United States

This system description paper presents a submission to the BEA 2026 Shared Task on Rubric-based Short Answer Scoring for German. It describes a multi-strategy scoring pipeline combining rubric-conditioned prompting, LoRA fine-tuning of Qwen-family models, and prediction aggregation. The system ranks second on three official tracks and third on the remaining track.

ACL 2026

Multi-Strategy Rubric-Based Short Answer Scoring for German

quadratic weighted kappa

ensemble aggregation

rubric-based scoring

automated short answer grading

lora fine-tuning

large language models

german

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.

**Online Registration Form**: https://acl.swoogo.com/acl2026

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.

workshop paper

This paper describes our submission to the BEA 2026 shared task on vocabulary item difficulty prediction in multilingual settings. We investigated whether transformer-based representations learned directly from item content can support the prediction of vocabulary item difficulty across different L1 groups. Our approach adopted a multilingual BERT-based architecture, specifically the mmBERT, with representation augmentation at both the layer and token levels, followed by a multi-task cascade learning that incorporates part-of-speech information as an auxiliary structural signal. Results showed that multi-task mmBERT consistently outperforms the shared-task XLM-RoBERTa baseline across languages, while gains from more complex aggregation are not uniform. The findings showed that strong multilingual representations provide a competitive foundation for vocabulary item difficulty prediction, while the benefits of additional architectural complexity depend on the language and training setting.

Precalibration of Vocabulary Item Difficulty with Multilingual Transformers and Multi-Task Learning

Humor is a complex form of communication that remains challenging for machines. Despite its broadness, most existing research on computational humor traditionally focused on modeling one specific type of humor. In this work, we wish to understand whether competence on specific humor tasks confers any ability to transfer to novel, unseen types; in other words, is this fragmentation inevitable? This question is especially timely as new humor types continuously emerge in online contexts (e.g., memes, anti-humor, AI fails). If LLMs are to keep up with this evolving landscape, they must be able to capture deeper, transferable mechanisms. 
To investigate this, we conduct a series of transfer learning experiments across four datasets, representing different humor tasks. We explore varied diversity settings (varying between 1-3 datasets in training, testing on a novel one). Experiments show that models are capable of some transfer, reaching up to 75% accuracy on binary unseen datasets; training on diverse sources improves transferability (1.88-4.05%) with minimal-to-no drop in in-domain performance. Somewhat surprisingly, the one dataset (Dad Jokes) emerges as the best enabler of transfer, but the hardest one to transfer to. We release data and code.

One Joke to Rule them All? On the (Im)possibility of Generalizing Humor Detection

Stand-up comedy, and humor in general, are often studied through their verbal content. Yet live performance relies just as much on embodied presence and audience feedback. We introduce TIC-TALK, a multimodal resource with 5,400+ temporally aligned topic segments capturing language, gesture, and audience response across 90 professionally filmed stand-up comedy specials (2015–2024). The pipeline combines BERTopic for 60 s thematic segmentation with dense sentence embeddings, Whisper-AT for 0.8 s laughter detection, a fine-tuned YOLOv8-cls shot classifier, and YOLOv8s-pose for raw keypoint extraction at 1 fps. Raw 17-joint skeletal coordinates are retained without prior clustering, enabling the computation of continuous kinematic signals—arm spread, kinetic energy, and trunk lean—that serve as proxies for performance dynamics. All streams are aligned by hierarchical temporal containment without resampling, and each topic segment stores its sentence-BERT embedding for downstream similarity and clustering tasks. As a concrete use case, we study laughter dynamics across 24 thematic topics: kinetic energy negatively predicts audience laughter rate (r=−0.75,N= 24), consistent with a stillness-before-punchline pattern; personal and bodily content elicits more laughter than geopolitical themes; and shot close-up proportion correlates positively with laughter (r= +0.28), consistent with reactive montage.

