Systems that group mentions of the same person, company or place across a stream of documents (cross-document coreference) mostly identify entities by their names. That works until the names change. Researchers move institutions, firms rebrand, cities are renamed, and over decades many unrelated people come to share a common name. A one-pass system with limited memory then fails in both directions. It splits one entity into several clusters because its surface form drifted, and it merges distinct entities because their names coincide. We propose a streaming coreference system that keeps, for each entity, a small fixed-size summary with three parts: what the entity is called; what its contexts look like once the name is hidden (entity masking, i.e. replacing the mention with a placeholder so a model must read the surrounding text rather than recognize the name); and which other entities it keeps appearing alongside across documents. A learned gate, driven by the time elapsed since the cluster was last seen, decides how much to trust each part. Recent evidence relies mostly on the name. Evidence across long gaps relies mostly on the masked context and the co-occurring entities.
The point is that one mechanism settles a tension visible in the existing literature. Streaming coreference benchmarks find that name features are indispensable: hybrid encoders win, and pure neural context encoders struggle with surface matching (Benchmarking Scalable Methods for Streaming Cross Document Entity Coreference). Work on evolving vocabulary finds that name features are exactly what breaks under drift (Covering Cracks in Content Moderation). Both findings are correct, but they hold at different time gaps. We make that dependence explicit and learnable, and we do it inside a memory budget that grows with the number of entities rather than the number of mentions.
Core Claim
Our hypothesis is that in a single chronological pass, under memory proportional to the number of distinct entities times a small constant, a coreference system whose per-entity evidence is split into a surface channel, a masked-context channel and a cross-document relational channel will recover substantially more correct links across long time gaps and renamings than the current hybrid streaming systems. On stationary streams it should lose essentially nothing. The new ingredient is not any single channel. It is the time-gap-dependent weighting between a channel that memorizes names and channels that cannot. That weighting borrows the agreement- and disagreement-decay idea from offline temporal record linkage and moves it into a learned, bounded-memory streaming setting.
How It Relates to Prior Work
Three lines of work each supply one piece and each leave out the others. The first is streaming cross-document coreference. Streaming Cross Document Entity Coreference Resolution clusters in one pass with a least-recently-used cache. Sequential Cross-Document Coreference Resolution compares each new mention to cluster prototypes in time linear in the number of clusters. Benchmarking Scalable Methods for Streaming Cross Document Entity Coreference formalizes memory-capped evaluation, with Bounded GRINCH (a clustering tree that can locally rearrange itself) and a Diversity-Cache that keeps varied representative mentions per cluster. These papers motivate the work and supply our clustering backbone. All of them assume stationary entities, use unmasked mentions, and treat each document's mentions independently of the other entities in that document.
The second line is temporal record linkage, which supplies the time-decay idea. Linking Temporal Records shows two effects. Disagreement between attribute values should be penalized less as the time gap grows. Agreement should be trusted less as the gap grows, because coincidental name collisions accumulate. Its ADJUST algorithm reports a 37–43% F-measure improvement over static linkage on DBLP. The same paper constrains our design: its greedy chronological variant (Early Binding) makes errors it cannot undo, and ADJUST needs repeated passes over the full dataset. We therefore cannot simply run ADJUST online. That is why we add a bounded repair buffer, described below.
The third line breaks the assumption that documents are independent, and it supplies the relational channel. Online Collective Entity Resolution shows that using co-author relations raises F1 on arXiv authors from 0.721 for attribute matching to 0.970 for collective resolution. However, it does so by query-time expansion over a static database. LinkBERT and CG-RAG exploit links between documents, but only over offline, fully indexed graphs. We keep only a bounded sketch of each cluster's frequent neighbor clusters, which costs memory per entity rather than per edge.
The masking channel follows the masked-context module of Covering Cracks in Content Moderation. That paper masks seed terms so a classifier learns the surrounding "contextual fingerprint", and doing so increases unique jargon coverage 4.23×. It operates on isolated sentences with no notion of identity over time; we use the same idea for identity tracking. Finally, KE-MHISTO motivates the setting. It documents entity linking F1 falling from above 0.80 on popular benchmarks to below 0.50 on long-tail historical entities, many of which have no knowledge-base entry. Linking to a modern knowledge base cannot help these entities, and a system that clusters without one can.
