Agreement Between Reconstructed and Native GDELT GKG Records: A Field-Level Baseline

Nikita A. Klimin

cs.CLOct 26, 2026

Agreement Between Reconstructed and Native GDELT GKG Records: A Field-Level Baseline

Authors: Nikita A. Klimin

Organizations: HSE University, Moscow, Russia

Abstract

The Global Knowledge Graph (GKG) of the Global Database of Events, Language, and Tone (GDELT) characterizes news articles using named entities, topics, quantitative and temporal features, quotations, tone, and dictionary-based content measures. Its codebook defines the record format, but the software that produces the records is not public, so researchers cannot generate compatible records for articles they collect themselves. We built an open-source reconstruction, gdelt-regkg, and measured how closely it agrees with GDELT's archived (native) records for 193,550 English articles from 8,729 publishers (2020-2025) and 2,500 articles in five other languages. Agreement depends strongly on the field. English tone, the percentage of positive minus negative words, has a mean absolute error (MAE) of 1.79 points and a concordance correlation coefficient of 0.75. Identity agreement (F1) is 0.65 for persons, 0.42 for organizations, and 0.65 for locations. It is lower for date components (0.42), complete counts with their objects and locations (0.07), and quotations (0.07), while GCAM category presence reaches micro F1 0.84 within the implemented dictionaries. In subsets where retrieved and native titles and lengths match, tone error falls to 1.47 and organization F1 rises to 0.48, yet most disagreement remains. A larger gazetteer raises coverage of native places from 89.5% to 97.1% with little change in location F1. Tone computed on English translations of non-English articles is less accurate (MAE 2.10-2.64) than for English articles. The results are conditional on article pages that could still be retrieved, and the correspondence subsets do not isolate the effect of input differences. Reconstructed records should therefore be assessed field by field and not treated as completely interchangeable with native ones.

Figures & tables

Field	Native GDELT	Reconstructed
Persons	boris johnson	boris johnson
Organizations	european union	european union
Location	U.K.; Brussels	U.K.; Brussels; U.K.
Themes	LEADER ; TAX_FNCACT ; …	LEADER ; TAX_FNCACT ; …
Tone	$-1.471$	$-1.538$
Token count	124	130

TABLE I: Example of Native and Reconstructed Fields

Year	Attempted URLs	Retained	Retention $\%$
2020	105,772	29,068	27.48
2021	92,123	25,144	27.29
2022	83,638	25,572	30.57
2023	107,833	36,957	34.27
2024	99,995	39,351	39.35
2025	87,860	37,458	42.63

TABLE II: Acquisition and Retention

Field	Reconstruction	Matched unit within an article
Persons, organizations, all names	Named-entity spans and name rules	Sets of normalized name identities
Themes	Phrase, wildcard, and context rules	Sets of exact theme codes
Locations	Place recognition and gazetteer resolution	Type, country, first-level administrative code (ADM1), and feature ID
Counts	Quantity–object–location templates	Multisets: label + quantity ; object additionally requires the attached object text; full tuple also requires location type, name, country, ADM1, latitude, longitude, and feature ID
Dates	Calendar-validated patterns	Resolution and available month, day, and year components
Amounts	Numeric/scale normalization and object attachment	Sets of numeric values ; value + object additionally requires the same normalized attached object text

TABLE III: Field Reconstruction and Comparison Units

Field / comparison	Active articles	Micro F1	Article-macro F1	Category-macro F1	Publisher-macro F1
Persons / normalized identities	172,393	0.646 [0.637, 0.656]	0.592 [0.582, 0.603]	—	0.589
Organizations / normalized identities	158,843	0.417 [0.411, 0.422]	0.325 [0.319, 0.331]	—	0.357
Locations / identities	167,617	0.654 [0.646, 0.662]	0.543 [0.532, 0.554]	—	0.586
Themes / all labels	192,585	0.617 [0.611, 0.622]	0.526 [0.519, 0.534]	0.028	0.529
Counts / label + quantity	28,143	0.285 [0.276, 0.295]	0.153 [0.147, 0.159]	—	0.179
Counts / object	28,143	0.166 [0.158, 0.176]	0.088 [0.085, 0.092]	—	0.104

TABLE IV: English Field Agreement

Fig. 1: Native and reconstructed tone for all 193,550 English articles. Color shows the number of articles per hexagon. The dashed line is identity.

