible.

Each Wikipedia article is represented using a Uniform Resource Identifier (URI), and information on these articles are also represented using URIs. Special predicate URIs are used to tie data points together to represent factual statements. Standards such as RDF and OWL are used to code these statements into machine readable formats.