Pith. sign in

REVIEW 3 major objections 4 minor 47 references

This paper introduces experiential intertextuality detection and shows that no annotation-free method exceeds r=0.38 against expert judgments, with a supervised hybrid reaching r=0.454.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 12:00 UTC pith:KEE3VUHK

load-bearing objection Solid new benchmark for experiential intertextuality, but the noise ceiling is overstated and the data is private. the 3 major comments →

arxiv 2607.29188 v1 pith:KEE3VUHK submitted 2026-07-31 cs.CL

Detecting Experiential Intertextuality Across Migration Routes: Beyond Surface Similarity in French Narratives

Sakayo Toadoum Sari , Nelly Robin , Michelle Auzanneau , Lakhdar Sais , Veronique Petit , Marie Veniard , Said Jabbour , Fabien Delorme This is my paper
classification cs.CL
keywords experiential intertextualitymigration narrativescross-route echo detectionannotation-free NLPzero-shot LLM scoringsentence-pair validationnoise ceilingFrench-language corpus
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces a new NLP task: detecting experiential intertextuality, the parallel lived experiences that migrants on different routes recount without sharing vocabulary. On 108 French migration narratives from the Trans-Saharan and Balkan corridors, the authors generate sentence pairs and score them with 16 annotation-free methods, validating against 816 expert judgments. They find that no surface, structural, or embedding method exceeds r=0.30, while a zero-shot 7B language model reaches r=0.375—about 57% of the estimated noise ceiling. A supervised hybrid of 31 features reaches r=0.454 and reveals that journey-phase position and themes such as life-threatening danger and labor exploitation are the strongest cross-route echo predictors. If right, this shows the task is real but hard, and that LLM-based scoring captures experiential signal that surface similarity misses.

Core claim

On its own terms, the paper establishes that experiential intertextuality—shared lived-experience echoes across geographically distinct migration narratives—is detectable by annotation-free scoring, but only weakly. The best single method, Qwen2.5-7B with a zero-shot prompt, correlates at r=0.375 with expert judgments (56.6% of the r=0.662 noise ceiling); a supervised ridge hybrid over all 31 features reaches r=0.454 (68.6% of ceiling). The paper further shows that matching journey phases, especially departure, yield the strongest echoes, that life-threatening danger and labor exploitation are the most universal cross-route themes, and that LLM scores largely subsume sentence-embedding signa

What carries the argument

The load-bearing mechanism is a two-stage evaluation scaffold. First, sentence pairs are automatically generated from 108 narratives across two corridors and assigned theme labels and positional context; second, 31 annotation-free scores from five families—lexical, POS-structural, context-aware, sentence embeddings, and LLM prompts—are correlated with 816 expert intertextuality ratings. The 'experiential intertextuality' construct itself, defined through thematic, functional, and pragmatic parallelism, is the central object; the noise ceiling (Spearman-Brown-derived r_max=0.662) is the yardstick that lets weak correlations be interpreted.

Load-bearing premise

The load-bearing premise is that the two expert raters are parallel measures of one true 'experiential intertextuality' construct, so the Spearman-Brown ceiling of 0.662 fairly represents what a perfect method could achieve; if their disagreements reflect multiple valid notions of shared experience, the ceiling is overstated.

What would settle it

Re-annotate a random subset of 200 pairs with three or more migration experts and recompute the noise ceiling; if the new ceiling falls below 0.454, the hybrid's claim to be at 68.6% of the attainable maximum collapses. Alternatively, run the two 7B models at full precision (no 4-bit quantization) or with a 70B model: if zero-shot Qwen's advantage over Mistral disappears, the reported model gap is an artifact of quantization or scale.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, automatic tools can flag candidate cross-route experiential echoes for social scientists, even though a fully automatic high-precision detector is not yet available.
  • Zero-shot LLM scoring, rather than embeddings or surface overlap, is the most promising annotation-free direction, and the gap between the two 7B models indicates model choice matters more than prompt crafting.
  • Departure-phase pairs and themes of mortal danger and labor exploitation are the most robust cross-corridor universals, giving HSS researchers concrete starting points for qualitative analysis.
  • The redundancy of embeddings relative to LLM scores suggests future pipelines can drop embedding features and focus on richer event-structure representations.
  • The weak POS results and the hard 'pragmatic intent' cases point to speech-act and event-role representations as the next bottleneck.

