24February2026
10:00 Master's Defense Room 85 of IC2
Topic on
Simplifying Commands in Portuguese for Voice Assistants through Language Models
Student
Fabrício Ferreira da Silva
Advisor / Teacher
Helio Pedrini
Brief summary
The increasing use of voice interfaces has transformed human-computer interaction, promising intuitive access to technology. However, a significant dissonance persists between the spontaneous natural language used by users (often verbose, contextual, and ambiguous) and the rigidity of traditional command recognition systems, which fail to interpret complex intentions. This work presents and evaluates a semantic rewriting architecture based on Large Language Models (LLMs), designed to act as an intermediate translation layer between user speech and task execution. The methodology was based on a comparative experimental design using the gpt-4o-mini model. The Command Engineering approach was compared against three supervised fine-tuning strategies. To this end, three training datasets with progressive volumes (50, 150, and 500 examples) were manually curated, structured to teach the model the desired simplification pattern—converting verbose requests into direct commands—without compromising its reasoning capabilities. An independent validation set, unseen during training, was specifically designed to test critical linguistic challenges such as logical negation, semantic inference, and parameter retention in sentences with multiple conditions. The LLM-as-a-Judge technique was employed to assign Likert scale scores (1-5) to the criteria of Preservation of Intention, Naturalness and Fluency, and Simplification Efficiency. Additionally, the applicability of the solution was validated in a realistic scenario: the validation sentences were audio-recorded and subjected to a controlled background noise environment (60-70 dB), and subsequently transcribed by an ASR (Google Speech-to-Text) system. The word error rate (WER) metric was calculated to measure the impact of transcription imperfections on the final performance of the rewriting models. The results revealed a fine-tuning paradox: increasing the data volume did not guarantee linear improvement. While the model with 150 examples reached the optimal point of semantic specialization, the model with 500 examples showed severe degradation, including hallucinations and loss of logical ability in negation tasks. The contributions of this study lie in the empirical evidence of the balance between general inference ability and model specialization, demonstrating that the curation of small datasets is more critical than volume for the alignment of virtual assistants.
Examination Board
Headlines:
| Hélio Pedrini | IC / UNICAMP |
| Rafael de Oliveira Werneck | IC / UNICAMP |
| César Henrique Córdova Quiroz | PUC-Campinas |
Substitutes:
| André Santanchè | IC / UNICAMP |
| Moacir Antonelli Ponti | ICMC / USP |