A Study of Temporal Fusion Strategies for Named Entity Recognition in Historical Texts

# A Study of Temporal Fusion Strategies for Named Entity Recognition in Historical Texts
Source: [https://arxiv.org/html/2606.27881](https://arxiv.org/html/2606.27881)
11institutetext:Digital Humanities Laboratory, EPFL, Lausanne, Switzerland###### Abstract

Temporal variation poses a unique challenge for named entity recognition \(NER\) in historical texts, where entities drift in surface form and salience across time\. While language models \(LMs\) have made progress in various NLP tasks, their ability to reason about temporality, especially in diachronic contexts, remains limited or at least, questionable\. In this paper, we systematically study how temporal metadata can be structurally embedded into NER models using a range of lightweight fusion strategies\. We experiment with both absolute and relative temporal representations, injected into Transformer\-based architectures via early or late fusion mechanisms such as cross\-attention, adapters, and concatenation\. Our evaluations on French and German historical datasets reveal that late fusion strategies yield more robust and temporally generalisable performance, particularly in early and noisy periods\.

## 1Introduction

Language is inherently temporal: its vocabulary, structures, and referents evolve across time\. Yet LMs, despite their generalization power, still struggle with temporal reasoning\[[7](https://arxiv.org/html/2606.27881#bib.bib29),[16](https://arxiv.org/html/2606.27881#bib.bib3),[20](https://arxiv.org/html/2606.27881#bib.bib38),[21](https://arxiv.org/html/2606.27881#bib.bib32),[23](https://arxiv.org/html/2606.27881#bib.bib33),[24](https://arxiv.org/html/2606.27881#bib.bib5),[39](https://arxiv.org/html/2606.27881#bib.bib35),[27](https://arxiv.org/html/2606.27881#bib.bib24),[36](https://arxiv.org/html/2606.27881#bib.bib25),[37](https://arxiv.org/html/2606.27881#bib.bib1)\]\. Studies show that even advanced \(generative\) models like GPT\-4 exhibit directionality biases\[[25](https://arxiv.org/html/2606.27881#bib.bib27)\], poor calibration over time\[[2](https://arxiv.org/html/2606.27881#bib.bib28)\], and difficulties retaining or reasoning over temporally anchored facts\[[6](https://arxiv.org/html/2606.27881#bib.bib4)\]\. This limitation is particularly problematic in tasks such as named entity recognition \(NER\) over historical texts, where entities evolve, drift, or vanish entirely across time\[[3](https://arxiv.org/html/2606.27881#bib.bib44),[8](https://arxiv.org/html/2606.27881#bib.bib10),[9](https://arxiv.org/html/2606.27881#bib.bib9),[10](https://arxiv.org/html/2606.27881#bib.bib21),pawłowski2024nlpfordigital,[28](https://arxiv.org/html/2606.27881#bib.bib12)\]\. While temporality has been studied more commonly in video\-based reasoning\[[18](https://arxiv.org/html/2606.27881#bib.bib36),[21](https://arxiv.org/html/2606.27881#bib.bib32)\], in NLP tasks such as QA\[[1](https://arxiv.org/html/2606.27881#bib.bib15),[4](https://arxiv.org/html/2606.27881#bib.bib34),[13](https://arxiv.org/html/2606.27881#bib.bib41),[17](https://arxiv.org/html/2606.27881#bib.bib40),[30](https://arxiv.org/html/2606.27881#bib.bib26),[33](https://arxiv.org/html/2606.27881#bib.bib31),[37](https://arxiv.org/html/2606.27881#bib.bib1)\], or retrieval augmentation\[[11](https://arxiv.org/html/2606.27881#bib.bib6)\], historical NER remains comparatively underexplored\.

Recent research has introduced temporal representations like time vectors\[[24](https://arxiv.org/html/2606.27881#bib.bib5)\], timestamp\-aware pretraining\[[6](https://arxiv.org/html/2606.27881#bib.bib4)\], temporal graphs\[[19](https://arxiv.org/html/2606.27881#bib.bib42),[22](https://arxiv.org/html/2606.27881#bib.bib43),[29](https://arxiv.org/html/2606.27881#bib.bib47),[32](https://arxiv.org/html/2606.27881#bib.bib49)\], and dynamic knowledge editing\[[38](https://arxiv.org/html/2606.27881#bib.bib2)\]to help models encode temporal signals\. Yet these remain largely disconnected from token\-level tasks like NER\. Interpretability studies such as probing\[[14](https://arxiv.org/html/2606.27881#bib.bib48),[34](https://arxiv.org/html/2606.27881#bib.bib46)\]and temporal diagnostic tests like TEMPLAMA\[[6](https://arxiv.org/html/2606.27881#bib.bib4)\]confirm that temporal information is often only weakly represented in model weights\.