Where Pith is reading between the lines

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

  • If the noise ceiling is as low as estimated, any future method that exceeds r=0.454 is already near the practical ceiling; the more interesting test is whether re-annotation with more experts raises or lowers that ceiling.
  • The paper's own flagged limitations—99 versus 9 narrative imbalance and dual-annotator alpha of 0.27—mean the method ranking is provisional; a balanced corpus or triple annotation could reshuffle the order of methods near the top.
  • The same annotation-free scaffold could be extended to other languages or to non-migration trauma narratives; the theme lexicon would need to be rebuilt, but the LLM-scoring and hybrid-validation machinery transfers directly.
  • The Qwen few-shot degradation suggests an anchoring effect; testing different demonstration sets or a single high-scoring example would clarify whether the degradation is calibration or prompt brittleness.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces experiential intertextuality detection, a new NLP task aimed at identifying shared lived experiences across migration narratives without annotated training data. From 108 French narratives, the authors automatically generate sentence pairs and score them with 16 annotation-free methods (lexical, POS, contextual, embedding, and LLM-based). They validate against 816 expert judgments, reporting that the best single method (Qwen2.5-7B zero-shot) attains r=0.375, a supervised hybrid reaches r=0.454, and journey phase predicts intertextuality. They estimate a noise ceiling of r_max=0.662 from inter-annotator correlation and claim that the methods achieve 56.6% and 68.6% of this ceiling, with 31.4% of the hybrid's gap to perfection being irreducible annotator noise.

Significance. The paper proposes a novel and societally relevant task, with a careful annotation protocol, a wide range of baselines, and a transparent evaluation including partial correlations and an ablation study. The finding that an LLM captures beyond-surface experiential similarity is interesting, and the public release of code supports reproducibility. However, the headline percentages depend on a miscalculated noise ceiling, and the significance tests ignore the clustered structure of the data. With these corrected, the qualitative ranking likely persists, but the interpretation as 'feasible but weak' becomes 'close to the achievable ceiling,' and some significance claims may not survive.

major comments (3)
  1. [Section 3 / Table 2] The noise ceiling is computed as r_max=0.662 from the Spearman-Brown formula for the 283 dual-annotated pairs, but this ceiling is then applied to all 816 validation pairs, of which 533 are single-annotated. For single-annotated pairs, the maximum correlation of a perfect method with the observed score is sqrt(r12)=0.530, not 0.662. Under classical test theory, the effective ceiling for the full sample is approximately 0.567. Thus the reported percentages (56.6% for Qwen, 68.6% for hybrid) are understated; the corrected values are about 66% and 80%. Consequently, the Limitations sentence '31.4% of the hybrid's gap to perfection is attributable to irreducible annotator noise' is also wrong: the 31.4% likely refers to (r_max-r_hybrid)/r_max, which is model shortfall, not noise. This does not change the r≤0.38 finding, but it materially changes the paper's interpretation of the results as l
  2. [Section 5 / Table 3] The significance tests treat the 816 validation pairs as independent. However, 815 of 816 pairs are cross-route with one sentence drawn from only 9 Balkan narratives; pairs sharing the same Balkan narrative are correlated. Fisher z CIs and Williams tests are therefore anti-conservative. For example, the claim that Qwen2.5 zero-shot significantly outperforms the theme lexicon (t=3.03, p=0.003) and the significance of the position-similarity feature (r=0.131, p<0.001) may not survive clustering. Please use cluster-robust correlation inference (e.g., cluster bootstrap by Balkan narrative) or a mixed-effects model that accounts for random effects of both source narratives.
  3. [Section 4.5 / Limitations] The paper's central claim is that the methods require 'no annotated training data.' Yet the few-shot variant uses three expert-annotated pairs as demonstrations, and the theme lexicon is curated from expert domain knowledge. While these are disclosed, the abstract and title's 'annotation-free' characterization is too strong. Consider qualifying the claim (e.g., 'no supervised fine-tuning') and distinguishing these supervision-light variants from zero-shot.
minor comments (4)
  1. [Section 4.6] The paper lists 31 features but the enumerated groups sum to 30. Please clarify the composition (e.g., whether 'pos_combined' is additional).
  2. [Abstract / Table 3] The abstract reports r=0.38 for the best single method, while Table 3 gives 0.375. Use consistent precision.
  3. [Figure 2] The y-axis starts at 0, making the Mistral few-shot improvement appear larger; consider starting at 0.1 for clarity.
  4. [Section 5] The Williams test is referenced but not defined. Please provide the formula or a citation.

