npsm logo

Download original image
Fig. 2. (Color online) The Bi-GRU structural model is trained using a triplet loss. The list of sentence embeddings created by Sentence-BERT is linked as residuals and used as input to the Bi-GRU model with Bahdanau attention. The learning parameters are adjusted to ensure that embeddings of documents with the same ID are closer together than those with different IDs.
New Phys.: Sae Mulli 2023;73:385~394