Morocco

Large Language Models (LLMs), even though exhibiting multiple capabilities on many NLP tasks, struggle with phonologically-grounded phenomena like rhyme detection and generation. When one moves to lower-resource languages such as Modern Greek, this is even more evident. In this paper, we present a hybrid neural-symbolic system that combines LLMs with deterministic phonological algorithms to achieve accurate rhyme identification and generation. We implement a comprehensive taxonomy of Greek rhyme types and employ an agentic generation pipeline with phonological verification. We use multiple prompting strategies (zero-shot, few-shot, Chain-of-Thought, and RAG-augmented) across several LLMs including Claude 3.7 and 4.5, GPT-4o, Gemini 2.0 and open-weight models like Llama 3.1 8B and 70B and Mistral Large. Results reveal a significant reasoning gap: while native-like models (Claude 3.7) perform intuitively (40% accuracy in identification), reasoning-heavy models (Claude 4.5) achieve state-of-the-art performance (54%) only when prompted with Chain-of-Thought. Most critically, pure LLM generation fails significantly (under 4% valid poems), while our hybrid verification loop restores performance to 73.1%. Along with the system presented, we further release a corpus of 40,000+ rhymes, derived from the Anemoskala and Interwar Poetry corpora, to support future research.

EACL 2026

LLMs Got Rhyme? Hybrid Phonological Filtering for Greek Poetry Rhyme Detection and Generation

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

workshop paper

#### *Message from the General Chair, Aline Villavicencio*
I’m delighted and honoured to welcome you to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), taking place in the beautiful city of Rabat, in Morocco, in March 24-29, 2026. EACL is the flagship European conference of the Association and EACL 2026 proudly continues our field’s tradition of excellence in scholarship, innovation, and inclusivity. I am deeply grateful to the many volunteers whose dedication, generosity, and tireless efforts have made this conference possible.
For the first time EACL is being hosted in the African continent. This is an important milestone for our community, and we are grateful to our Moroccan hosts for enabling this historic moment by bringing this edition of EACL to Rabat. We are also delighted that the Second Arabic NLP School is co-located with EACL. We hope attendees enjoy this wonderful opportunity to strengthen ties with the Computational Linguistics communities across the African continent. *[Read full message](https://drive.google.com/file/d/14NlmHvwM6fPJuMmOvVh7K0vtQbEyv3SZ/view?usp=sharing)* 

<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://underline.io/events/522/reception','_blank')">Go to Workshops and Tutorials Program</button></html>

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

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.eacl.org/registration/) first.

**Online Registration Form**: https://acl.swoogo.com/eacl2026

Registration Required

Welcome to the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL). EACL 2026 will be held in Rabat, Morocco, from March 24–29, 2026. 

This paper investigates how language models (LMs) reproduce the existing gender disparity found in German media discourse about artificial intelligence (AI). Building on a human-annotated corpus of quotations from German media discourse on AI, we first quantify the frequency with which male and female speakers are directly cited across domains and speaker roles. We then train LLäMmlein (Pfister et al., 2025), a state-of-the-art German-only language model, GBERT, and a logistic regression model using only the quoted text as input and without providing any gender cues to classify the quotation as originating from a male or female speaker. By comparing model predictions with corpus-based gold labels, we find that male voices dominate both the corpus and the model predictions. Balancing the data mitigates but does not fully eliminate this disparity, indicating that the strong male-default tendency of transformer models cannot be explained by corpus skew alone, but also by their priors from pretraining. The study contributes to the interpretability of language models' output for DH-related tasks, adaptation of NLP tools to domain-specific humanities corpora, and knowledge modelling in the humanities.

Invisible Speakers? Gender Disparity in German AI Discourse and Its Reflection in Language Models

Graph-based Retrieval-Augmented Generation (RAG) is increasingly used to explore long, heterogeneous, and weakly structured corpora, including historical archives. However, in such settings, naive full-corpus indexing is often computationally costly and sensitive to OCR noise, document redundancy, and topical dispersion. In this paper, we investigate corpus pre-targeting strategies as an intermediate layer to improve the efficiency and effectiveness of graph-based RAG for historical research. We evaluate a set of pre-targeting heuristics tailored to single-hop and multi-hop of historical questions on HistoriQA-ThirdRepublic, a French question-answering dataset derived from parliamentary debates and contemporary newspapers. Our results show that appropriate pre-targeting strategies can improve retrieval recall by 3–5% while reducing token consumption by 32–37% compared to full-corpus indexing, without degrading coverage of relevant documents. Beyond performance gains, this work highlights the importance of corpus-level optimization for applying RAG to large-scale historical collections, and provides practical insights for adapting graph-based RAG pipelines to the specific constraints of digitized archives.

