Expect the Unexpected? Testing the Surprisal of Salient Entities

Amir Zeldes; Jessica Lin

arxiv: 2604.10724 · v1 · submitted 2026-04-12 · 💻 cs.CL

Expect the Unexpected? Testing the Surprisal of Salient Entities

Jessica Lin , Amir Zeldes This is my paper

Pith reviewed 2026-05-10 16:07 UTC · model grok-4.3

classification 💻 cs.CL

keywords entity saliencesurprisaluniform information densitydiscoursegenre variationinformation distributionprompting method

0 comments

The pith

Globally salient entities show higher surprisal but reduce it for surrounding content.

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

This paper tests whether the overall importance of entities in a text influences how information is distributed, as captured by surprisal. It finds that entities marked as globally salient carry more surprise than others, even after accounting for where they appear, how long they are, and how deeply nested. At the same time, bringing these salient entities into the context makes the remaining text easier to predict on average. The pattern holds across many genres but is clearest in texts organized around a single topic and weaker in back-and-forth conversation. The work adds entity salience to the set of pressures that shape even distribution of information in discourse.

Core claim

Globally salient entities exhibit significantly higher surprisal than non-salient ones, even controlling for position, length, and nesting confounds. Moreover, salient entities systematically reduce surprisal for surrounding content when used as prompts, enhancing document-level predictability. This effect varies by genre, appearing strongest in topic-coherent texts and weakest in conversational contexts. Our findings refine the UID competing pressures framework by identifying global entity salience as a mechanism shaping information distribution in discourse.

What carries the argument

Global entity salience, identified through manual annotation of 70K mentions, tested via minimal-pair prompting that measures surprisal differences with and without the salient entity in context.

If this is right

Information density must be measured at both local and discourse-wide scales to capture how readers track important participants.
Salient entities function as anchors that lower the surprise of later material, supporting more efficient overall communication.
Genre differences show that information packaging adapts to the goals of the text, not just sentence-level rules.
Language models may improve document-level coherence by treating salient entities as high-priority context.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same salience-surprisal link could be tested in non-English languages to see whether discourse conventions differ.
Writers might deliberately front-load salient entities in some genres to help readers build a stable mental model early.
Automatic systems for detecting salience could be evaluated by checking whether they also predict where surprisal spikes occur.

Load-bearing premise

The manual labels for which entities count as globally salient are consistent and not shaped by the annotators already knowing the full text content.

What would settle it

Re-annotating salience on the same 70K mentions with new annotators and re-running the controlled surprisal comparisons yields no reliable difference between salient and non-salient entities.

Figures

Figures reproduced from arXiv: 2604.10724 by Amir Zeldes, Jessica Lin.

**Figure 3.** Figure 3: Mean surprisal scores across genres for Exper [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 2.** Figure 2: Decrease in mean surprisal relative to salience [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 4.** Figure 4: Decrease in mean surprisal relative to salience [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Mean surprisal scores across genres for Exper [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Mean surprisal (z-score) for salient vs. non-salient entities across genres. Bars show mean surprisal for [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: Surprisal contour in different genres. Blue lines represent the surprisal contour of the sentence followed [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

**Figure 8.** Figure 8: Change in mean surprisal for salience scores [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗

**Figure 9.** Figure 9: Change in mean surprisal for salience scores [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

**Figure 10.** Figure 10: Change in mean surprisal for salience scores [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗

read the original abstract

Previous work examining the Uniform Information Density (UID) hypothesis has shown that while information as measured by surprisal metrics is distributed more or less evenly across documents overall, local discrepancies can arise due to functional pressures corresponding to syntactic and discourse structural constraints. However, work thus far has largely disregarded the relative salience of discourse participants. We fill this gap by studying how overall salience of entities in discourse relates to surprisal using 70K manually annotated mentions across 16 genres of English and a novel minimal-pair prompting method. Our results show that globally salient entities exhibit significantly higher surprisal than non-salient ones, even controlling for position, length, and nesting confounds. Moreover, salient entities systematically reduce surprisal for surrounding content when used as prompts, enhancing document-level predictability. This effect varies by genre, appearing strongest in topic-coherent texts and weakest in conversational contexts. Our findings refine the UID competing pressures framework by identifying global entity salience as a mechanism shaping information distribution in discourse.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper adds a test of global entity salience to the UID framework using 70k annotations and a minimal-pair prompting trick, showing salient entities are more surprising but boost surrounding predictability, with genre variation.

