Systems that read long documents (court dossiers, archives, investigative reports) usually cut them into overlapping windows, extract events inside each window, and merge the results. Overlap keeps a name or a phrase from being split in half. It does nothing for two events that sit thirty pages apart: they never share a window, so no model ever judges whether one came before the other. The same failure appears one level up, in schema induction, the task of discovering from raw text what kinds of events exist and how they typically connect. Current auditable schema pipelines explicitly stop at within-event roles and leave links between events, such as "arrest typically precedes transfer by days," unaddressed.
We propose to give every extracted event a position on a single document-wide timeline, an estimated time interval with an explicit width for vagueness, and to use that position instead of token position as the coordinate for everything that happens after extraction. Events from different windows are merged, ordered, and related through a small model whose positional encoding is this timeline coordinate. Across a corpus, the resulting timelines yield candidate inter-event schema edges that carry typical time gaps. Human reviewers are asked to check only the edges where the timeline evidence and the language model disagree. The claim is that the global time coordinate, not a longer context window or more overlap, is what recovers long-range temporal relations and fine-grained sub-events. The claim is also that it concentrates scarce human attention on the graph-construction step, which earlier interactive schema work found to be the most expensive.
Core Claim
Long-document event extraction has two separate boundary problems. The first is span truncation, which overlapping windows solve well: overlapping fixed-length segmentation beat sentence-based segmentation 0.9267 to 0.7880 F1 on document alignment, as reported in the work described in InstructCoder. The second is relational fragmentation. Pairs of events that no window contains together can only be related by transitivity or by guesswork. MAVEN-ERE shows how much depends on this: 88.8% of its temporal relations are inferable through transitivity over the full document timeline (MAVEN-ERE). That makes a document-level timeline the natural object to reconstruct. Our first hypothesis is that an aggregation model that places events by estimated time, rather than by within-window token index, recovers cross-window temporal relations substantially better than overlap, compression, or feeding the whole document to a long-context model. Its advantage should grow with the token distance between the two events.
Our second hypothesis concerns schemas. Once events carry timeline coordinates, corpus-level statistics over open-typed event clusters give candidate edges of the form (type A, precedes, type B, gap distribution). These are exactly the inter-event links that SCOPE and SCION name as out of scope. Our third hypothesis is that conflict-triggered human review, where a person looks only at edges on which the timeline statistics and an LLM's judgment disagree, reaches the edge accuracy of full review at a fraction of the reviewer time.
Research Program
The pipeline has four stages. The novelty is in the third and fourth.
Stage 1: extraction. Documents are split into overlapping windows, following the 5,000-character block with 500-character overlap used by Díaz et al.. An open-vocabulary extractor proposes event mentions with evidence spans. We use the constrained editor style of SCION, in which the LLM edits mined candidates and must point to supporting text, because unconstrained LLM extraction is known to over-generate. TextEE reports zero-shot LLM event detection in the 30s F1 against the 70s for supervised systems.
Stage 2: timeline anchoring. For every mention, the model estimates an interval on a global axis. It uses explicit time expressions when they exist ("14 October 1973") and relative markers otherwise ("three days later," resolved against the nearest anchored event). Each estimate carries a width reflecting its granularity, so "late summer 1994" is a wide interval and "14:32 UTC" is a narrow one. This directly targets the mixed-granularity problem that periodic time encodings handle poorly.
Stage 3: cross-window aggregation. A small Transformer takes as input all event nodes from all windows of a document. Each node is represented by its mention embedding plus a temporal encoding of its interval:
$$ p(e) = \big[\,\phi(s_e),\ \phi(t_e),\ \phi(\log(1 + t_e - s_e))\,\big], $$
Here $s_e$ and $t_e$ are the interval's start and end on the global axis, and $\phi$ is a bank of sinusoids at geometrically spaced frequencies, as in continuous-time encodings for trajectories (Giuliari et al.). The log-width term tells the model how precise the estimate is. Attention between nodes is thus conditioned on elapsed time, following the interval-aware attention used in sequential recommendation (Chen et al.; Zhang et al.). Unlike those systems, it runs over events recovered from unstructured text rather than logged interactions.
