Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 29 of the 29 references it cites were found in a registry or an index; none of the 11 sources its claims cite is checked for support yet.

ATOM: AdapTive and OptiMized dynamic temporal knowledge graph construction using LLMs

Open access
Lairgi, Yassir, Moncla, Ludovic, Benabdeslem, Khalid, Cazabet, Rémy, Cléau, Pierre · 2026 · Findings of the Association for Computational Linguistics: EACL 2026

Abstract

In today's rapidly expanding data landscape, knowledge extraction from unstructured text is vital for real-time analytics, temporal inference, and dynamic memory frameworks.However, traditional static knowledge graph (KG) construction often overlooks the dynamic and time-sensitive nature of real-world data, limiting adaptability to continuous changes.Moreover, recent zero-or few-shot approaches that avoid domain-specific fine-tuning or reliance on prebuilt ontologies often suffer from instability across multiple runs, as well as incomplete coverage of key facts.To address these challenges, we introduce ATOM 1 (AdapTive and OptiMized), a few-shot and scalable approach that builds and continuously updates Temporal Knowledge Graphs (TKGs) from unstructured texts.ATOM splits input documents into minimal, self-contained "atomic" facts, improving extraction exhaustivity and stability.Then, it constructs atomic TKGs from these facts, employing a dual-time modeling that distinguishes between when information is observed and when it is valid.The resulting atomic TKGs are subsequently merged in parallel.Empirical evaluations demonstrate that ATOM achieves ∼ 18% higher exhaustivity, ∼ 33% better stability, and over 90% latency reduction compared to baseline methods, demonstrating a strong scalability potential for dynamic TKG construction.

Overview

CoChat Summary

TL;DR

ATOM is a few-shot, parallel approach that decomposes unstructured text into atomic facts, extracts temporal 5-tuples into atomic TKGs, and merges them with dual-time modeling to build and update dynamic temporal knowledge graphs. On a COVID-19 NYT subset, it reports higher exhaustivity and stability and large latency reductions over baselines.

Key findings

  • On 2020-COVID-NYT, atomic-fact extraction improved factual recall \(R_f\) by ∼31% (0.405 ± 0.150 for lead paragraphs vs. 0.720 ± 0.143 for atomic facts) and factual-temporal recall \(R_{f,t}\) by ∼18% (0.176 ± 0.123 vs. 0.354 ± 0.165); \(F_{1,f}\) rose from 0.504 to 0.638 and \(F_{1,f,t}\) from 0.220 to 0.313. Precision was lower for atomic facts: \(P_f\) 0.667 → 0.572 and \(P_{f,t}\) 0.292 → 0.281, described as a marginal ∼1% decrease in factual-temporal precision.
  • Stability across reruns was higher for atomic facts (Run 2: 0.552 ± 0.124; Run 3: 0.534 ± 0.12) than for lead paragraphs (Run 2: 0.212 ± 0.181; Run 3: 0.218 ± 0.195), a gain of ∼33%.
  • ATOM reduced latency by 93.8% compared with Graphiti and 95.3% compared with iText2KG; its Module-3 merge accounted for only 13% of total latency.
  • On DTKG construction, ATOM achieved entity-resolution \(F_1\) 0.994 and relation-resolution \(F_1\) 1, versus Graphiti (0.959, 0.902) and iText2KG (0.977, 0.989); temporal resolution was compared qualitatively, with ATOM extending validity periods and avoiding Graphiti’s treatment of observation time as \(t_{start}\).

Methods

  • ATOM decomposes documents into minimal, self-contained atomic facts, extracts 5-tuples in parallel to form atomic TKGs, then merges them via a parallel binary merge algorithm. Entity resolution uses exact name/label matches then cosine similarity with \(\theta_E = 0.8\); relation resolution uses \(\theta_R = 0.7\); temporal resolution merges observation and validity time sets. Dual-time modeling separates observation time from validity start/end, and merging avoids LLM calls.
  • Evaluation used the 2020-COVID-NYT subset of the NYT News dynamic and temporal dataset: 1,076 articles, 274 grouped by publication date, average 206 ± 156 tokens per group, date range 2020-01-09 to 2020-12-30; 4,223 atomic facts (2,037 with validity time, 2,186 without), 7,210 5-tuples, 4,223 atomic TKGs, and ∼2 5-tuples per atomic TKG. The subset was enriched with human-verified atomic facts and 5-tuples; publication dates were used as observation dates.
  • Models/settings: claude-sonnet-4-2025-01-31 for atomic fact decomposition, gpt-4.1-2025-04-14 for 5-tuple extraction, temperature 0, text-embedding-large-3 embeddings; baselines were Graphiti and iText2KG. Metrics included factual and factual-temporal recall, precision, and \(F_1\); stability via Jaccard similarity across runs; wall-clock latency; entity/relation resolution precision, recall, and \(F_1\); and qualitative temporal resolution.

