I have a simple question about Beam Search. I’m always under the impression that the scorer for Beam Search is typically some traditional n-gram based language model (such as KenLM). However, the scorer implemented by OpenNMT is GNMTGlobalScorer and referred to this paper (https://arxiv.org/pdf/1609.08144.pdf). Both didn’t seem to rely on outside language model!
It seems like the Scorer is more or less like a reweighting/re-normalization for the existing log probability generated by the decoder. Is it better than third-party LM (such as some neural LM or traditional LM)?