The model has three output heads: one that decides which mentions to merge as coreferent, one that predicts before/after/overlap for each pair, and one that assigns each event to a parent (subevent). We add the time-order loss of TSQA, which penalizes embeddings whose geometry contradicts chronology, since TSQA showed this is what fixes "before" versus "after" confusion. TSQA and TiPNN supply the method: relative time encoding and order constraints. Both operate on pre-built knowledge graphs, and our contribution is to apply them where the graph must first be built from segmented text.
Stage 4: temporal schema edges and conflict-triggered review. Event nodes are clustered into open types. We keep SCION's naming and evidence discipline, and use argument-sharing constraints in the spirit of Liberal Event Extraction. For each type pair, the corpus timelines give a precedence rate and an empirical gap distribution. Separately, an LLM is asked whether type A typically precedes type B. Edges on which the two sources agree with high confidence are accepted automatically. Edges on which they disagree, or where the gap distribution is multimodal, go to a human with the supporting evidence spans attached.
This reverses the design of Human-in-the-loop Schema Induction, which drafts scripts from scenario prompts with no source text and asks humans to repair the graph. That paper found human effort concentrated precisely at graph construction. It motivates our review policy, and the ungrounded drafting it uses is what we replace.
Baselines
Naive reference. The do-nothing reference is disjoint 512-token chunks. Relations are predicted only within a chunk, and the union of chunk outputs is taken, with every cross-chunk pair left unpredicted. It fixes the floor that the boundary effect imposes.
Overlap plus merge. This follows Díaz et al.: overlapping windows, window-local relation prediction with the same extractor, string-based entity/event merging, and then transitive closure over the local relations. It is the strongest cheap competitor because transitivity is known to cover most MAVEN-ERE temporal pairs.
Compression. TacoERE clusters sentences into topics, summarizes along event chains, and classifies from the compressed text. It reports 41.9 F1 with GPT-4 on MAVEN-ERE, up from 32.8 on raw text, and 34.1 with RoBERTa versus 32.3 for SIEF. It is the strongest published long-document event-relation method in the file, and its topic reordering plausibly discards chronology, which is what we test. We rerun it with the same backbone as our aggregator.
Whole-document long-context LLM. The same LLM used in Stage 1 receives the entire document and is asked for all events and temporal links. It gets the same prompt-tuning budget as our pipeline, and its token cost is reported alongside.
Supervised within-document reference. The RoBERTa-base baseline from MAVEN-ERE reports about 56.0 temporal F1, 81.4 coreference, and 27.5 subevent F1.
Beating the field means exceeding the best of these on cross-window temporal pairs by a margin larger than seed variance. The comparison holds when all systems share the extractor, the backbone size, and a matched hyperparameter search. It must also not fall below the MAVEN-ERE RoBERTa figure on within-window pairs.
Ablations. Three ablations isolate the key ingredient. The first replaces the timeline coordinate with global token offset, which is otherwise the same model. The second replaces the learned encoding with symbolic interval sorting and constraint propagation. The third feeds gold MAVEN-ERE timeline positions as an oracle to measure the ceiling. For schemas, SCION-lite reports 0.7518 Literal F1 and 0.7888 Graph F1 on SCOPE, and SCION-RL reports 0.8025 Literal F1 (SCOPE and SCION). Our pipeline must not degrade these within-event scores. For inter-event temporal schema edges there is no reported result found on Lacuna, so the reference points are the LLM-only edge judgment and the timeline-statistics-only edges. For human cost, the only figure available is the roughly one hour per manually authored schema cited in Human-in-the-loop Schema Induction. No reported per-edge review time was found on Lacuna, so we measure one.
Why not just…?
Why not just give the whole document to a long-context LLM? The evidence in the file points against it but is not conclusive for today's models. TacoERE found GPT-4 on raw long text at 32.8 F1, below the same model on compressed input. TextEE found LLM extraction in the 30s with hallucinated events. MuLD found long-context models still failing on ~90k-token narrative tasks. Newer models may close this gap, so the whole-document baseline is run in full. If it matches us on cross-window pairs at comparable cost, the aggregation contribution is not supported, and we would report that.
