0:00
/
0:00
Transcript

"A Rose by Any Other Name: LLM-Generated Explanations Are Good Proxies for Human Explanations to Collect Label Distributions on NLI"

Generated below podcast on this paper with Google's Illuminate.

Teaching AI to explain itself - turns out it's pretty good at it!

LLMs can generate high-quality explanations for Natural Language Inference tasks that match human explanations in effectiveness while reducing annotation costs.

-----

https://arxiv.org/abs/2412.13942

Original Problem 🤔:

Collecting human explanations for Natural Language Inference (NLI) tasks is expensive and time-consuming. While LLMs can approximate human judgment distributions using explanations, they still require human-written explanations.

-----

Solution in this Paper 🛠️:

→ The paper proposes using LLM-generated explanations instead of human explanations for NLI tasks

→ It introduces two strategies: Label-Free (one explanation per NLI label) and Label-Guided (explanations based on human annotations)

→ The system leverages multiple LLMs like Llama3-Chat-70b and GPT4 to generate diverse explanations

→ These explanations are combined with a few human labels to approximate human judgment distributions

-----

Key Insights 💡:

→ LLM explanations perform similarly to human explanations when guided by human labels

→ The method generalizes well to datasets without explanations

→ Explanation variability serves as an indicator for human label variation

→ The approach works effectively on out-of-distribution test sets

-----

Results 📊:

→ Achieves comparable KL divergence (0.234) to human explanations (0.238)

→ Shows robust performance on ANLI dataset with 35.1% F1 score

→ Demonstrates successful generalization to MNLI dataset without explanations

→ Maintains effectiveness across different LLM architectures