The easiest eay to measure hallucinations is to ask Turkers to count incorrect statements in a text. However reproduction papers published at the Human Evaluartion workshop suggest that this is not a reliable way to measure hallucination. Hopefully researchers will switch to better ways to measure hallucination!

Many people asume that “human evaluation” means asking people to rate or rank outputs. However there are many other types of human evaluation, most of which give more meaningful results than rating or ranking! I discuss some of these, including task-based evaluation, annotation-based evaluation, and real-world evaluation.

Four of my MSc students did projects evaluating chatGPT in different use cases: developers assist, translation, exam taking, emotional support. In every use case, chatGPT was often very impressive, but also had issues and limitations.

An ACL paper from the PhilHuman project looks at using experts vs non-experts in human evaluation. It concludes that the agreement between experts and non-experts is worse for texts from GPT3 than texts from GPT2; in other words, non-expert evaluation is less useful for high-quality texts produced by recent LLMs.

Evaluation advice in one sentence: plan your experiments in advance, including details; keep your experiment as simple and standard as possible; do a pilot experiment first to make sure everything works; and be very careful when you run the main experiment.

We may see a big change in NLG evaluation over the next few years, with LLM-based evaluation replacing metrics such as BLEU and BLEURT, and a renewed emphasis on high-quality human evaluation to assess semantic and pragmatic correctness. Would be a step forward if this happens!

I love getting questions about how to evaluate chatGPT, they are much more constructive than speculations about whether it is a threat to humanity. We need to understand what LLM technology can and cannot do, and rigorous experiments are the best way to do this. I give some advice and caveats about evaluating chatGPT in this blog, and am happy to answer questions from people who want to do high-quality evaluations.

CSL journal has just published a paper “Evaluating factual accuracy in complex data-to-text”, which summarises our work in this area. I strongly recommend the paper to anyone who is interested in evaluating the accuracy of texts produced by neural NLG systems.

An example from MedPaLM highlighted to me that generated texts can contain information which is factually accurate but still not appropriate, because (in this case) of its negative psychological impact. There are other such cases, and we should ensure that our evaluation criteria are sensitive to them.

I was very impressed by a recent paper that compared prompting-based MT to MT based on trained models. Results are very interesting; prompting-based MT generates fluent texts which however have accuracy problems. Also the paper itself is an excellent example of a high-quality NLP evaluation, and I recommd it to anyone who wants to do good NLP evaluations.