Readers, as well as journalists, are overwhelmed with the information available in online news articles, making it very di cult to verify and validate their content. An important tool to support readers in this task is that of named entity linking (NEL), i.e., semantically annotating entities mentioned in text with entities described in knowledge bases (KBs). In this poster, we introduce EL-NEL, the NEL component of DebateLab, a project with the more general objective of representing, mining and reasoning with online arguments. The models trained in EL-NEL are o ered as open source.
The current news ecosystem is characterised by a rapid spread of news and a vast increase in the quantity of information available. This abundance of information makes it di cult for readers to identify good-quality journalistic resources, which are set apart by the use of credible and justi ed information and opinions, in the form of well-formulated arguments based on facts. The problem is aggravated by the fact that the provided information and arguments are often con icting, and it is very di cult for a reader to evaluate their credibility without access to the relevant factual resources. This di culty often leads to unwanted behaviours, such as the placement of blind and unjusti ed trust to speci c opinions or sources, the generation of echo chambers, con rmation bias, and increased radicalisation and polarisation.
The DebateLab project1 is conducting research towards representing, mining and reasoning with online arguments. The goal of DebateLab is to o er a suite of tools and services that will assist both the work of the professional journalist in accomplishing everyday tasks (e.g., writing, archiving, retrieving articles), as well as the activity of the ordinary Web user (reader) who wishes to be well-informed about topics or entities of interest (e.g., persons, locations, events).
An essential component of the envisioned DebateLab tools is the task of identifying sub-strings (aka entity mentions ) that refer to real-world entities within an unstructured/textual argument, and mapping them to uniquely identi ed entity descriptions provided in target Knowledge Bases (KBs), a task known as named entity linking (NEL). NEL would allow the association of arguments with factual resources, thereby facilitating the validation of an argument, and the evaluation of its strength, veracity and applicability.
This work describes EL-NEL, the NEL component of DebateLab. Although
NEL is quite popular in the literature, and various pre-trained models exist,
the vast majority of those works are only applicable to the English language.
Indeed, many multilingual or language-agnostic approaches are usually based
on resources that are of poor quality [
More speci cally, this poster describes a pipeline for non-English NEL, which
employs alternative state-of-the-art tools (e.g., BERT [
Named Entity Recognition (NER). In the case of arguments only (and
not tags), we perform NER to extract entity mentions. We evaluated o
-theshelf NER tools (e.g., SpaCy3, polyglot4), but we also ne-tuned a Greek BERT
model [
Our preliminary results, presented in Table 1, indicate that the ne-tuned BERT model (with a micro-averaged F1 score 91%) yields a negligible improvement over a SpaCy model (90% micro-averaged F1) that we trained on the same Greek corpus (elNER), and a signi cant improvement over the pre-trained Greek SpaCy model (51% micro-averaged F1).
Candidate Generation. For each entity mention, we generate a list of
possible entities from a multitude of wiki-based KBs (Wikidata, DBpedia, Wikipedia,
YAGO). This component is mostly relying on existing lookup services o ered by
those KBs. For each entity mention, we perform several lookup queries to those
services and rank the generated candidates based on the employed service and
the pre-processing applied to each query (e.g., lemmatizing using TreeTagger [
Wikidata. We look up the entity mentions in Greek using the wbsearchentities service and retrieve the top-k results per query (k empirically set to 5). If we get no lookup results, we try again rst with the lemmatized and then the stemmed version of the original query (i.e., entity mention). The order of executing those lookup queries is maintained in the output rankings of the candidate entities.
Wikipedia and DBpedia. We extract o ine alternative names and redirect pages from the Greek Wikipedia, which we add to the list of candidate entities for each entity mention. In addition, the search mechanism of Wikipedia was also employed and the high ranked results were added to the candidate set. For DBpedia candidates, we simply use the Wikipedia su xes.
Wiki er. We get additional Wikipedia candidates by calling the Wiki er [
Disambiguation. We distinguish between two cases for entity disambiguation: (i) NER has detected a single entity mention, and we return the results obtained by a single ranked list of candidates, and (ii) more than one entity mentions are detected, so we consider the cohesion of the candidates, returning those whose entity embeddings are more similar to each other.
We examine three di erent types of entity embeddings, two context-independent
(fastText [
In our preliminary experimental results, we observe that BERT performs best for the case of single entity mentions, since it is the only context-dependent method, while fastText performs better when more than one entity mentions appear in an argument, since in that case we can also consider the other entity mentions as context.
Extending this approach to other languages. The proposed pipeline
can be extended to other languages too, as long as a medium-resource coverage
(i.e., NER corpora and/or pre-trained NER models, POS taggers, a relevant
sizeable Wikipedia corpus) is available in the target language. The adoption
of the approach in low-resource languages is also applicable and can bene t
from cross-lingual transfer learning approaches [
In this poster, we introduced EL-NEL, a Greek named entity linking
component which is employed in the DebateLab project. We have brie y described
the modular architecture and alternative options for each component of
ELNEL, as well as some preliminary ndings. In our ongoing and future work, we
plan to explore KGs beyond the Wiki-based ones. We also plan to incorporate
graph embeddings (e.g., ED-GNN [
This project has received funding from the Hellenic Foundation for Research and Innovation (HFRI) and the General Secretariat for Research and Technology (GSRT), under grant agreement No 4195.