United States

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.

ACL 2026

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

an lushan rebellion

emotion embeddings

tang poetry

prompt sensitivity

digital humanities

emotion analysis

stages 

COLING, the International Conference on Computational Linguistics, is one of the premier conferences for natural language processing and computational linguistics. Often grouped within the field of artificial intelligence, but actually pre-dating the development of artificial intelligence, advances in computational linguistics and natural language processing are now some of the major drivers behind the use of artificial intelligence for commercial and social applications – for example, on-line search, machine translation and with voice-assisted conversational devices.

First established in 1965, the biennial COLING conference is held in diverse parts of the globe and attracts participants from both top-ranked research centers and emerging countries. Today, the most important developments in our field are taking place not only in universities and academic research institutes, but also in industrial research departments and in technological startups. COLING conferences provide opportunities for all these communities to showcase their exciting developments.

COLING 2020

COLING, the International Conference on Computational Linguistics, is one of the premier conferences for natural language processing and computational linguistics. Often grouped within the field of artificial intelligence, but actually pre-dating the development of artificial intelligence, advances in computational linguistics and natural language processing are now some of the major drivers behind the use of artificial intelligence for commercial and social applications – for example, on-line search, machine translation and with voice-assisted conversational devices.

workshop paper

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.

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.

Beyond Genre Categories: How Narrative Pattern Coherence and Spanning Distance Shape Film Success

Nearly all studies on web registers – online text varieties associated with distinct social contexts and linguistic features – use full documents as the unit of analysis. However, web documents often contain sections in different registers. A cooking blog post, for instance, may combine personal storytelling, recipe instructions, user comments, and promotional text within a single URL. This internal variation raises doubts about the validity of document-level register labeling. In this paper, we propose an LLM-based segmentation approach that identifies register-homogeneous segments within documents and apply it to a 10,000-document English sample from HPLT 3.0. We show that segmentation addresses persistent problems in register analysis, including low inter-annotator agreement and category fuzziness. Strikingly, it also reveals that most web documents contain more than one register, making register mixing the norm rather than the exception on the web.

Register Mixing Is the Norm on the Web

Identifying intertextual parallels is central to philology, traditionally requiring labor-intensive manual analysis. While digitized historical corpora enable automated approaches using semantic sentence embeddings, training such models requires large annotated datasets, which are scarce for low-resource languages. We address this challenge by introducing a scalable automatic annotation pipeline for training semantic embedding models for Classical Tibetan. Our method combines unsupervised contrastive bootstrapping with iterative pair mining, generating silver-standard similarity labels through two complementary annotation strategies: (1) an ensemble of embedding models and rerankers, and (2) an LLM-as-a-judge committee using best–worst scaling. 
When combined with a domain‑specific, gold‑standard annotated dataset for sequential fine‑tuning, the resulting text-similarity model achieves a state‑of‑the‑art Spearman correlation of 0.864 on the STS task. This enables effective semantic search in Classical Tibetan and provides a framework for automatic supervision in low-resource languages used in digital humanities. We will make our code, dataset, and trained model publicly available upon publication.

Scaling Sentence Similarity for Classical Tibetan with Automatic Annotations

Historical corpora for Tagalog remain limited, particularly texts produced during the Martial Law period under the dictatorship of Ferdinand Marcos Sr. (1972–1986). Much of this material remains undigitized, restricting computational analysis of a significant period in Philippine political history. To support research on historical Tagalog texts, we introduce PHMartialLawNER, a gold-standard named entity recognition corpus constructed from newspapers and underground publications of the Martial Law era. The corpus includes approximately 13k extracted sentence segments (362,000 tokens), consolidated into 8k annotated text spans through a semi-automatic pipeline with manual validation. The reliability of the annotation is measured using Cohen’s $\kappa$, reaching 0.86 on all tokens and 0.72 on annotated tokens, with a pairwise F1-score of 0.74. The schema defines historically relevant entity categories including Person (Individual, Collective), Organization (Political, Government, Other), Event (Local, International), Production (Media, Government, Doctrine), as well as Time, Numerical Statistics, Location, and Object entities, specifically identifying weapon artifacts. We establish baseline performance using GLiNER variants, calamanCy models, and transformer-based architectures under zero-shot and few-shot settings. The PHMartialLawNER corpus will be publicly released to support Tagalog NLP, historical text processing, and digital humanities research.

PHMartialLawNER: A Tagalog Named Entity Recognition Corpus for the Philippine Martial Law Era

Literary translation requires balancing target-language fluency with faithfulness to the source. Recent large language models (LLMs) often produce fluent translations, but it remains unclear whether fluency corresponds to semantic preservation in literary text. We examine this relationship using 130,486 translated paragraphs from 106 novels in 16 source languages, including human, Google Translate, and TranslateGemma translations. Fluency is measured as original-likeness with a translationese classifier trained on paragraph part-of-speech n-grams, and faithfulness with the automatic translation evaluation metric COMET-KIWI. We control for paragraph length and find a consistent negative correlation between fluency and faithfulness. The pattern appears for both human and Google Translate, but is weaker and often non-significant for TranslateGemma. These results show that segment length matters for automatic evaluation and suggest a tradeoff between fluency and faithfulness in literary translation.

Downloads

Next from ACL 2026

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

Similar lecture

An Iterative Emotion Interaction Network for Emotion Recognition in Conversations

.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

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

Similar lecture

An Iterative Emotion Interaction Network for Emotion Recognition in Conversations

Downloads