Limitations

  • Stated: atomic fact decomposition can propagate errors—the LLM may generate inferred facts not present in the source, increasing hallucination, and may fail to assign temporal information, lowering factual-temporal precision. Distance-metric merging can occasionally merge semantically distinct but highly similar named entities (e.g., gpt-5:model and gpt-3.5:model). ATOM also costs more than direct extraction due to the added decomposition step and increased output tokens.
  • Evident from the text: evaluation is limited to one news subset (2020-COVID-NYT) and uses lead paragraphs rather than full articles; temporal resolution is evaluated qualitatively, with quantitative evaluation listed as future work; stability improves but remains imperfect (atomic-fact Jaccard ∼0.55); and the approach relies on proprietary LLM APIs and thresholds estimated from prior work or generated pairs.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.18653/v1/2026.findings-eacl.49

Cites

29 of 29 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 ATOM's architecture, running in parallel, ensuring scalability, speed, and continuous updates. Unstructured texts observed at time t are denoted by D t , the i -th temporal atomic fact observed at time t is denoted by f t,i , the i -th atomic TKG observed at time t is denoted by G t i , the TKG snapshot observed at time t is denoted by G t s , and the updated TKG at time t is denoted by G t .
Figure 1 ATOM's architecture, running in parallel, ensuring scalability, speed, and continuous updates. Unstructured texts observed at time t are denoted by D t , the i -th temporal atomic fact observed at time t is denoted by f t,i , the i -th atomic TKG observed at time t is denoted by G t i , the TKG snapshot observed at time t is denoted by G t s , and the updated TKG at time t is denoted by G t .CC-BY · From the PDF
Figure 2 Example overview of ATOM 's pipeline. It begins with atomic fact decomposition, followed by the extraction of atomic TKGs from these facts, which are then merged in parallel. When an incoming update arrives, ATOM handles the temporal resolution by transforming the end action into the affirmative part while modifying only the t end , then merges the resulting atomic TKG with the existing DTKG.
Figure 2 Example overview of ATOM 's pipeline. It begins with atomic fact decomposition, followed by the extraction of atomic TKGs from these facts, which are then merged in parallel. When an incoming update arrives, ATOM handles the temporal resolution by transforming the end action into the affirmative part while modifying only the t end , then merges the resulting atomic TKG with the existing DTKG.CC-BY · From the PDF
Figure 3 Exhaustivity vs. token count as context for (a) the atomic fact decomposition (b) the 5-tuple extraction. The x-axis displays the context size (lead paragraphs) only; total input size includes additional system prompt tokens ( ∼ 1.4k for (a) and ∼ 1.7k for (b)).
Figure 3 Exhaustivity vs. token count as context for (a) the atomic fact decomposition (b) the 5-tuple extraction. The x-axis displays the context size (lead paragraphs) only; total input size includes additional system prompt tokens ( ∼ 1.4k for (a) and ∼ 1.7k for (b)).CC-BY · From the PDF
Figure 4 Latency comparison of the baseline methods as a function of the number of atomic facts as input.
Figure 4 Latency comparison of the baseline methods as a function of the number of atomic facts as input.CC-BY · From the PDF
Figure F.1: Two DTKGs constructed using ATOM and Graphiti from 09-01-2020 (in UNIX, 1578524400) to 23-012020 (in UNIX, 1579734000) from 2020-COVID-NYT dataset. Left ( ATOM ): Preserves observation times ( t obs ) separately from validity periods, with timestamps encoded in UNIX format to eliminate overhead associated with string parsing operations and timezone conversion calculations. Right (Graphiti): Treats observation time as validity start time. valid _ at corresponds to t start in Graphiti's time modeling. The highlighted fact 'The mysterious respiratory virus spread to at least 10 other countries' is observed on 23-01-2020, but this does not guarantee the spread occurred at that time. ATOM 's dual-time modeling prevents such temporal misattribution.
Figure F.1: Two DTKGs constructed using ATOM and Graphiti from 09-01-2020 (in UNIX, 1578524400) to 23-012020 (in UNIX, 1579734000) from 2020-COVID-NYT dataset. Left ( ATOM ): Preserves observation times ( t obs ) separately from validity periods, with timestamps encoded in UNIX format to eliminate overhead associated with string parsing operations and timezone conversion calculations. Right (Graphiti): Treats observation time as validity start time. valid _ at corresponds to t start in Graphiti's time modeling. The highlighted fact 'The mysterious respiratory virus spread to at least 10 other countries' is observed on 23-01-2020, but this does not guarantee the spread occurred at that time. ATOM 's dual-time modeling prevents such temporal misattribution.CC-BY · From the PDF
Figure F.2: Temporal resolution comparison between ATOM and Graphiti. Two atomic facts observed on January 28, 2020, report death counts from January 24 (26 deaths) and January 27 (at least 80 deaths). Left ( ATOM ) : performs temporal resolution by detecting similar relations and extending their validity period history ( t end in the figure). Right (Graphiti) : creates separate relations for each atomic fact, resulting in duplication. Moreover, Graphiti misinterprets 'By January 24, 2020' and 'By January 27, 2020' as validity start times rather than validity end times, leading to temporal misattribution.
Figure F.2: Temporal resolution comparison between ATOM and Graphiti. Two atomic facts observed on January 28, 2020, report death counts from January 24 (26 deaths) and January 27 (at least 80 deaths). Left ( ATOM ) : performs temporal resolution by detecting similar relations and extending their validity period history ( t end in the figure). Right (Graphiti) : creates separate relations for each atomic fact, resulting in duplication. Moreover, Graphiti misinterprets 'By January 24, 2020' and 'By January 27, 2020' as validity start times rather than validity end times, leading to temporal misattribution.CC-BY · From the PDF
Figure F.3: Temporal resolution comparison between ATOM and Graphiti. Two atomic facts observed on different dates (April 16 and April 19, 2020) describe protest activities during two time periods ( 'the week of April 13' and 'the week of April 19' ). Left ( ATOM ) : merges similar relations and extend their validity periods ( t start and t end in the figure). Right (Graphiti) : maintains separate relations for each atomic fact. Moreover, Graphiti fails to translate 'In the week of April 13, 2020' and 'In the week of April 19, 2020' into proper validity periods as ATOM does.
Figure F.3: Temporal resolution comparison between ATOM and Graphiti. Two atomic facts observed on different dates (April 16 and April 19, 2020) describe protest activities during two time periods ( 'the week of April 13' and 'the week of April 19' ). Left ( ATOM ) : merges similar relations and extend their validity periods ( t start and t end in the figure). Right (Graphiti) : maintains separate relations for each atomic fact. Moreover, Graphiti fails to translate 'In the week of April 13, 2020' and 'In the week of April 19, 2020' into proper validity periods as ATOM does.CC-BY · From the PDF

