LIT: Label-Informed Transformers on Token-Based Classification

Transformer-based language models have led to the investigation of various embedding and modeling techniques for several downstream natural language processing tasks. Nevertheless, the comprehensive exploration of semantic information about the label from encoder and decoder components in these tasks is yet to be fully realized. In this paper, we propose LIT, an end-to-end pipeline architecture that integrates the transformer’s encoder-decoder mechanism with an additional label semantic to token classification tasks (i.e., historical named entity recognition (NER) and automatic term extraction (ATE)). Our findings demonstrate that LIT outperforms the benchmark in F1 with a maximal rise of 9.5% points in the historical NER task and 11.2% points in the ATE task for the gold standard excluding named entities.