read the letter

The main thing to know is that salient entities show higher surprisal than non-salient ones after basic controls, and prompting with them lowers surprisal for nearby content, with the pattern clearest in topic-focused genres and weaker in conversation. This is framed as a refinement to the competing pressures account of information density rather than a full overhaul. The work does a few things cleanly. The annotation effort covers 70k mentions across 16 genres, which is a decent sample for discourse work. The minimal-pair prompting setup is a direct way to probe the salience effect without relying solely on observational correlations. They report controls for position, length, and nesting, and the abstract claims the results survive those. That gives the central claim some empirical grounding beyond prior UID papers. The soft spots are mostly in the reporting layer. The abstract gives no numbers on effect sizes, exact statistical tests, or inter-annotator agreement for the salience labels, so it is hard to judge how robust the differences really are or whether annotation noise could be driving part of the signal. The prompting method itself is new enough that any model-specific artifacts would need checking in the full methods section. If those details are handled well in the paper, the concern stays minor; if not, it becomes the main thing to fix. This is aimed at researchers already working on UID, surprisal modeling, or entity salience in computational discourse. A reader who follows that literature will see a clear incremental step and can decide whether the genre split is worth following up. It is not transformative on its own, but the data scale and the prompting angle make it worth a referee's time rather than a desk reject. I would send it out for review with a request for fuller stats and annotation diagnostics.

Referee Report

3 major / 2 minor

Summary. The manuscript investigates the relationship between global entity salience and surprisal in discourse, drawing on 70K manually annotated mentions across 16 English genres and a novel minimal-pair prompting procedure. It claims that salient entities exhibit significantly higher surprisal than non-salient ones after controlling for position, length, and nesting; that salient entities reduce surprisal for surrounding content when used as prompts; and that this effect is strongest in topic-coherent genres and weakest in conversational contexts. The work positions these findings as refining the Uniform Information Density hypothesis by identifying salience as an additional competing pressure on information distribution.

Significance. If the empirical results prove robust, the paper contributes a new mechanism—global entity salience—to the UID competing-pressures framework, supported by large-scale manual annotation and a controlled prompting design. The scale of the annotated data and the attempt to isolate salience via minimal pairs are clear strengths that could inform future discourse modeling and predictability research.

major comments (3)

[Section 3] The annotation protocol (Section 3) does not report inter-annotator agreement, number of annotators per mention, or detailed guidelines for determining global salience; given that the central claims rest on the reliability of the 70K annotations, this omission is load-bearing for interpreting the surprisal differences.
[Section 4] Results (Section 4) report statistical significance but omit effect sizes, exact test statistics, and a full account of how the controls for position, length, and nesting were implemented in the regression models; without these, the strength of the 'significantly higher surprisal' claim cannot be fully evaluated.
[Section 5] The minimal-pair prompting method (Section 5) lacks sufficient specification of pair construction, prompt templates, and checks for model-specific artifacts; this is critical because the claim that salient entities enhance document-level predictability depends on the method cleanly isolating global salience.

minor comments (2)

[Abstract] The abstract uses '70K' while the main text should adopt consistent numeric formatting (e.g., 70,000) for readability.
[Figures] Figure captions and axis labels in the genre-variation plots could more explicitly state the surprisal metric and model used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. The comments identify important gaps in reporting that we agree need to be addressed to strengthen the manuscript. We respond to each major comment below and commit to the corresponding revisions.

read point-by-point responses

Referee: [Section 3] The annotation protocol (Section 3) does not report inter-annotator agreement, number of annotators per mention, or detailed guidelines for determining global salience; given that the central claims rest on the reliability of the 70K annotations, this omission is load-bearing for interpreting the surprisal differences.

Authors: We agree that the current description of the annotation protocol is insufficient. In the revised manuscript we will add a new subsection to Section 3 that (a) reproduces the key portions of the annotation guidelines used to determine global salience (entity persistence, topical centrality, and discourse role), (b) states that three annotators independently labeled each of the 70K mentions with majority vote for final labels, and (c) reports inter-annotator agreement (Fleiss' kappa) computed on a held-out double-annotated subset. These additions will allow readers to evaluate the reliability of the salience distinctions that underpin the surprisal results. revision: yes
Referee: [Section 4] Results (Section 4) report statistical significance but omit effect sizes, exact test statistics, and a full account of how the controls for position, length, and nesting were implemented in the regression models; without these, the strength of the 'significantly higher surprisal' claim cannot be fully evaluated.

