Token-level QE model for partial QE and integrating it during decoding to improve translation quality.