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

1 Introduction

• Static KGs provide snapshots without temporal dynamics, whereas TKGs associate timestamps or intervals with facts.The introduction notes TKG use in modeling dynamic information and agent memory.Source: §1 · p. 1
• Sequential approaches can become prohibitive bottlenecks when continuously arriving information must be incorporated in real time.Source: §1 · p. 1
• Few-shot KG methods can omit key relationships, vary across runs, overlook time, and struggle with dynamic data and scalability.Traditional methods may also depend on predefined ontologies and supervised learning requiring extensive human annotation.Source: §1 · pp. 1–2
• ATOM decomposes text into atomic facts, then uses parallel 5-tuple extraction and parallel merging to construct dynamic TKGs.The paper presents this as a few-shot, scalable approach intended to improve stability and exhaustivity.Source: §1 · p. 2

2 Related work

• iText2KG constructs KGs incrementally using embeddings and cosine similarity, but is described as non-exhaustive, unstable, non-temporal, and difficult to scale.Source: §2 · p. 2
• Graphiti uses LLM calls for entity, relation, and time resolution, creating cost and scalability challenges while leaving exhaustivity and stability unaddressed.Source: §2 · p. 2
• AriGraph combines semantic and episodic memory, but its entity resolution can cause semantic drift and scalability problems as unstructured data grows.Source: §2 · p. 2
• ATOM targets incomplete coverage on long texts, run-to-run instability, and lack of scalable dynamic updates with atomic decomposition and parallel merging.Source: §2 · p. 2

3.1 Problem statement

• A dynamic temporal KG tracks when information is observed separately from the validity period expressed by a fact.A temporal atomic fact is a short, self-contained snippet conveying one piece of information and may omit an explicit validity period.Source: §3.1 · pp. 2–3

3.2 ATOM’s Framework

• The framework seeks exhaustive snapshots that capture semantically present 5-tuples and stable results across repeated runs on the same text.Source: §3.2 · p. 4

3.2.1 Module-1: Atomic fact decomposition

• Atomic decomposition is intended to reduce information loss and output variance by giving the LLM clear, single-fact contexts.The approach uses an experimentally selected chunk size; processing hundreds or thousands of facts motivates parallel extraction and merging.Source: §3.2.1 · p. 4

3.2.2 Module-2: Atomic TKGs construction

• The LLM extracts 5-tuples from each atomic fact in parallel and preprocesses end-actions as affirmative facts by changing only tend.Observation time is supplied as context for interpreting relative temporal expressions.Source: §3.2.2 · p. 4

3.2.3 Module-3: Parallel atomic merge of TKGs and DTKG update

• Pairwise merging resolves entities and relations, then combines observation and validity times for matching relations without LLM calls.Entity and relation matches use exact matching or cosine-similarity thresholds; temporal resolution aligns end-actions with beginning facts.Source: §3.2.3 · pp. 4–5
• ATOM merges pairs of atomic TKGs in parallel until one snapshot remains, then merges that snapshot with the previous DTKG.An odd TKG carries forward to the next iteration; the parallel strategy scales with available threads.Source: Figure 2 · p. 5, §3.2.3 · p. 5

4.1 Metrics

• Evaluation tracks factual and factual-temporal 5-tuple quality, run-to-run Jaccard stability, and total wall-clock latency for DTKG construction.Factual-temporal true positives require both the fact and its validity boundaries to be correct.Source: §4.1 · pp. 5–6

4.2 Datasets and baseline methods

• Experiments use 1,076 COVID-19 NYT articles from 2020, with publication dates as observation times; Graphiti is the primary comparator because it also handles temporal data.The subset comes from a dataset spanning two million lead paragraphs from 2000 onward.Source: §4.2 · p. 6

4.3 Exhaustivity deterioration in longer contexts

• As context grows, exhaustivity falls across models; Claude remains strongest on atomic facts but degrades on 5-tuples, while <400-token chunks keep decomposition above 0.8.Figure 3 plots exhaustivity against context size in lead paragraphs; total input also includes system-prompt tokens.Source: Figure 3 · p. 7, §4.3 · p. 7

4.4 TKG stability

• ∼33% gain in stability: atomic-fact 5-tuple sets had higher Jaccard overlap across reruns than lead-paragraph sets.Table 1 evaluates reruns against Run 1 for both input types.Source: Table 1 · p. 7, §4.4 · p. 7

4.5 The exhaustivity and quality of the 5-tuples

• ∼31% factual recall Rf and ∼18% factual-temporal recall Rf,t gains came with ∼9% lower factual precision Pf and ∼1% lower factual-temporal precision Pf,t.The authors attribute the factual-precision decrease to plausible inferred facts absent from the gold standard, and the temporal-precision decrease to missing temporal information during decomposition.Source: §4.5 · pp. 7–8, Table 2 · p. 8

4.6 ATOM’s time complexity

• ATOM reduces latency by 93.8% compared to Graphiti and 95.3% compared to iText2KG through parallel LLM calls and merging.Figure 4 compares latency as the number of input atomic facts grows.Source: Figure 4 · p. 8, §4.6 · p. 8

4.7 Consistency of ATOM’s DTKG construction

• ATOM matches iText2KG on entity/relation resolution and improves over Graphiti; unlike Graphiti, it extends validity histories and separates observation time from validity periods.This dual-time treatment can avoid assigning a validity start date when the source provides only an observation date.Source: §4.7 · pp. 8–9, Table 3 · p. 9