Authors: We accept that the statistical reporting in Section 4 is incomplete. The revision will include (i) effect sizes (standardized regression coefficients and partial R^{2} for the salience predictor), (ii) full test statistics (t-values, degrees of freedom, and exact p-values), and (iii) the complete model specification: a linear mixed-effects regression with fixed effects for salience, log(position), log(length), nesting (binary), and random intercepts for document and genre. We will also add a supplementary table with the full coefficient table and model diagnostics. revision: yes
Referee: [Section 5] The minimal-pair prompting method (Section 5) lacks sufficient specification of pair construction, prompt templates, and checks for model-specific artifacts; this is critical because the claim that salient entities enhance document-level predictability depends on the method cleanly isolating global salience.

Authors: We agree that the minimal-pair procedure requires fuller documentation. In the revised Section 5 we will (a) describe the exact criteria and matching procedure used to construct each salient/non-salient pair (frequency, syntactic role, and linear position), (b) reproduce the full prompt templates together with an example pair, and (c) report robustness checks across two additional models plus controls for prompt length and lexical overlap. These details will demonstrate that the observed reduction in surrounding surprisal is attributable to global salience rather than model-specific or surface-form artifacts. revision: yes

Circularity Check

0 steps flagged

Empirical study with no circular derivation

full rationale

The paper's central claims rest on new empirical data from 70K manually annotated entity mentions across 16 genres of English plus a novel minimal-pair prompting method. These support the reported findings on surprisal differences for salient entities and their effects on surrounding content, after controlling for position, length, and nesting. No load-bearing step reduces by construction to self-citations, fitted parameters renamed as predictions, or self-definitional equations; the derivation chain consists of data collection, statistical controls, and genre-specific observations that are independently falsifiable outside the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract provides limited detail on assumptions; relies on standard NLP domain assumptions about surprisal as a valid information measure and the reliability of manual salience annotations, with no free parameters or invented entities explicitly introduced.

axioms (2)

domain assumption Surprisal metrics derived from language models serve as a reliable proxy for information content and predictability in natural discourse.
The study uses surprisal to test UID without additional validation of the metric's alignment with human information processing.
domain assumption Manually annotated entity salience labels accurately capture global importance independent of local context.
The central comparison depends on these annotations being valid and consistent across the 70K mentions.

pith-pipeline@v0.9.0 · 5460 in / 1413 out tokens · 111197 ms · 2026-05-10T16:07:08.779352+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 1 internal anchor

[1]

Jennifer E Arnold. 2010. How speakers refer: The role of accessibility. Language and Linguistics Compass, 4(4):187--203

work page 2010
[2]

Jennifer E Arnold. 2025. Why does recency guide pronoun comprehension? I t’s not just topicality, attention, or predictability. Discourse Processes, pages 1--23

work page 2025
[3]

Vincent Boswijk and Matt Coler. 2020. https://doi.org/doi:10.1515/opli-2020-0042 What is salience? Open Linguistics, 6(1):713--722

work page doi:10.1515/opli-2020-0042 2020
[4]

Wallace Chafe. 1994. Discourse, Consciousness, and Time. The Flow and Displacementof Conscious Experiencein Speaking and Writing. University of Chicago Press, Chicago & London

work page 1994
[5]

Thomas Hikaru Clark, Clara Meister, Tiago Pimentel, Michael Hahn, Ryan Cotterell, Richard Futrell, and Roger Levy. 2023. https://doi.org/10.1162/tacl_a_00589 A cross-linguistic pressure for U niform I nformation D ensity in word order . Transactions of the Association for Computational Linguistics, 11:1048--1065

work page doi:10.1162/tacl_a_00589 2023
[6]

Thomas Hikaru Clark, Ethan Gotlieb Wilcox, Edward Gibson, and Roger P. Levy. 2022. Evidence for availability effects on speaker choice in the R ussian comparative alternation. In Proceedings of the 44th Annual Meeting of the Cognitive Science Society, pages 3044--3050

work page 2022
[7]

