Character Iconicity vs. Arbitrariness: An Arabic NLP Perspective
Arabic script can represent vowels through letters that share common shapes, but readability may not rely solely on visual distinctions. This paper explores whether keeping or remapping these letter distinctions matters for NLP performance.
T-TAMER: Provably Taming Trade-offs in ML Serving
This work presents T-TAMER, a framework formalizing multi-model human evaluation as a best-arm identification problem, addressing trade-offs in accuracy, latency, and resource usage in ML serving with practical evaluations.
HUKUKBERT: Domain-Specific Language Model for Turkish Law
Introducing HukukBERT, a Turkish legal language model that leverages a hybrid DAPT methodology, demonstrating strong performance in legal term predictions and structural segmentation of court decisions.
Does Machine “know” interpersonal pragmatics? Evidence from MARBERT’s learning of emoji pragmatics in Arabic digital discourse
This study explores whether Transformer-based models, like MARBERT, can understand emoji pragmatics in Arabic digital discourse, finding that while it captures interpersonal functions effectively, it struggles with more implicit meanings.
PICTURE: Enhancing Theory-of-Mind in Large Language Models by Revealing, Not Hiding, Characters’ Lack of Knowledge
This research proposes PICTURE, enhancing Theory of Mind in LLMs by making characters’ lack of knowledge explicit rather than using event hiding, facilitating better performance in reasoning tasks.
Multi-Task GRPO: Reliable LLM Reasoning Across Tasks
This paper presents Multi-Task GRPO (MT-GRPO), a novel algorithm that optimizes worst-task performance in multi-task settings, effectively improving task balance.
VIVID: A Culturally Grounded Benchmark Exposing the Figurative Language Gap in Vietnamese NLP
VIVID presents the first benchmark focused on figurative language in Vietnamese, evaluating and revealing gaps in current NLP models’ capabilities.
Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization
This work introduces Multi-Agent Perspectivist Preference Optimization (MAP-PO) to improve detection of sexism in language, emphasizing the importance of diverse perspectives in annotation.
Automatic Annotation of Ancient Greek Vowel Length
This paper presents a general-purpose macronizer for Ancient Greek texts to automatically annotate vowel lengths, crucial for improving NLP tasks related to Ancient Greek.