Shapeshifter and the impact of representation in classification

Abstract

This work explores the representation format of text during the classification process. We defined eight types of representations using graphs to study the impact of the representation in a model. We build the graphs based on the dependency tree and input them into the UGformer to classify the documents. As a result, we observed that the best result is always at least two percentage points higher than the worst result.

Tasks

Reproductions