In the domain of NER, earlier efforts to address time drift focused on sampling or data augmentation in high\-churn environments like different platforms of social media\[[5](https://arxiv.org/html/2606.27881#bib.bib8),[28](https://arxiv.org/html/2606.27881#bib.bib12),[35](https://arxiv.org/html/2606.27881#bib.bib18)\]\. Meanwhile, historical NER introduces compounding challenges: diachronic drift, OCR degradation, and multilingual variation\. Benchmarks such as HIPE\[[8](https://arxiv.org/html/2606.27881#bib.bib10),[9](https://arxiv.org/html/2606.27881#bib.bib9),[10](https://arxiv.org/html/2606.27881#bib.bib21)\]have laid the groundwork, and newer work has begun exploring temporally aware grounding through context retrieval\[[29](https://arxiv.org/html/2606.27881#bib.bib47)\], temoral knowledge graphs injection\[[12](https://arxiv.org/html/2606.27881#bib.bib14)\], or LLM\-based inference\[[15](https://arxiv.org/html/2606.27881#bib.bib17)\]\. While this is a good start, none have systematically compared architectural fusion strategies or directly assessed in practice whether models internalise temporal information\.

In this paper, we \(1\) systematically inject temporal information into a Transformer architecture using explicit year embeddings, \(2\) design and compare a suite of modular, interpretable fusion strategies that incorporate time at different points in the model \(e\.g\., early vs\. late\), and \(3\) benchmark their impact across decades and languages, while probing whether the models genuinely internalise temporal signals\. We hope that this study will contribute to a clearer understanding of how time can be structurally integrated into token\-level models and inform future work in \(practical\) historical NLP and temporally\-aware sequence modeling\.

## 2Incorporating Temporality into NER

##### Task Formulation\.

We treat historical named entity recognition \(NER\) as a straightforward token classification task, just with a temporal twist\. Each input consists of a sequence of tokensX=\(x1,x2,…,xn\)X=\(x\_\{1\},x\_\{2\},\.\.\.,x\_\{n\}\), along with the document’s publication yeary​e​a​r∈ℕyear\\in\\mathbb\{N\}\. The goal is to assign each tokenxix\_\{i\}a labellil\_\{i\}, selecting from a standard entity taxonomy or marking it as non\-entity\. We use a Transformer\-based architecture, where an encoder produces contextualized token representationsH=Encoder​\(X\)∈ℝT×dH=\\text\{Encoder\}\(X\)\\in\\mathbb\{R\}^\{T\\times d\}, withTTthe number of tokens andddthe hidden size\. Each labellil\_\{i\}is then predicted fromhi∈ℝdh\_\{i\}\\in\\mathbb\{R\}^\{d\}, the contextualised representation of tokenxix\_\{i\}\.

##### Temporal Fusion Strategies\.

To enable temporal adaptation of token classification models, we incorporate a temporal fusion module that integrates temporal context into token representations\. This module fuses contextualised encoder outputs with a year\-specific embedding using one of several strategies\. We categorise them intotwo fusion types:

- •early fusion, where temporal information is injected before or during encoding; and
- •late fusion, where temporal information is applied to the encoder output\.

We explore these strategies intwo modes of encoding temporal information:

- •absolutemode, the embedding index corresponds directly to the publication year \(e\.g\., 1889\); and
- •time\-distancemode, we instead compute the number of years between the document’s publication date and a fixed reference year, namely 2025, assigning lower indices to more recent documents\.

More specifically, lety=Emb​\(y​e​a​r\)∈ℝdy=\\text\{Emb\}\(\{year\}\)\\in\\mathbb\{R\}^\{d\}denote the embedding of the document’s publication year\.

##### Baseline\.

This strategy skips temporal fusion entirely, i\.e\.,H~t=Ht\\tilde\{H\}\_\{t\}=H\_\{t\}, and serves as a control condition\.