• ATOM is a few-shot approach for constructing and dynamically updating TKGs from unstructured text.Source: §5 · p. 9
• Atomic fact decomposition addresses exhaustivity and stability challenges often observed in LLM-based TKG construction.Source: §5 · p. 9
• ATOM’s parallel architecture accelerates TKG construction and supports scalability for larger unstructured texts; future work includes evaluating temporal resolution and fine-tuning decomposition.Source: §5 · p. 9

• Atomic fact decomposition may introduce inferred facts absent from the source or misassign temporal information, reducing factual-temporal precision.Source: §6 · p. 9
• Distance-metric merging can occasionally combine semantically distinct named entities with high similarity.The passage gives gpt-5:model and gpt-3.5:model as an example.Source: §6 · p. 9
• ATOM costs more than direct extraction because decomposition adds a step and maintaining high exhaustivity requires more output tokens.Source: §6 · p. 9

• AI assistants were used exclusively for grammatical corrections and text polishing.Source: §7 · p. 9
• The authors state that the scientific claims, experimental designs, and original ideas are theirs.Source: §7 · p. 9

A ATOM’s algorithms

• The appendix presents one algorithm for merging pairs of TKGs and another for parallelizing the merge of TKG lists.Source: A ATOM’s algorithms · p. 11

B Example of the atomic fact decomposition

• The example decomposes an observed news passage about Real Madrid’s Champions League victory into separate atomic facts.The facts separately describe the win, its score, and fans’ celebration.Source: B Example of the atomic fact… · p. 11

C Estimating the merging thresholds

• λ = 0.8, β = 0.2, and θE = 0.8 were selected using cosine similarity over 1,200 typology-aware similar-entity pairs; θR = 0.7 was retained.The hybrid measure combines entity-name and entity-label embeddings; the previous θR estimate was based on 1,500 similar entity and relation-name pairs.Source: C Estimating the merging thresholds · p. 11

D Observation time modeling

• For retrospective text streams, observation time should be the original publication time, not the later processing or ingestion time.This distinction preserves event ordering and supports relative-time inference.Source: D Observation time modeling · p. 11
• For continuously monitored sources, ingestion time can represent observation time because it is the earliest feasible moment the system could know the information.Observation-time granularity depends on the application: COVID-19 news was simulated with daily snapshots, while social media may require per-post snapshots.Source: D Observation time modeling · p. 11

E Annotation guidelines for 2020-COVID-NYT

• The 2020-COVID-NYT ground truth used a two-step semi-manual process with Claude Sonnet 4 and human annotators supervising its annotations.The model was selected after outperforming other available SOTA LLMs on a small subset of lead paragraphs.Source: E Annotation guidelines for… · p. 11

E.1 Phase 1: atomic fact decomposition

• Phase 1 decomposed complex paragraphs into atomic, self-contained, temporally grounded statements, which annotators checked for exhaustivity, hallucinations, and validity-time correctness.Compound sentences were split so each fact contained one piece of information, with redundancies removed.Source: §E.1 · pp. 11–12
• Binary merging maps matching entities, updates relation endpoints and names using similarity thresholds, and merges the times of similar relations.The procedure unions start times, end times, and observation times for similar relations.Source: §E.1 · p. 12
• Facts were made understandable in isolation by replacing pronouns with entity names and adding necessary modifiers.Relative time references were converted to absolute dates based on observation time; role or action endings were captured with their specific times.Source: §E.1 · p. 12

E.2 Phase 2: 5-tuple extraction

• Phase 2 extracts temporal 5-tuples (es, r, eo, tstart, tend) from atomic facts, with annotators checking exhaustivity, factual accuracy, and temporal validity.Validity information is assigned to tstart and tend using the absolute dates resolved in Phase 1.Source: §E.2 · pp. 12–13
• Entities are distinct concepts with dates excluded from entity mentions, and relations use canonical present-tense names regardless of event time.Dates are reserved for temporal arguments; examples of present-tense relation names include is_CEO and works_at.Source: §E.2 · pp. 12–13
• Parallel merging repeatedly pairs temporal knowledge graphs, applies binary merging to each pair in parallel, and carries an unpaired graph forward when needed.The process continues until one merged graph remains.Source: §E.2 · p. 13
• Affirmative actions populate tstart, while end actions populate tend and retain the relation name in affirmative form.For example, an end action such as “is no longer the CEO” populates tend while the relation remains is_CEO.Source: §E.2 · p. 13
• ATOM preserves observation time separately from validity time, avoiding the assumption that a fact was valid when it was observed.In the example, a virus-spread fact observed on January 23 does not establish that the spread occurred on that date.Source: Figure F.1 · p. 15
• For two January death-count facts, ATOM extends similar relations’ validity history, while Graphiti creates separate relations and misreads “By” dates as validity starts.The facts observed January 28 refer to January 24 (26 deaths) and January 27 (at least 80 deaths).Source: Figure F.2 · p. 16
• For protest facts observed on different dates, ATOM merges similar relations and validity periods, whereas Graphiti keeps separate relations and fails to translate the week references into proper validity periods.The facts refer to activities during the weeks of April 13 and April 19, 2020.Source: Figure F.3 · p. 17