Circularity Check

0 steps flagged

No significant circularity: annotation-free scorers are fixed functions validated against independent expert labels; the only supervised model is explicitly labeled an upper bound.

full rationale

The derivation chain is non-circular. Pair generation from raw narratives is automatic and label-free. Each annotation-free scorer in Sections 4.1-4.5 is a fixed function of sentence content or context; no equation is defined in terms of the expert validation scores, and no fitted parameter is renamed as a prediction. The hybrid model (Eq. 6) is the only method trained on expert labels and is explicitly presented as a supervised upper bound, so its r=0.454 does not masquerade as annotation-free performance. The theme lexicon has a dual role as a feature and for pair labeling, but the paper discloses this and provides an empirical check: the number of theme labels per pair is negatively correlated with expert scores (r=-0.148), which rules out the concern that thematic stratification forces high-scoring pairs. Few-shot LLM prompting uses three expert demonstrations, but the paper discloses this, and the central 'no annotation-free method exceeds r=0.38' claim is carried by zero-shot Qwen (r=0.375). The author self-citations (Robin 2014; Ing et al. 2025) motivate the corpus and the HSS phenomenon; they do not supply a mathematical constraint used to derive any score, so they are not load-bearing. The noise-ceiling calculation in Section 3 is a reliability correction based on annotator agreement, not a reduction of the methods to their inputs; whether it is correctly applied to single-annotated pairs is a statistical validity concern, not circularity. No step in the paper's own equations reduces to its inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The paper introduces a new construct and a hand-built lexicon, but the main free parameters are not fitted to the target (expert scores). The reliance on expert judgments as ground truth is the central epistemic load: all correlations are conditional on the reliability of those judgments, which is low.

free parameters (4)
  • Theme lexicon (130 terms, 8 categories)
    Curated from the migration narratives themselves; used in Eq. 1 to score thematic co-activation. The selection of which terms/categories are discriminative is a hand choice, not a fitted number.
  • 15 thematic categories for pair labeling
    Expert-defined labels used to stratify pair generation and annotation; related to the lexicon's 8 subsets. They shape what counts as a theme but were not fitted to expert scores.
  • Journey-phase quartile mapping = 4 quartiles
    The mapping phi: [0,1] -> {1,2,3,4} in Eq. 5 is an arbitrary discretization of narrative position; results depend on this choice.
  • Few-shot demonstrations (3 examples) = low/med/high scores
    Three expert-annotated pairs selected by hand to cover low, medium, and high scores; this selection affects Mistral's marked improvement in few-shot.
axioms (4)
  • domain assumption Expert intertextuality ratings are a valid operationalization of the construct.
    The entire evaluation treats the mean of two expert scores as ground truth; but alpha=0.27 indicates low agreement. If the construct is not reliably measured, all correlations are attenuated.
  • standard math Spearman-Brown prophecy formula applies to the two annotators.
    The noise ceiling r_max=sqrt(rho) assumes the two expert scores are parallel measures and the average is more reliable; this is standard but depends on assumptions about error structure.
  • domain assumption Sentence embeddings and LLM outputs in French are meaningful for migration narratives.
    The methods assume pretrained models capture semantic and experiential similarity in non-native, oral French; no external validation on migration text.
  • domain assumption spaCy's French dependency parser works adequately on non-native French.
    POS and dependency features rely on fr_core_news_sm; the authors acknowledge migrant French is highly variable, which they cite as a reason POS features underperform.
invented entities (1)
  • Experiential intertextuality no independent evidence
    purpose: The target construct: shared lived experience across narratives, distinct from textual similarity or event identity.
    Defined and operationalized only through the authors' expert ratings and examples; no external validation or previously established measure exists.

