Two Publications Accepted to Findings of ACL at NAACL 2025

We are pleased to announce that two papers authored by members of our research group have been accepted for publication in the Findings of ACL at the 2025 North American Chapter of the Association for Computational Linguistics (NAACL) conference, to be held in New Mexico.

1. Playing with Voices: Tabletop Role-Playing Game Recordings as a Diarization Challenge

This paper addresses the complexities of speaker diarization within the context of tabletop role-playing games (TTRPGs), which are characterized by frequent speaker turns and voice modulations as players assume different characters. The study introduces a novel dataset and benchmarks it against established corpora such as AMI and ICSI.

Key findings include:

• TTRPGs present unique diarization challenges due to rapid speaker changes and voice conversions.

• Traditional diarization models exhibit difficulty in distinguishing between actual speakers and their role-played personas.

• Evaluation of two diarization models indicates increased confusion rates on TTRPG data, with the wespeaker model significantly underestimating the number of distinct speakers.

Access the full paper here.

2. Large Language Models as Grammatical Feature Taggers for African American English

This work evaluates the capability of rule-based models, transformer architectures, and large language models (LLMs) to identify key grammatical features of African American English (AAE), focusing on phenomena such as Habitual Be and Multiple Negation.

Key findings include:

• LLMs outperform traditional baseline models but remain susceptible to biases, including recency effects and issues related to formality mismatches.

• The grammatical complexity inherent in AAE poses significant challenges for natural language processing models, underscoring the need for enhanced training datasets and architectural refinements.

Access the full paper here.