Research Program
Each incoming mention $m$ carries a timestamp and three encodings. The first is a character-and-token surface encoding, as in the hybrid encoders of Logan et al. The second is a context encoding computed after replacing the mention with a typed placeholder such as [PERSON]. The third is the set of other mentions in the same document, each carrying the cluster identifier already assigned to it.
Each cluster $c$ keeps a fixed-size summary. That summary holds a small diverse set of surface forms, a running mean and a small reservoir of masked-context vectors, a top-$k$ heavy-hitter sketch of co-occurring cluster identifiers, and the time $t_c$ at which the cluster was last observed. The score for attaching $m$ to $c$ is
$$s(m,c) = g_s(\Delta t)\,\mathrm{sim}_{\text{surf}} + g_x(\Delta t)\,\mathrm{sim}_{\text{mask}} + g_r(\Delta t)\,\mathrm{sim}_{\text{rel}} - \lambda(\Delta t)\,\mathbb{1}[\text{type conflict}],$$
where $\Delta t$ is the gap between $m$'s timestamp and $t_c$. The three similarity terms compare $m$'s surface, masked-context and co-occurrence encodings to the cluster summary. The gates $g_s, g_x, g_r$ are small monotone networks of $\Delta t$ trained on held-out drift data, and $\lambda(\Delta t)$ is a decaying penalty for incompatible entity types.
We expect but do not impose that the surface gate falls with the gap, mirroring agreement decay, while the other two gates stay flat or rise. Learning the gates, rather than fixing them, tests whether the decay shape found offline in structured records carries over to text.
Greedy commitment remains a problem in a single pass. We address it with a repair buffer of fixed total size $B$ holding recent low-margin decisions, meaning mentions whose best and second-best cluster scores were close. When a later mention scores highly against two clusters that both hold buffered low-margin members, the system tests a merge. This lets a renamed entity's old and new clusters be chained together by an intermediate "bridging" mention, which is the mechanism Li et al. exploit offline. Total memory is $O(n \cdot k)$ for $n$ live clusters with summary size $k$, plus $B$. This fits the semi-streaming spirit of per-entity rather than per-mention storage. We do not claim formal approximation guarantees, because the matching and clustering lower bounds in the file, such as Query Lower Bounds for Correlation Clustering Under Memory Constraints, concern objectives far from ours. We report accuracy against the memory budget instead.
The main construction effort is a drift-heavy streaming benchmark, since none of AIDA, MedMentions or Zeshel models multi-decade change. We build two streams.
The first comes from Developing a Temporal Bibliographic Data Set for Entity Resolution. It contains 2,931,038 records for 81,177 authors from 1936–2018 with affiliation histories and heavy homonymy (69 authors named "Wei Wang"). We order each author record by year and render it as a short text of name, affiliation, title and co-authors.
The second comes from WikiLinkGraphs. From it we take anchor-text mentions in timestamped Wikipedia revisions, with gold identity given by the dataset's redirect histories, which map historical article titles to stable entities. We use its template-filtered links, so navigation boilerplate does not flood the relational channel. This stream contains true renamings, and we tag each mention by whether its surface form ever changed and by whether its entity has a modern knowledge-base page.
Baselines
The naive reference point is static, time-blind streaming clustering: greedy nearest-neighbor attachment with a surface-plus-context encoder, as in Logan et al., with no memory cap. This shows how much drift costs before any memory pressure is applied.
The strongest existing method is the hybrid encoder with Bounded GRINCH and Diversity-Cache from Benchmarking Scalable Methods for Streaming Cross Document Entity Coreference. That paper reports 91.1 CoNLL F1 on AIDA and stays near peak while caching 10% of mentions. It is the strongest competitor because it already handles surface variation and bounded memory well, and the open question is whether drift needs anything beyond it. We run it on the same chronologically ordered streams, with the same encoder backbone and training data and the same memory budget measured in stored vectors, using the same hyperparameter search budget as our method. We also give it a learned time-gap feature concatenated into its scorer, so that it can in principle learn decay. Beating this time-augmented version means a statistically significant gain in CoNLL F1 on the renamed and long-gap slices at equal memory, while staying within one point of it on AIDA and MedMentions.
The offline upper reference is ADJUST from Linking Temporal Records, run with unlimited passes on the bibliographic stream. That paper reports only a relative 37–43% F-measure gain on DBLP and no absolute figure that we can reuse, so we re-run it. The target is to close most of the gap between the streaming baseline and ADJUST while staying in one pass.