### Early Fusion

##### Cross\-Attention Fusion \(early\-cross\-attention\)\.

Temporal information is injected*before*encoding via cross\-attention between the token embeddings and the year embedding:

H~=H\+MultiHeadAttention​\(Q=H,K=y,V=y\),\\tilde\{H\}=H\+\\text\{MultiHeadAttention\}\(Q=H,K=y,V=y\),whereHHdenotes the input token embeddings andyyis the year embedding, broadcast to match the input length\. This mechanism allows each token to attend directly to the temporal context during encoding\.

### Late Fusion

##### Adapter Fusion \(adapter\)\.

A lightweight MLP \(adapter\) processes the year embedding and adds the result to each token:

H~t=Ht\+MLP​\(y\),MLP:ℝd→ℝd\.\\tilde\{H\}\_\{t\}=H\_\{t\}\+\\text\{MLP\}\(y\),\\quad\\text\{MLP\}:\\mathbb\{R\}^\{d\}\\rightarrow\\mathbb\{R\}^\{d\}\.

##### Concatenation Fusion \(concat\)\.

Generic fusion technique in many tasks, The year embedding is concatenated to each token vector and projected back to the original dimensionality:

H~t=W⋅\[Ht;y\],W∈ℝ2​d×d\.\\tilde\{H\}\_\{t\}=W\\cdot\[H\_\{t\};y\],\\quad W\\in\\mathbb\{R\}^\{2d\\times d\}\.

##### Relative Temporal Fusion \(relative\)\.

A nonlinear encoder transforms the year embedding into a relative temporal representation, which is used in a feature\-wise linear modulation \(FiLM\)\-like modulation\[[26](https://arxiv.org/html/2606.27881#bib.bib7)\]:

y′=LayerNorm​\(SiLU​\(W​y\)\),H~t=γ​\(y′\)⊙Ht\+β​\(y′\),where:y^\{\\prime\}=\\text\{LayerNorm\}\(\\text\{SiLU\}\(Wy\)\),\\tilde\{H\}\_\{t\}=\\gamma\(y^\{\\prime\}\)\\odot H\_\{t\}\+\\beta\(y^\{\\prime\}\),\\text\{where:\}SiLU​\(x\)=x⋅σ​\(x\)​is the sigmoid linear unit and​σ​\(x\)​is the logistic sigmoid\.\\text\{SiLU\}\(x\)=x\\cdot\\sigma\(x\)\\text\{ is the sigmoid linear unit and \}\\sigma\(x\)\\text\{is the logistic sigmoid\.\}

##### Cross\-Attention Fusion \(late\-cross\-attention\)\.

Temporal information is fused with the encoder output using cross\-attention, similar to theearly fusionone but*after*encoding:

H~=H\+MultiHeadAttention​\(Q=H,K=y,V=y\)\.\\tilde\{H\}=H\+\\text\{MultiHeadAttention\}\(Q=H,K=y,V=y\)\.

## 3Experimental Setup

##### Datasets\.

Our experiments are based on thehipe2020dataset, as included in the HIPE\-2022 shared task\[[10](https://arxiv.org/html/2606.27881#bib.bib21)\]\. We focus exclusively on the French and German subsets, which include publication year metadata required for temporal modeling \(the English subset was excluded due to missing training data\)\. We use the coarse\-grained entity taxonomy \(loc, org, pers, time, prod\) and retain all documents regardless of their temporal span\. The French data comprises 10,923 annotated mentions across 1798–2018 with an average OCR noise rate of≈\\approx33%, while the German subset contains 6,584 mentions spanning 1798–1950 with≈\\approx43% OCR noise\. While all splits cover wide temporal ranges, our goal is not to simulate chronological generalisation but to analyse the structural inclusion of time in the models\.

##### Evaluation & Hyperparameters\.

We evaluate all models using micro\-averaged F1 scores, computed at the entity level\. All models are fine\-tuned using the standard Transformer architecture, with the multilingual historical variant as base model111[https://huggingface\.co/dbmdz/bert\-base\-historic\-multilingual\-cased](https://huggingface.co/dbmdz/bert-base-historic-multilingual-cased)\[[31](https://arxiv.org/html/2606.27881#bib.bib45)\], with a maximum sequence length of 512 tokens\. Models are trained using a batch size of 16, for 5 epochs, with a fixed seed \(2025\) for reproducibility\.

