I was very disappointed to realise that the evaluation suite for Amazon Nova (and I assume for other LLMs) has poor coverage of NLG tasks. Which is surprising since LLMs are largely used to generate texts; shouldnt they be evaluated, at least in part, on their ability to do this well?

I am very happy to see that the MT community is adopting the annotation-based MQM protocol as a gold-standard evalution technique. Having such a gold standard both strengthens evaluation and also supports exciting new research in evaluation.

In NLG we focus on quantitative evaluation, but qualitative techniques can also be used. Quantatitive hypothesis testing is essential, but its also really useful to ask people what they think of an NLG system in an open-ended way.

In early Sept I ran a one-day class on evaluation. I summarise what I did in this class and give links to my presentations, in case this is useful to other people.

I list five challenges to evaluating LLMs, which unfortunately seem to be ignored by many researchers. Which means that many published LLM evaluations cannot be trusted. This blog is based on a recent workshop talk.

I suspect we may be reaching the point where the most common type of human evaluation in NLG (ratings/rankings by crowdworkers or students) are less meaningful than evaluations using LLMs. But better forms of human evaluation, based on annotation or impact, are still very useful and give insights which we cannot get from LLMs.

In human evaluation, it is absolutely essential that subjects understand what they are supposed to do; otherwise evaluations will not be meaningful or replicable. This may sound obvious, but it was repeatedly raised as a concern in the replication shared task in the 2024 Human Evaluation workshop.

I present some suggestions for doing good evaluations, which are based on previous blogs I have written.

Data contamination (testing and evaluating LLMs using test data which is known the the LLM) may be a huge problem in NLP, leading to a lot of invalid scientific claims. Unfortunately, many NLP researchers ignore the problem, which is really worrying.

It is very rare to see evaluations in the NLP research literature which are based on measuring the impact of systems on real-world users. I’d love to see more such evaluations, and describe some ways of doing this, along with a few examples.