LG AI Research Proposes an Advanced Question-Answering (QASA) Approach and Benchmark for Scientific Papers

The process of reading scientific papers goes beyond simple information acquisition, inducing deep questions that require associative thinking and logical reasoning. Existing QA tasks have limitations in performing such complex reasoning and rich evidence extraction.

To address this, the QASA (Question Answering on Scientific Articles) approach is proposed. This methodology decomposes the QA task into three sub-tasks.

• Associative Selection: Extracts paragraphs related to the question to gather knowledge that will serve as the answer and its rationale.
• Rationale Generation: Generates rationale containing the core answer, detailed explanation, and supplementary information from the selected paragraphs.
• Systematic Composition: Systematically combines all generated rationale to construct a final answer that is easy to read without redundancy.

In addition, the QASA benchmark, built with the participation of experts in the AI/ML field, was also unveiled. This benchmark divides question types into three levels—Surface, Testing, and Deep—and contains a total of 1,798 high-quality question-answer pairs.