#### 3\.0\.1NER Performance Across Temporal Strategies\.

To evaluate the effectiveness of temporal conditioning, we plot F1 scores across publication years for each fusion strategy under two temporal modes:absoluteandtime\-distancein Figure[1](https://arxiv.org/html/2606.27881#S3.F1)\. At first glance, we might not be able to see big improvements, but we do observe several slight temporal patterns across both languages:

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/yearly_f1_by_strategy_type_row_dbmdz_bert_base_historic_hipe-fr.png)

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/yearly_f1_by_strategy_type_row_dbmdz_bert_base_historic_hipe-de.png)

Figure 1:F1 scores over time for French \(top\) and German \(bottom\) subsets of HIPE\-2020 under two temporal modes:absolute\(left\) andtime\-distance\(right\)\.- •1800–1850: Early periods exhibit high variability in F1 scores, likely due to OCR noise and sparse annotations\. Late fusion strategies demonstrate notable gains in robustness, particularly under thetime\-distancemode, outperforming both baseline and early fusion\.
- •1850–1900: Performance stabilizes across models\. While all strategies benefit from improved data quality, late fusion still maintains a slight edge, especially in French\. Early fusion appears more sensitive to temporal encoding choices\.
- •1900–1950: F1 scores fluctuate again, particularly in German, with drops around 1940–1950\. This may be attributed to document scarcity or inconsistencies in historical orthography\. Late fusion again proves more resilient\.
- •1950–2000: Baseline models catch up, but late fusion strategies retain superiority, especially in the German subset\. The performance gap narrows, suggesting a a a diminishing marginal benefit from temporal conditioning in modern decades\.
- •2000–2018: All models improve steadily due to better OCR and more standardised data\. However, late fusion strategies still outperform slightly, reflecting their capacity to generalise across time even when temporal drift is lower\.

Generally, we observe that all temporal fusion strategies, particularly late fusion ones, consistently improve NER performance across both languages with some benefits most pronounced in early or noisy periods, but before establishing the significance of these results, we analyse next other possible influencing factors\.

#### 3\.0\.2Absolute Time versus Distance\-based Encoding\.

We compare the impact ofabsoluteversustime\-distancetemporal encoding by computing the mean F1 score difference per strategy \(Figure[2](https://arxiv.org/html/2606.27881#S3.F2)\)\.

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/delta_f1_by_temporal_mode_hipe-fr.png)\(a\)French
![Refer to caption](https://arxiv.org/html/2606.27881v1/images/delta_f1_by_temporal_mode_hipe-de.png)\(b\)German

Figure 2:Average F1 score difference betweentime\-distanceandabsolutetemporal modes, computed for each fusion strategy\. Positive values indicate improved performance\.We see that, in German, strategies likeconcat,relative, andadapterbenefit fromtime\-distanceencoding \(up to \+3 F1\), suggesting improved temporal generalisation\. In French, however, effects are less consistent: whileconcatgains slightly, others such asadapterandlate\-cross\-attentionperform better withabsoluteencoding\. These results could imply that while the choice of temporal mode is secondary to the fusion strategy, it can still influence outcomes and should be tuned per language and setup\.

#### 3\.0\.3Entity Length Sensitivity\.

To explore whether temporal strategies differentially impact entity mentions of varying surface complexity, we categorize entities by their character length: those with 10 characters or fewer are considered*short*, those between 11 and 20 as*medium*, and those exceeding 20 characters as*long*\. For each group, we compute average F1 scores and analyze the performance gap between long and short entities, denoted asΔ\\DeltaF1 = F1long\{\}\_\{\\text\{long\}\}\- F1short\{\}\_\{\\text\{short\}\}\.

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/delta_f1_long_short_by_strategy_and_mode_hipe-fr.png)

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/delta_f1_long_short_by_strategy_and_mode_hipe-de.png)

Figure 3:Difference in F1 score for French \(top\) and German \(bottom\) between long and short entity mentions \(Δ\\DeltaF1 = F1long\{\}\_\{\\text\{long\}\}\- F1short\{\}\_\{\\text\{short\}\}\), across decades and fusion strategies for each temporal mode\.Figure[3](https://arxiv.org/html/2606.27881#S3.F3)presents the length sensitivity analysis across French \(left\) and German \(right\) subsets of the HIPE\-2020 corpus\. We observe that late fusion strategies tend to show a more stable or slightly positive gain for longer entities across both languages, particularly in earlier decades where surface forms tend to be longer or more structurally complex\. The effect is more pronounced under thetime\-distancetemporal mode, where relative temporal encoding appears to support generalisation over long spans\. Baseline and early fusion strategies, by contrast, exhibit more unpredictability or minimal difference\. These results suggest that injecting temporal signals at later stages of the model helps preserve surface\-level distinctions critical for accurately identifying long entities\.

