Compositional Neural Machine Translation by Removing the Lexicon from Syntax
This work addresses the challenge of compositional understanding in NLP, potentially benefiting translation and parsing tasks, though it appears incremental as it builds on existing LSTM frameworks.
The paper tackled the problem of separating lexical and syntactic knowledge in neural machine translation by proposing neural units that enforce this constraint over LSTM encoder-decoder models, achieving competitive performance in semantic parsing, syntactic parsing, and English-to-Mandarin translation with improvements over standard LSTM architectures on many metrics.
The meaning of a natural language utterance is largely determined from its syntax and words. Additionally, there is evidence that humans process an utterance by separating knowledge about the lexicon from syntax knowledge. Theories from semantics and neuroscience claim that complete word meanings are not encoded in the representation of syntax. In this paper, we propose neural units that can enforce this constraint over an LSTM encoder and decoder. We demonstrate that our model achieves competitive performance across a variety of domains including semantic parsing, syntactic parsing, and English to Mandarin Chinese translation. In these cases, our model outperforms the standard LSTM encoder and decoder architecture on many or all of our metrics. To demonstrate that our model achieves the desired separation between the lexicon and syntax, we analyze its weights and explore its behavior when different neural modules are damaged. When damaged, we find that the model displays the knowledge distortions that aphasics are evidenced to have.