Recommendation in Online Q&A Communities Based on BERT Pre-training Technique

Open-source and online Q&A communities make use of tags and keywords for indexing, classification, and thematic search. In this study, we propose TagBERT, a novel model for recommending tags on new questions, which makes use of a combination of deep learning and BERT models. In this model, first, the processed sentences are converted into numerical vectors by means of the BERT tokenizer. Next, the attributes are extracted using the CNN network, and, afterward, the DNN network is trained on the extracted attributes in order to recommend tags. To evaluate our model, we used four datasets, i.e. Free-code, UNIX, WordPress, and Software Engineering. Our proposed model obtained the highest precision score over baseline deep-learning and conventional methods. In contrast to previous studies in which precision was significantly reduced as a result of increased recommended tags, the precision of our model did not remarkably vary with an increase in the number of tags.