=Paper=
{{Paper
|id=Vol-3683/sample-1col
|storemode=property
|title=GeoEDdA: A Gold Standard Dataset for Geo-semantic Annotation of Diderot & d'Alembert's Encyclopédie
|pdfUrl=https://ceur-ws.org/Vol-3683/paper1.pdf
|volume=Vol-3683
|authors=Ludovic Moncla,Denis Vigier,Katherine McDonough
|dblpUrl=https://dblp.org/rec/conf/ecir/MonclaVM24
}}
==GeoEDdA: A Gold Standard Dataset for Geo-semantic Annotation of Diderot & d'Alembert's Encyclopédie==
https://ceur-ws.org/Vol-3683/paper1.pdf
GeoEDdA: A Gold Standard Dataset for Geo-semantic
Annotation of Diderot & d’Alembert’s Encyclopédie
Ludovic Moncla1,* , Denis Vigier2 and Katherine McDonough3
1
INSA Lyon, CNRS, UCBL, LIRIS, UMR 5205, F-69621
2
Université Lumière Lyon 2, ICAR, UMR 5142, Lyon, France
3
Lancaster University, UK
Abstract
This paper describes the methodology for creating GeoEDdA, a gold standard dataset of geo-semantic
annotations from entries in Diderot and d’Alembert’s eighteenth-century Encyclopédie. Aiming to explore
spatial information beyond toponyms identified with the commonly used Named Entity Recognition
(NER) task, we test the newer span categorization task as an approach for retrieving complex references
to places, generic spatial terms, other entities, and relations. We test an active learning method, using the
Prodigy web-based tool to iteratively train a machine learning span categorization model. The resulting
dataset includes labeled spans from 2,200 paragraphs. As a preliminary experiment, a custom spaCy
spancat model demonstrates strong overall performance, achieving an F-score of 86.42%. Evaluations
for each span category reveal strengths in recognizing spatial entities and persons (including nominal
entities, named entities and nested entities).
Keywords
Geo-semantic annotations, Spatial role labeling, Gold standard dataset, Span categorization, Spatial
humanities
1. Introduction
This paper presents an annotation schema and active learning method for creating a gold
standard dataset of geo-semantic annotations from entries in the Encyclopédie ou Dictionnaire
raisonnée des sciences, des arts et des métiers, a key text of the Enlightenment printed between
1751 and 1772 (in French). Geo-semantic annotation, or spatial role labeling [1], involves
the identification and labeling of place-specific information and semantic classes in text. By
combining geospatial information with semantic context, the GeoEDdA dataset aims to facilitate
multi-scale and -type text analysis. This enables research that depends on evidence of the
interconnection between Enlightenment ideas, historical events, people, generic spatial forms,
and specific places [2].
Like projects such as Living with Machines and Space Time Narratives, we seek to diversify
what we capture when we collect spatial information from historical documents. For the former,
GeoExT 2024: Second International Workshop on Geographic Information Extraction from Texts at ECIR 2024, March 24,
2024, Glasgow, Scotland
*
Corresponding author.
$ ludovic.moncla@insa-lyon.fr (L. Moncla); denis.vigier@univ-lyon2.fr (D. Vigier); k.mcdonough@lancaster.ac.uk
(K. McDonough)
https://ludovicmoncla.github.io (L. Moncla)
0000-0002-1590-9546 (L. Moncla); 0009-0006-0836-0985 (D. Vigier); 0000-0001-7506-1025 (K. McDonough)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
the focus was on improving recognition of formal, theoretically locatable place names at multiple
scales (populated places, but also streets, buildings, or other landmarks) [3]. On the other hand,
the latter project has - similar to our work here - addressed the issue of conceptualizing and
annotating generic place names such as street, lake (e.g. what they term “locales”, in contrast
to “locations”, which are usually toponyms [4]). (Generic place names may also be embedded
within toponyms, of course.) Thus a critical aspect of this computational approach to annotation
lies in prioritizing digital humanities concerns about information retrieval over earlier concepts
of what was defined as “spatial” within the confines of Natural Language Processing (NLP)
and Geographic Information Retrieval tasks such as NER [5] and Spatial Role Labeling [6]. For
historical research focused on the interplay between ideas, discourse, and spatial references, the
commonly used token classification task has proved more restrictive than a span categorization
task.