Manning, Joakim Nivre, and Daniel Zeman

Marie-Catherine de Marneffe, Christopher D. Manning, Joakim Nivre, and Daniel Zeman. 2021. https://doi.org/10.1162/coli_a_00402 U niversal D ependencies . Computational Linguistics, 47(2):255--308

work page doi:10.1162/coli_a_00402 2021
[8]

Milan Dojchinovski, Dinesh Reddy, Tom \'a s Kliegr, Tom \'a s Vitvar, and Harald Sack. 2016. Crowdsourced corpus with entity salience annotations. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), pages 3307--3311

work page 2016
[9]

Jesse Dunietz and Dan Gillick. 2014. A new entity salience task with millions of training examples. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, volume 2: Short Papers, pages 205--209

work page 2014
[10]

August Fenk and Gertraud Fenk. 1980. Konstanz im K urzzeitgedächtnis -- K onstanz im sprachlichen I nformationsfluß? Zeitschrift für experimentelle und angewandte Psychologie, 27(3):400--414

work page 1980
[11]

Michael Gamon, Tae Yano, Xinying Song, Johnson Apacible, and Patrick Pantel. 2013. Identifying salient entities in web pages. In Proceedings of the 22nd ACM international conference on Information & Knowledge Management, pages 2375--2380

work page 2013
[12]

Matteo Gay, Coleman Haley, Mario Giulianelli, and Edoardo Ponti. 2026. https://doi.org/10.18653/v1/2026.eacl-long.178 Is information density uniform when utterances are grounded on perception and discourse? In Proceedings of the 19th Conference of the E uropean Chapter of the A ssociation for C omputational L inguistics (Volume 1: Long Papers) , pages 382...

work page doi:10.18653/v1/2026.eacl-long.178 2026
[13]

T. Givón. 1983. https://www.jbe-platform.com/content/books/9789027280251 Topic Continuity in Discourse . John Benjamins

work page arXiv 1983
[14]

Grosz, Aravind K

Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. 1995. https://aclanthology.org/J95-2003/ C entering: A framework for modeling the local coherence of discourse . Computational Linguistics, 21(2):203--225

work page 1995
[15]

Grosz and Candace L

Barbara J. Grosz and Candace L. Sidner. 1986. https://aclanthology.org/J86-3001/ Attention, intentions, and the structure of discourse . Computational Linguistics, 12(3):175--204

work page 1986
[16]

Florian Jaeger and Roger Levy

T. Florian Jaeger and Roger Levy. 2006. https://proceedings.neurips.cc/paper_files/paper/2006/file/c6a01432c8138d46ba39957a8250e027-Paper.pdf Speakers optimize information density through syntactic reduction . In Advances in Neural Information Processing Systems, volume 19. MIT Press

work page 2006
[17]

Jessica Lin and Amir Zeldes. 2021. https://doi.org/10.18653/v1/2021.law-1.18 W iki GUM : Exhaustive entity linking for wikification in 12 genres . In Proceedings of the Joint 15th Linguistic Annotation Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 170--175, Punta Cana, Dominican Republic. Association for Computational Linguistics

work page doi:10.18653/v1/2021.law-1.18 2021
[18]

Jessica Lin and Amir Zeldes. 2025. https://doi.org/10.18653/v1/2025.findings-acl.24 GUM - SAGE : A novel dataset and approach for graded entity salience prediction . In Findings of the Association for Computational Linguistics: ACL 2025, pages 438--455, Vienna, Austria. Association for Computational Linguistics

work page doi:10.18653/v1/2025.findings-acl.24 2025
[19]

Yang Janet Liu, Tatsuya Aoyama, Wesley Scivetti, Yilun Zhu, Shabnam Behzad, Lauren Elizabeth Levine, Jessica Lin, Devika Tiwari, and Amir Zeldes. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.684 GDTB : Genre diverse data for E nglish shallow discourse parsing across modalities, text types, and domains . In Proceedings of the 2024 Conference on Empiri...

work page doi:10.18653/v1/2024.emnlp-main.684 2024
[20]

Yang Janet Liu and Amir Zeldes. 2023. https://doi.org/10.18653/v1/2023.findings-acl.593 GUMS um: Multi-genre data and evaluation for E nglish abstractive summarization . In Findings of the Association for Computational Linguistics: ACL 2023, pages 9315--9327, Toronto, Canada. Association for Computational Linguistics

work page doi:10.18653/v1/2023.findings-acl.593 2023
[21]

