Named Entity Recognition (NER) stands out as one of the primary applications of Enterprise AI. Particularly in data processing, the synergy between NER and transcription is evident – after analyzing documents, key terms can be extracted. Additionally, Relation Extraction (RE) and Event Extraction (EE) complement NER in these applications.
Before the surge of Large Language Models (LLMs), training a language model, typically a transformer, required laborious efforts. Each entity necessitated the painstaking process of labeling a few hundred examples, assessing accuracy, and undergoing multiple iterations until the desired accuracy for every label was achieved.
LLMs simplify this process significantly. Depending on the task’s complexity, a zero-shot natural language prompt can be employed to extract key entities, such as extracting names and addresses from a document.
In dealing with more intricate problems, few-shot learning becomes more beneficial. This involves presenting the LLM with 4-5 examples of documents and their extracted entities, tasking it with extracting similar entities from new documents. Few-shot learning with LLM prompting proves more straightforward than fine-tuning NER models.
If these approaches fall short, resorting to supervised fine-tuning (SFT) becomes necessary. SFT is particularly effective when dealing with specific (dense) extractions and is complemented by techniques like using code for entity extraction and data augmentation.
A comprehensive exploration of these methods and techniques is presented in the survey paper titled “Large Language Models for Generative Information Extraction: A Survey” (link provided).
Several existing models in production exhibit subpar accuracy in NER, necessitating human-in-the-loop supervision and correction – a costly and labor-intensive pipeline. It is now imperative to transition to LLMs, which, in many cases, can eliminate the need for human intervention and streamline the extraction process.