Timing In stand-up Comedy: Text, Audio, Laughter, Kinesics (TIC-TALK): Pipeline and Database for the Multimodal Study of Comedic Timing

Humor is one of the most culturally embedded and socially significant dimensions of human communication, yet it remains largely unexplored as a dimension of Large Language Model (LLM) alignment. In this study, five frontier language models play the same Cards Against Humanity games (CAH) as human players. The models select the funniest response from a slate of ten candidate cards across 9,894 rounds. While all models exceed the random baseline, alignment with human preference remains modest. More striking is that models agree with each other substantially more often than they agree with humans. We show that this preference is partly explained by systematic position biases and content preferences, raising the question whether LLM humor judgment reflects genuine preference or structural artifacts of inference and alignment.

Cards Against LLMs: Benchmarking Humor Alignment in Large Language Models

We present exploratory experiments in the comedic roasting capabilities of GPT4o. Specifically, @ComedyCentral roasts were scraped to design a survey in which participants blindly evaluated snippets of human and AI roasts, and had to predict the author (AI/human) in a second round of reviewing. The results show that there is no significant difference in how the barbs in human- and AI-generated roasts are rated. Further, a qualitative analysis showed that although the model utilizes specific recurrent phrases to imitate the style of human comedians, both generative LLM detectors and humans performed suboptimally in predicting the true author of the roasts.

The Roast of GPT4o: Experiments in Generating, Detecting and Evaluating Celebrity Roast Comedy

We investigate whether scaling model parameters improves humor generation through a controlled ablation study. Using five Qwen3 variants (8B–235B, dense and MoE), we generate jokes across 50 themes. Beyond evaluating humor scaling, this work serves as an empirical study into the nature of LLM versus human evaluations on highly subjective creative tasks. While an automated judge yields a perfect monotonic ranking between parameter count and win rate, human annotators find no significant aggregate difference in humor quality. Restricting to themes where annotators agree reveals a significant preference for the largest model (p = 0.039), suggesting scaling effects exist but are masked by a "quality floor." Crucially, our analysis of bias characteristics shows that the automated judge exhibits severe positional and length biases compared to human evaluators, further suggesting that LLMs may systematically distort quality differences on subjective tasks.

Does Bigger Mean Funnier? Evaluating Humor Generation Across the Qwen3 Model Family

Large language models (LLMs) can generate jokes that rival human comedy in quality, yet their originality remains uncertain. This study develops and validates automated, corpus-based methods for quantifying joke originality by focusing on core "topic handles" – the key nouns or short noun phrases that capture a joke's script opposition and logical mechanism (per the General Theory of Verbal Humor). Using a reference corpus of one million English jokes from Reddit, we compute Pointwise Mutual Information (PMI) on handle pairs in three variants (raw co-occurrence, semantic-cluster smoothing, and word- decomposition) and two embedding-based measures (handle-level conceptual distance and full-text corpus novelty via Sentence-BERT). We tested the measures on 400 LLM- generated jokes (200 from GPT-4o and 200 from GPT-5.4 prompted for original two-sentence jokes) and 80 benchmark jokes from the Witscript-powered JEST benchmark (including non- humorous controls). Three professional comedians rated all items on 0–3 scales of originality and funniness. Automated measures, especially raw PMI average (ρ = .386) and corpus novelty (ρ = .372), correlated significantly with human originality ratings but showed no relation to funniness. Benchmark jokes were rated most original, and the originality– funniness relationship differed sharply across sources (positive for benchmark, negative for GPT-4o). The results demonstrate that handle-based PMI and semantic novelty metrics offer practical, quantitative tools for assessing originality in both human and AI-generated humor, advancing objective evaluation of computational creativity while highlighting distinct patterns in how general-purpose LLMs versus specialized humor systems produce novel jokes.

Navigating the Joke Space: Towards Automated Originality Assessment of AI-Generated Humor

Computational mental health (CMH) classifiers often degrade under distribution shift because human annotators and distant-supervision pipelines reward different linguistic signals. 
 
