Replace USE embeddings as basis for semantic edges

Universal Sentence Encoder (USE) embeddings are not granular enough in their semantic discrimination for creating accurate semantic edges between titles in all cases. In some cases USE embeddings reduce to keyword similarity (misleading keyword similarities in boldface), e.g.:

• Title: Enhancing Domain-Specific Supervised Natural Language Intent Classification with a Top-Down Selective Ensemble Model

  • Matched (wrong) topic: Time Series Classification Algorithms

• Title: Multivariate analyses of affix productivity in translated English

  • Matched (wrong) topic: Construction Labor Productivity and Neural Networks