Core Concepts
A unified biomedical named entity recognition solution that leverages instruction tuning on large language models and sequence labeling to handle diverse entities across multiple datasets.
Abstract
The paper introduces VANER, a novel biomedical named entity recognition (BioNER) model that integrates large language models (LLMs) to achieve versatility and adaptability.
Key highlights:
VANER employs instruction tuning to train on multiple BioNER datasets, allowing it to extract various types of entities. This overcomes the limitations of traditional BioNER methods that are task-specific and demonstrate poor generalizability.
VANER utilizes the open-source LLaMA2 model as the backbone and removes the causal mask during training and inference to enhance performance on sequence labeling tasks.
To address the lack of specialized medical knowledge in the backbone LLM, VANER integrates external entity knowledge bases and employs instruction tuning to compel the model to densely recognize curated entities.
VANER significantly outperforms previous LLM-based BioNER models and surpasses the majority of conventional state-of-the-art BioNER systems, achieving the highest F1 scores across three datasets.
VANER demonstrates robust domain adaptation capabilities, accurately recognizing entities in unseen datasets.
The model is resource-efficient, requiring only a single 4090 GPU for training and inference.
Stats
VANER achieves the highest F1 scores across three datasets: BC4CHEMD (93.18%), BC5CDR-chem (94.34%), and Linnaeus (94.06%).
On average, VANER outperforms all other models except BioLinkBERT.
VANER's Dense Bioentities Recognition (DBR) method significantly boosts performance, with an average F1 score improvement of 2.2 points compared to the model without DBR.
Quotes
"VANER, a novel versatile and adaptive BioNER model that integrates LLMs."
"VANER employs instruction tuning to train on multiple BioNER datasets, allowing it to extract various types of entities."
"VANER significantly outperforms previous LLM-based BioNER models and surpasses the majority of conventional state-of-the-art BioNER systems."