#### 3\.0\.4Entity Type Gains from Temporal Fusion\.

To assess which entity types benefit most from temporal fusion, we compute the gain in prediction frequency for each surface–type pair, defined as the increase in count compared to the baseline\. Figure[4](https://arxiv.org/html/2606.27881#S3.F4)shows thatlocentities exhibit the highest variability and occasional large gains, suggesting that temporal conditioning improves their recall, likely due to historical drift and ambiguity\. Other types such asorg,pers,prod, andtimeshow more modest and consistent distributions, indicating that improvements are generally limited in magnitude\.

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/entity_type_gain_vs_baseline_fr.png)\(a\) French
![Refer to caption](https://arxiv.org/html/2606.27881v1/images/entity_type_gain_vs_baseline_de.png)\(b\) German

Figure 4:Distribution of gain over baseline for each entity type, measured as the difference in surface form frequency between temporal models and the baseline\.
#### 3\.0\.5Do the Time\-based Models Really Learn Time?

To evaluate the extent to which our models encode temporal information internally, we adopt a linear probing strategy\. LethCLS∈ℝdh\_\{\\text\{CLS\}\}\\in\\mathbb\{R\}^\{d\}denote the final hidden representation of the\[CLS\]token\. We train a linear classifier of the form:

y^=arg⁡maxi⁡Wi⊤​hCLS\+bi,\\hat\{y\}=\\arg\\max\_\{i\}\\;W\_\{i\}^\{\\top\}h\_\{\\text\{CLS\}\}\+b\_\{i\},whereW∈ℝd×YW\\in\\mathbb\{R\}^\{d\\times Y\},b∈ℝYb\\in\\mathbb\{R\}^\{Y\}, andYYis the number of discrete publication years\. To ensure that the probing task evaluates*latent*temporal knowledge rather than reflecting direct access to input metadata, we modify the forward pass by injecting a randomly sampled publication yeary∈ℕy\\in\\mathbb\{N\}during inference\. This disables architectural conditioning on the true document year \(whetherabsoluteortime\-distance\)\. To account for randomness and obtain a more stable signal, we repeat the probing process five times and report the average accuracy across runs\.

![Refer to caption](https://arxiv.org/html/2606.27881v1/images/probing_accuracy_comparison.png)Figure 5:Probing accuracy across models\. Left: grouped by fusion type\. Right: grouped by fusion strategy\.Figure[5](https://arxiv.org/html/2606.27881#S3.F5)reports the average prediction accuracy, grouped by fusion type and strategy, and we notice that the late fusion strategies consistently yield higher accuracy than early fusion or baseline models, confirming that injecting temporal information after contextualization better preserves temporal signals in the latent space\. Among them, late cross\-attention, adapter, and concatenation lead to the best results\. In contrast, the baseline and early\-cross\-attention models show minimal temporal encoding\. We could even say that probing shows that time\-aware architectures, especially late\-fusion models, encode temporality even when gold\-year metadata is removed\. This suggests that structural fusion mechanisms lead to genuine internalisation of temporal context, rather than relying on surface\-level cues\.

#### 3\.0\.6Are the Improvements Really Significant?

To assess whether temporal fusion strategies offer statistically significant improvements over the baseline, we conducted paired t\-tests across yearly F1 scores for each strategy and temporal mode\. We noticed that most fusion strategies do not yield statistically significant gains at thep<0\.05p<0\.05threshold with the exception of thelate\-cross\-attentionstrategy under theabsolutetemporal mode demonstrates a significant difference compared to the baseline \(p=0\.041p=0\.041\)\. This could suggest that, while temporal fusion generally improves model performance, these improvements are often subtle and not uniformly consistent across years\.

