Improvement of Neural Reverse Dictionary by Using Cascade Forward Neural Network

: A reverse dictionary maps a description to the word specified by the description. The neural reverse dic- tionary (NRD) learns to map word embeddings for an input definition into an embedding of the word defined by the definition using neural networks. Such a function encodes phrasal semantics and bridges the gap between them and lexical semantics. However, previous NRD has a limitation in accuracy due to its insu ffi cient capacity. To solve this problem, we used novel combinations of neural networks, which are e ff ective in neural machine translation and image processing, with su ffi cient capacities. We found that, an LSTM output adjustment by using a multi-layer fully connected network with bypass structures (CFNN) was more e ff ective for reverse dictionary tasks than using more complicated LSTM. BiLSTM + CFNN was comparable to the commercial system OneLook Reverse Dictionary in some metrics, and noised biLSTM + CFNN which we tuned by a noising data augmentation outperformed OneLook Reverse Dictionary in almost all metrics. We also examined the reasons for the success of biLSTM + CFNN and revealed that a bypass structure of the CFNN and balance in the capacity of LSTM and the CFNN contribute to the improved performance of the NRD.