Neural hidden Markov model for machine translation
Wang, Weiyue; Ney, Hermann (Thesis advisor); van Genabith, Josef (Thesis advisor); Hoos, Holger Hendrik (Thesis advisor)
Aachen : RWTH Aachen University (2023)
Dissertation / PhD Thesis
Dissertation, RWTH Aachen University, 2023
Recently, neural machine translation systems have shown promising performance, outperforming phrase-based systems, which were the state-of-the-art in statistical machine translation for more than 10 years. Regardless of whether a recurrent neural network with long short-term memory or a convolutional neural network or a self-attentive transformer network is used, the attention mechanism is always one of the key components that all modern neural machine translation systems contain. In this work, we propose a completely novel neural architecture for machine translation, which is referred to as a direct hidden Markov model, as an alternative to attention-based systems.An attention component helps an encoder-decoder model attend to specific positions on the source side to produce a translation. In this way, the translation performance is significantly improved. This mechanism captures the correspondence between the source and target hidden states, and has a similar functionality to an alignment model in a phrase-based machine translation system. However, recent studies have found that using attention weights straight out of the box to align words results in poor alignment quality. This inspires us to introduce an explicit alignment model into the neural architecture in order to improve the alignment and thus also the translation quality of the overall system. To this end, we propose to use the concept of the hidden Markov model from statistical machine translation, which is made up of a lexicon model and an alignment model. In the neural hidden Markov model, the lexicon and alignment probabilities are modeled by neural networks, and the alignment is modeled from target to source direction so that it can be used directly in the forward translation as an attention component. The two models are trained jointly with the forward-backward algorithm and the end-to-end training process of a neural machine translation system is not violated. Various neural network architectures can be used to model the lexicon and the alignment probabilities. We start with simple feedforward neural networks and apply our first model to re-rank n-best lists generated by phrase-based systems and observe significant improvements. In order to build a monolithic direct hidden Markov model, the more powerful recurrent neural networks with long short-term memory are applied to the architecture, and a standalone decoder is implemented. By replacing the attention mechanism with a first-order alignment model, we achieve comparable performance to the baseline attention model while significantly improving the alignment quality. The improvements in alignment quality do not lead to an improvement in translation performance, one possible reason is that the attention mechanism is aimed at finding a relevant context for estimating the next target word and therefore source words with high attention weights are not necessarily translation equivalents of the target word. To keep pace with the development of neural machine translation, we also study the possibility of applying the transformer architecture to the direct hidden Markov model. Contrary to the recurrent neural network case, here we do not completely replace the attention mechanism with the alignment model. Instead, we combine the alignment information obtained by the hidden Markov model factorization with the attention mechanism. The experimental results show that applying the concept of the direct hidden Markov model significantly improves the performance of the state-of-the-art self-attentive transformer architecture in terms of TER and CHARACTER scores. In addition to the work on the direct hidden Markov model, we propose two novel metrics for machine translation evaluation, called CHARACTER and EED. These are easy-to-use and publicly available to everyone. They perform promisingly in the annual WMT metrics shared tasks and are always among the front runners.
- Department of Computer Science 
- Chair of Computer Science 6 (Machine Learning and Reasoning)