On the Hallucination in Simultaneous Machine Translation

TL;DR

This paper analyzes hallucination issues in Simultaneous Machine Translation, revealing how target-side information influences hallucination and suggesting that reducing reliance on target context can mitigate the problem.

Contribution

It provides a comprehensive analysis of hallucination in SiMT, focusing on distribution and target-side context, and proposes a method to alleviate hallucination by limiting target information usage.

Findings

Hallucination words are influenced by target-side context.

Reducing target-side information decreases hallucination.

Understanding hallucination distribution aids in developing better SiMT models.

Abstract

It is widely known that hallucination is a critical issue in Simultaneous Machine Translation (SiMT) due to the absence of source-side information. While many efforts have been made to enhance performance for SiMT, few of them attempt to understand and analyze hallucination in SiMT. Therefore, we conduct a comprehensive analysis of hallucination in SiMT from two perspectives: understanding the distribution of hallucination words and the target-side context usage of them. Intensive experiments demonstrate some valuable findings and particularly show that it is possible to alleviate hallucination by decreasing the over usage of target-side information for SiMT.