Named entity recognition (NER) is usually the basis of transforming unstructured data (text) into structured data (say, in JSON format). The current best approach for NER in ChatGPT is few-shot-learning, where we give a few examples and asks for it to solve ONE instance of the task. The problem with this approach is the cost. We need to use a lot of tokens from examples in order to predict a single instance. I'm trying to overcome this issue by giving more than one instance for prediction. Something like this:
``` <instruction> <1 - example> <2 - example> <3 - new instance> <4 - new instance>
<answer for 1> <answer for 2> ```
Wishing the generator to predict answers for instances 3 and 4.
From my experience, results are worst than predicting a single instance.
Does anyone have better results in this task?
Thanks!
0