Week 4

Finished

Reading

Ten Lectures on Mind and Language
- Conceptual Integration
  - Mental spaces (input space, generic space, blended space, etc)
  - Cross-space mappings
- Types of conceptual integration
  - Simplex, Mirror, Simple-scope, Double-scope
Frame Blending

Generate Examples with Llama 2 - Zero-shot

It seems that llama2 react poorly to the zero-shot prompting, because it is not really blending the frames, but just using the frames.

zero-shot

Generate Examples with Llama 2 - One-shot

Here, I provided a example, “Time is money.”, to let llama2 understand more the intended output. Also, the chain of thought (input space - cross-space mapping - blended space analysis) is provided in the example so that llama2 can have a similar train of thought when generating the output.

one-shot_prompt

“Health is wealth.”, generated by llama2.

one-shot_1

“He is worth millions.”, generated by llama2.

one-shot_2

Additionally, I asked llama2 to generate an example blending the frame “physics” and the frame “family”, here is the result:

“The kids were running wild in the house like energy is loose.”

one-shot_physics_family

RAG with Llama-index

Without RAG, when we ask llama2 about specific questions about the FrameNet dataset, it would reply with seemingly related content, but not exactly consistent with FrameNet dataset. That would be an issue when we want to generate frame blending examples align with FrameNet dataset.

no_rag

Therefore, I built a llama-index application, taking the directory that contains all the .xml frame files as the input, and replying frame-specific questions with solid reference.

Now the reply aligns with the content of the Execute_plan frame in the FrameNet dataset.

rag

Other Messages From The Weekly Meeting

To-do List For Next Week

Weekly Meeting Slides Link

Red Hen Meeting - Jun 19

Week 3

Week 5