## 4Insights & Conclusions

By structurally injecting time into Transformer\-based architectures using modular fusion strategies, we demonstrate that temporal conditioning yields modest yet consistent gains for historical NER across languages, decades, and entity types\. Late fusion strategies, particularlylate\-cross\-attention, perform most robustly, especially in early, noisy periods, and help improve recognition of longer entities and temporally variable types like locations\. Thus, based on our findings, we recommend: \(1\) adopting late fusion for integrating time; \(2\) testing bothabsoluteandtime\-distanceencodings, as their impact is context\-dependent; and \(3\) using temporal fusion as a lightweight enhancement for diachronic or noisy corpora\. While our approach is not novel and we acknowledge the growing utility of generative LLMs, we emphasize that real\-world historical corpora often impose constraints: they may be large, private, or governed by restrictive policies\. In such cases, structured methods that leverage metadata such as time or publication dates remain an important source of exploitable information for interpretable models\.

## Limitations

While our study presents a systematic comparison of temporal fusion strategies for historical NER, there still remain several limitations\. First, we only consider year\-level granularity, which may be insufficient for domains requiring finer temporal resolution\. Second, our experiments are confined to the HIPE\-2020 dataset’s French and German subsets, results may not generalise to other languages or genres of historical text\. Third, the probing analysis focuses solely on linear decodability of year embeddings and may underestimate more subtle forms of temporal encoding\. Finally, our models are evaluated under controlled conditions using a single backbone architecture; real\-world applications with noisy or missing metadata may yield different results\.