Like many digital humanities annotation tasks, the act of labeling historical documents
(whether texts, images, or other media) creates an interpretative “world” which researchers
develop claims about, reducing complexity enough to think with and establish relations between
specific parts of documents [7]. The rise of models in digital history has somehow rarely
translated into historians (digital or not) openly talking about modeling. Here, however, we
are nevertheless “opening the black box of interpretation” [8] through the formalization of
modeling in activities like the ones documented in this paper. In this work, as with the examples
from Living with Machines and Space Time Narratives, we move beyond NLP-based traditions
of linguistic annotation to define token-level classes that reflect the needs of spatially-driven
historical research.
Before describing our the methodology, we clarify how we classify tokens. In token classifi-
cation, the focus is on annotating individual tokens within the text such as in Named Entity
Recognition (NER) [9]. Token classification assigns a class to each token, and any token is
associated with at most one entity. This method allows for precise identification of specific terms
or concepts, but may fall short in capturing the holistic meaning that arises from the interaction
of multiple tokens. Approaches in the digital humanities that depend on token-based NER to
identify spatial information in texts often miss complex expressions of spatial information , and
are not designed to recognize generic place names. The former can appear as sequences (“city
in France near the Atlantic Ocean”), embedded in non-spatial entities (“duchess of Brittany”),
while the latter are simply not proper names (“park”). On the other hand, span categorization
extends its purview to encompass such phrases or “nested” expressions [10, 11], aiming to
capture contextual information. This broader approach not only accounts for the interplay
between words but also considers the spatial relationships and semantic (or geo-semantic)
content contained within a defined span. By addressing both token-level intricacies and broader
spans, GeoEDdA strikes a balance between granularity and contextual richness, enabling a
comprehensive exploration of the geo-semantic landscape in the Encyclopédie, and potentially
in other historical texts. This methodological duality ensures that the dataset provides a robust
foundation for research that goes beyond simple toponyms as the dominant unit of spatial
analysis and embraces the flexibility of spans for spatial humanities research.
2. Annotation Schema
In order to facilitate the geo-semantic annotation process, we propose a well-defined tagset
that refines the XML-TEI schema introduced in [12] and adapts it to Encyclopédie entries. The
annotation classes reflect spatial information in the text, including named and unnamed (e.g.
common/generic) entities. We also annotate non-spatial entities, thereby capturing formal and
informal references to places in context alongside people, events, and objects. The tagset is as
follows (with text in French followed by English translations as needed):
• NC-Spatial: a common noun that identifies a spatial entity (nominal spatial entity) includ-
ing natural features, e.g.ville/city, la rivière/river, royaume/kingdom.
• NP-Spatial: a proper noun identifying the name of a place (spatial named entities), e.g.
France, Paris, la Chine/China.
• ENE-Spatial: nested spatial entity , e.g. ville de France/city in France, royaume
de Naples/kingdom of Naples, la mer Baltique/Baltic Sea.
• Relation: spatial relation, e.g. dans/in, sur/on, à 10 lieues de/10 leagues from.
• Latlong: geographic coordinates, e.g. Long. 19. 49. lat. 43. 55. 44.
• NC-Person: a common noun that identifies a person or persons (nominal spatial entity),
e.g. roi/king, l’empereur/emperor, les auteurs/authors.
• NP-Person: a proper noun identifying the name of a person or persons (person named
entities), e.g. Louis XIV, Pline/Pliny, les Romains/Romans.
• ENE-Person: nested person entity, e.g. le czar Pierre/Tsar Peter, roi de
Macédoine/king of Macedonia.
• NP-Misc: a proper noun identifying spans not classified as spatial or person span, e.g.
l’Eglise/the Church, 1702, Pélasgique/Pelasgian.
• ENE-Misc: nested named entity not classified as spatial or person entity, e.g. l’ordre
de S. Jacques/Order of Santiago, la déclaration du 21 mars 1671/the
declaration of March 21, 1671.
• Head: name of article, e.g. Rennes
• Domain-Mark: words indicating the knowledge domain, provided by the editors, e.g.
Géographie, Geog., en Anatomie.
