SLP-LMNMT: Source Language Prediction in Bilingual and Many-to-One Neural Machine Translation Tasks
Dongxing Li, Dongdong Guo · 2024
This paper proposes a novel approach, Source Language Prediction-Language Model for Neural Machine Translation (SLP-LMNMT), based on the UNILM’s sequence-to-sequence (seq2seq) model. This model is specifically designed for bilingual and many-to-one translation tasks, with a focus on predicting the source language type. It reframes NMT as a natural language understanding task by using a unified encoder and decoder structure, departing from the conventional encoder2decoder setup. Similar to BERT, SLP-LMNMT combines a restricted masked language model (R-MLM) to determine whether the masked tokens are from the source and target sequences, along with a Source Language Prediction (SLP) task to identify the source language type. Experiments demonstrate the efficiency of SLP-LMNMT, show significant improvements in translation performance compared to unidirectional models.