 We introduce TSS (Triple-Stream Stress probe), a multi-channel diagnostic framework that decomposes text into (A) lexical character n-grams, (B) a small, mostly content-free morpho-syntactic channel, and (C) a 154-feature psycholinguistic style channel. Across four English datasets (N = 12,906), TSS reveals a lexical interference effect: adding lexical features to the style channel reduces Macro-F1 on human-labeled data (mean drop 0.072, p < 10⁻⁴) but not on auto-labeled data. 
 
 We propose Degree of Divergence (DoD), a difference-in-differences statistic adapted from econometrics for label-source auditing, with instance-level bootstrap inference; the headline estimate is DoD(BC–A) = 0.0374, 95% CI [0.0097, 0.0651], p = 0.0032. 
 
 A platform-stratified Twitter-only DoD (which removes the Reddit vs. Twitter contrast) reproduces the pattern with bootstrap inference: DoDᵀᵥ,BC–A = +0.096 (p < 0.001) and DoDᵀᵥ,AC–A = −0.089 (p < 0.001). Interventional masking (pos_only) retains ∼95–99% of Channel C's performance after destroying content words on human datasets, indicating that the style channel does not rely primarily on lexical surface form. 
 
 TSS is positioned as a diagnostic audit framework, not a clinical screening tool: it flags label-source-specific shortcut learning before generalization claims are made.

The Divergence Hypothesis: Unmasking Lexical Interference and Label Bias in Mental Health NLP

Large language models achieve strong performance on biomedical question answering and summarization benchmarks, yet traditional evaluation metrics often fail to detect clinically significant factual errors. We introduce a unified evaluation framework that combines reference-based measures with evidence-grounded factuality verification to assess biomedical text generation. Evaluating four open-source models across three benchmarks (BioASQ, PubMedQA, MedLFQA), we find that 13.4–24.7% of generated claims are contradicted and 23–41% are unsupported, despite high lexical overlap scores. Our proposed Fact-Aligned Score (FAS) correlates strongly with claim-level verifiability (rho=0.68), substantially outperforming ROUGE-L (rho=0.41). We release an open-source toolkit with model outputs and analysis scripts to support reproducible factuality evaluation and safer deployment of biomedical LLMs.

Towards Unified Factuality Evaluation for Biomedical QA and Summarization: Aligning Metrics with Clinical Use-Cases

Seizure freedom is a key clinical outcome for people with epilepsy (PWE) yet it is primarily recorded in free-text notes and letters in the United Kingdom, making it difficult to aggregate and track at scale. This paper introduces a generative LLM-based pipeline boosted by synthetic data to identify a PWE's seizure freedom status in clinicians' records. We fine-tuned seven different LLMs with between 4-14 billion parameters using LoRA to compare models trained on synthetic records against those trained on expert annotated records. The best performing configuration, based on Qwen-2.5-14B, was trained entirely on synthetic records and used chain-of-thought (CoT) reasoning (both generated by GPT-5). This achieved an F1 score of 0.90±0.02 on double-annotated test data and outperformed the equivalent model trained on authentic clinician records, which achieved 0.87±0.04. The synthetically trained models also have the benefit of outputting their CoT reasoning process for greater decision-making transparency and can also make use of the unused supervised training data for significantly increased test examples. This work has implications for monitoring a key treatment outcome for PWE automatically and at scale.

Downloads

Next from ACL 2026

Precalibration of Vocabulary Item Difficulty with Multilingual Transformers and Multi-Task Learning

Similar lecture

IWM-DKM at BEA 2026 Shared Task 2: Supplementing Supervised Fine-Tuning for Rubric-Based Short Answer Scoring

.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

Precalibration of Vocabulary Item Difficulty with Multilingual Transformers and Multi-Task Learning

Similar lecture

IWM-DKM at BEA 2026 Shared Task 2: Supplementing Supervised Fine-Tuning for Rubric-Based Short Answer Scoring

Downloads