Multilingual Datasets for Custom Input Extraction and Explanation Requests Parsing in Conversational XAI Systems

Conversational AI systems struggle to understand user requests across languages: The paper identifies a critical gap where current explainable AI systems work well for English but fail to generalize to other languages due to insufficient training data, limiting access to AI explanations for non-English speakers.

Large language models show promise but inconsistent performance across languages: When tested on multilingual datasets, three different LLMs demonstrated varying abilities to understand user requests in five different languages, suggesting that model size alone doesn't guarantee reliable multilingual performance for explanation requests.

Custom user inputs remain largely unsupported in conversational AI systems: Most current systems only work with pre-configured data, but the paper highlights that users often want to input their own custom data and receive explanations—a capability that is "largely limited" in existing systems and needs development.

New multilingual datasets are essential for responsible AI development: The paper introduces two new datasets (MultiCoXQL and Compass) covering five languages including low-resource languages, addressing the data scarcity problem that prevents fair and equitable access to explainable AI across different linguistic communities.

Evaluation methods matter for assessing fairness across languages: The research demonstrates that testing AI systems using monolingual, cross-lingual, and multilingual evaluation approaches reveals different performance patterns, suggesting that responsible AI assessment requires comprehensive testing beyond single-language benchmarks.