Classifying Unreliable Narrators with Large Language Models

Abhisheik Sharma; Anneliese Brei; Katharine Henry; Shashank Srivastava; Snigdha Chaturvedi

REVIEW 4 major objections 4 minor 14 references

Classifying Unreliable Narrators with Large Language Models

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper argues that large language models can begin to classify unreliable narrators across fiction, blogs, Reddit, and reviews, and that the more abstract the cue, the worse they do.

desk verdict First expert-labeled dataset for unreliable narrators, carefully built and honestly evaluated; the paper's own tables undercut one headline claim, but the resource and task definition survive. read the letter →

arxiv 2506.10231 v1 pith:Z444C44E submitted 2025-06-11 cs.CL

Anneliese Brei , Katharine Henry , Abhisheik Sharma , Shashank Srivastava , Snigdha Chaturvedi This is my paper · ORCID

classification cs.CL

keywords unreliablenarratornarratologylargelanguagemodelscurriculumlearningtextclassificationfirst-personnarrativeexpertannotationdomaintransfer

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 introduces TUNA, an expert-annotated collection of first-person narratives from fiction, blogs, Reddit, and hotel reviews, and defines three increasingly abstract classification tasks taken from literary narratology: intra-narrational unreliability (surface verbal tics), inter-narrational unreliability (contradiction by another character or an unchanged unreliable past self), and inter-textual unreliability (matching the narrator to literary tropes such as the naïf, madman, pícaro, or clown). It then evaluates six large language models in zero-shot, few-shot, fine-tuned, and curriculum-learning settings. The central empirical claim is that the task is learnable but far from solved: expert agreement on the labels is substantial (Cohen's kappa 0.71–0.75), the best fine-tuned 8B model reaches macro-F1 around 57 on the surface-cue task across domains, and performance falls to about 19 on trope matching. The paper also argues that learning on fiction snippets transfers to real-world text domains, and that the difficulty of the three tasks forms a clear gradient tied to how abstract the required inference is.

What carries the argument

The machinery is a trio: the narratological taxonomy of Hansen (2007) and Riggan (1978), which supplies the label sets and the definitions annotators apply; the TUNA dataset itself, with expert annotations, resolved disagreements, and written justifications for every label; and the curriculum-learning training setup, which orders fiction training samples by the number of plausible candidate labels an LLM detects, training first on unambiguous samples. The taxonomy defines the task, the dataset provides the ground truth the paper argues is reliable, and the curriculum ordering is the device that most consistently improves small-model performance, yielding the paper's best results on the surface-cue task.

What would settle it

Re-annotate a random sample of TUNA texts with five ordinary readers who are not literature experts, giving them only the definitions from Section 3 and the same within-narrative instruction, and compare their resolved labels with the expert labels; if Cohen's kappa falls below about 0.4, the dataset captures an expert-specific convention rather than a general property of text. A second, complementary observation: a model that scores higher on inter-textual than on intra-narrational unreliability would contradict the paper's difficulty-gradient finding.

Watch

Extended reading notes

Core claim

The paper's central claim is that unreliable narrator identification, borrowed from literary theory, can be operationalized as three separate classification tasks and that large language models can learn them, although none of the tasks is solved. The authors contend that expert annotators with degrees in English literature can reliably apply Hansen's taxonomy to short first-person texts, with Cohen's kappa between 0.71 and 0.75 for the three tasks. On the learning side, they find that fine-tuning and curriculum learning on fiction snippets consistently outperform zero-shot prompting, that an 8B model (Llama3.1-8B) with curriculum learning achieves macro-F1 of about 57 for intra-narrational unreliability averaged over domains, about 34 for inter-narrational, and about 19 for inter-textual, and that these out-of-domain results are broadly comparable to in-domain fiction results. The paper also reports that curriculum learning helps smaller models but not larger ones, that few-shot learning suffices for the 70B model, and that BERT-scale classifiers generalize poorly to out-of-domain texts.

