The Roles of Language Models and Hierarchical Models in Neural Sequence-to-Sequence Prediction

With the advent of deep learning, research in many areas of machine learning is converging towards the same set of methods and models. For example, long short-term memory networks (Hochreiter and Schmidhuber, 1997) are not only popular for various tasks in natural language processing (NLP) such as speech recognition, machine translation, handwriting recognition, syntactic parsing, etc., but they are also applicable to seemingly unrelated fields such as bioinformatics (Min et al., 2016). Recent advances in contextual word embeddings like BERT (Devlin et al., 2019) boast with achieving state-of-the-art results on 11 NLP tasks with the same model. Before deep learning, a speech recognizer and a syntactic parser used to have little in common as systems were much more tailored towards the task at hand. At the core of this development is the tendency to view each task as yet another data mapping problem, neglecting the particular characteristics and (soft) requirements that tasks often have in practice. This often goes along with a sharp break of deep learning methods with previous research in the specific area. This thesis can be understood as an antithesis to the prevailing paradigm. We show how traditional symbolic statistical machine translation (Koehn, 2009) models can still improve neural machine translation (Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014; Bahdanau et al., 2015, NMT) while reducing the risk of common pathologies of NMT such as hallucinations and neologisms. Other external symbolic models such as spell checkers and morphology databases help neural models to correct grammatical errors in text.