Clara Meister, Tiago Pimentel, Patrick Haller, Lena J \"a ger, Ryan Cotterell, and Roger Levy. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.74 Revisiting the U niform I nformation D ensity hypothesis . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 963--980, Online and Punta Cana, Dominican Republic. ...

work page doi:10.18653/v1/2021.emnlp-main.74 2021
[22]

Tiago Pimentel, Ryan Cotterell, and Brian Roark. 2021. https://doi.org/10.18653/v1/2021.eacl-main.3 Disambiguatory signals are stronger in word-initial positions . In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 31--41, Online. Association for Computational Linguistics

work page doi:10.18653/v1/2021.eacl-main.3 2021
[23]

Hannah Rohde and Andrew Kehler. 2014. Grammatical and information-structural influences on pronoun production. Language, Cognition and Neuroscience, 29(8):912--927

work page 2014
[24]

Péter Rácz. 2013. https://doi.org/doi:10.1515/9783110305395 Salience in Sociolinguistics. A Quantitative Approach . De Gruyter Mouton, Berlin, Boston

work page doi:10.1515/9783110305395 2013
[25]

Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In NeurIPS EMC 2 Workshop

work page 2019
[26]

Eleftheria Tsipidi, Franz Nowak, Ryan Cotterell, Ethan Wilcox, Mario Giulianelli, and Alex Warstadt. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1047 Surprise! U niform I nformation D ensity isn`t the whole story: Predicting surprisal contours in long-form discourse . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Pro...

work page doi:10.18653/v1/2024.emnlp-main.1047 2024
[27]

Schumacher

Klaus von Heusinger and Petra B. Schumacher. 2019. https://doi.org/10.1016/j.pragma.2019.07.025 Discourse prominence: Definition and application . Journal of Pragmatics, 154:117--127

work page doi:10.1016/j.pragma.2019.07.025 2019
[28]

Chuan Wu, Evangelos Kanoulas, Maarten de Rijke, and Wei Lu. 2020. WN-Salience : A corpus of news articles with entity salience annotations. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 2095--2102

work page 2020
[29]

Alessandra Zarcone, Marten Van Schijndel, Jorrig Vogels, and Vera Demberg. 2016. Salience and attention in surprisal-based accounts of language processing. Frontiers in Psychology, 7(844)

work page 2016
[30]

Amir Zeldes. 2017. The GUM corpus: Creating multilayer resources in the classroom. Language Resources and Evaluation, 51(3):581--612

work page 2017
[31]

Amir Zeldes. 2022. https://doi.org/10.5210/dad.2022.102 Can we fix the scope for coreference? P roblems and solutions for benchmarks beyond OntoNotes . Dialogue & Discourse, 13(1):41--62

work page doi:10.5210/dad.2022.102 2022
[32]

Amir Zeldes, Katherine Conhaim, and Lauren Levine. 2026. https://arxiv.org/abs/2603.27358 Not worth mentioning? A pilot study on salient proposition annotation . ArXiv preprint 2603.27358

work page internal anchor Pith review arXiv 2026
[33]

Yilun Zhu, Sameer Pradhan, and Amir Zeldes. 2021. https://doi.org/10.18653/v1/2021.acl-short.59 O nto GUM : Evaluating contextualized SOTA coreference resolution on 12 more genres . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2...

work page doi:10.18653/v1/2021.acl-short.59 2021
[34]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

work page
[35]

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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

work page

[1] [1]

Jennifer E Arnold. 2010. How speakers refer: The role of accessibility. Language and Linguistics Compass, 4(4):187--203

work page 2010

[2] [2]

Jennifer E Arnold. 2025. Why does recency guide pronoun comprehension? I t’s not just topicality, attention, or predictability. Discourse Processes, pages 1--23

work page 2025

[3] [3]

Vincent Boswijk and Matt Coler. 2020. https://doi.org/doi:10.1515/opli-2020-0042 What is salience? Open Linguistics, 6(1):713--722

work page doi:10.1515/opli-2020-0042 2020

[4] [4]

Wallace Chafe. 1994. Discourse, Consciousness, and Time. The Flow and Displacementof Conscious Experiencein Speaking and Writing. University of Chicago Press, Chicago & London

work page 1994

[5] [5]