Load-bearing premise

The whole evaluation presupposes that a taxonomy developed for literary fiction transfers validly to short real-world narratives, and that expert readers applying only within-narrative evidence can reliably identify abstract phenomena such as trope matching; if the expert labels reflect shared literary conventions rather than properties ordinary readers perceive, every reported F1 score measures agreement with an artifact of the annotation procedure.

Editorial extensions

If this is right

Expert annotators can apply the three-part taxonomy of unreliability to short real-world first-person texts with substantial agreement, making the task empirically tractable.
Training on fiction snippets with curriculum ordering lets an 8B model beat zero-shot prompting, reaching macro-F1 around 57 on verbal-tics detection over blogs, Reddit, reviews, and fiction.
For every tested model, the three tasks form a difficulty gradient from surface verbal tics to abstract trope matching, implying that model performance tracks the abstractness of the required inference.
Knowledge learned from fiction transfers to real-world domains, with out-of-domain scores comparable to in-domain ones for the same task.
At the largest model scale tested, few-shot learning matches fine-tuning, suggesting that fewer examples suffice as model size grows.

Reading between the lines

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

One could test whether the difficulty gradient is driven by data sparsity rather than inference depth: the inter-textual classes are rare in some domains (for example, only 3 naïf labels in 100 reviews), so a class-balanced or synthetic augmentation experiment would separate data availability from task abstractness.
Because annotations were resolved through discussion among literature experts, the labels likely encode one interpretive community's norms; measuring label stability across lay readers, writers, and other languages would stress-test the construct validity of the dataset.
The paper's default that a narrator is reliable until cues appear suggests a complementary control task: prompting a model to justify why a narrator is reliable could expose systematic over-triggering on hedging language, which the paper's own error analysis already hints at.
The transfer design treats fiction as the source domain; inverting the direction and training on reviews or Reddit to detect fiction tropes would reveal whether the abstract cues are domain-general or literature-bound.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Jotting this down after reading arXiv:2506.10231. The short version: this is a genuinely new resource paper. TUNA is the first labeled dataset for unreliable narrator classification, built with care — ten English-literature experts, Cohen's kappa 0.71–0.75, disagreements resolved, full prompts and checkpoints in the appendix. The task decomposition into intra-, inter-narrational, and inter-textual unreliability comes straight from Hansen and Riggan and is a sensible operationalization. I believe the authors when they say no prior automated work exists; the related work on character understanding is adjacent but not on point. The evaluation is a clean held-out comparison against human labels. That is real and useful.

The soft spots are real but mostly mechanical. First, there is no majority-class baseline anywhere. The label distributions are heavily skewed (e.g., 3 vs 63 in some review classes), and macro-F1 without a majority baseline makes it hard to know whether a score like 19.30 on inter-textual is meaningful. A simple always-majority baseline would settle that. Second, the paper's second takeaway — 'methods using training samples outperform zero-shot' — is contradicted by its own Table 4 for Llama3.3-70B, Mistral-7B, and Phi3-medium, where zero-shot is competitive or better. The claim may hold for Llama3.1-8B, but as written it is overstated. Third, the inter-narrational definition in the prompt appendix adds a condition ('who has demonstrated at least one form of intra-narrational unreliability') that is not in the main-text definition; someone replicating from the appendix will run a slightly different task. Fourth, test classes with 2 to 7 positive examples (Reviews inter-narrational, Reviews inter-textual A and D) make those macro-F1 numbers noisy; the paper does not discuss this.

On the construct validity worry: the stress-test note worries that expert labels may not reflect ordinary readers' perceptions. That is a fair limitation, and the paper does not test it. But I do not think it is fatal. The dataset is explicitly an expert-annotated resource, and the paper's own claims are modest ('potential'). If someone wants to use this for real-world credibility assessment, they would need to validate against non-expert judgments. As a resource paper, it is fine.

Bottom line: the resource and task definition are solid, and the benchmark is useful despite the reporting flaws. I would send this to peer review. The fixes are straightforward: add majority baselines, qualify the training-vs-zero-shot claim, reconcile the inter-narrational definition, and flag the tiny classes.

Who it is for: computational narrative understanding and anyone building datasets for subjective literary constructs. I would cite it if I worked on narrative NLP. Bring to reading group? Maybe.