How to Efficiently Explore Noisy Historical Data? Leveraging Corpus Pre-Targeting to Enhance Graph-based RAG

Armenian epigraphy, one of the world's oldest and most diverse inscriptional traditions, remains largely absent from digital research infrastructures due to a lack of basic linguistic and conceptual resources. No machine-readable corpus, standardized terminology, or controlled vocabulary exists for describing Armenian inscription types, preventing indexing and interoperability. This paper addresses this gap by constructing the first dataset of Armenian inscription-type terminology and by developing a computational pipeline for analyzing it at scale. We digitize and preprocess a broad corpus of authoritative printed publications; curate a culturally grounded terminology list; and train transformer-based NER models to identify both attested inscription types and potential terminological variants across unseen texts. The resulting resources form the first empirical foundation for modelling Armenian epigraphic concepts needed for further developing a SKOS vocabulary aligned with, yet culturally distinct from, existing international epigraphic ontologies.

From Corpus to Concept Scheme: Developing a SKOS Vocabulary for Armenian Epigraphic Heritage

This paper presents the first application of an automatic token-classification approach for detecting reported speech spans in Classical Latin using transformer-based neural architectures. Focusing on Seneca the Elder's Declamatory Anthology, the study addresses the text's highly polyphonic nature, resulting from the use of reported speech. Instead of relying exclusively on sentence-level syntactic information, the proposed approach treats reported speech detection as a token-level sequence labeling problem. This enables the identification of reported speech spans extending across multiple sentences. We fine-tune three Latin neural language models —LatinBERT, LaBERTa, and PhilBERTa— for binary token-level classification and conduct experiments both with and without punctuation. The results show that RoBERTa-based models effectively identify reported speech, with LaBERTa achieving the best performance (F1 scores above 0.90).

Detecting reported speech as a token classification task: an application to Classical Latin?

We introduce CroCoSyn, a controlled, cross-lingual and cross-model corpus of 25,920 LLM-generated film synopses in English and French. Each synopsis is generated under systematically varied conditions, including model type, temperature, genre, protagonist gender, and narrative constraints, and enriched with structured metadata capturing characters and their relationships. Comparing Mistral and Llama across different model temperature degrees, CroCoSyn enables fine-grained analysis of narrative content, style, and character representation across models and languages. The corpus supports research on gender and cultural biases and story generation evaluation, and provides a foundation for comparative studies between LLM-generated and human-written narratives.

CroCoSyn: A Cross-Lingual and Cross-Model Corpus of LLM-Generated Film Synopses

Armenian epigraphy is extensively documented in printed scholarly corpora, yet lacks machine-readable editions that support interoperability or computational analysis. In this paper, we present Armenian AutoEpiDoc, a system that automatically converts expert-verified Armenian inscription records into EpiDoc-compliant TEI/XML files. Operating on curated and domain-validated data, AutoEpiDoc maps Armenian-specific metadata to EpiDoc structures through rule-based templates and schema-aware validation. The workflow significantly reduces manual encoding effort and provides a scalable path toward producing digital editions and integrating Armenian inscriptions into international epigraphic infrastructures.

Armenian AutoEpiDoc: Automated Extraction and Encoding of Armenian Inscriptions into EpiDoc TEI/XML

Transformer models have revolutionized NLP, yet many morphologically rich languages remain underrepresented in large-scale pre-training efforts. With SindBERT, we set out to chart the seas of Turkish NLP, providing the first large-scale RoBERTa-based encoder for Turkish. Trained from scratch on 312 GB of Turkish text (mC4, OSCAR23, Wikipedia), SindBERT is released in both base and large configurations, representing the first large-scale encoder-only language model available for Turkish. We evaluate SindBERT on part-of-speech tagging, named entity recognition, offensive language detection, and the TurBLiMP linguistic acceptability benchmark. Our results show that SindBERT performs competitively with existing Turkish and multilingual models, with the large variant achieving the best scores in two of four tasks but showing no consistent scaling advantage overall. This flat scaling trend, also observed for XLM-R and EuroBERT, suggests that current Turkish benchmarks may already be saturated. At the same time, comparisons with smaller but more curated models such as BERTurk highlight that corpus quality and diversity can outweigh sheer data volume. Taken together, SindBERT contributes both as an openly released resource for Turkish NLP and as an empirical case study on the limits of scaling and the central role of corpus composition in morphologically rich languages. The SindBERT models are released under the MIT license and made available in both fairseq and Huggingface formats.