## References

- \[1\]P\. Agarwal, J\. Strötgen, L\. del Corro, J\. Hoffart, and G\. Weikum\(2018\)DiaNED: time\-aware named entity disambiguation for diachronic corpora\.External Links:[Link](https://www.aclweb.org/anthology/P18-2109/)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[2\]H\. Beniwal, D\. Patel, K\. N\. D, H\. Ladia, A\. Yadav, and M\. Singh\(2024\)Remember this event that year? assessing temporal information and reasoning in large language models\.External Links:[Link](https://arxiv.org/abs/2402.11997)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[3\]E\. Boros, A\. Hamdi, E\. L\. Pontes, L\. Cabrera\-Diego, J\. G\. Moreno, N\. Sidere, and A\. Doucet\(2020\)Alleviating digitization errors in named entity recognition for historical documents\.InProceedings of the 24th conference on computational natural language learning,pp\. 431–441\.Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[4\]H\. Chang, C\. Ye, Z\. Tao, J\. Wu, Z\. Yang, Y\. Ma, X\. Huang, and T\. Chua\(2024\)A comprehensive evaluation of large language models on temporal event forecasting\.External Links:[Link](https://arxiv.org/abs/2407.11638)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[5\]S\. Chen, L\. Neves, and T\. Solorio\(2021\-06\)Mitigating temporal\-drift: a simple approach to keep NER models crisp\.InProceedings of the Ninth International Workshop on Natural Language Processing for Social Media,Online,pp\. 163–169\.External Links:[Link](https://www.aclweb.org/anthology/2021.socialnlp-1.14),[Document](https://dx.doi.org/10.18653/v1/2021.socialnlp-1.14)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[6\]J\. R\. Cole\(2022\)Time\-aware language models as temporal knowledge bases\.External Links:[Link](https://aclanthology.org/2022.tacl-1.15/),[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00459)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1),[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[7\]X\. Ding and L\. Wang\(2024\)Do language models understand time?\.External Links:[Link](https://arxiv.org/abs/2412.13845)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[8\]M\. Ehrmann, M\. Romanello, S\. Bircher, and S\. Clematide\(2020\)Introducing the clef 2020 hipe shared task: named entity recognition and linking on historical newspapers\.\.External Links:[Link](https://doi.org/10.1007/978-3-030-45442-5%5C_68),[Document](https://dx.doi.org/10.1007/978-3-030-45442-5%5F68)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1),[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[9\]M\. Ehrmann, M\. Romanello, A\. Doucet, and S\. Clematide\(2022\)Introducing the hipe 2022 shared task: named entity recognition and linking in multilingual historical documents\.\.External Links:[Link](https://doi.org/10.1007/978-3-030-99739-7%5C_44),[Document](https://dx.doi.org/10.1007/978-3-030-99739-7%5F44)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1),[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[10\]M\. Ehrmann, M\. Romanello, S\. Najem\-Meyer, A\. Doucet, and S\. Clematide\(2022\)Extended overview of HIPE\-2022: Named Entity Recognition and Linking in Multilingual Historical Documents\.InProceedings of the Working Notes of CLEF 2022 \- Conference and Labs of the Evaluation Forum,G\. Faggioli, N\. Ferro, A\. Hanbury, and M\. Potthast \(Eds\.\),Vol\.3180\.External Links:[Document](https://dx.doi.org/10.5281/zenodo.6979577),[Link](http://ceur-ws.org/Vol-3180/paper-83.pdf)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1),[§1](https://arxiv.org/html/2606.27881#S1.p3.1),[§3](https://arxiv.org/html/2606.27881#S3.SS0.SSS0.Px1.p1.2)\.
- \[11\]A\. Gade and J\. Jetcheva\(2024\)It’s about time: incorporating temporality in retrieval augmented language models\.External Links:[Link](https://arxiv.org/abs/2401.13222)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[12\]C\. González\-Gallardo, E\. Boros, E\. Giamphy, A\. Hamdi, J\. G\. Moreno, and A\. Doucet\(2023\)Injecting temporal\-aware knowledge in historical named entity recognition\.\.External Links:[Link](https://doi.org/10.1007/978-3-031-28244-7%5C_24),[Document](https://dx.doi.org/10.1007/978-3-031-28244-7%5F24)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[13\]R\. Gruber, A\. Abdallah, M\. Färber, and A\. Jatowt\(2024\)ComplexTempQA: a large\-scale dataset for complex temporal question answering\.External Links:[Link](https://arxiv.org/abs/2406.04866)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[14\]W\. Gurnee and M\. Tegmark\(2024\)Language models represent space and time\.External Links:[Link](https://openreview.net/forum?id=jE8xbmvFin)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[15\]T\. Hiltmann, M\. Dröge, N\. Dresselhaus, T\. Grallert, M\. Althage, P\. Bayer, S\. Eckenstaler, K\. Mendi, J\. M\. Schmitz, P\. Schneider, W\. Sczeponik, and A\. Skibba\(2025\)NER4all or context is all you need: using llms for low\-effort, high\-performance ner on historical texts\. a humanities informed approach\.External Links:[Link](https://arxiv.org/abs/2502.04351)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[16\]R\. Jain, D\. Sojitra, A\. Acharya, S\. Saha, A\. Jatowt, and S\. Dandapat\(2023\)Do language models have a common sense regarding time? revisiting temporal commonsense reasoning in the era of large language models\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.418/)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[17\]Z\. Jia, A\. Abujabal, R\. S\. Roy, J\. Strötgen, and G\. Weikum\(2018\)TempQuestions: a benchmark for temporal question answering\.\.External Links:[Link](https://doi.org/10.1145/3184558.3191536),[Document](https://dx.doi.org/10.1145/3184558.3191536)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[18\]D\. Ko, J\. S\. Lee, W\. Kang, B\. Roh, and H\. J\. Kim\(2023\)Large language models are temporal and causal reasoners for video question answering\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.261/)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[19\]K\. Liang, L\. Meng, M\. Liu, Y\. Liu, W\. Tu, S\. Wang, S\. Zhou, X\. Liu, and F\. Sun\(2022\)A survey of knowledge graph reasoning on graph types: static\.Dynamic, and Multimodal\.Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[20\]L\. Liu, S\. Yu, R\. Wang, Z\. Ma, and Y\. Shen\(2024\)How can large language models understand spatial\-temporal data?\.External Links:[Link](https://arxiv.org/abs/2401.14192)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[21\]R\. Liu, C\. Li, H\. Tang, Y\. Ge, Y\. Shan, and G\. Li\(2024\)ST\-llm: large language models are effective temporal learners\.Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[22\]Y\. Lu, Y\. Zhou, J\. Li, Y\. Wang, X\. Liu, D\. He, F\. Liu, and M\. Zhang\(2025\)Knowledge editing with dynamic knowledge graphs for multi\-hop question answering\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.39,pp\. 24741–24749\.Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[23\]P\. Nako and A\. Jatowt\(2025\)Navigating tomorrow: reliably assessing large language models performance on future event prediction\.External Links:[Link](https://arxiv.org/abs/2501.05925)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[24\]K\. Nylund, S\. Gururangan, and N\. A\. Smith\(2023\)Time is encoded in the weights of finetuned language models\.External Links:[Link](https://arxiv.org/abs/2312.13401)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1),[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[25\]V\. Papadopoulos, J\. Wenger, and C\. Hongler\(2024\)Arrows of time for large language models\.External Links:[Link](https://openreview.net/forum?id=UpSe7ag34v)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[26\]E\. Perez, F\. Strub, H\. De Vries, V\. Dumoulin, and A\. Courville\(2018\)Film: visual reasoning with a general conditioning layer\.InProceedings of the AAAI conference on artificial intelligence,Vol\.32\.Cited by:[§2](https://arxiv.org/html/2606.27881#S2.SSx2.SSS0.Px3.p1.1)\.
- \[27\]Y\. Qiu, Z\. Zhao, Y\. Ziser, A\. Korhonen, E\. M\. Ponti, and S\. B\. Cohen\(2023\)Are large language models temporally grounded?\.External Links:[Link](https://arxiv.org/abs/2311.08398)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[28\]S\. Rijhwani and D\. Preotiuc\-Pietro\(2020\)Temporally\-informed analysis of named entity recognition\.External Links:[Link](https://www.aclweb.org/anthology/2020.acl-main.680/)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1),[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[29\]G\. D\. Rosin, I\. Guy, and K\. Radinsky\(2022\)Time masking for temporal language models\.InProceedings of the fifteenth ACM international conference on Web search and data mining,pp\. 833–841\.Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1),[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[30\]A\. G\. Ruiz, T\. de la Rosa, and D\. Borrajo\(2025\)On the temporal question\-answering capabilities of large language models over anonymized data\.External Links:[Link](https://arxiv.org/abs/2504.07646)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[31\]S\. Schweter, L\. März, K\. Schmid, and E\. Çano\(2022\)HmBERT: historical multilingual language models for named entity recognition\.External Links:[Link](https://arxiv.org/abs/2205.15575)Cited by:[§3](https://arxiv.org/html/2606.27881#S3.SS0.SSS0.Px2.p1.1)\.
- \[32\]R\. Song, S\. He, S\. Gao, L\. Cai, K\. Liu, Z\. Yu, and J\. Zhao\(2023\-07\)Multilingual knowledge graph completion from pretrained language models with knowledge constraints\.InFindings of the Association for Computational Linguistics: ACL 2023,A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 7709–7721\.External Links:[Link](https://aclanthology.org/2023.findings-acl.488/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.488)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[33\]Q\. Tan, H\. T\. Ng, and L\. Bing\(2023\)Towards benchmarking and improving the temporal reasoning capability of large language models\.External Links:[Link](https://arxiv.org/abs/2306.08952)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[34\]S\. Thukral, K\. Kukreja, and C\. Kavouras\(2021\-11\)Probing language models for understanding of temporal expressions\.InProceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP,J\. Bastings, Y\. Belinkov, E\. Dupoux, M\. Giulianelli, D\. Hupkes, Y\. Pinter, and H\. Sajjad \(Eds\.\),Punta Cana, Dominican Republic,pp\. 396–406\.External Links:[Link](https://aclanthology.org/2021.blackboxnlp-1.31/),[Document](https://dx.doi.org/10.18653/v1/2021.blackboxnlp-1.31)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[35\]A\. Ushio, F\. Barbieri, V\. Sousa, L\. Neves, and J\. Camacho\-Collados\(2022\)Named entity recognition in twitter: a dataset and analysis on short\-term temporal shifts\.External Links:[Link](https://aclanthology.org/2022.aacl-main.25/)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p3.1)\.
- \[36\]J\. Wallat, A\. Jatowt, and A\. Anand\(2024\)Temporal blind spots in large language models\.External Links:[Link](https://arxiv.org/abs/2401.12078)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[37\]S\. Xiong, A\. Payani, R\. Kompella, and F\. Fekri\(2024\)Large language models can learn temporal reasoning\.External Links:[Link](https://aclanthology.org/2024.acl-long.563/)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.
- \[38\]X\. Yin, J\. Jiang, L\. Yang, and X\. Wan\(2023\)History matters: temporal knowledge editing in large language model\.External Links:[Link](https://arxiv.org/abs/2312.05497)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p2.1)\.
- \[39\]L\. N\. Zheng, C\. G\. Dong, W\. E\. Zhang, L\. Yue, M\. Xu, O\. Maennel, and W\. Chen\(2024\)Understanding why large language models can be ineffective in time series analysis: the impact of modality alignment\.External Links:[Link](https://arxiv.org/abs/2410.12326)Cited by:[§1](https://arxiv.org/html/2606.27881#S1.p1.1)\.