If I train "lstm1" instead of "bidi", it is almost always missing the last character in the output sentence. Is it expected behaviour?