Verification

Claims the paper makes

0 Valid · 4 Not valid · 11 Not checked · 53 Own results and interpretations

21 more claims

Create a free account to see all 68 claims and to verify the 10 references still open.

See all claims — free account
  • Probably not a substantial claim. 21
  • It may state two separate claims as one. 29

Create a free account to see these claims.

Citations

Cites

Show all 29 references

Cited by

See all 7

Metadata

Authors
Lairgi, Y., Moncla, L., Benabdeslem, K., Cazabet, R., & Cléau, P.
Affiliations
Lyon 1 Université (FR)
Venue
Findings of the Association for Computational Linguistics: EACL 2026
Published
2026
Pages
950-966
OpenAlex
W7140084193
Cited by
7
References
29
Citations by year
2026: 7
FWCI
6.53
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • Findings of the Association for Computational Linguistics: EACL (publishedVersion, cc-by)
  • HAL (submittedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Advanced Graph Neural Networks
  • Physical Sciences › Computer Science › Artificial Intelligence › Semantic Web and Ontologies
Keywords
temporal knowledge graphs, knowledge graph construction, few-shot learning, large language models, atomic facts, knowledge graph updating

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-18, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-18 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: ATOM: AdapTive and OptiMized dynamic temporal knowledge graph construction using LLMs. https://app.cochat.ai/papers/10.18653/v1/2026.findings-eacl.49
The paper
Lairgi, Y., Moncla, L., Benabdeslem, K., Cazabet, R., & Cléau, P. (2026). ATOM: AdapTive and OptiMized dynamic temporal knowledge graph construction using LLMs. Findings of the Association for Computational Linguistics: EACL 2026. https://doi.org/10.18653/v1/2026.findings-eacl.49

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.