Chaining Event Spans for Temporal Relation Grounding

Dohyeon Lee; Jongho Kim; Minsoo Kim; Seung-won Hwang

REVIEW 2 major objections 5 minor 26 references

Chaining Event Spans for Temporal Relation Grounding

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Chaining answers to questions about the same event into a predicted timeline grounds temporal relations better than answer-overlap supervision.

desk verdict A genuine new mechanism for temporal reading comprehension with solid empirical support; the main soft spot is a notational ambiguity in Eq. 6, not the idea. read the letter →

arxiv 2506.14213 v1 pith:7NIELHB7 submitted 2025-06-17 cs.CL

Jongho Kim , Dohyeon Lee , Minsoo Kim , Seung-won Hwang This is my paper

classification cs.CL

keywords temporalreadingcomprehensionrelationextractiontimelinereasoningeventspanscross-timeattentionspuriousoverlapTORQUETB-Dense

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that a standard training signal for temporal reading comprehension—using answer overlaps between similar questions as a proxy for semantic similarity—is unreliable because unrelated questions can coincidentally share answers. To fix this, it proposes the Timeline Reasoning Network (TRN), which first answers each question individually, then chains all answers from a group of questions about the same event into a predicted relative timeline of event spans. The chained timeline serves as extra grounding for the per-question answers, allowing the model to ignore spurious overlaps. On TORQUE, TRN achieves state-of-the-art test F1 (76.9), exact-match (52.8), and consistency (38.1), and it improves TB-Dense micro-F1 to 65.8 without any external supervision.

What carries the argument

The central mechanism is the 'cross-time attention' module inside the evidence-chaining transformer layer. It operates data-wise rather than sequence-wise: for each passage-token position k, it lets the representation from each question-passage pair attend to the representations of the same token position in all other pairs in the question group, thereby chaining individual answer evidences into a coherent event time span. The paper also constructs a syntax-aware graph network (dependency trees of the question and passage, connected at root nodes and co-mentioned event words) to supply the per-question evidence, but the chaining step is what carries the inductive reasoning. The two-step loss is the mean of extraction and chaining losses.

What would settle it

Probe the cross-time attention weights on a TORQUE question group whose questions differ in length: the aligned tokens should consistently correspond to the same passage event; if they refer to different events, the timeline chaining is misaligned. Alternatively, re-run TRN with passage-relative positional indices and check whether the reported gains persist.

Watch

Extended reading notes

Core claim

The central discovery is that span-based timeline grounding—obtained by inductively chaining the answers to multiple questions about the same event—yields more reliable temporal understanding than point-wise supervision based on answer overlap. In the TRN two-step process, evidence extraction answers each question using a pre-trained language model and a syntax-aware graph network; evidence chaining then takes the collected answer representations for a question group and exchanges information across questions at matching passage-token positions via a cross-time attention module. The output is a predicted timeline of event spans, which is used to refine every answer in the group. The authors show this mitigates spurious overlaps (dissimilar questions sharing coincidental answers) and leads to state-of-the-art results on both temporal reading comprehension (TORQUE) and temporal relation extraction (TB-Dense).

Load-bearing premise

Cross-time attention assumes that the same token position (k) in every question-passage sequence refers to the same passage word across different questions, even though question lengths differ; if this positional alignment is broken, the chained timeline is meaningless.

