Human-centric Dialog Training via Offline Reinforcement Learning
Autor: | Noah Jones, Natasha Jaques, Craig Ferguson, Asma Ghandeharioun, Shixiang Shane Gu, Judy Hanwen Shen, Agata Lapedriza, Rosalind W. Picard |
---|---|
Jazyk: | angličtina |
Rok vydání: | 2020 |
Předmět: |
FOS: Computer and information sciences
Vocabulary Computer Science - Machine Learning Computer science media_common.quotation_subject 050801 communication & media studies 010501 environmental sciences Machine learning computer.software_genre 01 natural sciences Machine Learning (cs.LG) Laughter 0508 media and communications Similarity (psychology) Reinforcement learning Dialog box 0105 earth and related environmental sciences media_common Class (computer programming) Computer Science - Computation and Language business.industry 05 social sciences Action (philosophy) Language model Artificial intelligence business Computation and Language (cs.CL) computer Generative grammar |
Zdroj: | EMNLP (1) |
Popis: | How can we train a dialog model to produce better conversations by learning from human feedback, without the risk of humans teaching it harmful chat behaviors? We start by hosting models online, and gather human feedback from real-time, open-ended conversations, which we then use to train and improve the models using offline reinforcement learning (RL). We identify implicit conversational cues including language similarity, elicitation of laughter, sentiment, and more, which indicate positive human feedback, and embed these in multiple reward functions. A well-known challenge is that learning an RL policy in an offline setting usually fails due to the lack of ability to explore and the tendency to make over-optimistic estimates of future reward. These problems become even harder when using RL for language models, which can easily have a 20,000 action vocabulary and many possible reward functions. We solve the challenge by developing a novel class of offline RL algorithms. These algorithms use KL-control to penalize divergence from a pre-trained prior language model, and use a new strategy to make the algorithm pessimistic, instead of optimistic, in the face of uncertainty. We test the resulting dialog model with ratings from 80 users in an open-domain setting and find it achieves significant improvements over existing deep offline RL approaches. The novel offline RL method is viable for improving any existing generative dialog model using a static dataset of human feedback. To appear in EMNLP 2020 (long paper) |
Databáze: | OpenAIRE |
Externí odkaz: |