2021/10/11 by Rose E. Wang, Wang, Rose E., Julia White +5
Computer Science · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Multiagent Systems (cs.MA) #Natural Language Processing Techniques #Speech and dialogue systems #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2110.05422
openalex publication_date 2021/10/11 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
To be good conversational partners, natural language processing (NLP) systems\nshould be trained to produce contextually useful utterances. Prior work has\ninvestigated training NLP systems with communication-based objectives, where a\nneural listener stands in as a communication partner. However, these systems\ncommonly suffer from semantic drift where the learned language diverges\nradically from natural language. We propose a method that uses a population of\nneural listeners to regularize speaker training. We first show that language\ndrift originates from the poor uncertainty calibration of a neural listener,\nwhich makes high-certainty predictions on novel sentences. We explore ensemble-\nand dropout-based populations of listeners and find that the former results in\nbetter uncertainty quantification. We evaluate both population-based objectives\non reference games, and show that the ensemble method with better calibration\nenables the speaker to generate pragmatic utterances while scaling to a large\nvocabulary and generalizing to new games and listeners.\n