The last results of the pipeline, based on character embeddings, are disappointing (the definitions almost systematically start with "A small tree of the..."). It might be necessary to plug external embeddings in the decoder, or to include a language model somehow, especially seeing that generated sequences include ungrammatical subsequences like "of of of a a".
The last results of the pipeline, based on character embeddings, are disappointing (the definitions almost systematically start with "A small tree of the..."). It might be necessary to plug external embeddings in the decoder, or to include a language model somehow, especially seeing that generated sequences include ungrammatical subsequences like "of of of a a".