Each category within the tagset is carefully curated to capture the diverse nature of the
content, ranging from nominal ("common", or generic) entities (i.e., NC-*) and named entities
(i.e., NP-*) to nested entities (i.e., ENE-*) and spatial relations (see examples above). As noted
earlier, nominal entity mentions [13] can be used as a co-reference to a named entity, or may
have the ability to refer to a unique object or place, or, alternatively, to a generic type of place.
Nested spans aim to capture and structure nested entities (also known as extended named
entities) [11].
By delineating the semantic content of spans such as spatial features, places, persons, and
their relationships, our annotation guidelines provide clarity and consistency in the annotation
process. This schema not only addresses the intricacies of spatial and person spans but also
incorporates additional elements such as geographic coordinates, miscellaneous spans, domain
markers, and entry names ensuring a holistic and nuanced approach to the geo-semantic
annotation of the corpus. Working with spans enables us to extend annotation to spatial
references and relations that would be undetectable in an entity-driven approach. This tagset
therefore represents a first step towards a spatial-historical interpretation of Encyclopédie content
that is informed by a more thorough semantic analysis of simple and complex toponyms as well
as “extra-toponymic" spatial information (e.g. the unnamed, generic words describing places, as
well as spatial relations).
3. Active learning for dataset labeling
The process of labeling data can be time-consuming, as it is usually done manually. For
humanities research, annotation is often performed by experts or by students trained by experts.
Tools like Recogito [14] facilitate “semi-automatic" annotation by suggesting likely labels, but
for large corpora and for tasks requiring hundreds of examples for each label, machine-learning-
based active learning methods can now aid with annotation. This section describes our use of
such techniques to optimize the labeling process. Active learning involves an intelligent selection
of data samples for annotation, emphasizing the acquisition of the most informative instances
that contribute significantly to model performance. In the humanities, active learning is still a
relatively new approach being explored for reducing the time required to annotate training or
evaluation data from texts, but it is already showing promising results for reducing the number
of annotations required to improve performance on tasks like NER [15]. By leveraging iterative
model-human interaction, active learning not only enhances the efficiency of dataset labeling
but also minimizes the annotation burden on human annotators (in our case, the research team).
3.1. Methodology
To execute our geo-semantic annotation process effectively, we adopted an iterative methodology
using Prodigy1 , a web-based annotation tool developed by ExplosionAI that supports creating
labeled data for machine learning tasks including NER and span categorization.
The initial dataset is composed of 74,000 Encylopédie entries provided by the ARTFL Project2 .
During annotation, data are labeled on a paragraph-by-paragraph basis rather than by whole
entries. This approach enables the subsequent distribution of paragraphs from lengthy articles
across training and testing datasets, ensuring a more granular and representative annotation. By
annotating at the paragraph level, we enhance the flexibility in constructing datasets for training
and evaluation, allowing for a finer understanding of model performance and generalization on
diverse textual contexts.
Initially, a small dataset was manually annotated using Prodigy by the project team mem-
bers, allowing annotators to contribute their expertise in identifying and labeling spatial and
person spans according to the predefined tagset. Subsequently, a first machine learning span
categorization model was trained—specifically, the spaCy3 spancat model embedded in the
Prodigy training pipeline—on this annotated subset. Although initial evaluation scores were
1
https://prodi.gy
2
https://artfl-project.uchicago.edu
3
https://spacy.io
Figure 1: Example of the Prodigy interface for manual validation. Annotators can see basic metadata
for the paragraph as well as the options for the spancat tags.
low, an iterative loop was established. In this loop, the trained model predicted annotations
on additional paragraphs. Human annotators then interacted with these predictions through
the Prodigy interface, correcting and refining the model outputs (see Figure 1). This iterative
process progressively refines both the model’s predictive capabilities and the overall dataset
quality. By linking the strengths of human expertise with machine learning algorithms, our
methodology aims to achieve an effective synergy in the geo-semantic annotation pipeline,
ensuring the creation of a robust and accurate gold standard dataset. Through this iterative
approach, we continuously evaluate the model’s performance, correcting misclassifications,
and reinforcing correct categorizations. This methodology not only contributes to building a
gold standard dataset but also simultaneously trains a machine learning model, completing two
goals at once.
The GeoEDdA gold standard dataset is composed of 2,200 randomly selected paragraphs
from 2,001 entries. All paragraphs were written in French (mid eighteenth-century) and are
distributed among the Encyclopédie knowledge domains as shown on Table 1. Knowledge
domains were assigned to each paragraph using a BERT-based supervised text classification
model trained on Encyclopédie entries (these represent a simplified, composite set of domains
compared to the original domains noted in the entries) [16].
