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

Tags are how Stack Overflow routes a question to people who can answer it, so bad tags are a retrieval problem for everyone downstream. Earlier recommenders treated a post as a single text field; PTM4Tag instead encodes its title, description, and code with three independent pre-trained models and combines them.

Keeping the components separate matters because they are different languages: prose and source code reward different pre-training. The paper reports which PTM to use for which component, and shows the triplet design outperforming single-encoder baselines on tag recommendation.