Active learning is an ML technique where the model picks the most informative examples to label, cutting training data costs. See how it works in chatbots.
More about Active Learning
Active learning is a machine learning technique where the model itself decides which examples are most worth labelling, instead of being handed a fixed batch of labelled training data. The model looks at a large pool of unlabelled data, picks the items it is most unsure about, and sends only those to a human for labelling. The labelled results go back into training, the model updates, and the cycle repeats.
The payoff is efficiency. A well-designed active learning loop can reach the same accuracy with a fraction of the labelled data that a conventional supervised pipeline needs. For teams paying human annotators by the hour, that directly maps to savings.
How Active Learning Works
A standard active learning loop has five steps:
- Seed: train an initial model on a small labelled set.
- Score: run the current model across the unlabelled pool and score each example for informativeness.
- Select: pick the top-scoring examples, often the ones the model is least certain about.
- Label: send those to a human (or an oracle) to annotate.
- Retrain: add the new labels to the training set and update the model.
The selection step is where most of the design work happens. Common strategies include:
- Uncertainty sampling: pick examples where the model's predicted probability is closest to 50/50.
- Query-by-committee: train several models and pick examples where they disagree.
- Expected model change: pick examples that would shift the model's weights the most.
- Diversity sampling: cover the input space broadly instead of clustering around one uncertain region.
Active Learning for Chatbots
For AI chatbots and conversational agents, active learning fits naturally into the conversation stream. Every user message that the bot handles becomes a potential training signal. Instead of labelling all of them, the system picks the ones where:
- Confidence in intent classification is low.
- The retrieved answer looks irrelevant based on reranker scores.
- The user follows up with a correction or a frustrated reply, suggesting the bot got it wrong.
- Two similar questions produced very different answers.
These cases get surfaced to a reviewer, often through a human-in-the-loop interface, who corrects the label, the answer, or both. Over time the chatbot gets better at exactly the things it was failing at.
SiteSpeak gathers transcripts and per-message feedback so teams can run this kind of loop on top of their own chatbot, feeding corrections back into the knowledge base and improving retrieval quality without having to retrain the underlying model.
Active Learning vs. Passive Learning
Passive (or conventional supervised) learning uses whatever labelled data is available, regardless of how useful each example is. That works well when data is cheap and plentiful. It breaks down when labels are expensive, which is true for almost all real-world NLP work.
Active learning shines when:
- Unlabelled data is abundant but labelling costs are high.
- The domain is specific enough that public datasets are not a good match.
- A small team wants to keep the labelling workload sustainable.
It is less useful when the labelling task is trivial or when you already have millions of labelled examples.
Limitations
Active learning is not free:
- Cold start: you still need enough labelled data to train an initial model that produces meaningful uncertainty estimates.
- Biased sampling: always picking the "hardest" examples can ignore easy failure modes and overfit to edge cases.
- Overhead: running inference on the whole unlabelled pool every iteration is expensive.
In practice, teams mix active learning with random sampling to keep the training set balanced.
