添加并删除 utterance 后LUIS意图分数未复原，此现象是否正常？

Great question—this behavior is totally normal for LUIS, and it all comes down to how its machine learning models work. Let me break it down simply:

Why the score doesn’t revert to the original 0.53 #

LUIS relies on statistical machine learning models to classify intents, and these models aren’t "reversible" in a strict sense. Here’s what’s happening behind the scenes:

• When you first trained your intent with the initial small set of utterances, the model learned basic patterns from that limited data, resulting in the 0.53 score.
• Adding a new utterance gave the model more context to learn from, so it adjusted its internal parameters to better fit the expanded dataset—hence the jump to 0.82.
• Even after deleting that extra utterance, your dataset is back to its original size, but the model doesn’t just "undo" the prior training. Each training run optimizes the model’s parameters based on the current dataset, and there’s inherent randomness in this process (like minor variations in weight initialization or optimization steps). This means the second training run (post-deletion) won’t produce exactly the same model as the first, leading to the 0.62 score instead of 0.53.

Why the score is higher than the initial 0.53 #

It’s also common to see a slightly elevated score after this cycle because the model already went through one round of training with more data. Even though you removed the extra utterance, the model’s parameters have already been tuned to better recognize patterns related to your intent, making it more stable than it was in the initial training with fewer examples.

Key takeaway #

Small score fluctuations when working with limited training data are totally expected in LUIS. To get more consistent results, focus on adding a diverse set of utterances for each intent—this gives the model more robust patterns to learn from, reducing the impact of minor dataset changes.

内容的提问来源于stack exchange，提问作者xiang xing