Abstract
When learning a language, people can quickly expand their understanding ofthe unknown content by using compositional skills, such as from two words "go"and "fast" to a new phrase "go fast." In recent work of Lake and Baroni (2017),modern Sequence-to-Sequence(seq2seq) Recurrent Neural Networks (RNNs) can makepowerful zero-shot generalizations in specifically controlled experiments.However, there is a missing regarding the property of such stronggeneralization and its precise requirements. This paper explores this positiveresult in detail and defines this pattern as the synonymous generalization, anability to recognize an unknown sequence by decomposing the difference betweenit and a known sequence as corresponding existing synonyms. To betterinvestigate it, I introduce a new environment called Colorful Extended CleanupWorld (CECW), which consists of complex commands paired with logicalexpressions. While demonstrating that sequential RNNs can perform synonymousgeneralizations on foreign commands, I conclude their prerequisites forsuccess. I also propose a data augmentation method, which is successfullyverified on the Geoquery (GEO) dataset, as a novel application of synonymousgeneralization for real cases.