Referee Report

4 major / 4 minor

Summary. The paper introduces TUNA, a new expert-annotated dataset of first-person narratives drawn from fiction, blog posts, subreddit posts, and hotel reviews, labeled for three types of narrator unreliability (intra-narrational, inter-narrational, inter-textual) borrowed from the narratology of Hansen (2007) and Riggan (1978). It defines binary and multi-class classification tasks and evaluates several open-weight and proprietary LLMs under zero-shot, few-shot, LoRA fine-tuning, and curriculum learning. The core claims are that the dataset provides reliable expert labels (Cohen's kappa 0.71-0.75), that the tasks form a difficulty gradient from verbal tics to abstract trope matching, that training on fiction can transfer to real-world domains with macro-F1 up to about 57, and that fine-tuned models generally outperform zero-shot prompting, with curriculum learning often beating plain fine-tuning.

Significance. If the dataset and task definition are valid, this is a novel and potentially useful resource: it is, to the authors' knowledge, the first dataset and experimental benchmark for automated unreliable-narrator detection, and the annotation protocol is careful, with expert annotators, reported agreement, disagreement resolution, and fully disclosed prompts and code. The main weakness is that the paper's central motivation is about real-world readers and real-world texts, but the labels are produced by English-literature experts applying a literary-fiction taxonomy, with no validation that ordinary readers perceive the annotated phenomena. The reported F1 scores may therefore measure agreement with an expert-convention artifact rather than a property that transfers to naturalistic use. The paper also overstates several experimental conclusions that its own tables do not support.

major comments (4)

[Section 3, Section 3.4, Section 8] Construct validity is load-bearing for the abstract and introduction's real-world framing. Section 3 assumes that 'given a text, it contains explicit or implicit information that can be leveraged to ascertain the narrator's unreliability' and that a narrator is reliable 'until the reader notices' unreliability. Yet the dataset is annotated solely by English-literature experts applying Hansen's and Riggan's literary taxonomy, and Section 8 lists only length, language, gender, and dataset-size limitations, with no check that non-expert or ordinary readers perceive the same unreliability. If expert labels do not generalize to the readers of blogs, reviews, and Reddit posts who motivate the task, every macro-F1 in Tables 3-5 measures agreement with an annotation artifact. Please add a validation study with non-expert readers, or at minimum report agreement between expert labels and an independent set of lay annotations and temper the real-world transfer claims accordingly.
[Section 5.1, Tables 3-4] The blanket claim that 'methods using training samples (i.e., CL, fine-tuning, few-shot) outperform the zero-shot method' is contradicted by the paper's own tables. In Table 3, for Llama3.1-8B the zero-shot inter-narrational Subreddit score is 30.97, higher than CL's 30.91 and fine-tuning's 25.59; the inter-textual Blog score is 24.23 for zero-shot versus 19.58 for CL and 18.99 for fine-tuning. In Table 4, Llama3.3-70B zero-shot intra-narrational (54.20) and inter-textual (28.52) exceed both CL and fine-tuned scores. The claim should be restricted to specific models, tasks, and domains, or replaced by a systematic significance comparison across all conditions.
[Section 5.1, Table 3] The claim that curriculum learning outperforms fine-tuning 'for most cases' is not supported by Table 3. For example, inter-narrational Fiction (34.59 CL vs. 34.63 FT), inter-narrational Review (35.29 vs. 36.59), inter-textual Fiction (27.42 vs. 28.59), and inter-textual Review (16.72 vs. 17.54) all show fine-tuning at least as good as CL. Because the table reports row-wise significance comparisons only against columns to the right, it is unclear whether the CL-vs-FT differences are statistically reliable; please report the relevant paired comparisons and adjust the conclusion accordingly.
[Appendix D.3.3 vs. Section 3.2] The evaluation prompt for inter-narrational unreliability adds a condition not present in the task definition: '<B>: Other character contradiction: Another character contradicts narrator who has demonstrated at least one form of intra-narrational unreliability.' Section 3.2 defines other-character contradiction without requiring prior intra-narrational unreliability, and the annotator instructions in Appendix B.1 do not include this requirement either. Since the annotation labels and the LLM prompts may therefore operationalize different definitions, the reported F1 scores could understate or overstate model performance in a way that is not attributable to the task. Please justify this extra condition or remove it, and verify that annotation and evaluation definitions coincide.