Table 1
Distribution of the annotated paragraphs among a simplified set of Encyclopédie knowledge domains.
Knowledge domain Paragraphs Knowledge domain Paragraphs
Geography (Géographie) 1,096 Literature (Belles-lettres) 65
History (Histoire) 259 Military (Militaire) 62
Law (Droit Jurisprudence) 113 Commerce 48
Physics (Physique) 92 Fine arts (Beaux-arts) 44
Professions (Métiers 92 Agriculture 36
Medicine (Médecine) 88 Hunting (Chasse) 31
Philosophy (Philosophie) 69 Religion 23
Natural history (Histoire naturelle) 65 Music (Musique) 17
(a) in paragraphs classified as Géographie (b) in paragraphs not classified as Géographie
Figure 2: Spans distribution
Figure 2 shows the spans distribution within paragraphs classified under Géographie or not.
80% of spans in paragraphs classified under Géographie refer to spatial spans (i.e., *-Spatial,
Relation, Latlong) (see Figure 2a) against only 25% in other paragraphs (see Figure 2b). This goes
to 25% and 42% for person spans in paragraphs classified under Géographie, and paragraphs
not classified as Géographie, respectively.
Table 2
Distribution of spans (tokens and paragraphs) across the datasets.
Span Train Validation Test Span Train Validation Test
NC-Spatial 3,252 358 355 NP-Person 1,599 170 150
NP-Spatial 4,707 464 519 ENE-Person 492 49 57
ENE-Spatial 3,043 326 334 NP-Misc 948 108 96
Relation 2,093 219 226 ENE-Misc 255 31 22
Latlong 553 66 72 Head 1,261 142 153
NC-Person 1,378 132 133 Domain-Mark 1,069 122 133
Paragraphs 1,800 200 200 Spans 20,650 2,187 2,250
Tokens 132,398 14,959 13,881
With the aim of utilizing this dataset for training and evaluating span categorization al-
gorithms, a train/val/test split was performed. The validation and test sets each consist of
200 paragraphs: 100 classified under Géographie and 100 from another knowledge domain.
The datasets can be downloaded from the HuggingFace Hub4 and Zenodo5 [17] and also from
4
https://huggingface.co/datasets/GEODE/GeoEDdA
5
https://zenodo.org/records/10530177
the Github project repository6 . They are available in the JSONL format provided by Prodigy
and the binary spaCy format (ready to use with the spaCy train pipeline). Table 2 shows the
distribution of each span class and the total number of paragraphs, tokens, and spans across
the datasets. As already observed in Figure 2, Table 2 shows that Spatial spans (NC-Spatial,
NP-Spatial, ENE-Spatial, Relation and Latlong) are over represented in comparison to Person
or Miscellaneous spans. This highlights the importance of geographical information even on
paragraphs not classified as Géographie, but we also note the near absence non-Spatial spans
within paragraphs classified as Géographie, e.g. NC-Spatial, NP-Spatial, ENE-Spatial, Relation
and Latlong cover more than 75% of all labeled spans. Witnessing this particularly strong spatial
signature of Géographie paragraphs - even for this small annotated dataset - points to the
importance of distinguishing between types of content at different levels: word-, paragraph-,
and article-level spatial information is not equally distributed across the Encyclopédie.
4. Training a Span Categorization Model
4.1. Custom spaCy spancat model
As a first experiment, a machine learning model was trained and evaluated using the proposed
Gold Standard dataset. Developed within the spaCy natural language processing library, this
model is tailored to our unique dataset and annotation schema requirements. It goes beyond
traditional NER tasks by specifically categorizing spans (including longer phrases and nested
entities), allowing for a more nuanced understanding of the relationships between entities
within the text. Subsequently, we employed the spaCy span categorizer pipeline, comprising
a suggester function that proposes candidate spans, potentially with overlaps, and a labeler
model that predicts zero or more labels for each candidate. The model can be downloaded and
installed directly from the HuggingFace Hub7 and executed using the spaCy Python library.
