Disentangled Retrieval and Reasoning for Implicit Question Answering

Area of Science:

• Artificial Intelligence
• Natural Language Processing

Background:

• Existing open-domain question answering (QA) methods primarily address explicit questions.
• Implicit QA, where reasoning steps are not stated in the question, presents significant challenges in evidence retrieval and answer inference due to limited overlap and latent reasoning strategies.

Purpose of the Study:

• To propose a novel systematic solution, DisentangledQA, for implicit open-domain question answering.
• To effectively disentangle latent elements within implicit questions to guide both evidence retrieval and reasoning processes.

Main Methods:

• DisentangledQA separates topic and attribute information from implicit questions to refine evidence retrieval.
• A disentangled reasoning model is employed for answer prediction, utilizing retrieved evidence and the latent reasoning strategy representation.

Main Results:

• The proposed method significantly improved performance on the StrategyQA dataset, with a 31.7% increase in evidence retrieval and a 4.5% increase in answer inference.
• DisentangledQA achieved state-of-the-art results on the StrategyQA dataset's official leaderboard.
• The method also demonstrated strong performance on the EntityQuestions dataset, highlighting its general applicability to open-domain QA tasks.

Conclusions:

• DisentangledQA offers an effective framework for tackling implicit question answering by focusing on disentangling latent question elements.
• The approach enhances representation learning for each module, leading to superior performance in both evidence retrieval and answer inference.
• This work advances the capabilities of open-domain QA systems in handling more complex and implicit queries.