minor comments (4)

[Section 4.1] In the task definition for inter-narrational unreliability, '<A>' is described as 'same reliable character over time', but Section 3.2 and Figure 2 call this 'same unreliable narrator over time'; the Section 4.1 wording appears to be a typo and should be corrected.
[Table 9 caption] The caption states 'For inter-tex unreliabilities: ... (C) → none', but the table's inter-textual columns are (A) through (E), and the 'none' class is (E). The caption label should be corrected to avoid ambiguity.
[Section 5.1, takeaway 4] The phrasing 'out-of-domain performances ... are not better but good compared to in-domain performances' is unclear; the intended meaning appears to be 'not worse' or 'comparable', and the sentence should be rewritten to state the direction and magnitude of the comparison precisely.
[Section 3.4] The paper reports Cohen's kappa as a single value per task but does not explain whether this is the average over annotator pairs, a Fleiss-style multirater variant, or a value from a particular pair. Please specify the exact computation method.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the evaluation is a held-out comparison against externally produced expert labels, and the authors' self-citations are contextual rather than load-bearing.

full rationale

The paper's derivation chain is an empirical benchmark construction: it borrows narratological definitions from external sources (Hansen 2007; Riggan 1978), builds an expert-annotated dataset, and evaluates LLMs against held-out human labels. None of the reported F1 scores are derived from model outputs or from the paper's own prior claims; they compare model predictions to independent expert annotations. The curriculum-learning difficulty split is generated by an LLM's trait counts, but the final evaluation uses human gold labels, so the ranking does not force any prediction. Self-citations (Brahman and Chaturvedi 2020; Chaturvedi et al. 2017; Srivastava et al. 2016; Iyyer et al. 2016; Huang et al. 2021; Vijjini et al. 2024) appear only in related-work motivation or as a data source for the AITA corpus, and they do not supply any premise from which the central results follow. The paper's definitions of unreliability are explicit assumptions, not hidden redefinitions of the target result. The limitations section discusses dataset size, language, text length, and gender categories, none of which reveal a circular step. Construct validity of the expert labels is a legitimate external concern, but it is a question of whether the task measures what the motivation claims, not a circularity within the paper's own evaluation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

No new entities, forces, or conserved quantities are postulated. The task labels (verbal tics, same-unreliable-character-over-time, other-character-contradiction, naif, madman, picaro, clown) are borrowed from the cited narratological literature, and the paper adds no invented theoretical objects. The dataset TUNA is a new resource but a collected and annotated corpus rather than a postulated entity. The free-parameter ledger reflects the hand-tuned inference hyperparameters that directly shape the reported F1 numbers.

free parameters (3)

temperature = 0.7
Generation temperature for all LLM experiments, reported as chosen after experimentation with higher and lower values (Appendix D.2). It affects all reported F1 numbers.
top-p = 0.9
Nucleus sampling parameter for all LLM experiments, chosen by the authors after experimentation (Appendix D.2).
LoRA training epochs = 3
Fine-tuning duration with default PEFT configuration for all open-weight models (Appendix D.2); affects the fine-tuned and curriculum learning results.

assumptions (6)

