Every word has synonyms in specific context.
For instance:
The athlete runs really “fast” => The athlete runs really "quick"
Is there a technical way using OpenNMT to make the decoder take this in consideration. I can use phrase_table to replace one by the other. But I prefer taking both choices and predict the best possible output .
Did you experience the decoder producing words that are not correct in the context? Usually this is what NMT is good at.
Actually I’m working on abstractive text summarization. I workin on making the output to be more abstractive (not totally constrained by the input). So I integrate lexical DB (WordNet). The two possible ways is applying the changes on the input or the output of the model (encoder-decoder).
But having a model independently capable of generating more alternative hypotheses (like replacing word by its synonym using an external DB) would be great (cleaner and faster) and this is my objective now.