Editorial extensions

If this is right

TRN sets a new state of the art on TORQUE test (F1 76.9, EM 52.8, C 38.1), surpassing both answer-overlap-supervised (OTR-QA) and externally-supervised (ECONET, UBA) baselines without external supervision.
The method transfers to the TB-Dense temporal-relation-extraction benchmark, improving micro-F1 to 65.8 from ECONET's 64.8, despite using no external corpus.
Passage-level consistency (Cp) rises from 8.3 with contrastive learning to 11.7 with TRN, indicating better cross-event temporal coherence.
Ablations show the evidence-chaining step (and within it, cross-time attention) is the main driver of the gain; removing it drops F1 from 77.6 to 76.0, while removing only the graph network drops to 76.1.
The improvement over contrastive learning grows with question-group size (from 1.5pp for groups of 1–3 to 3.6pp for groups of 7+), consistent with the timeline being more informative when more questions are chained.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If cross-time attention really chains the timeline by absolute token index, the paper's own description leaves an open question: the k-th token of a question-passage sequence occupies a different absolute position when question lengths differ, so the alignment may pair unrelated tokens; the reader should check whether the reported gains survive a passage-relative positional alignment.
The idea of chaining evidence spans to build a timeline may generalize beyond temporal relations to other multi-question reasoning tasks that require consistent global structure, such as causal or logical reasoning—the paper hints at this in the limitations but does not implement it.
A testable extension would be to apply TRN's cross-time attention to large language models with question groups on other temporal benchmarks, potentially mitigating the known weakness of LLMs in temporal ordering without needing human annotations.
The proposed passage-level consistency metric (Cp) could serve as a model-agnostic diagnostic for temporal coherence of any reading-comprehension system, not just TRN.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read it. The short version: there is a real new mechanism here and the experiments back it up. The cross-time attention that chains question-answer pairs from the same TORQUE question group into a predicted timeline is genuinely different from the contrastive and decomposition baselines. The ablations allocate the gain to that chaining step, and the TB-Dense result suggests the idea transfers beyond reading comprehension. Code is public and the benchmarks are standard and external.

The stress-test worry about Eq. 6 is not the fatal flaw that it might look like. The paper explicitly says 'each passage token pk attends to the same positioned token from related data,' and the answer representations in Section 3.1 are per passage token, so the intended reading is passage-relative indexing. That said, the formula itself never states whether k is passage-relative or absolute in the concatenated [Q,P] input, and the whole mechanism rests on alignment. The notation should be corrected and the code checked; this is a clarity defect, not evidence that the model is doing something broken.

The softer spots are the usual ones. Significance tests are only run against RoBERTa, DeBERTa, and ECONET, not against OTR-QA or UBA. The authors explain that those baselines are not reproducible, which is a fair reason, but it does mean the 'state of the art' claim depends on point estimates. Test-set numbers are best-of-run, consistent with the baselines, though averages would be more informative. The predicted timeline is an intermediate with no direct supervision; the newly proposed passage-level consistency metric is a reasonable proxy, but it is not a direct verification that the model actually builds a coherent timeline.

There is also a small overstatement in the abstract: 'outperforms previous methods' across the board, when the dev-set consistency score is one point behind OTR-QA. The discussion handles this honestly, so it is minor.

Overall: this is a solid contribution to temporal reasoning. A serious referee should see it. The main revision request is to clarify Eq. 6, run the significance tests against the remaining baselines if possible, and report seed-averaged test numbers.

Referee Report

2 major / 5 minor

Summary. The paper proposes the Timeline Reasoning Network (TRN) for temporal reading comprehension and temporal relation extraction. In a first 'evidence extraction' step, a PLM and a syntax-aware graph network produce per-passage-token answer scores for each question. In a second 'evidence chaining' step, a transformer with a cross-time attention module aligns answer representations across the questions in a group and predicts a timeline used to re-ground answers. The authors evaluate on TORQUE and TB-Dense, reporting state-of-the-art TORQUE test F1 of 76.9 and an improved TB-Dense F1 of 65.8, and they introduce a passage-level consistency metric Cp to argue that TRN mitigates spurious answer overlaps.

Significance. If the results are reproducible, TRN is a useful step beyond contrastive answer-overlap methods, offering a way to use the structure of question groups without external temporal supervision. Strengths include evaluation on standard external benchmarks, three-way ablations isolating the chaining step and cross-time attention, a new consistency diagnostic, and public code. The main concerns are the under-specified alignment in the core cross-time attention equation and an unsupported significance claim against the principal baseline; both are local but load-bearing and should be fixed before publication.

major comments (2)

[§3.2, Eq. (6)] The paper must specify whether the token index k in CrossTimeAttention(p_i^k, p_j^k, p_j^k) is relative to the passage (the k-th token of the Pv slice) or absolute in the concatenated [Q, P] sequence. The preceding text says the chaining input is the stacked answer representations \hat A_i obtained from Pv, which suggests passage-relative indexing, but Eq. (6) is written with the generic p^k_i and no accompanying statement. If the implementation uses the raw PLM output positions, then for a question group whose members have different lengths, the same passage word sits at different absolute positions across the l sequences, and the attention would pair unrelated tokens, making the predicted timeline meaningless. Please state the indexing convention explicitly and confirm that the Pv slice is used as the chaining input.
[§4.4, Table 1, and §4.3] The claim that 'TRN significantly beats both DeBERTa-v3-large and OTR-QA' is not backed by the significance test described in §4.3, which was run only against RoBERTa, DeBERTa, and ECONET. Table 1's superscripts (r, d, e) confirm that OTR-QA and UBA are not included, and the text in §4.3 explicitly says these tests were omitted because of reproducibility concerns. Since OTR-QA is the main baseline in the paper, the sentence overstates the result. Either run paired significance tests against OTR-QA (and ideally UBA) or remove the word 'significantly' and rephrase the claim as a numerical improvement. This is load-bearing for the paper's central SOTA claim.

minor comments (5)

[Table 4 caption] The caption reads 'Comparison of CL and TDN on the dev set of TORQUE.' The model is TRN, not TDN; this typo should be corrected.
[§4.3] 'Spacy' should be 'spaCy' when referring to the library.
[§4.3 and §4.4] The test-set reporting procedure uses the best score over runs while the dev set reports the average. This is stated and motivated by the baselines' practices, but it would strengthen the paper to also report the average and standard deviation over seeds on the test set, especially since the paired t-tests are computed over instances.
[Figure 4] The y-axis label 'F1 gap' should specify that the gap is in percentage points, and the caption should explain the binning of question-group sizes more fully.
[§7 Limitation] The first limitation states that the predicted timeline 'still has a chance of error' but does not describe what errors can arise or how they might affect the chained predictions; a short expansion would make the limitation concrete.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core results are external-benchmark F1 scores and the chaining step is a supervised multi-step transformation, not a tautology.

full rationale

TRN's central claims are evaluated on TORQUE and TB-Dense, whose labels are not produced by the authors; the reported F1/EM/C improvements are external empirical results. The evidence chaining step (Eq. 6) takes the first-step answer representations and produces new logits trained with Lchain on the same answer labels; this is a standard multi-step model, not a definitional equivalence. The 'predicted timeline' is an unsupervised intermediate, so attributing gains to timeline reasoning is an interpretation rather than a circular derivation. The new Cp metric is author-defined and computed from the model's own predictions, which makes it weaker evidence for the mechanism claim, but it does not feed into the main F1 comparison or reduce any prediction to its input. No load-bearing self-citation or imported uniqueness theorem appears. A technical ambiguity remains about whether token index k in Eq. 6 is passage-relative, but that is a specification/correctness issue, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method does not introduce physical entities or fitted constants; it relies on standard attention and a set of hyperparameters. The key assumptions are about the interpretability of answer predictions as timeline evidence and the alignment of cross-time attention.

free parameters (4)

Graph iteration count T = 2 or 3
Hyperparameter searched on the dev set; affects the graph network depth but not the conceptual mechanism.
Chaining layer count T' = 2 or 3
Hyperparameter searched on the dev set for the evidence chaining transformer.
Hidden size of chaining transformer = 1024
Set for the cross-time attention module; standard model size choice.
Cp threshold = 80% F1
Chosen following the convention from Gardner et al. (2020) and Ning et al. (2020); not central to the main results.

assumptions (4)

domain assumption Answers to questions about the same event provide evidence for a relative timeline of events.
Section 3.2 states 'each prediction can be interpreted as temporal evidence like when one event occurred relative to the asked event.' This is the core modeling assumption of the chaining step.
domain assumption Cross-time attention aligns passage tokens across different question-passage sequences by token index.
Equation (6) defines CrossTimeAttention = Attention(p_i^k, p_j^k, p_j^k) with k as token index, implicitly assuming the same passage token occupies the same position across sequences of different question lengths.
domain assumption Syntactic dependency graphs improve temporal reasoning.
Section 3.1 builds syntax-aware graph networks, citing prior TRE work; this is a design choice inherited from the literature.
standard math Standard transformer attention and backpropagation are valid.
The model relies on the transformer architecture (Vaswani et al., 2017) and standard neural network training.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Chaining Event Spans for Temporal Relation Grounding." pith.science (2026). https://pith.science/paper/7NIELHB7

