United States

This study proposes a quantitative framework for profiling LLM dispositions as stable, model-specific regularities in output under repeated, controlled elicitation. Using a structured narrative constraint-selection task administered across six frontier models and three instruction types, we operationalize disposition through two dimensions: &quot;consistency&quot;, measured as cross-replication selection overlap via Jaccard similarity, and &quot;diversity&quot;, measured as dispersion across options via the inverse Simpson index. We further introduce Narrative Landscape, a PCA-based visualization that maps each model&#39;s selection profile into a shared space for direct comparison. Results reveal a clear rigidity-exploration spectrum across model families and show that instruction types shift the geometry of selection spaces even when scalar metrics appear similar, indicating that comparable scores can mask qualitatively distinct selection topologies.

ACL 2026

Narrative Landscape: Mapping Narrative Dispositions Across LLMs

selection topology

pca visualization

replication-based evaluation

model dispositions

large language models

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. <br>
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) <br><br><br>
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>
<br><br><br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

This study proposes a quantitative framework for profiling LLM dispositions as stable, model-specific regularities in output under repeated, controlled elicitation. Using a structured narrative constraint-selection task administered across six frontier models and three instruction types, we operationalize disposition through two dimensions: "consistency", measured as cross-replication selection overlap via Jaccard similarity, and "diversity", measured as dispersion across options via the inverse Simpson index. We further introduce Narrative Landscape, a PCA-based visualization that maps each model's selection profile into a shared space for direct comparison. Results reveal a clear rigidity-exploration spectrum across model families and show that instruction types shift the geometry of selection spaces even when scalar metrics appear similar, indicating that comparable scores can mask qualitatively distinct selection topologies.

workshop paper

We present a new publicly available corpus of 100,502 movie reviews from Kazakhstan collected from kino.kz, spanning 2001–2025 and covering 4,943 unique film titles. The dataset is multilingual, consisting mainly of Russian reviews alongside Kazakh and code-switched texts. Reviews are manually annotated for language and sentiment polarity, and 11,309 reviews additionally contain explicit user-provided ratings. We define two sentiment tasks—three-way polarity classification and five-class score classification—and benchmark classical BoW/TF–IDF baselines against multilingual transformer models (mBERT, XLM-RoBERTa, RemBERT). Experimental results show that transformer models consistently outperform classical baselines on polarity classification, while score classification remains challenging under leakage-controlled evaluation due to severe class imbalance and subtle distinctions between adjacent rating levels.

100,000+ Movie Reviews from Kazakhstan: Russian, Kazakh, and Code-Switched Texts

The Kṛṣṇa Yajurveda survives in multiple recensions that share substantial ritual content, yet the degree and distribution of textual overlap across recensions have never been quantified systematically. This paper presents a computational analysis of text reuse across three recensions---the Maitrāyaṇī Saṃhitā (MS), the Kāṭhaka Saṃhitā (KS), and the Taittirīya Saṃhitā (TS)---for two ritual sections (Agnyupasthāna and Punarādhāna), using ICoMa (Intertextuality Collation Machine), a new web-based multi-algorithm collation tool. Five independent similarity algorithms consistently rank MS--KS as the most closely related pair, corroborating the philological consensus. Crucially, the two ritual sections exhibit strikingly different reuse profiles: Punarādhāna shows near-identical MS--KS overlap (up to 93.5%) with sharp divergence from TS, while Agnyupasthāna displays moderate, broadly distributed similarity across all three pairs. These contrasting patterns provide quantitative evidence that different ritual categories followed distinct paths of textual transmission within the Yajurvedic tradition. ICoMa and the experimental data are freely available.

