End-to-end SLU is having a moment, but treating sequence labeling as sequence prediction quietly throws away decades of well-understood token-level tagging machinery. This paper proposes a compositional end-to-end SLU design that keeps the simplicity of E2E while giving you back the token-level formulation, complete with CRF-style globally normalized losses and clean per-component performance monitoring. For teams weighing cascaded ASR-plus-NLU pipelines against direct E2E, it is a pragmatic middle path.

The trick is inserting an intermediate ASR-trained decoder that converts speech into token-level representations before the NLU tagger runs, so you inherit compatibility with pretrained ASR and NLU components without paying the error-propagation tax of a hard cascade. On named entity recognition across SLU benchmarks, the compositional models beat both cascaded and direct E2E baselines. The deep-dive walks through the architecture, the training signal design, how globally normalized CRF losses slot back into the picture, and why per-component performance monitoring matters when you actually have to debug this thing in production. If you are shipping voice assistants, call-center analytics, or any product where slot filling and intent detection sit on top of ASR, this is worth a watch when compositional E2E is on your roadmap.