A model that assigns one or more labels to an input — the fundamental pattern behind query intent models, spam detectors, sentiment models, and content detection.
A classifier is a model trained to assign a label — or set of labels — to an input. Given a search query, it outputs an intent category. Given a block of text, it outputs a sentiment score. Given a web page, it outputs whether links on it are organic or paid. Classification is the most common task in applied machine learning, and most of DEJAN's production models are classifiers of one form or another.
Binary classifiers output one of two labels. Multi-class classifiers output one label from a set of three or more. Multi-label classifiers can assign several labels simultaneously — a query might be both informational and local, for example. Token classifiers assign a label to each individual token in a sequence rather than to the whole input; LinkBERT and LinkjeBERT work this way, labelling each token as O (not a link position) or LINK.
Modern text classifiers are almost always built by fine-tuning a pre-trained transformer encoder. A classification head — typically a single linear layer — is added on top of the encoder's output representation, and the whole model is trained on labelled examples. The encoder learns to produce representations that make the classification head's job easy; the head learns the decision boundary.
DEJAN's classifier portfolio covers query intent (ALBERT-based, 8 labels), query well-formedness (ALBERT-based, binary), AI content detection (DeBERTa-v3, binary), link spam detection (binary), sentiment (7-point multi-label), and the grounding classifier (DeBERTa-based, binary: deserves grounding vs. does not). Each is a specialist — trained on one task, deployed for that task only.
