NER - Best Approach for Extracting Physical Dimensions from Logistics Texts: BERT or LLM?

AI Discussions
,
1

Hello everyone,

I’m developing an application to extract physical dimensions from short, open-field logistics texts at my company. These texts typically have a median length of about 50 characters, although they can go up to 500 characters. Due to the variability in how these dimensions are described, I’m considering whether to fine-tune a BERT model or to use a large language model (LLM) for this task. Since fine-tuning an LLM isn’t an option due to resource constraints, I would appreciate your insights on which approach might be more effective.

Has anyone here conducted a comparison of Named Entity Recognition (NER) capabilities between BERT and LLMs for similar tasks? Any shared experiences or guidance would be very helpful.

Thanks in advance!

2

Hi!
I have used BERT variants for NER tasks. IMO, its always better to start with smaller model to make it work with your use case. The experiments are more controllable as well. Not much work has been done on NER using LLMs to concretely say they are much superior. You will find plenty of resources to get NER done using BERT variants as well, but not for LLMs.