research/study on OL and prompt engineering for the project

[HamedBabaei]

• [x] Halard Sack course (only section 5 about OL ) : https://www.youtube.com/playlist?list=PLoOmvuyo5UAcBXlhTti7kzetSsi1PpJGR
• [x] paper lecture Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing : https://www.youtube.com/watch?v=TE6urdkTR4I

[HamedBabaei]

The diagram for the project. Arrows represent steps for ontology learning and since this is an investigation of LLMs I just put question marks and tasks that we should employ LLMs. This repository is for RQ2 and RQ3.2. I didn't notice that all could be in the same repository. So, I just put the diagram here for your reference @jd-coderepos we can talk about it at Friday or Monday's meeting.

• TaskA: answer RQ1
• TaskB: answer RQ2
• TaskC: answer RQ3.1
• TaskD: answer RQ3.2

[HamedBabaei]

Interesting Question or Finding that I think is worth talking about it:

According to the diagram and RQ2 (this is the same that we discussed at the last meeting about hierarchy -- subclasses detection). ‌In the end, RQ2 output would be generation hierarchies automatically. So we will have the hierarchies using classes from the last levels (here it is level 3).

Considering RQ1, the goal was to identify terms classes in representative levels (the same that this [1] paper does). Now if we form the hierarchies automatically, then I think we don't need to go to investigate entity classes in higher levels (levels 2, and 1 are higher levels and level 3 is a lower level). So we might only consider the classes in the lower level (level 3) and do the term class identification. And we know the hierarchies (according to the RQ2) so this is fine for us to move on with only entity class detection using level 3.

[HamedBabaei]

The main aim of Zero-shot learning is to gain the ability to predict the results without any training samples; the machine has to recognize the objects from classes that are not trained during training.

This with the OpenPrompt library will save us in the experimentation part if we consider only zero-shot inference for without fine-tuning part of our experiments to see how much LM knows the classes!