4.2. Evaluation
To assess the performance of the geo-semantic span categorization model, we evaluated it using
the test set described above. The overall model performance (on the Test set) demonstrates
strong precision, recall, and F-score values, attaining 93.98%, 79.82%, and 86.33%, respectively.
The model’s performance by span category is presented in Table 3. Notably, the model exhibits
high precision, recall, and F-score for crucial categories such as NC-Spatial, NP-Spatial,
and ENE-Spatial. The model also excels in accurately categorizing Head and Domain-mark
entities, achieving high precision and recall values. However, certain categories like Latlong
and ENE-Misc are not recognized at all. This can be explained by the low number of examples
(see Table 2) and the high diversity of forms (or values) of these two categories of spans. Only
255 ENE-Misc spans are present in the training set compared to 3,043 for ENE-Spatial for
instance. Furthermore, the Latlong spans consist of numerical values that might not be as
effectively represented as words in the embeddings [18].
6
https://github.com/GEODE-project/ner-spancat-edda
7
https://huggingface.co/GEODE/fr_spacy_custom_spancat_edda
Table 3
Model performance by span (Test set)
Tag Precision Recall F-score Tag Precision Recall F-score
NC-Spatial 96.50 93.24 94.84 NP-Person 92.47 90.00 91.22
NP-Spatial 92.74 95.95 94.32 ENE-Person 92.16 82.46 87.04
ENE-Spatial 91.67 95.51 93.55 NP-Misc 93.24 71.88 81.18
Relation 97.33 64.60 77.66 ENE-Misc 0.00 0.00 0.00
Latlong 0.00 0.00 0.00 Head 97.37 24.18 38.74
NC-Person 93.07 70.68 80.34 Domain-mark 99.19 91.73 95.31
5. Conclusion
In this paper, we presented the creation of the GeoEDdA gold standard dataset for geo-semantic
annotation of the Encyclopédie, including a review of the annotation schema and the active
learning approach to labeling spans in the text. GeoEDdA includes over 20,000 annotated spans
across 2,200 paragraphs. As an initial experiment, we trained and evaluated a spaCy span
categorization model, yielding insights into its strengths and areas for improvement. This
evaluation provides a comprehensive understanding of the model’s effectiveness in handling
specific entities. It also lays a foundation for future refinements and broader applications in
geo-semantic annotation across a wider variety of types of spatial and non-spatial information
than traditional NER.
As further work, we intend to expand the GeoEDdA dataset by annotating additional data,
aiming to enhance the model’s performance across all span categories, with particular focus on
improving geographical coordinates and miscellaneous spans categorization (e.g. new deposits
will be added to the Zenodo record). Next, we plan to compare span categorization algorithms,
further advancing our understanding and refining the capabilities of next generation, post-NER
geo-semantic annotation models. The final step is using inferred results that depend on this
dataset as training data for historical research. This will allow us to examine spatial language
across all of the Encyclopédie alongside additional information about, for example, knowledge
domains. Ultimately, this documentation of our thinking about spatial information reflects a
commitment to open historical research and promotes a more flexible approach to recognizing
spatial language.
Acknowledgments
The authors, all members of the GEODE team, are grateful to the ASLAN project (ANR-10-
LABX-0081) of the Université de Lyon, for its financial support within the French program
“Investments for the Future” operated by the National Research Agency (ANR).
References
[1] J. Zhou, W. Xu, End-to-end learning of semantic role labeling using recurrent neural
networks, in: Proceedings of the 53rd Annual Meeting of the ACL and the 7th Interna-
tional Joint Conference on NLP, Beijing, China, 2015, pp. 1127–1137. doi:10.3115/v1/
P15-1109.
[2] K. McDonough, L. Moncla, M. Van de Camp, Named entity recognition goes to old regime
france: geographic text analysis for early modern french corpora, International Journal of
Geographical Information Science 33 (2019) 2498–2522. doi:10.1080/13658816.2019.
1620235.
[3] M. C. Ardanuy, F. Nanni, K. Beelen, L. Hare, The past is a foreign place: Improving toponym
linking for historical newspapers, in: Proceedings of the Computational Humanities
Research Conference 2023, Paris, France, December 6-8, 2023, volume 3558 of CEUR
Workshop Proceedings, CEUR-WS.org, 2023, pp. 368–390.
