Cost-effective speech-act extraction during training

Develop more cost-effective supervised or alternative speech-act extraction methods that maintain the accuracy of language-model-based extraction while reducing the time and computational costs incurred during training.

Background

Speech-act extraction is used only during training, which avoids inference-time latency, but the paper reports substantial computational cost for extracting speech acts with gpt-oss-120b. This remains a practical limitation for scaling the approach to larger datasets or resource-constrained settings.

The authors explicitly identify the development of lower-cost extraction alternatives, such as supervised speech-act extraction models, as unresolved future work, with the requirement that extraction accuracy be maintained.