DOI: 10.18485/judig.2026.2.ch3 ISSN:

srWOZ: Towards Task-Oriented Dialogue Modelling in Serbian

Anđelka Zečević, Jovana Rađenović, Milica Gašić, Songbo Hu, Ranka Stanković, Mihailo Škorić
Show PDF Cite
The rapid adoption of large language models and agentic frameworks has increased the demand for high-quality conversational datasets for the development and evaluation of dialogue systems. Among these, task-oriented dialogue systems have emerged as a particularly important class of conversational systems, assisting users in accomplishing specific goals through natural language interaction. Despite their growing popularity and practical importance, research and development of task-oriented dialogue systems for Serbian remain limited due to the scarcity of adequate language resources and annotated dialogue datasets.

In this paper, we present srWOZ , a task-oriented multi-turn conversational dataset for Serbian. Its creation is inspired by the existing Multi 3 WOZ dataset, a multilingual multi-domain multi-parallel dataset of culturally adapted task-oriented dialogues. In order to create srWOZ , we followed the original Multi 3 WOZ language adaptation methodology. First, we adapted the underlying ontology to the Serbian cultural and institutional context, covering seven domains: attractions, hospitals, police services, hotels, restaurants, taxis, and public transportation, while also performing slot-value normalization. Second, we adapted outline-based dialogue generation component by adapting dialogue schemas and dialogue writing modules. Finally, we collected dialogues from native Serbian speakers with a bachelor degree in Serbian language using a bottom-up outline-based method. This process resulted in an initial dataset comprising 812 dialogues.

By addressing the lack of conversational resources for Serbian, srWOZ contributes to the broader effort of extending dialogue technologies to low-resource linguistic settings by highlighting the process and the difficulty and best practices of such an endeavour.

More from our Archive