Why not just normalize time expressions and sort the events? This is the most serious threat, and it is why symbolic sorting with interval propagation is an ablation rather than a strawman. Sorting should win when documents are densely dated. The learned encoding can only justify itself on events with vague, relative, or missing anchors, where a soft model of width and context must fill gaps. We therefore stratify every result by anchor type. If symbolic sorting matches the learned encoding everywhere, we adopt it and the claim narrows to "a global time coordinate, however computed."
Why not just rely on overlap plus transitive closure? Transitivity only propagates what window-local models got right, and one local error spreads through the closure. It also cannot relate two events that share no chain of in-window neighbors. The distance-stratified comparison measures exactly where closure runs out.
Why not just review nothing, or review everything? Fully automatic edges inherit LLM sycophancy and sprawl, which motivated SCION's constraints and is documented by BrokenMath. Full review is the cost we are trying to cut. The review experiment below compares conflict-triggered review against both extremes and against a random sample of edges of equal size.
Proposed Experiments
Cross-window temporal relations. The first experiment uses MAVEN-ERE. We process each document with deliberately small windows (128 and 256 tokens) so that a large share of gold temporal pairs cross window boundaries. Temporal, coreference, and subevent F1 are reported separately for within-window and cross-window pairs, and further bucketed by token distance. The decisive check is whether our aggregator's advantage over overlap-plus-closure and TacoERE grows with distance while the token-offset ablation does not share it. If the token-offset ablation performs as well, time is not the active ingredient. The gold-timeline oracle shows how much of the remaining gap is timeline estimation error.
Temporal schema induction. The second experiment builds a gold standard of type-level temporal schema edges from MAVEN-ERE by aggregating gold temporal relations over MAVEN event types across documents. Induced edges are scored by edge precision and recall after type alignment, and by rank correlation of predicted versus gold precedence rates. We also report SCOPE Literal and Graph F1 to confirm that within-event schema quality is unchanged relative to SCION.
Sub-event granularity on the Lonquén archive. The third experiment reruns the setup of Díaz et al. on the Lonquén archive, where GPT-4o found 18 of 38 gold events and collapsed 20 sub-events into macro-events. The check is whether separating events by timeline interval recovers sub-events without inventing individuals. The gold set is small, so we treat this as a qualitative validation, not a headline number.
Human review. The fourth experiment has annotators review induced temporal edges on a held-out slice of MAVEN-ERE types under four policies: full review, conflict-triggered review, a random sample matched in size to the conflict set, and no review. We record minutes per accepted edge and final edge accuracy against the derived gold. The claim holds if conflict-triggered review reaches full-review accuracy within a small tolerance at substantially lower time, and clearly beats the size-matched random sample. The random-sample comparison is what separates targeted review from simply reviewing less.
References
- InstructCoder: Instruction Tuning Large Language Models for Code Editing
- MAVEN-ERE: A Unified Large-scale Dataset for Event Coreference, Temporal, Causal, and Subevent Relation Extraction
- SCOPE and SCION: A Benchmark and an Auditable Reference Pipeline for Schema Induction and Fusion from Text
- Automatic Knowledge-Graph Creation from Historical Documents: The Chilean Dictatorship as a Case Study
- TextEE: Benchmark, Reevaluation, Reflections, and Future Challenges in Event Extraction
- Transformer Networks for Trajectory Forecasting
- Exploring Periodicity and Interactivity in Multi-Interest Framework for Sequential Recommendation
- Learning Long- and Short-Term User Literal-Preference with Multimodal Hierarchical Transformer Network for Personalized Image Caption
- Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs
- Temporal inductive path neural network for temporal knowledge graph reasoning
- Liberal Event Extraction and Event Schema Induction
- Human-in-the-loop Schema Induction
- TacoERE: Cluster-aware Compression for Event Relation Extraction
- MuLD: The Multitask Long Document Benchmark
- BrokenMath: A Benchmark for Sycophancy in Theorem Proving with LLMs