Tree Transduction as a Means for Efficient Sentence Simplification

The use of natural language descriptions as a medium for artists and designers to implement text-based interfaces shows promising results and, as such, this paper examines the possibility of using sentence simplification to accomplish that specific goal. In particular, it has the aim of examining the effectiveness of using tree transduction to remove visually irrelevant tokens from a given sentence. Said goal is accomplished through running trials using three LSTM models - for monolingual machine translation, POS tags and tree transduction. The results show that the tree transduction model has the best performance for the metrics of simplicity, grammatical correctness and preservation of meaning, while the POS tags model performs best in terms of efficiency.