Thomas Hikaru Clark, Clara Meister, Tiago Pimentel, Michael Hahn, Ryan Cotterell, Richard Futrell, and Roger Levy. 2023. https://doi.org/10.1162/tacl_a_00589 A cross-linguistic pressure for U niform I nformation D ensity in word order . Transactions of the Association for Computational Linguistics, 11:1048--1065

work page doi:10.1162/tacl_a_00589 2023

[6] [6]

Thomas Hikaru Clark, Ethan Gotlieb Wilcox, Edward Gibson, and Roger P. Levy. 2022. Evidence for availability effects on speaker choice in the R ussian comparative alternation. In Proceedings of the 44th Annual Meeting of the Cognitive Science Society, pages 3044--3050

work page 2022

[7] [7]

Manning, Joakim Nivre, and Daniel Zeman

Marie-Catherine de Marneffe, Christopher D. Manning, Joakim Nivre, and Daniel Zeman. 2021. https://doi.org/10.1162/coli_a_00402 U niversal D ependencies . Computational Linguistics, 47(2):255--308

work page doi:10.1162/coli_a_00402 2021

[8] [8]

Milan Dojchinovski, Dinesh Reddy, Tom \'a s Kliegr, Tom \'a s Vitvar, and Harald Sack. 2016. Crowdsourced corpus with entity salience annotations. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), pages 3307--3311

work page 2016

[9] [9]

Jesse Dunietz and Dan Gillick. 2014. A new entity salience task with millions of training examples. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, volume 2: Short Papers, pages 205--209

work page 2014

[10] [10]

August Fenk and Gertraud Fenk. 1980. Konstanz im K urzzeitgedächtnis -- K onstanz im sprachlichen I nformationsfluß? Zeitschrift für experimentelle und angewandte Psychologie, 27(3):400--414

work page 1980

[11] [11]

Michael Gamon, Tae Yano, Xinying Song, Johnson Apacible, and Patrick Pantel. 2013. Identifying salient entities in web pages. In Proceedings of the 22nd ACM international conference on Information & Knowledge Management, pages 2375--2380

work page 2013

[12] [12]

Matteo Gay, Coleman Haley, Mario Giulianelli, and Edoardo Ponti. 2026. https://doi.org/10.18653/v1/2026.eacl-long.178 Is information density uniform when utterances are grounded on perception and discourse? In Proceedings of the 19th Conference of the E uropean Chapter of the A ssociation for C omputational L inguistics (Volume 1: Long Papers) , pages 382...

work page doi:10.18653/v1/2026.eacl-long.178 2026

[13] [13]

T. Givón. 1983. https://www.jbe-platform.com/content/books/9789027280251 Topic Continuity in Discourse . John Benjamins

work page arXiv 1983

[14] [14]

Grosz, Aravind K

Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. 1995. https://aclanthology.org/J95-2003/ C entering: A framework for modeling the local coherence of discourse . Computational Linguistics, 21(2):203--225

work page 1995

[15] [15]

Grosz and Candace L

Barbara J. Grosz and Candace L. Sidner. 1986. https://aclanthology.org/J86-3001/ Attention, intentions, and the structure of discourse . Computational Linguistics, 12(3):175--204

work page 1986

[16] [16]

Florian Jaeger and Roger Levy

T. Florian Jaeger and Roger Levy. 2006. https://proceedings.neurips.cc/paper_files/paper/2006/file/c6a01432c8138d46ba39957a8250e027-Paper.pdf Speakers optimize information density through syntactic reduction . In Advances in Neural Information Processing Systems, volume 19. MIT Press

work page 2006

[17] [17]

Jessica Lin and Amir Zeldes. 2021. https://doi.org/10.18653/v1/2021.law-1.18 W iki GUM : Exhaustive entity linking for wikification in 12 genres . In Proceedings of the Joint 15th Linguistic Annotation Workshop (LAW) and 3rd Designing Meaning Representations (DMR) Workshop, pages 170--175, Punta Cana, Dominican Republic. Association for Computational Linguistics

work page doi:10.18653/v1/2021.law-1.18 2021

[18] [18]

Jessica Lin and Amir Zeldes. 2025. https://doi.org/10.18653/v1/2025.findings-acl.24 GUM - SAGE : A novel dataset and approach for graded entity salience prediction . In Findings of the Association for Computational Linguistics: ACL 2025, pages 438--455, Vienna, Austria. Association for Computational Linguistics

work page doi:10.18653/v1/2025.findings-acl.24 2025

[19] [19]

