It HAS to be Subjective: Human Annotator Simulation via Zero-shot Density Estimation

TL;DR

This paper presents a novel zero-shot density estimation framework for human annotator simulation that models human variability, enabling efficient generation of human-like annotations and better mimicking human judgment diversity.

Contribution

It introduces a meta-learning approach with new model classes, conditional integer flows and softmax flows, to simulate human annotations considering ordinal and categorical data.

Findings

Outperforms existing methods in predicting human annotation distributions

Effectively captures inter-annotator disagreements

Demonstrates superior efficiency on real-world tasks

Abstract

Human annotator simulation (HAS) serves as a cost-effective substitute for human evaluation such as data annotation and system assessment. Human perception and behaviour during human evaluation exhibit inherent variability due to diverse cognitive processes and subjective interpretations, which should be taken into account in modelling to better mimic the way people perceive and interact with the world. This paper introduces a novel meta-learning framework that treats HAS as a zero-shot density estimation problem, which incorporates human variability and allows for the efficient generation of human-like annotations for unlabelled test inputs. Under this framework, we propose two new model classes, conditional integer flows and conditional softmax flows, to account for ordinal and categorical annotations, respectively. The proposed method is evaluated on three real-world human evaluation tasks and shows superior capability and efficiency to predict the aggregated behaviours of human annotators, match the distribution of human annotations, and simulate the inter-annotator disagreements.