Mapping to Vector Representations Based on Distributed Semantics for Waka

This study proposes a framework for mapping waka poetry into vector representations based on distributional semantics, to evaluate the cultural similarity between anthologies at a macroscopic level. We compared these restructured results using the following three embeddings: i) fastText pre-trained on general Japanese, ii) fastText retrained on our proposed corpus, iii) a shallow encoder applying phrase-specific convolutional layers and 1 inter-phrase self-attention layer at the inter-phrase level. Each waka poem was collected into a poetry vector using methods such as simple averaging of word (or character) vectors. At the source level, poem vectors were averaged to define a center of gravity. The geometric structure was verified using the cosine similarity matrix of the collection and a two-dimensional projection.