PTM4Tag+: Tag Recommendation of Stack Overflow Posts with Pre-trained Models

A Stack Overflow post is not one blob of text: the title, the prose description, and the code block carry different signals, and flattening them into a single sequence throws that structure away. PTM4Tag+ keeps them separate, encoding each with its own pre-trained model in a triplet architecture before fusing the representations for tag prediction.

This extended version broadens the earlier ICPC study across a wider set of pre-trained models — BERT-family encoders alongside code-aware models such as CodeT5 — and reports which architectural choices actually pay for themselves. It also examines the cost side, since a triplet of PTMs is three times the inference budget of a single one.