pith-pipeline@v1.3.0-daily-deepseek · 12753 in / 8659 out tokens · 90020 ms · 2026-08-03T12:00:58.886388+00:00 · methodology

0 comments
read the original abstract

Migrants traversing geographically distinct routes such as the Trans-Saharan and Balkan corridors often recount strikingly parallel lived experiences: police violence, smuggler exploitation, dangerous crossings, and family separation. We introduce the task of experiential intertextuality detection: automatically identifying shared experiential echoes across migration narratives without requiring annotated training data. From 108 French migration narratives spanning both corridors, we automatically generate sentence pairs and score them using annotation-free methods: lexical baselines, sentence embeddings, POS-based structural features, a migration-specific theme lexicon, context-aware narrative features, and zero-shot LLM scoring with Qwen2.5-7B and Mistral-7B under three prompting strategies. We validate all methods against 816 expert-annotated intertextuality judgments (inter-annotator Krippendorff's $\alpha = 0.27$). Our results reveal that all surface, structural, and embedding methods correlate only weakly with expert judgments ($r \leq 0.30$); Qwen2.5-7B zero-shot achieves the best single-method correlation ($r = 0.38$); few-shot examples degrade Qwen but dramatically improve Mistral; narrative position significantly predicts intertextuality, with departure-phase pairs showing the highest experiential echoes; and a supervised hybrid combining all 31 features achieves $r = 0.45$, a 21% improvement over the best individual method.

Figures

Figures reproduced from arXiv: 2607.29188 by Fabien Delorme, Lakhdar Sais, Marie Veniard, Michelle Auzanneau, Nelly Robin, Said Jabbour, Sakayo Toadoum Sari, Veronique Petit.

Figure 1
Figure 1. Figure 1: Pipeline overview. Left: sentence pairs are automatically generated from raw narratives with context localization. Middle: five annotation-free method families produce 31 scores per pair. Right: scores validated against expert judgments; supervised hybrid as upper-bound. ⋆ = novel. detection, argumentation mining, and pragmatic interpretation (Gilardi et al., 2023). Qwen Team (2024) show strong multilingua… view at source ↗
Figure 3
Figure 3. Figure 3: Mean expert intertextuality by journey-phase [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Prompting strategy comparison. Few-shot degrades Qwen but dramatically improves Mistral. This asymmetry suggests that Qwen has stronger zero-shot French comprehension and migration-relevant world knowledge, while Mis￾tral requires calibration to understand the task. The few-shot degradation for Qwen may reflect score anchoring: the three demonstration exam￾ples bias the model toward their score distribu￾ti… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 2 linked inside Pith

  1. [1]

    Telematics and Informatics , volume=

    Sentiment analysis on. Telematics and Informatics , volume=. 2018 , publisher=

    2018

  2. [2]

    Analyzing Shift in Narratives Regarding Migrants in

    Hussain, Muhammad Nihal and Bandeli, Kevin K and Al-khateeb, Samer and Agarwal, Nitin , booktitle=. Analyzing Shift in Narratives Regarding Migrants in. 2018 , publisher=

    2018

  3. [3]

    Proceedings of ECML-PKDD , year=

    Text Mining from Migration Narratives , author=. Proceedings of ECML-PKDD , year=

  4. [4]

    Proceedings of SemEval , pages=

    Cer, Daniel and Diab, Mona and Agirre, Eneko and Lopez-Gazpio, I. Proceedings of SemEval , pages=. 2017 , publisher=

    2017

  5. [5]

    2016 , publisher=

    Agirre, Eneko and Banea, Carmen and Cer, Daniel and Diab, Mona and Gonzalez-Agirre, Aitor and Mihalcea, Rada and Rigau, German and Wiebe, Janyce , booktitle=. 2016 , publisher=

    2016

  6. [6]

    Proceedings of ACL , pages=

    Martin, Louis and Muller, Benjamin and Su. Proceedings of ACL , pages=. 2020 , publisher=

