Curriculum-based transfer learning for an effective end-to-end spoken\n language understanding and domain portability

We present an end-to-end approach to extract semantic concepts directly from\nthe speech audio signal. To overcome the lack of data available for this spoken\nlanguage understanding approach, we investigate the use of a transfer learning\nstrategy based on the principles of curriculum learning. This approach allows\nus to exploit out-of-domain data that can help to prepare a fully neural\narchitecture. Experiments are carried out on the French MEDIA and PORTMEDIA\ncorpora and show that this end-to-end SLU approach reaches the best results\never published on this task. We compare our approach to a classical pipeline\napproach that uses ASR, POS tagging, lemmatizer, chunker... and other NLP tools\nthat aim to enrich ASR outputs that feed an SLU text to concepts system. Last,\nwe explore the promising capacity of our end-to-end SLU approach to address the\nproblem of domain portability.\n