domain assumption A given text contains explicit or implicit information that can be leveraged to ascertain the narrator's unreliability.
Stated at the start of Section 3 as a foundational premise for the annotation and classification tasks.
domain assumption A narrator is reliable until the reader notices explicit or implicit information indicating unreliability.
Stated in Section 3, following Wall (1994); this sets the default label and shapes label distributions.
domain assumption The Hansen (2007) taxonomy and Riggan (1978) trope list are valid operationalizations of narrator unreliability for both fiction and real-world text.
Adopted without empirical validation in Sections 1 and 3; the transfer of literary-theoretic categories to blogs, Reddit, and reviews is the paper's core proposal.
domain assumption The three forms of unreliability can be classified independently even when textual cues overlap.
Section 1 states the forms may overlap but are classified and determined separately; annotators pick one label per form.
standard math Bootstrapping over 5 inference runs estimates F1 variability adequately.
Appendix D.2 resamples 1000 predictions from 5 runs; this captures model stochasticity but not annotator or sample variation.
ad hoc to paper Inter-narrational other-character-contradiction requires the narrator to have demonstrated intra-narrational unreliability.
This condition appears only in the prompt definitions (Appendix D.3.3, inter-narrational B) and not in Sections 3.2, 4.1, or Appendix B.1, creating an internal inconsistency.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Classifying Unreliable Narrators with Large Language Models." pith.science (2026). https://pith.science/paper/Z444C44E

@misc{pith2026250610231,
  author       = {Pith},
  title        = {Pith review of: Classifying Unreliable Narrators with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z444C44E}},
  note         = {Machine review of arXiv:2506.10231}
}

read the original abstract

Often when we interact with a first-person account of events, we consider whether or not the narrator, the primary speaker of the text, is reliable. In this paper, we propose using computational methods to identify unreliable narrators, i.e. those who unintentionally misrepresent information. Borrowing literary theory from narratology to define different types of unreliable narrators based on a variety of textual phenomena, we present TUNa, a human-annotated dataset of narratives from multiple domains, including blog posts, subreddit posts, hotel reviews, and works of literature. We define classification tasks for intra-narrational, inter-narrational, and inter-textual unreliabilities and analyze the performance of popular open-weight and proprietary LLMs for each. We propose learning from literature to perform unreliable narrator classification on real-world text data. To this end, we experiment with few-shot, fine-tuning, and curriculum learning settings. Our results show that this task is very challenging, and there is potential for using LLMs to identify unreliable narrators. We release our expert-annotated dataset and code and invite future research in this area.

Figures

Figures reproduced from arXiv: 2506.10231 by the authors.

**Figure 2.** Distribution of resolved labels. For intra-nar [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Breakdown of correctly predicted (green) vs. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Number of characters vs. number of samples. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 7.** Figure 7: Breakdown of correctly predicted (green) vs. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]

**Figure 8.** Figure 8: Number of characters vs. number of samples (Llama3.1-8B experiments). All Samples (solid black) is the distribution of all narratives with respect to the number of characters. Blue, green, and orange solid lines show correct predictions, and corresponding dashed lines…

**Figure 9.** Figure 9: Number of characters vs. number of samples [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

[1]

In Column C, read the narrative

work page
[2]

In Column D, list any observed verbal tics (this is a space for notes, or entering the corre- sponding letter): (a) Admission of fault/bias (b) Defensive tone (c) Digressions (d) Hedging Language (e) Inconsistencies (f) Selective Memory (g) Statement of Potential Disbelief

work page
[3]

In Column E, select the best option for intra- narrational unreliabilities (choose the corre- sponding letter): (a) Verbal tic(s) present (b) None: intra-narrationally reliable

work page
[4]

In Column F, select the best option for inter- narrational unreliabilities (choose the corre- sponding letter): (a) Same unreliable character over time (b) Other character contradiction (c) None: inter-narrationally reliable

work page
[5]

In Column G, select the best option for inter- textual unreliabilities (choose the correspond- ing letter): (a) Naïf (b) Madman (c) Pícaro (d) Clown (e) None: inter-textually reliable

work page
[6]

This is a space for any additional comments you might have

In Column H, please leave annotator’s notes. This is a space for any additional comments you might have. Feel free to use it or leave it blank. C Annotated Label Statistics We report the distribution of labels chosen by the human annotators for each text domain for intra- narrational, inter-narrational, and inter-textual un- reliabilities in Table 6. D Ex...

work page
[7]

Where did you learn your French anyway?

