BERT Based Embedding Model

AbstractMany pairwise classification tasks, such as paraphrase detection and open-domain question answering, naturally have extreme label imbalance (e.g., 99.99% of examples are negatives). In contrast, many recent datasets heuristically choose …...