基于词频效应控制的神经机器翻译用词多样性增强方法

Neural machine translation (NMT) optimized by maximum likelihood estimation is prone to problems such as unargmaxable tokens or poor accuracy of low-frequency words, which leads to the lack of word-level diversity in the generated translations. The unbalanced distribution of word frequency on the training data is one of the reasons for the above phenomenon. This paper aims to alleviate the above problems by limiting the impact of word frequency on the estimated probability when decoding NMT. Specifically, we adopt a denoising framework of Half-Sibling Regression based on causal inference theory, combined with the adaptive denoising coefficient proposed in this paper to control the effect of word frequency on estimated probability, in order to obtain more accurate model estimated probability, and enrich the diversity of the words used in NMT translations. The experiments in this paper are carried out on four translation tasks representing different resource scales: Uyghur-Chinese, Chinese-English, English-German and English-French. In addition, the proposed method is model-agnostic and interpretable.