@misc{pith2026250614213,
  author       = {Pith},
  title        = {Pith review of: Chaining Event Spans for Temporal Relation Grounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NIELHB7}},
  note         = {Machine review of arXiv:2506.14213}
}

read the original abstract

Accurately understanding temporal relations between events is a critical building block of diverse tasks, such as temporal reading comprehension (TRC) and relation extraction (TRE). For example in TRC, we need to understand the temporal semantic differences between the following two questions that are lexically near-identical: "What finished right before the decision?" or "What finished right after the decision?". To discern the two questions, existing solutions have relied on answer overlaps as a proxy label to contrast similar and dissimilar questions. However, we claim that answer overlap can lead to unreliable results, due to spurious overlaps of two dissimilar questions with coincidentally identical answers. To address the issue, we propose a novel approach that elicits proper reasoning behaviors through a module for predicting time spans of events. We introduce the Timeline Reasoning Network (TRN) operating in a two-step inductive reasoning process: In the first step model initially answers each question with semantic and syntactic information. The next step chains multiple questions on the same event to predict a timeline, which is then used to ground the answers. Results on the TORQUE and TB-dense, TRC and TRE tasks respectively, demonstrate that TRN outperforms previous methods by effectively resolving the spurious overlaps using the predicted timeline.

Figures

Figures reproduced from arXiv: 2506.14213 by the authors.

**Figure 2.** The illustration of (a) point-wise timeline grounding and (b) span-based one. (a) The model brings similar [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Overview of TRN. (a) The Evidence Extraction Step answers each question with semantic (PLM) and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Plot of the relationship between the question [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Qualitative analysis of contrastive learning and TRN. Events in the passage are highlighted in bold. In [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 14 canonical work pages

[1]

Taylor Cassidy, Bill McDowell, Nathanael Chambers, and Steven Bethard. 2014. An annotation framework for dense event ordering. In Annual Meeting of the Association for Computational Linguistics

work page 2014
[2]

Chunkit Chan, Jiayang Cheng, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2023. http://arxiv.org/abs/2304.14827 Chatgpt evaluation on sentence level relations: A focus on temporal, causal, and discourse relations

arXiv 2023
[3]

Kunlong Chen, Weidi Xu, Xingyi Cheng, Zou Xiaochuan, Yuyu Zhang, Le Song, Taifeng Wang, Yuan Qi, and Wei Chu. 2020. Question directed graph attention network for numerical reasoning over text. In Conference on Empirical Methods in Natural Language Processing

work page 2020
[4]

Wenhu Chen, Xinyi Wang, and William Yang Wang. 2021. A dataset for answering time-sensitive questions. arXiv preprint arXiv:2108.06314

arXiv 2021
[5]

Fei Cheng and Yusuke Miyao. 2017. Classifying temporal relations by bidirectional lstm over dependency paths. In Annual Meeting of the Association for Computational Linguistics

work page 2017
[6]

Tianqing Fang, Zhaowei Wang, Wenxuan Zhou, Hongming Zhang, Yangqiu Song, and Muhao Chen. 2023. http://arxiv.org/abs/2305.14970 Getting sick after seeing a doctor? diagnosing and mitigating knowledge conflicts in event temporal reasoning

work page Pith review arXiv 2023
[7]

Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, et al. 2020. Evaluating models’ local decision boundaries via contrast sets. Findings of Empirical Methods in Natural Language Processing

work page 2020
[8]

Rujun Han, Xiang Ren, and Nanyun Peng. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.436 ECONET : Effective continual pretraining of language models for event temporal reasoning . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5367--5380, Online and Punta Cana, Dominican Republic. Association for Compu...

work page doi:10.18653/v1/2021.emnlp-main.436 2021

Show all 26 references

[9]

Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543

2021 arXiv
[10]

Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2022. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. In The Eleventh International Conference on Learning Representations

2022
[11]

Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. https://doi.org/10.5281/zenodo.1212303 spacy: Industrial-strength natural language processing in python

2020 doi
[12]

Hao Huang, Xiubo Geng, Guodong Long, and Daxin Jiang. 2022. https://doi.org/10.18653/v1/2022.naacl-main.28 Understand before answer: Improve temporal reading comprehension via precise question understanding . In Proceedings of the 2022 Conference of the North American Chapter ...

