Why are accuracy and precision values better with entities than with intent?

I’m currently doing a article about nlu evaluation and i got curious about the results. Because entity accuracy, precision and f1 score is way bigger than with intents. Anyone have thoughts about the motive?


Hi @juliamrc, welcome to the forum :slight_smile: What you observe could happen because in your concrete dataset, intents are more difficult to detect than the entities. For example, overlapping intents that make use of similar vocabulary in different classes, would be one such issue. Without looking at the concrete data and your pipeline, it’s difficult to investigate though. If you’re able to share them, could you post your config.yml and your data files (nlu.yml, etc.)? Please enclose them in code blocks (using ```). Thanks!

1 Like