Grammatically-Interpretable Learned Representations in Deep NLP Models

NIPS 2017, Workshop |

We introduce two architectures, the Tensor Product Recurrent Network (TPRN)
and the Tensor Product Generation Network (TPGN). In the application of TPRN,
internal representations — learned by end-to-end optimization in a deep neural network
performing a textual QA task — are interpretable using basic concepts from
linguistic theory. This interpretability is achieved without paying a performance
penalty. In another application, image-to-text generation or image captioning,
TPGN gives better results than the state-of-the-art long short-term memory (LSTM)
based approaches. Learned internal representations in the TPGN can also be
interpreted as containing grammatical-role information.