Evaluating character understanding of large language models via character profiling from fictional works. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8015–8036. A Examples of Unreliable Narrator Types Samples from the training set are given in Sec- tions A.1, A.2, A.3. Samples from the testings sets are...

work page 2024
[419]

Monika Fludernik

IEEE. Monika Fludernik. 2000. Unreliable narration. stu- dien zur theorie und praxis unglaubwürdigen erzäh- lens in der englischsprachigen literatur. Poetica, 32(1/2):251–255. Tommaso Fornaciari, Federico Bianchi, Massimo Poe- sio, Dirk Hovy, et al. 2021. Bertective: Language models and contextual information for deception detection. In Proceedings of the...

work page 2000

Show all 14 references

[2013]

In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 352–361

Learning latent personas of film characters. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 352–361. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Pr...

2009
[2017]

In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31

Unsupervised learning of evolving relation- ships between literary characters. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31. Alex Sebastião Constâncio, Denise Fukumi Tsunoda, Helena de Fátima Nunes Silva, Jocelaine Martins da Silveira, and Debora...

2023
[2020]

Same unreliable character over time

library. API-based GPT models are obtained Intra-nar Inter-nar Inter-tex Corpus (A) (R) (A) (B) (R) (A) (B) (C) (D) (R) Fiction 264 235 48 38 413 49 76 64 75 235 Train/Valid180 193 41 31 301 37 59 49 50 178 Test84 42 7 7 112 12 17 15 25 57 Blog posts 75 31 36 6 64 11 23 10 19 ...

2024
[2021]

let your characters tell their story

“let your characters tell their story”: A dataset for character-centric narrative understanding. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1734–1752. Judee K Burgoon and David B Buller. 1994. Interper- sonal deception: Iii. effects of dece...

2021
[2022]

The remains of the day

Lora: Low-rank adaptation of large language models. ICLR. Tenghao Huang, Faeze Brahman, Vered Shwartz, and Snigdha Chaturvedi. 2021. Uncovering im- plicit gender bias in narratives through common- sense inference. In Findings of the Association for Computational Linguistics: E...

2021
[2024]

To tell the truth: Language of deception and language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (V olume1: Long Papers), pages 8506–8520, Mexico City, Mexico. Associ...

2024

Pith tools

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

[1] [1]

In Column C, read the narrative

work page

[2] [2]

In Column D, list any observed verbal tics (this is a space for notes, or entering the corre- sponding letter): (a) Admission of fault/bias (b) Defensive tone (c) Digressions (d) Hedging Language (e) Inconsistencies (f) Selective Memory (g) Statement of Potential Disbelief

work page

[3] [3]

In Column E, select the best option for intra- narrational unreliabilities (choose the corre- sponding letter): (a) Verbal tic(s) present (b) None: intra-narrationally reliable

work page

[4] [4]

In Column F, select the best option for inter- narrational unreliabilities (choose the corre- sponding letter): (a) Same unreliable character over time (b) Other character contradiction (c) None: inter-narrationally reliable

work page

[5] [5]

In Column G, select the best option for inter- textual unreliabilities (choose the correspond- ing letter): (a) Naïf (b) Madman (c) Pícaro (d) Clown (e) None: inter-textually reliable

work page

[6] [6]

This is a space for any additional comments you might have

In Column H, please leave annotator’s notes. This is a space for any additional comments you might have. Feel free to use it or leave it blank. C Annotated Label Statistics We report the distribution of labels chosen by the human annotators for each text domain for intra- narrational, inter-narrational, and inter-textual un- reliabilities in Table 6. D Ex...

work page

[7] [7]

Where did you learn your French anyway?

Evaluating character understanding of large language models via character profiling from fictional works. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 8015–8036. A Examples of Unreliable Narrator Types Samples from the training set are given in Sec- tions A.1, A.2, A.3. Samples from the testings sets are...

work page 2024

[8] [419]

Monika Fludernik

IEEE. Monika Fludernik. 2000. Unreliable narration. stu- dien zur theorie und praxis unglaubwürdigen erzäh- lens in der englischsprachigen literatur. Poetica, 32(1/2):251–255. Tommaso Fornaciari, Federico Bianchi, Massimo Poe- sio, Dirk Hovy, et al. 2021. Bertective: Language models and contextual information for deception detection. In Proceedings of the...

work page 2000