Fig. 2: Tone differences against the pair mean for all 193,550 English articles. Color shows articles per hexagon. The solid line is mean bias (0.22); dashed lines are bias

\pm 1.96

sample SD of the differences (

-4.67

, 5.10).

Subset	Articles	Tone MAE	Tone CCC	Person F1	Org. F1	Location F1
Full corpus	193,550	1.79 [1.76, 1.82]	0.749 [0.736, 0.759]	0.646	0.417	0.654
Title-available	193,100	1.79 [1.76, 1.82]	0.749 [0.737, 0.759]	0.647	0.417	0.654
Title-matched	117,217	1.77 [1.73, 1.81]	0.759 [0.742, 0.772]	0.664	0.435	0.671
Title-and-length ( $\pm 5\%$ ) subset	23,121	1.48 [1.46, 1.51]	0.827 [0.822, 0.833]	0.736	0.473	0.720
Title-and-length ( $\pm 10\%$ ) subset	54,485	1.47 [1.46, 1.49]	0.830 [0.826, 0.834]	0.748	0.483	0.727
Title-and-length ( $\pm 20\%$ ) subset	81,668	1.51 [1.49, 1.52]	0.830 [0.827, 0.833]	0.742	0.475	0.723

TABLE V: Agreement by Input Correspondence and Deduplication

Fig. 3: Monthly agreement for the full corpus and the Title-and-length (

\pm 10

%) subset: (a) tone MAE; (b) GCAM implemented presence micro F1; (c) count WAPE for four dictionaries, full corpus; (d) the same, Title-and-length (

\pm 10

%) subset. Points are monthly estimates without uncertainty. Shading marks June 2021–September 2022; the dotted line marks June 2025.

Fig. 4: Date-component micro F1 by year (rows) and calendar month (columns): (a) all date components; (b) day–month dates without a year. Each month is sampled on two days.

Fig. 5: Pooled GCAM lexical counts by category. The dashed line is identity; logarithmic axes add one to retain zeros.

	Coverage (%)	Location F1	Count full-tuple F1
BF	89.5	0.654 [0.646, 0.661]	0.066 [0.062, 0.072]
BT	89.5	0.725 [0.719, 0.729]	0.068 [0.063, 0.074]
EF	97.1	0.654 [0.646, 0.662]	0.070 [0.066, 0.076]
ET	97.1	0.727 [0.722, 0.731]	0.073 [0.067, 0.078]

TABLE VI: Geographic Agreement

	Paired $n$	MAE	CCC
EN*	193,550	1.79 [1.76, 1.82]	0.749 [0.736, 0.759]
FR	500	2.40 [2.22, 2.58]	0.515 [0.443, 0.581]
ES	500	2.10 [1.95, 2.27]	0.630 [0.569, 0.681]
DE	500	2.61 [2.39, 2.84]	0.446 [0.332, 0.543]
RU	500	2.56 [2.34, 2.80]	0.593 [0.524, 0.657]
ZH	499	2.64 [2.37, 2.94]	0.513 [0.402, 0.620]

TABLE VII: Translated Tone

	Presence F1		Happiness
	Micro	Macro	MAE	CCC
FR	0.613 [0.585, 0.639]	0.509	0.039 [0.033, 0.044]	0.826 [0.765, 0.878]
ES	0.771 [0.736, 0.803]	0.756	0.040 [0.035, 0.045]	0.902 [0.869, 0.926]
DE	0.679 [0.646, 0.712]	0.527	0.083 [0.076, 0.093]	0.464 [0.389, 0.542]
RU a	0.860 [0.833, 0.886]	0.860	0.060 [0.054, 0.067]	0.802 [0.753, 0.843]
ZH b	1.000	1.000	0.098 [0.087, 0.112]	0.671 [0.578, 0.742]

