Seminar in Computational Linguistics

I will talk about my recent work on understanding neural machine translation (NMT). The first one is understanding NMT by simplification. We explore the importance of each component in NMT and focus on encoder-free models. We investigate what makes the performance drop and the word alignment in encoder-free models. The second one is about the process of disambiguating word senses in NMT. We feed encoder/decoder hidden states into a word sense disambiguation (WSD) classification task. We find that encoders achieve much better result than the baseline which uses word embeddings. In the end, I will talk about the ongoing work, the analysis on character-level NMT and on hidden states in NMT.