2022 doi
[13]

Yinya Huang, Meng Fang, Yu Cao, Liwei Wang, and Xiaodan Liang. 2021. Dagn: Discourse-aware graph network for logical reasoning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pa...

2021
[14]

Thomas Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. ArXiv, abs/1609.02907

2016 arXiv
[15]

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

2019 arXiv
[16]

Puneet Mathur, Rajiv Jain, Franck Dernoncourt, Vlad Morariu, Quan Hung Tran, and Dinesh Manocha. 2021. https://doi.org/10.18653/v1/2021.acl-short.67 TIMERS : Document-level temporal relation extraction . In Proceedings of the 59th Annual Meeting of the Association for Computat...

2021 doi
[17]

Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.88 TORQUE : A reading comprehension dataset of temporal ordering questions . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language...

2020 doi
[18]

Qiang Ning, Hao Wu, and Dan Roth. 2018. A multi-axis annotation scheme for event temporal relations. In Annual Meeting of the Association for Computational Linguistics

2018
[19]

Qiu Ran, Yankai Lin, Peng Li, Jie Zhou, and Zhiyuan Liu. 2019. N um N et: Machine reading comprehension with numerical reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan...

2019
[20]

Chao Shang, Peng Qi, Guangtao Wang, Jing Huang, Youzheng Wu, and Bowen Zhou. 2021. Open temporal relation extraction for question answering. In Conference on Automated Knowledge Base Construction

2021
[21]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

2017
[22]

Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio’, and Yoshua Bengio. 2017. Graph attention networks. ArXiv, abs/1710.10903

2017 arXiv
[23]

Shuaicheng Zhang, Qiang Ning, and Lifu Huang. 2022. https://doi.org/10.18653/v1/2022.findings-naacl.29 Extracting temporal event relation with syntax-guided graph transformer . In Findings of the Association for Computational Linguistics: NAACL 2022, pages 379--390, Seattle, U...

2022 doi
[24]

going on a vacation

Ben Zhou, Daniel Khashabi, Qiang Ning, and Dan Roth. 2019. “going on a vacation” takes longer than “going for a walk”: A study of temporal commonsense understanding. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Internati...

2019
[25]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...
[26]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Taylor Cassidy, Bill McDowell, Nathanael Chambers, and Steven Bethard. 2014. An annotation framework for dense event ordering. In Annual Meeting of the Association for Computational Linguistics

work page 2014

[2] [2]

Chunkit Chan, Jiayang Cheng, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2023. http://arxiv.org/abs/2304.14827 Chatgpt evaluation on sentence level relations: A focus on temporal, causal, and discourse relations

arXiv 2023

[3] [3]

Kunlong Chen, Weidi Xu, Xingyi Cheng, Zou Xiaochuan, Yuyu Zhang, Le Song, Taifeng Wang, Yuan Qi, and Wei Chu. 2020. Question directed graph attention network for numerical reasoning over text. In Conference on Empirical Methods in Natural Language Processing

work page 2020

[4] [4]

Wenhu Chen, Xinyi Wang, and William Yang Wang. 2021. A dataset for answering time-sensitive questions. arXiv preprint arXiv:2108.06314

arXiv 2021

[5] [5]

Fei Cheng and Yusuke Miyao. 2017. Classifying temporal relations by bidirectional lstm over dependency paths. In Annual Meeting of the Association for Computational Linguistics

work page 2017

[6] [6]

Tianqing Fang, Zhaowei Wang, Wenxuan Zhou, Hongming Zhang, Yangqiu Song, and Muhao Chen. 2023. http://arxiv.org/abs/2305.14970 Getting sick after seeing a doctor? diagnosing and mitigating knowledge conflicts in event temporal reasoning

work page Pith review arXiv 2023

[7] [7]

Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, et al. 2020. Evaluating models’ local decision boundaries via contrast sets. Findings of Empirical Methods in Natural Language Processing

work page 2020

[8] [8]

Rujun Han, Xiang Ren, and Nanyun Peng. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.436 ECONET : Effective continual pretraining of language models for event temporal reasoning . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5367--5380, Online and Punta Cana, Dominican Republic. Association for Compu...

work page doi:10.18653/v1/2021.emnlp-main.436 2021