On the Benchmarking of LLMs for Open-Domain Dialogue Evaluation

TL;DR

This paper critically examines current benchmarks for open-domain dialogue evaluation, revealing that they rely on outdated datasets and fail to accurately assess modern chatbot capabilities, especially in detecting deficiencies.

Contribution

It highlights the limitations of existing evaluation benchmarks and provides empirical evidence that current LLM evaluators struggle to identify issues in recent chatbot dialogues.

Findings

Existing benchmarks use outdated datasets and metrics.

LLM evaluators like GPT-4 have difficulty detecting deficiencies.

Current evaluation methods do not fully reflect modern chatbot capabilities.

Abstract

Large Language Models (LLMs) have showcased remarkable capabilities in various Natural Language Processing tasks. For automatic open-domain dialogue evaluation in particular, LLMs have been seamlessly integrated into evaluation frameworks, and together with human evaluation, compose the backbone of most evaluations. However, existing evaluation benchmarks often rely on outdated datasets and evaluate aspects like Fluency and Relevance, which fail to adequately capture the capabilities and limitations of state-of-the-art chatbot models. This paper critically examines current evaluation benchmarks, highlighting that the use of older response generators and quality aspects fail to accurately reflect modern chatbot capabilities. A small annotation experiment on a recent LLM-generated dataset (SODA) reveals that LLM evaluators such as GPT-4 struggle to detect actual deficiencies in dialogues generated by current LLM chatbots.