| Nome do aluno | Thiago Santana Vaillant |
|---|---|
| Título do trabalho | From Entailment to Understanding: A Domain-Agnostic Approach to Intent Recognition and Slot Filling Without Fine-Tuning |
| Resumo do trabalho |
Voice user interfaces promise hands-free operation in real environments. Nevertheless, deploying them across many domains remains costly and hard to maintain because conventional intent–slot models depend on labeled data and repeated fine-tuning. In this work, we present a domain-agnostic approach that requires no additional training on target domains, intents, or slots, framing both intent recognition and slot filling as natural language inference and enabling strict zero-shot transfer with reusable pre-trained components. The approach was evaluated on public benchmarks and on a new evaluation-only dataset derived from interviews with field professionals, designed to capture realistic, time-pressured commands. Across these settings, the approach showed strong intent selection and competitive slot extraction for concrete slots. In addition, performance is mainly shaped by semantic overlap among slot definitions and by constraints inherent to NLI-based tasks such as extractive QA and sentence entailment. Taken together, the results indicate that an entailment-first formulation can serve as a practical alternative to supervised adaptation for VUIs that must generalize across domains. |
| Orientador | Eduardo Almeida |
| Membro Titular 1 (com afiliação) | Nathalia Nascimento (Penn State University) |
| Link para o curriculum lattes |
https://greatvalley.psu.edu/ |
| Membro Titular 2 (com afiliação) | Tatiane Rios (IC-UFBA) |
| Link para o curriculum lattes |
http://lattes.cnpq.br/ |
| Suplente 1 (com afiliação) | Frederico Araujo Durao (IC-UFBA) |
| Suplente 2 (com afiliação) | Nelio Cacho (UFRN) |
| Data do exame | 26 Mar, 2026 |
| Horário do exame | 10:00 AM |
Data da Defesa:
26/03/2026 - 10:00
Tipo de Defesa:
Qualificação de Mestrado