Unsupervised Learning of Task-Specific Tree Structures with Tree-LSTMs

For years, recursive neural networks (RvNNs) have shown to be suitable for representing text into fixed-length vectors and achieved good performance on several natural language processing tasks. However, the main drawback of RvNN is that it requires explicit tree structure (e.g. parse tree), which makes data preparation and model implementation hard. In this paper, we propose a novel tree-structured long short-term memory (Tree-LSTM) architecture that efficiently learns how to compose task-specific tree structures only from plain text data. To achieve this property, our model uses Straight-Through (ST) Gumbel-Softmax estimator to decide the parent node among candidates and to calculate gradients of the discrete decision. We evaluate the proposed model on natural language interface and sentiment analysis and show that our model outperforms or at least comparable to previous Tree-LSTM-based works. Especially in the natural language interface task, our model establishes the new state-of-the-art accuracy of 85.4%. We also find that our model converges significantly faster and needs less memory than other models of complex structures.