Quantifying Text Reuse Across Three Kṛṣṇa Yajurveda Recensions: Using Multi-Algorithm Computational Collation

Ancient and endangered languages pose a unique challenge for NLP: their datasets are inherently scarce, difficult to expand, and built from formulaic corpora---making data-quality issues especially consequential yet rarely audited. Motivated by the need to understand what current NMT can realistically achieve for such languages, we investigate hieroglyphic-to-German translation, where a recent study reported 61.5 BLEU using fine-tuned M2M-100. Our reproduction yields only 37.0 BLEU with the released model. Investigating this gap, we find \textbf{32\% of test targets appear identically in training} (16/50; 50\% under 8-gram overlap at 70\% threshold). This contamination inflates scores dramatically: contaminated samples achieve up to 83.8 BLEU / 0.924 COMET-22 versus 30.9--39.2 BLEU / 0.622--0.676 COMET-22 on clean samples across five model configurations spanning two architectures. Document-level decontamination reduces contaminated BLEU by only 4.6 points because 8/16 targets persist via other source documents---target-level deduplication is required. We release a decontaminated 34-sample test set and establish corrected baselines (30.9--39.2 BLEU), providing a realistic assessment of NMT capability for this endangered writing system.

Data Contamination in Neural Hieroglyphic Translation: A Reproducibility Study

Many Tang-poetry emotion studies still rely on coarse labels (e.g., positive/negative), while recent LLM-based attempts face a practical problem: one-word emotion outputs are highly sensitive to prompt wording. When labels shift with phrasing, historical interpretation becomes hard to reproduce and hard to trust. Focusing on Tang poetry around the An Lushan Rebellion (安史之乱), we propose a fine-grained sentence-level workflow centered on emotion embeddings: we use continuous hidden-state vectors, run automatic clustering, and then consolidate labels for interpretation. On the same 3,198 emotional sentences, one-word outputs show only 50.3% A/B exact agreement, while embedding-based clustering remains stable and well distributed ($H_{norm}=0.989$; $20/20$ active clusters). On 7,195 labeled sentences, a char-based baseline reaches 0.446 micro-F1 and 0.395 macro-F1. This multi-stage label-construction path supports historically grounded findings, including the emotional turning point around 762, and also reveals layered patterns that are less visible in coarse setups. These results suggest that stable representation is a prerequisite for turning computational outputs into credible evidence for humanities interpretation.

Beyond Prompt-Sensitive Emotion Words: Stable Embeddings for Tang Poetry Analysis

Large language models (LLMs) can convincingly imitate human writing styles, yet it remains unclear how much stylistic information is encoded in embeddings from any language model and retained after LLM rewriting. We investigate these questions in French, using a controlled literary dataset to quantify the effect of stylistic variation via changes in embedding dispersion. We observe that embeddings reliably capture authorial stylistic features and that these signals persist after rewriting, while also exhibiting LLM-specific patterns. These analytical results offer promising directions for authorship imitation detection in the era of language models.

Measuring Embedding Sensitivity to Authorial Style in French: Comparing Literary Texts with Language Model Rewritings

This research explores the intersection of cultural heritage and Generative AI (Gen-AI), examining AI-generated historical image reconstructions as a potential tool for visualising multiple perspectives in heritage interpretation. In critical heritage studies, the concept of multivocality or polyvocality advocates for representing diverse, often underrepresented, perspectives in how heritage is understood and communicated. We evaluated three prominent AI image generation models across three heritage test cases. A total of 13 user prompts generated 39 images, which underwent both linguistic analysis of intermediate prompt transformations and systematic visual assessment by heritage experts for historical accuracy and cultural sensitivity. The findings revealed both strengths and limitations of the models. While the models produced visually compelling outputs and, in some cases, meaningfully distinct depictions across perspectives, they also exhibited representation imbalances, neutralisation and amplification tendencies, inconsistencies in human portrayal, and misinterpretations introduced during the linguistic transformation of user inputs. Based on these findings, we propose initial guidelines for structured prompt construction that target the specific failure patterns identified. The research suggests that generative AI could serve as a supplementary tool, not a definitive historical source, for exploring multivocal heritage interpretation, particularly in museum and visitor engagement contexts, provided it is used critically and in conjunction with expert input.

