Improving Sequence-to-Sequence Learning via Optimal Transport
This addresses the issue of semantic coherence in sequence generation for NLP practitioners, though it is incremental as it builds on existing sequence-to-sequence frameworks.
The paper tackled the problem of sequence-to-sequence models failing to capture long-range semantic structure due to word-level maximum likelihood training, and introduced a novel supervision method based on optimal transport to impose global sequence-level guidance, resulting in consistent improvements across NLP tasks like machine translation, text summarization, and image captioning.
Sequence-to-sequence models are commonly trained via maximum likelihood estimation (MLE). However, standard MLE training considers a word-level objective, predicting the next word given the previous ground-truth partial sentence. This procedure focuses on modeling local syntactic patterns, and may fail to capture long-range semantic structure. We present a novel solution to alleviate these issues. Our approach imposes global sequence-level guidance via new supervision based on optimal transport, enabling the overall characterization and preservation of semantic features. We further show that this method can be understood as a Wasserstein gradient flow trying to match our model to the ground truth sequence distribution. Extensive experiments are conducted to validate the utility of the proposed approach, showing consistent improvements over a wide variety of NLP tasks, including machine translation, abstractive text summarization, and image captioning.