Exploring the Correlation between Human and Machine Evaluation of Simultaneous Speech Translation

TL;DR

This paper investigates how well automatic metrics reflect human judgments in evaluating simultaneous interpretation, a task characterized by non-linearity and subjective expectations. It compares sentence-embedding based semantic similarity methods and GPT-3.5 prompting against human assessments across a small English-to-Spanish corpus, examining both human and machine translations. The key finding is that GPT-3.5 with direct prompting shows the strongest, most consistent correlation with human judgments, particularly when larger context windows are used, while traditional embedding methods display more variability. The work highlights practical potential for real-time evaluation and QA workflows, but acknowledges limitations such as low inter-rater reliability and limited data, suggesting avenues for future refinement and broader testing.

Abstract

Assessing the performance of interpreting services is a complex task, given the nuanced nature of spoken language translation, the strategies that interpreters apply, and the diverse expectations of users. The complexity of this task become even more pronounced when automated evaluation methods are applied. This is particularly true because interpreted texts exhibit less linearity between the source and target languages due to the strategies employed by the interpreter. This study aims to assess the reliability of automatic metrics in evaluating simultaneous interpretations by analyzing their correlation with human evaluations. We focus on a particular feature of interpretation quality, namely translation accuracy or faithfulness. As a benchmark we use human assessments performed by language experts, and evaluate how well sentence embeddings and Large Language Models correlate with them. We quantify semantic similarity between the source and translated texts without relying on a reference translation. The results suggest GPT models, particularly GPT-3.5 with direct prompting, demonstrate the strongest correlation with human judgment in terms of semantic similarity between source and target texts, even when evaluating short textual segments. Additionally, the study reveals that the size of the context window has a notable impact on this correlation.

Figures (4)

• Figure 1: Correlations among machine evaluation methods
• Figure 2: Correlations for Translation H and Translation M
• Figure 3: Correlations for Translation H according to window size
• Figure 4: Correlation for Translation M according to window size