Hierarchical Pre-training for Sequence Labelling in Spoken Dialog
This work addresses sequence labeling tasks like Dialog Act and Emotion/Sentiment identification for spoken dialog systems, presenting an incremental improvement with a new benchmark and method.
The authors tackled the problem of learning generic representations for sequence labeling in spoken dialog by proposing a hierarchical encoder based on transformers, pre-trained on OpenSubtitles with 2.3 billion tokens, and evaluated it on a new benchmark called SILICONE with 10 datasets, achieving competitive results with fewer parameters compared to state-of-the-art models.
Sequence labelling tasks like Dialog Act and Emotion/Sentiment identification are a key component of spoken dialog systems. In this work, we propose a new approach to learn generic representations adapted to spoken dialog, which we evaluate on a new benchmark we call Sequence labellIng evaLuatIon benChmark fOr spoken laNguagE benchmark (\texttt{SILICONE}). \texttt{SILICONE} is model-agnostic and contains 10 different datasets of various sizes. We obtain our representations with a hierarchical encoder based on transformer architectures, for which we extend two well-known pre-training objectives. Pre-training is performed on OpenSubtitles: a large corpus of spoken dialog containing over $2.3$ billion of tokens. We demonstrate how hierarchical encoders achieve competitive results with consistently fewer parameters compared to state-of-the-art models and we show their importance for both pre-training and fine-tuning.