Prompting the Past: Linguistic Transformations and Cultural Accuracy in AI-Generated Image Reconstructions for Multivocal Cultural Heritage

Languages change over time. Computational models can be trained to recognize such changes enabling them to estimate the publication date of texts. Despite recent advancements in Large Language Models (LLMs), their performance on automatic dating of texts, also known as Temporal Text Classification (TTC), has not been explored. This study provides the first systematic evaluation of leading proprietary (Claude 3.5, GPT-4o, Gemini 1.5) and open-source (LLaMA 3.2, Gemma 2, Mistral, Nemotron 4) LLMs on TTC using three historical corpora, two in English and one in Portuguese. We test zero-shot and few-shot prompting, and fine-tuning settings. Our results indicate that proprietary models perform well, especially with few-shot prompting. They also indicate that fine-tuning substantially improves open-source models but that they still fail to match the performance delivered by proprietary LLMs.

Temporal Text Classification with Large Language Models

Cross-lingual detection of intertextuality and translation in Latin and Ancient Greek through computational approaches is of great interest for classical studies.
While several systems exist for parallel sentence detection, based on general multilingual or specific models for Latin-Ancient Greek, they have not been compared against each other. 
Therefore, we present a synthetic benchmark to evaluate the performance of language models regarding cross-lingual Ancient Greek and Latin parallel sentence mining. 
We first compare six language models to encode sentences and then further improve the cross-lingual alignment through post-processing, fine-tuning, and knowledge distillation. 
We find that the whitening transformation in combination with knowledge distillation provides excellent results. 
Specifically, SPhilBERTa, a trilingual language model for Ancient Greek and Latin, benefits the most from the improvements and achieves an almost perfect retrieval score of 97.6 on our benchmark.

Evaluating Latin and Ancient Greek Sentence Alignment through Parallel Sentence Mining

The Aramaic proclitic *dalet*, widely used in historical Hebrew texts, serves two distinct grammatical functions: as a subordinating conjunction and as a possessive preposition. Because these functions are orthographically identical and no annotated resources exist for this task, large-scale computational analysis of their usage has previously been infeasible. In this paper we introduce a new BERT model for historical Hebrew in which all prefixes are segmented and encoded as independent tokens. This representation allows the model to evaluate proclitics directly and enables a zero-shot method for determining the grammatical role of the *dalet* clitic using masked language modeling predictions. We evaluate the approach on a manually annotated dataset drawn from historical Hebrew literature spanning multiple regions and historical periods, achieving over 94% accuracy. Applying the method to a corpus of more than 300 million words of historical Hebrew texts, we conduct large-scale stylistic analyses of the choice between the Aramaic *dalet* and available Hebrew alternatives. The results reveal geographic and diachronic trends and identify distinct sociolectal patterns within the corpus. The prefix-segmented model and annotated dataset are released for public use.

Modeling the "Dalet" Clitic in Historical Hebrew Texts: A New Prefix-Segmented BERT Model and Stylistic Analysis

Prior research on cultural markets has relied on genre labels to distinguish products, overlooking the specific content features that differentiate films within the same genre. We address this gap using tropes as building blocks of narrative structure. From a dataset of 30k tropes across 18k films (TVTropes.org), we identify 29 narrative patterns via community detection and characterize each film by two measures: coherence (how concentrated its tropes are within a few patterns) and spanning distance (how far apart the patterns it combines are). Regression analyses show that coherence improves both audience evaluations and attention, while spanning distance increases evaluations but reduces attention. These findings extend category-spanning theory from genre labels to the internal narrative composition of films, demonstrating how stories are constructed and shape audience responses.

Downloads

Next from ACL 2026

100,000+ Movie Reviews from Kazakhstan: Russian, Kazakh, and Code-Switched Texts

Similar lecture

Style over Story: Measuring LLM Narrative Preferences via Structured Selection

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

100,000+ Movie Reviews from Kazakhstan: Russian, Kazakh, and Code-Switched Texts

Similar lecture

Style over Story: Measuring LLM Narrative Preferences via Structured Selection

Downloads