SindBERT, the Sailor: Charting the Seas of Turkish NLP

Large Language Models (LLMs) achieve strong performance on many tasks, but they still struggle with morphologically rich, low-resource languages such as Turkish. This difficulty stems from Turkish being an agglutinative language and underrepresented in multilingual training data, which causes current models to often fail at capturing its morphology, flexible word order, and formal registers. In this paper, we introduce MODA (Model Adapted for Domain Applications), a Turkish-specialized LLM built via a modular pipeline that combines continual pre-training, parameter-efficient fine-tuning, and model merging. Starting from Qwen2.5-7B as the base model, we first perform large-scale continual pre-training on a Turkish web corpus to improve grammatical and morphological representations. We then apply parameter-efficient supervised fine-tuning on task-oriented instruction data, and finally merge specialized variants into a single production-ready model. We evaluate MODA on TurkishMMLU, the Turkish subset of EXAMS, and TRCLAIM-19, where it consistently outperforms both the base and instruction-tuned Qwen2.5-7B models. Our results support a training strategy that explicitly separates linguistic acquisition from task alignment when adapting LLMs to morphologically rich, underrepresented languages under realistic hardware constraints.

Building a Turkish Large Language Model via Continual Pre-Training and Parameter-Efficient

Text-to-SQL systems have achieved strong performance on English benchmarks, yet their behavior in morphologically rich, low-resource languages remains largely unexplored. We introduce BIRDTurk, the first Turkish adaptation of the BIRD benchmark, constructed through a controlled translation pipeline that adapts schema identifiers to Turkish while strictly preserving the logical structure and execution semantics of SQL queries and databases. Translation quality is validated using a CLT-based statistical framework, achieving 98.15% accuracy on human-evaluated samples. Using BIRDTurk, we evaluate inference-based prompting, agentic multi-stage reasoning, and supervised fine-tuning. Our results reveal that Turkish introduces consistent performance degradation—driven by both structural linguistic divergence and underrepresentation in LLM pretraining—while agentic reasoning demonstrates stronger cross-lingual robustness. Supervised fine-tuning remains challenging for standard multilingual baselines but scales effectively with modern instruction-tuned models. BIRDTurk provides a controlled testbed for cross-lingual Text-to-SQL evaluation under realistic database conditions. We release the training and development splits to support future research.

BIRDTurk: A Large-Scale Adaptation of the Text-to-SQL Dataset to Turkish

Most large language models (LLMs) are trained on massive datasets that include private infor- mation, which may be disclosed to third-party users in output generation. Developers put de- fences to prevent the generation of harmful and private information, but jailbreaking methods can be used to bypass them. Machine unlearn- ing aims to remove information that may be private or harmful from the model’s genera- tion without retraining the model from scratch. While machine unlearning has gained some popularity to counter the removal of private information, especially in English, little to no attention has been given to Turkish unlearn- ing paradigms or existing benchmarks. In this study, we introduce TUNE (Turkish Unlearn- ing Evaluation), the first benchmark dataset for Turkish unlearning task for personal infor- mation. TUNE consists of 9842 input-target text pairs about 50 fictitious personalities with two training task types: (1) Q&A and (2) In- formation Request. We fine-tuned the mT5 base model to evaluate various unlearning meth- ods, including our proposed approach. We find that while current methods can help unlearn unwanted private information in Turkish, they also unlearn other information we want to re- tain in the model.

Downloads

Next from EACL 2026

Invisible Speakers? Gender Disparity in German AI Discourse and Its Reflection in Language Models

Similar lecture

PhonologyBench: Evaluating Phonological Skills of Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.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 EACL 2026

Invisible Speakers? Gender Disparity in German AI Discourse and Its Reflection in Language Models

Similar lecture

PhonologyBench: Evaluating Phonological Skills of Large Language Models

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads