- NLI datasets must be carefully designed to avoid biased or inaccurate results.
- Lexical ambiguity, human judgment disagreements, and lack of knowledge in models are common issues in existing NLI datasets.
- Pragmatic inferences are essential for understanding language usage and should be incorporated into NLI datasets.
- Diverse and well-designed datasets can lead to more accurate and reliable machine learning models.
- The authors propose a new approach to NLI dataset design that prioritizes pragmatic inferences.
Computation and Language, Computer Science