Abstract
While learning models are typically studied for inputs in the form of a fixeddimensional feature vector, real world data is rarely found in this form. Inorder to meet the basic requirement of traditional learning models, structuraldata generally have to be converted into fix-length vectors in a handcraftedmanner, which is tedious and may even incur information loss. A common form ofstructured data is what we term "semantic tree-structures", corresponding todata where rich semantic information is encoded in a compositional manner, suchas those expressed in JavaScript Object Notation (JSON) and eXtensible MarkupLanguage (XML). For tree-structured data, several learning models have beenstudied to allow for working directly on raw tree-structure data, However suchlearning models are limited to either a specific tree-topology or a specifictree-structured data format, e.g., synthetic parse trees. In this paper, wepropose a novel framework for end-to-end learning on generic semantictree-structured data of arbitrary topology and heterogeneous data types, suchas data expressed in JSON, XML and so on. Motivated by the works in recursiveand recurrent neural networks, we develop exemplar neural implementations ofour framework for the JSON format. We evaluate our approach on several UCIbenchmark datasets, including ablation and data-efficiency studies, and on atoy reinforcement learning task. Experimental results suggest that ourframework yields comparable performance to use of standard models withdedicated feature-vectors in general, and even exceeds baseline performance incases where compositional nature of the data is particularly important. The source code for a JSON-based implementation of our framework along withexperiments can be downloaded at https://github.com/EndingCredits/json2vec.