Source of the 'gt' field in the Redial experimental data

[BIG9MONKEY]

Hello! Your work is of great reference value to me! May I ask how the 'gt' field in the Redial dataset (input_processed_*.jsonl) used in your experiments was derived from the original dataset? Good luck!

[kookeej]

Hello,
Thank you for your interest in our work.

We followed the preprocessing procedure described in the paper "Large Language Models as Zero-Shot Conversational Recommenders" (CIKM 2023). In that work, the ground-truth items were constructed by removing movies that had already been mentioned earlier in the dialogue.

We adopted the same approach. Specifically, we used the dataset available at:
https://huggingface.co/datasets/ZhankuiHe/redial_cikm

For each record, we removed any movie IDs in the rec field that also appeared in prev_entity. We then mapped the remaining movie IDs to their corresponding movie titles in the format Movie Name (Year).

Please feel free to reach out if you have any further questions😊