The sequence to sequence Transformer model that generates translations can sometimes output gibberish, especially when it has a difficult time interpreting the input. We’ve experimented with different strategies for translating symbols and training models on more and better data should also improve this issue.
1 Like