@inproceedings{ogawa-etal-2026-designing,
    title = "Designing {LLM} Agents for User-Centered Language Service Selection",
    author = "Ogawa, Ryoichiro  and
      Lin, Donghui  and
      Uwano, Fumito",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://preview.aclanthology.org/paragraph-normalization/2026.lrec-1.42/",
    doi = "10.63317/59qnh9w46bv6",
    pages = "599--608",
    abstract = "With the rapid expansion of language resources and services across repositories and platforms, users face an overwhelming number of options. While this diversity promises flexibility, non-experts struggle to compose appropriate resource pipelines and select services that satisfy both functional and non-functional requirements. We propose a user-centered framework of LLM agents that interprets natural-language requests and performs end-to-end language service selection. The agents extract functional requirements to form coherent task compositions and select suitable language services for each component by interpreting non-functional quality aspects embedded in contextual cues. To ensure reliable and explainable decisions, we employ a four-step structured reasoning procedure that combines Few-Shot exemplars and Chain-of-Thought reasoning: extracting functional requirements, inducing non-functional evaluation axes, applying these axes as constraints in candidate retrieval, and determining a final composition. We construct a benchmark dataset pairing diverse user requests with standardized language service profiles containing metadata and quality indicators, and evaluate our framework against representative prompting-based baselines. Results show consistent gains in Precision, Recall, and F1-score, demonstrating improved capture of both functional intent and quality preferences. These findings demonstrate that structured LLM agents can bridge natural-language user intents and language service configurations, enabling end-to-end selection and composition in a transparent and user-centered manner."
}