LG AI Research Proposes QASA: A Novel AI Benchmark Dataset And A Computational Approach

Humans are good at reasoning, and this differentiates us from other living beings. Reasoning involves associative thinking and logical reasoning. One trivial way of reasoning is asking questions like what, when, where, and why. This reasoning can lead one to new discoveries and innovative ideas. 

Now, Imagine yourself stuck writing your own scientific paper and facing difficulty in asking the right curious questions. Due to growing volumes of scientific papers and professional articles, the traditional process is no longer feasible as it is time-consuming. Reading scientific articles raises questions and includes testing and deep questioning, which require full-stack reasoning. To answer such naturally advanced questions, researchers at LG propose a Question Answering on Scientific Articles (QASA) approach, which involves full-stack cognitive reasoning. 

Researchers designed a 3 step scheme to guide readers and authors to ask questions while reading the whole scientific paper rather than just the abstract. The first is to allow the reader to ask advanced surface, testing, and deep questions. Secondly, these Questions and Answers are further collected and compared with the questions asked by the expert readers. Finally, the readers and authors are invited to propose their multifaceted long-form answers to the collected questions. 

Researchers claim that QASA contains 1798 QA pairs on AI/ML papers, which regular readers asked for. On average, each paper has 15.1 to 29 questions and 39.4% of deep reasoning level questions. Their QASA approach involves associative selection to extract relevant information from paragraphs, evidential rationale generation to grasp only evidential rationale from each extracted paragraph, and systematic composition to relate evidential rationales to a comprehensive answer. 

In order to ensure realistic questions, the questioner is allowed to choose papers of their choice and select whether they want to read all the sections called deep reading or one particular section called skim reading and prepare questions that didn’t contain the answers. The answerers were also given the choice to choose papers from the papers that the questioners worked on to provide relevant answers. The answerers are guided to answers as a comprehensive passage based on their own-generated evidential rationales from the selected paragraphs. 

Researchers conducted a pairwise evaluation scheme where evaluators compare two answers to the same question. They provided two responses to the evaluators, one from the QASA scheme and the other from InstructGPT. The answers from the full-stack QA tend to be more complete and grounded than those from InstructGPT.

QASA approach involves modeling each subtask with pre-trained Language Models (LM) with multi-task instructions. Public and synthetic data could serve as the test bed for QASA, providing full-stack cognitive reasoning on scientific articles and manuscripts. This will ease the effort in retrieving and reranking the relevant information to read and restrict the useful information manually. 

Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 27k+ ML SubRedditDiscord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

The post LG AI Research Proposes QASA: A Novel AI Benchmark Dataset And A Computational Approach appeared first on MarkTechPost.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *