HILANCO

Hungarian Intelligent Language Applications Consortium

« Back to Home Page

HIL-SBERT

Sentence-BERT (SBERT) models are fine-tuned BERT networks aimed at obtaining high-quality sentence embeddings. SBERT was introduced in the original paper by Reimers and Gurevych (2019). We used multilingual knowledge distillation proposed by Reimers and Gurevych (2020) for creating a Hungarian model. The pre-trained hubert-base-cc (Nemeskey 2021) was fine-tuned on the Hunglish 2.0 parallel corpus (Varga et al. 2005) to mimic the bert-base-nli-stsb-mean-tokens model provided by UKPLab. Sentence embeddings were obtained by applying mean pooling to the huBERT output. The data was split into training (98%) and validation (2%) sets. By the end of the training, a mean squared error of 0.106 was computed on the validation set. Our code was based on the Sentence-Transformers library. Our model was trained for 2 epochs on a single GTX 1080Ti GPU card with the batch size set to 32. The training took approximately 15 hours.

To DOWNLOAD the models, please fill out the registration form: » REGISTRATION FORM «

References


More Language Models