Fast Domain Adaptation For Goal-Oriented Dialogue Using A Hybrid Generative-Retrieval Transformer

2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) |

Goal-oriented dialogue systems are now widely adopted in industry, where practical aspects of using them becomes of key importance. As such, it is expected from such systems to fit into a rapid prototyping cycle for new products and domains. For data-driven dialogue systems (especially those based on deep learning) that amounts to maintaining production-level performance having been provided with a few ‘seed’ dialogue examples, normally referred to as data efficiency. With extremely data-dependent deep learning methods, the most promising way to achieve practical data efficiency is transfer learning—i.e., leveraging a greater, highly represented data source for training a base model, then fine-tuning it to available in-domain data. In this paper, we present a hybrid generative-retrieval model that can be trained using transfer learning. By using GPT-2 as the base model and fine-tuning it to the multidomain MetaLWOz dataset, we obtain a robust dialogue model able to perform both response generation and ranking 1 . Combining both, it outperforms several competitive generative-only and retrieval-only baselines, measured by language modeling quality on MetaLWOz as well as in goal- oriented metrics (Intent/Slot Fl-scores) on the MultiWoz corpus.

Publication Downloads

MetaLWOz: A Dataset of Multi-Domain Dialogues for the Fast Adaptation of Conversation Models

July 29, 2019

We introduce the Meta-Learning Wizard of Oz (MetaLWOz) dialogue dataset for developing fast adaptation methods for conversation models. This data can be used to train task-oriented dialogue models, specifically to develop methods to quickly simulate user responses with a small amount of data. Such fast-adaptation models fall into the research areas of transfer learning and meta learning. The dataset consists of 37,884 crowdsourced dialogues recorded between two human users in a Wizard of Oz setup, in which one was instructed to behave like a bot, and the other a true human user. The users are assigned a task belonging to a particular domain, for example booking a reservation at a particular restaurant, and work together to complete the task. Our dataset spans 47 domains having 227 tasks total. Dialogues are a minimum of 10 turns long.