KI-BERT: Infusing Knowledge Context for Better Language and Domain Understanding
This addresses the limitation of contextualized models in handling conceptual and ambiguous entities for NLP applications, though it is an incremental improvement over existing knowledge-aware methods.
The paper tackles the problem of transformer-based language models lacking knowledge context by proposing KI-BERT, which infuses knowledge from graphs like ConceptNet and WordNet during fine-tuning, significantly outperforming BERT and other variants on GLUE subtasks and domain-specific tasks.
Contextualized entity representations learned by state-of-the-art transformer-based language models (TLMs) like BERT, GPT, T5, etc., leverage the attention mechanism to learn the data context from training data corpus. However, these models do not use the knowledge context. Knowledge context can be understood as semantics about entities and their relationship with neighboring entities in knowledge graphs. We propose a novel and effective technique to infuse knowledge context from multiple knowledge graphs for conceptual and ambiguous entities into TLMs during fine-tuning. It projects knowledge graph embeddings in the homogeneous vector-space, introduces new token-types for entities, aligns entity position ids, and a selective attention mechanism. We take BERT as a baseline model and implement the "Knowledge-Infused BERT" by infusing knowledge context from ConceptNet and WordNet, which significantly outperforms BERT and other recent knowledge-aware BERT variants like ERNIE, SenseBERT, and BERT_CS over eight different subtasks of GLUE benchmark. The KI-BERT-base model even significantly outperforms BERT-large for domain-specific tasks like SciTail and academic subsets of QQP, QNLI, and MNLI.