TABLE VIII: Original-Language GCAM

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Dictionary / inventory	Observed categories	Native assignments	Micro F1	Category-macro F1	WAPE (%)
GCAM (all observed)	2,442	178,070,983	0.824	0.651	—
GCAM (implemented)	2,359	171,168,198	0.840	0.674	—
Forest Values [ 4 ]	4	148,939	0.857	0.858	38.6
General Inquirer [ 45 ]	228	31,872,896	0.927	0.893	41.4
Lexicoder Sentiment [ 56 ]	4	412,145	0.971	0.922	42.7
Lexicoder Topics [ 1 ]	28	857,526	0.638	0.626	95.8

Appendix

TABLE 01: English GCAM Agreement

ID	Resource	Cat.	Counts	Scores
1	Forest Values [ 4 ]	4	4	0
2	General Inquirer [ 45 ]	228	228	0
3	Lexicoder Sentiment [ 56 ]	4	4	0
4	Lexicoder Topics [ 1 ]	28	28	0
5	LIWC 2007 [ 38 ]	62	62	0
6	Loughran–McDonald [ 31 ]	6	6	0

Appendix

TABLE A1: GCAM Dictionaries and Implemented Coverage

KG Construction

Knowledge graph construction turns text, tables, images, and databases into structured entities and relationships that machines can query and reason over.

See all 195 papers

Geolocating News about Extreme Climate Events: A Comparative Analysis of Off-the-Shelf Tools for Toponym Identification in German
May 5, 2026Brielen Madureira, Mariana Madruga de Brito, Andreas NieklerNamed Entity Recognition
LeipzigLab - Climate Discourse, Leipzig University, Leipzig, Germany · Helmholtz Centre for Environmental Research - UFZ, Leipzig, Germany · Leipzig University, Leipzig, Germany +1
RefinedGAW: A Refined German Academic Web Dataset
Oct 25, 2026Maximilian Idahl, Oleh Astappiev, Benedikt Droste +5
ellamind · L3S Research Center, Leibniz University Hannover · Berlin School of Library and Information Science, Humboldt-Universität zu Berlin
EDDA-Coordinata: An Annotated Dataset of Historical Geographic Coordinates
Feb 27, 2026Ludovic Moncla, Pierre Nugues, Thierry Joliveau +1Historical Document Processing Document Information Extraction
INSA Lyon, CNRS, Lyon 1 Université, LIRIS, UMR5205, 69621 Villeurbanne, France · Lund University, Lund Sweden · Université de Saint-Etienne, CNRS, EVS UMR5600, Saint-Etienne, France +1

May 5, 2026J/K move · Enter open · S save

Geolocating News about Extreme Climate Events: A Comparative Analysis of Off-the-Shelf Tools for Toponym Identification in German

Brielen Madureira, Mariana Madruga de Brito, Andreas Niekler

LeipzigLab - Climate Discourse, Leipzig University, Leipzig, Germany · Helmholtz Centre for Environmental Research - UFZ, Leipzig, Germany · Leipzig University, Leipzig, Germany+1

Named Entity Recognition

Open paper AskPDF arXiv

Determining the geolocation of extreme climate events and disasters in texts is a common problem in climate impact and adaptation research. Named-entity recognition (NER) tools are typically used to identify a pool of toponyms that serve as candidate event locations. In this study, we conduct a comparative analysis of three off-the-shelf NER tools, namely Flair, Spacy and Stanza. We describe and quantify differences between their outputs for German news articles and evaluate them extrinsically based on three methods to determine the country where events took place. We show how their contrasts are propagated into downstream tasks and can yield distinct decisions about a document's geographical focus, which, in turn, can impact conclusions about countries' prominence in German media.

Abstract

KG Construction

Explore similar work

Geolocating News about Extreme Climate Events: A Comparative Analysis of Off-the-Shelf Tools for Toponym Identification in German

RefinedGAW: A Refined German Academic Web Dataset

EDDA-Coordinata: An Annotated Dataset of Historical Geographic Coordinates