Yang Janet Liu, Tatsuya Aoyama, Wesley Scivetti, Yilun Zhu, Shabnam Behzad, Lauren Elizabeth Levine, Jessica Lin, Devika Tiwari, and Amir Zeldes. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.684 GDTB : Genre diverse data for E nglish shallow discourse parsing across modalities, text types, and domains . In Proceedings of the 2024 Conference on Empiri...

work page doi:10.18653/v1/2024.emnlp-main.684 2024

[20] [20]

Yang Janet Liu and Amir Zeldes. 2023. https://doi.org/10.18653/v1/2023.findings-acl.593 GUMS um: Multi-genre data and evaluation for E nglish abstractive summarization . In Findings of the Association for Computational Linguistics: ACL 2023, pages 9315--9327, Toronto, Canada. Association for Computational Linguistics

work page doi:10.18653/v1/2023.findings-acl.593 2023

[21] [21]

Clara Meister, Tiago Pimentel, Patrick Haller, Lena J \"a ger, Ryan Cotterell, and Roger Levy. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.74 Revisiting the U niform I nformation D ensity hypothesis . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 963--980, Online and Punta Cana, Dominican Republic. ...

work page doi:10.18653/v1/2021.emnlp-main.74 2021

[22] [22]

Tiago Pimentel, Ryan Cotterell, and Brian Roark. 2021. https://doi.org/10.18653/v1/2021.eacl-main.3 Disambiguatory signals are stronger in word-initial positions . In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 31--41, Online. Association for Computational Linguistics

work page doi:10.18653/v1/2021.eacl-main.3 2021

[23] [23]

Hannah Rohde and Andrew Kehler. 2014. Grammatical and information-structural influences on pronoun production. Language, Cognition and Neuroscience, 29(8):912--927

work page 2014

[24] [24]

Péter Rácz. 2013. https://doi.org/doi:10.1515/9783110305395 Salience in Sociolinguistics. A Quantitative Approach . De Gruyter Mouton, Berlin, Boston

work page doi:10.1515/9783110305395 2013

[25] [25]

Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In NeurIPS EMC 2 Workshop

work page 2019

[26] [26]

Eleftheria Tsipidi, Franz Nowak, Ryan Cotterell, Ethan Wilcox, Mario Giulianelli, and Alex Warstadt. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1047 Surprise! U niform I nformation D ensity isn`t the whole story: Predicting surprisal contours in long-form discourse . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Pro...

work page doi:10.18653/v1/2024.emnlp-main.1047 2024

[27] [27]

Schumacher

Klaus von Heusinger and Petra B. Schumacher. 2019. https://doi.org/10.1016/j.pragma.2019.07.025 Discourse prominence: Definition and application . Journal of Pragmatics, 154:117--127

work page doi:10.1016/j.pragma.2019.07.025 2019

[28] [28]

Chuan Wu, Evangelos Kanoulas, Maarten de Rijke, and Wei Lu. 2020. WN-Salience : A corpus of news articles with entity salience annotations. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 2095--2102

work page 2020

[29] [29]

Alessandra Zarcone, Marten Van Schijndel, Jorrig Vogels, and Vera Demberg. 2016. Salience and attention in surprisal-based accounts of language processing. Frontiers in Psychology, 7(844)

work page 2016

[30] [30]

Amir Zeldes. 2017. The GUM corpus: Creating multilayer resources in the classroom. Language Resources and Evaluation, 51(3):581--612

work page 2017

[31] [31]

Amir Zeldes. 2022. https://doi.org/10.5210/dad.2022.102 Can we fix the scope for coreference? P roblems and solutions for benchmarks beyond OntoNotes . Dialogue & Discourse, 13(1):41--62

work page doi:10.5210/dad.2022.102 2022

[32] [32]

Amir Zeldes, Katherine Conhaim, and Lauren Levine. 2026. https://arxiv.org/abs/2603.27358 Not worth mentioning? A pilot study on salient proposition annotation . ArXiv preprint 2603.27358

work page internal anchor Pith review arXiv 2026

[33] [33]

Yilun Zhu, Sameer Pradhan, and Amir Zeldes. 2021. https://doi.org/10.18653/v1/2021.acl-short.59 O nto GUM : Evaluating contextualized SOTA coreference resolution on 12 more genres . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2...

work page doi:10.18653/v1/2021.acl-short.59 2021

[34] [34]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

work page

[35] [35]

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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

work page