Exploring influence of language structures on neural networks
Speaker:
Vojtěch Hudeček
Abstract:
Common approaches assume that the neural model can find patterns and relationships in the data itself and is also capable of representing them reasonably. Our goal is to explore how much can information about linguistic structure help neural networks. It can be syntactic or morphological information. We thus want to create such networks that would take data's structure into account or reflect it directly in their architecture. First, we focus on models that are able to derive the structure information by themselves. To achieve this, we use tree structured data sets that provide us with the necessary information.