Tweet-Inspired Intelligent Subselection of Semantically-Related Lyrical Training Data

Abstract

A current challenge in AI research is enabling AI systems to be inspired by external sources. We present a method for subselecting portions of a training corpus based on relevance to an external inspiring source. Our system takes an external, text-based inspiring source (e.g., tweet), extracts weighted lexical topics contained in the inspiring source, and uses these weighted topics to rank training instances in a corpus of song lyrics according to their relevance to the inspiring source. The system extends on the capabilities of the Empath framework by automatically generating domain-specific categories and mapping functions. The system offers a novel approach toward improved lexical semantic analyses for comparative corpus ranking.