Generating Coherent Narratives with Subtopic Planning to Answer How-to Questions.

Area of Science:

• Natural Language Processing
• Artificial Intelligence

Background:

• Answering how-to questions is a significant challenge in question answering (QA).
• Many narrow, long-tail how-to questions are not well-served by traditional search engines.
• Limited annotated data exists for training systems to answer these types of questions.

Purpose of the Study:

• To develop a novel architecture for generating coherent, long-form answers to how-to questions.
• To address the lack of annotated data for this specific QA task.
• To improve the quality and diversity of generated answers.

Main Methods:

• Proposed a new architecture comprising passage retrieval, subtopic planning, and narrative generation.
• Implemented a subtopic planning module to create diverse subtopics for answer structure.
• Consolidated multiple relevant passages into a single, coherent, explanatory answer.

Main Results:

• Demonstrated the difficulty of generating high-quality narratives for how-to questions.
• Showcased that the proposed architecture with subtopic planning yields improved answer quality.
• Evaluated generated narratives using automatic metrics and human assessments, confirming diversity and quality.

Conclusions:

• Generating long-form answers for how-to questions is a complex but achievable task.
• Subtopic planning is crucial for enhancing topic coherence and answer quality in narrative generation.
• The proposed architecture offers a promising approach for tackling long-tail how-to question answering.