Reducing Conversational Agents’ Overconfidence Through Linguistic Calibration
نویسندگان
چکیده
Abstract While improving neural dialogue agents’ factual accuracy is the object of much research, another important aspect communication, less studied in setting dialogue, transparency about ignorance. In this work, we analyze to what extent state-of-the-art chit-chat models are linguistically calibrated sense that their verbalized expression doubt (or confidence) matches likelihood model’s responses factually incorrect correct). We find these poorly calibrated, yet show correctness can accurately be predicted. By incorporating such metacognitive features into training a controllable generation model, obtain agent with greatly improved linguistic calibration.
منابع مشابه
Reducing Overconfidence in Spreadsheet Development
Despite strong evidence of widespread errors, spreadsheet developers rarely subject their spreadsheets to post-development testing to reduce errors. This may be because spreadsheet developers are overconfident in the accuracy of their spreadsheets. This conjecture is plausible because overconfidence is present in a wide variety of human cognitive domains, even among experts. This paper describe...
متن کاملOverconfidence by Bayesian-Rational Agents
T paper derives two mechanisms through which Bayesian-rational individuals with differing priors will tend to be relatively overconfident about their estimates and predictions, in the sense of overestimating the precision of these estimates. The intuition behind one mechanism is slightly ironic: In trying to update optimally, Bayesian agents overweight information of which they overestimate the...
متن کاملDebiasing Expert Overconfidence: A Bayesian Calibration Model
In a decision and risk analysis, experts may provide subjective probability distributions that encode their beliefs about future uncertain events. For continuous variables, experts often provide these judgments in the form of quantiles of the distribution (e.g., 5th, 50th, and 95th percentiles). Psychologists have shown, though, that such subjective distributions tend to be too narrow, represen...
متن کاملConversational Agents
This paper describes a conversational agent for language learning. Our goal is to provide an environment that allows learners to practice language skills by interacting with other learners and artificial conversational agents using instant messaging and email. The learners take part in an adventure story and must use dialogues that are typically found in basal language textbooks to move forward...
متن کاملConversational Agents
Conversational agents integrate computational linguistics techniques with the communication channel of the Web to interpret and respond to statements made by users in ordinary natural language. Web-based conversational agents deliver high-volumes of interactive text-based dialogs. Recent years have seen significant activity in enterprise-class conversational agents. This chapter describes the p...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
ژورنال
عنوان ژورنال: Transactions of the Association for Computational Linguistics
سال: 2022
ISSN: ['2307-387X']
DOI: https://doi.org/10.1162/tacl_a_00494