    2020

  7. [7]

    Kristeva, Julia , year=. S

  8. [8]

    La fabrique du parcours migratoire sur la route des

    Bacon, Lucie , year=. La fabrique du parcours migratoire sur la route des

  9. [9]

    Migrations, observatoire et droit: Complexit

    Robin, Nelly , booktitle=. Migrations, observatoire et droit: Complexit. 2014 , publisher=

    2014

  10. [10]

    EPJ Data Science , volume=

    The emotional arcs of stories are dominated by six basic shapes , author=. EPJ Data Science , volume=. 2016 , publisher=

    2016

  11. [11]

    Proceedings of ACL , pages=

    Unsupervised Learning of Narrative Event Chains , author=. Proceedings of ACL , pages=. 2008 , publisher=

    2008

  12. [12]

    Proceedings of NAACL-HLT , pages=

    A Corpus and Cloze Evaluation for Deeper Understanding of Commonsense Stories , author=. Proceedings of NAACL-HLT , pages=. 2016 , publisher=

    2016

  13. [13]

    Proceedings of the National Academy of Sciences , volume=

    Gilardi, Fabrizio and Alizadeh, Meysam and Kubli, Ma. Proceedings of the National Academy of Sciences , volume=

  14. [14]

    Sentence-

    Reimers, Nils and Gurevych, Iryna , booktitle=. Sentence-. 2019 , publisher=

    2019

  15. [15]

    Proceedings of ACL , pages=

    Unsupervised Cross-lingual Representation Learning at Scale , author=. Proceedings of ACL , pages=. 2020 , publisher=

    2020

  16. [16]

    Text , volume=

    Rhetorical Structure Theory: Toward a functional theory of text organization , author=. Text , volume=

  17. [17]

    Prasad, Rashmi and Dinesh, Nikhil and Lee, Alan and Miltsakaki, Eleni and Robaldo, Livio and Joshi, Aravind and Webber, Bonnie , booktitle=. The

  18. [18]

    Proceedings of ACL , pages=

    Learning Latent Personas of Film Characters , author=. Proceedings of ACL , pages=. 2013 , publisher=

    2013

  19. [19]

    The Event

    Caselli, Tommaso and Vossen, Piek , booktitle=. The Event. 2017 , publisher=

    2017

  20. [20]

    Proceedings of LREC , pages=

    Using a Sledgehammer to Crack a Nut? Lexical Diversity and Event Coreference Resolution , author=. Proceedings of LREC , pages=

  21. [21]

    Digital Scholarship in the Humanities , volume=

    Modeling the Scholars: Detecting Intertextuality through Enhanced Word-Level N-Gram Matching , author=. Digital Scholarship in the Humanities , volume=. 2015 , publisher=

    2015

  22. [23]

    Mistral 7

    Jiang, Albert Q and Sablayrolles, Alexandre and Mensch, Arthur and Bamford, Chris and Chaplot, Devendra Singh and de las Casas, Diego and Bressand, Florian and Lengyel, Gianna and Lample, Guillaume and Saulnier, Lucile and others , journal=. Mistral 7

  23. [24]

    Proceedings of ACL , pages=

    Multilingual End to End Entity Linking , author=. Proceedings of ACL , pages=. 2023 , publisher=

    2023

  24. [25]

    Eneko Agirre, Carmen Banea, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Rada Mihalcea, German Rigau, and Janyce Wiebe. 2016. S em E val-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In Proceedings of SemEval, pages 497--511. ACL

    2016

  25. [26]

    Lucie Bacon. 2022. La fabrique du parcours migratoire sur la route des B alkans: Co-construction des r \'e cits et \'e critures (carto)graphiques . Ph.D. thesis, Universit \'e de Poitiers

    2022

  26. [27]

    David Bamman, Brendan O'Connor, and Noah A Smith. 2013. Learning latent personas of film characters. In Proceedings of ACL, pages 352--361. ACL

    2013

  27. [28]

    Tommaso Caselli and Piek Vossen. 2017. The event S tory L ine corpus: A new benchmark for causal and temporal relation extraction. In Proceedings of the Events and Stories in the News Workshop, pages 77--86. ACL

