Text this: P-BERT: Toward Long Sequence Modeling by Enabling Language Representation With Prefix Sequence Compression, Soft Position Embedding, and Data Augmentation for Patent Relevance Assessment