[4] I. Ezeani, P. Rayson, I. Gregory, E. Haris, A. Cohn, J. Stell, T. Cole, J. Taylor, D. Boden-
hamer, N. Devadasan, E. Steiner, Z. Frank, J. Olson, Towards an extensible framework
for understanding spatial narratives, in: Proceedings of the 7th ACM SIGSPATIAL Inter-
national Workshop on Geospatial Humanities, ACM, Hamburg, Germany, 2023, p. 1–10.
doi:10.1145/3615887.3627761.
[5] C. B. Jones, R. S. Purves, Geographical information retrieval, International Journal of Geo-
graphical Information Science 22 (2008) 219–228. doi:10.1080/13658810701626343.
[6] P. Kordjamshidi, M. Van Otterlo, M.-F. Moens, Spatial role labeling: Towards extraction
of spatial relations from natural language, ACM Transactions on Speech and Language
Processing (TSLP) 8 (2011) 1–36.
[7] D. Gerstorfer, E. Gius, J. Jacke, Working on and with Categories for Text Analysis: Chal-
lenges and Findings from and for Digital Humanities Practices, Digital Humanities Quar-
terly 17 (2023).
[8] S. Schwandt, Opening the black box of interpretation: Digitla history practices as models
of knowledge, History and Theory 61 (2022) 77–85. doi:10.1111/hith.12281.
[9] M. Ehrmann, A. Hamdi, E. L. Pontes, M. Romanello, A. Doucet, Named entity recognition
and classification in historical documents: A survey, ACM Computing Surveys 56 (2023)
1–47. doi:10.1145/3604931.
[10] M. Gaio, L. Moncla, Extended named entity recognition using finite-state transducers:
An application to place names, in: Proceedings of the 9th international conference on
advanced geographic information systems, applications, and services (GEOProcessing
2017), Nice, France, 2017.
[11] Y. Wang, H. Tong, Z. Zhu, Y. Li, Nested named entity recognition: a survey, ACM
Transactions on Knowledge Discovery from Data (TKDD) 16 (2022) 1–29. doi:10.1145/
3522593.
[12] L. Moncla, M. Gaio, A multi-layer markup language for geospatial semantic annotations,
in: Proceedings of the 9th Workshop on Geographic Information Retrieval, Paris, France,
2015, pp. 1–10. doi:10.1145/2837689.2837700.
[13] A. Medad, M. Gaio, L. Moncla, S. Mustière, Y. Le Nir, Comparing supervised learning
algorithms for spatial nominal entity recognition, in: Proceedings of the 23rd AGILE
Conference on Geographic Information Science, Chania, Greece, 2020, pp. 1–18. doi:10.
5194/agile-giss-1-15-2020.
[14] R. Simon, E. Barker, L. Isaksen, P. De Soto Cañamares, Linked data annotation without the
pointy brackets: Introducing recogito 2, Journal of Map & Geography Libraries 13 (2017)
111–132. doi:10.1080/15420353.2017.1307303.
[15] A. Erdmann, D. J. Wrisley, B. Allen, C. Brown, S. Cohen-Bodénès, M. Elsner, Y. Feng,
B. Joseph, B. Joyeux-Prunel, M.-C. de Marneffe, Practical, efficient, and customizable active
learning for named entity recognition in the digital humanities, in: Proceedings of the 2019
Conference of the North American Chapter of the ACL: Human Language Technologies,
ACL, Minneapolis, Minnesota, 2019, pp. 2223–2234. doi:10.18653/v1/N19-1231.
[16] A. Brenon, L. Moncla, K. McDonough, Classifying encyclopedia articles: Comparing
machine and deep learning methods and exploring their predictions, Data & Knowledge
Engineering 142 (2022) 102098. doi:10.1016/j.datak.2022.102098.
[17] L. Moncla, D. Vigier, K. McDonough, GeoEDdA: A Gold Standard Dataset for Named Entity
Recognition and Span Categorization Annotations of Diderot & d’Alembert’s Encyclopédie,
2024. doi:10.5281/zenodo.10530178.
[18] D. Sundararaman, S. Si, V. Subramanian, G. Wang, D. Hazarika, L. Carin, Methods for
numeracy-preserving word embeddings, in: Proceedings of the 2020 Conference on
Empirical Methods in Natural Language Processing (EMNLP), ACL, 2020, pp. 4742–4753.
doi:10.18653/v1/2020.emnlp-main.384.