    2017

  28. [29]

    Daniel Cer, Mona Diab, Eneko Agirre, I \ n igo Lopez-Gazpio, and Lucia Specia. 2017. S em E val-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation. In Proceedings of SemEval, pages 1--14. ACL

    2017

  29. [30]

    Nathanael Chambers and Dan Jurafsky. 2008. Unsupervised learning of narrative event chains. In Proceedings of ACL, pages 789--797. ACL

    2008

  30. [31]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of ACL, pages 8440--8451. ACL

    2020

  31. [32]

    Agata Cybulska and Piek Vossen. 2014. Using a sledgehammer to crack a nut? lexical diversity and event coreference resolution. In Proceedings of LREC, pages 4545--4552

    2014

  32. [33]

    Christopher Forstall, Walter Scheirer, David Bamman, and Gregory Crane. 2015. Modeling the scholars: Detecting intertextuality through enhanced word-level n-gram matching. Digital Scholarship in the Humanities, 30(4):503--515

    2015

  33. [34]

    Fabrizio Gilardi, Meysam Alizadeh, and Ma \"e l Kubli. 2023. ChatGPT outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30):e2305016120

    2023

  34. [35]

    Muhammad Nihal Hussain, Kevin K Bandeli, Samer Al-khateeb, and Nitin Agarwal. 2018. Analyzing shift in narratives regarding migrants in E urope via blogosphere. In Proceedings of the International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction, pages 181--190. Springer

    2018

  35. [36]

    David Ing, Fabien Delorme, Said Jabbour, Nelly Robin, and Lakhdar Sais. 2025. Text mining from migration narratives. In Proceedings of ECML-PKDD

    2025

  36. [37]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, and 1 others. 2023. Mistral 7 B . arXiv preprint arXiv:2310.06825

    Pith/arXiv arXiv 2023

  37. [38]

    Julia Kristeva. 1969. S \'e m \'e iotik \`e : Recherches pour une s \'e manalyse . Seuil, Paris

    1969

  38. [39]

    William C Mann and Sandra A Thompson. 1988. Rhetorical structure theory: Toward a functional theory of text organization. Text, 8(3):243--281

    1988

  39. [40]

    Louis Martin, Benjamin Muller, Pedro Javier Ortiz Su \'a rez, Yoann Dupont, Laurent Romary, \'E ric Villemonte de la Clergerie, Djam \'e Seddah, and Beno \^i t Sagot. 2020. C amem BERT : a tasty F rench language model. In Proceedings of ACL, pages 7203--7219. ACL

    2020

  40. [41]

    Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of NAACL-HLT, pages 839--849. ACL

    2016

  41. [42]

    Nihan \"O zt \"u rk and Serkan Ayvaz. 2018. Sentiment analysis on T witter: A text mining approach to the S yrian refugee crisis. Telematics and Informatics, 35(1):136--147

    2018

  42. [43]

    Mikhail Plekhanov, Nora Kassner, Kashyap Popat, Louis Martin, Simone Merello, Boris Kozlovskii, Fabio A Dreyer, and Nicola Cancedda. 2023. Multilingual end to end entity linking. In Proceedings of ACL, pages 3512--3527. ACL

    2023

  43. [44]

    Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The P enn discourse T ree B ank 2.0. In Proceedings of LREC, pages 2961--2968

    2008

  44. [45]

    Qwen Team . 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115

    Pith/arXiv arXiv 2024

  45. [46]

    Andrew J Reagan, Lewis Mitchell, Dilan Kiley, Christopher M Danforth, and Peter Sheridan Dodds. 2016. The emotional arcs of stories are dominated by six basic shapes. EPJ Data Science, 5(1):1--12

    2016

  46. [47]

    Nils Reimers and Iryna Gurevych. 2019. Sentence- BERT : Sentence embeddings using siamese BERT -networks. In Proceedings of EMNLP-IJCNLP, pages 3982--3992. ACL

    2019

  47. [48]

    Nelly Robin. 2014. Migrations, observatoire et droit: Complexit \'e du syst \`e me migratoire ouest-africain. In HdR. Universit \'e de Poitiers

    2014