Inspired by KeyBERT, KeyBERTVi implements a similar keyword extraction technique that leverages the embeddings of PhoBERT and minimal linguistics properties to extract keywords and keyphrases that are most similar to the document.

This implementation took inspiration from the simple yet intuitive and powerful method of KeyBERT, applied for the Vietnamese language. PhoBERT are used to generate both document-level embeddings and word-level embeddings for extracted N-grams. Cosine similarity is then used to compute which N-grams are most similar to the document-level embedding, thus can be perceived as most representative of the document. Preprocessing catered to the Vietnamese language was applied.

Test with your own documents at KeyBERTVi Space.

You can use existing pre-trained models in the repo or download your own and put them in pretrained-models folder.

Note: dir_path is the absolute path to the repo.

As PhoBERT requires VnCoreNLP as part of pre-processing, the folder pretrained-models/vncorenlp is required. To download your own:

More information needed

More information needed