Four ablations isolate the ingredients. Fixed, time-independent gates test whether the time gating matters. Removing the masked channel, and alternatively unmasking it so the context encoder sees the name, tests whether masking itself matters or merely adding a context channel. Removing the relational sketch tests the cost of treating documents independently. Setting $B=0$ tests the repair buffer.
Why not just…?
Why not just add a timestamp feature to the existing hybrid encoder? This is the real threat. A supervised scorer given $\Delta t$ could learn to discount name agreement across long gaps without any masking. The trouble is that it can only discount the name evidence. It still has to find positive evidence somewhere, and its context encoder is trained with the name visible, so its context vectors may themselves encode the old name and fail to match the new one. Our "unmasked context channel" ablation and the time-augmented baseline test exactly this. If the time-augmented hybrid matches us on the renamed slice, we will report that masking is unnecessary and that the contribution reduces to relational sketches plus repair.
Why not just normalize names with redirect or alias tables first? For head entities with Wikipedia histories this is cheap and probably strong, and the alias-normalized baseline will be included. However, alias tables exist only for entities a knowledge base already covers. KE-MHISTO shows that long-tail historical entities are often missing (the "NIL problem"), and bibliographic affiliation changes are not renamings that any alias table records. Reporting results separately for entities with and without knowledge-base coverage will show where alias tables suffice and where they do not.
Why not just run ADJUST periodically on windows of the stream? Windowed batch re-clustering does repair errors. However, it either keeps all mentions in each window, which costs memory per mention, or it cannot reach back across long gaps, which is where drift lives. We will run windowed ADJUST at the same memory budget as our system. If it wins, the repair buffer is the wrong design and should be replaced with periodic summary-level re-clustering.
Why not just use a long-context language model over retrieved documents? CRAG reports GPT-4 Turbo at about 34% accuracy alone and about 44% with straightforward retrieval on dynamic queries (CRAG). KE-MHISTO reports roughly 0.31–0.32 strict question-answering accuracy for LLaMA 3.3 70B on long-tail historical entities. This also does not address the core problem: clustering millions of mentions in a single pass needs a persistent memory, not a context window. We do not run this as a full baseline. We do run a small diagnostic in which a language model adjudicates the repair buffer's merge candidates, to see whether it beats the learned score on those hard cases.
Proposed Experiments
The primary metrics are CoNLL F1 and B³ F1, reported overall and on four slices: renamed entities, gaps longer than ten years, homonym-heavy names, and entities without knowledge-base coverage. We also report the Maximum Active Entities statistic of Logan et al., which measures how many mentions a system must store concurrently to cluster without loss, and plot F1 as a curve against memory budget.
The first decisive check is whether, on the WikiLinkGraphs renamed slice and the DBLP long-gap slice, our system beats the time-augmented hybrid baseline at matched memory. Failure on this check would falsify the core claim. The second is whether the learned surface gate actually decays with $\Delta t$, and whether its shape resembles the agreement decay Li et al. estimated on structured records. If the gate stays flat, the time-gating story is wrong even if accuracy improves. The third is whether we stay within one CoNLL F1 point of the reported 91.1 on AIDA under a 10% mention cache, which confirms that the drift machinery does no harm when nothing drifts. The fourth is a stress test on the European patent data used by Li et al., replaying the stream in random order as well as chronological order. If the gains survive only in chronological order, the method depends on the order of arrival more than on its representations, and we will say so.
References
- Benchmarking Scalable Methods for Streaming Cross Document Entity Coreference
- Covering Cracks in Content Moderation: Delexicalized Distant Supervision for Illicit Drug Jargon Detection
- Streaming Cross Document Entity Coreference Resolution
- Sequential Cross-Document Coreference Resolution
- Linking Temporal Records
- Online Collective Entity Resolution
- LinkBERT: Pretraining Language Models with Document Links
- CG-RAG: Research Question Answering by Citation Graph Retrieval-Augmented LLMs
- KE-MHISTO: Towards a Multilingual Historical Knowledge Extraction Benchmark for Addressing the Long-Tail Problem
- Query Lower Bounds for Correlation Clustering Under Memory Constraints
- Developing a Temporal Bibliographic Data Set for Entity Resolution
- WikiLinkGraphs: A Complete, Longitudinal and Multi-Language Dataset of the Wikipedia Link Networks
- CRAG - Comprehensive RAG Benchmark