Hierarchical Pre-training for Sequence Labelling in Spoken Dialog

TL;DR

This paper introduces a hierarchical transformer-based pre-training approach for sequence labeling in spoken dialog, evaluated on the new SILICONE benchmark, showing competitive performance with fewer parameters.

Contribution

It proposes a novel hierarchical encoder with extended pre-training objectives for spoken dialog, and introduces the SILICONE benchmark for evaluation.

Findings

Hierarchical encoders achieve competitive results with fewer parameters.

Pre-training on OpenSubtitles improves sequence labeling performance.

Hierarchical models are effective for both pre-training and fine-tuning in spoken dialog tasks.

Abstract

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.