Building a Discourse-Argument Hybrid System for Vietnamese Why-Question Answering

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Deep learning models achieve high F1 scores on question answering tasks, but perform poorly on why-questions.
• Existing models struggle to provide explanations, achieving F1 scores of 0.57-0.7 on SQuAD v1.1 for why-questions.
• Why-questions require explanations, which can be arguments or subjective opinions, a capability lacking in current models.

Purpose of the Study:

• To propose and evaluate a new approach for answering why-questions.
• To enhance the ability of systems to provide explanations for complex queries.
• To improve the answer rate for why-questions beyond the capabilities of current deep learning models.

Main Methods:

• Utilizing discourse analysis to identify explicit arguments and opinions.
• Applying natural language inference to detect implicit arguments and calculate sentence similarity.
• Combining discourse analysis and natural language inference to generate answer candidates for why-questions.

Main Results:

• The proposed system demonstrated a higher answer rate (77.0%) compared to a deep learning model (61.0%) on a Vietnamese translated SQuAD v1.1 why-question dataset.
• While the F1 score did not surpass deep learning models, the system successfully answered a greater number of why-questions.
• The approach effectively identifies both explicit and implicit information relevant to explaining why-questions.

Conclusions:

• The integration of discourse analysis and natural language inference offers a promising direction for improving why-question answering.
• The developed system shows potential in providing more comprehensive answers to explanation-seeking queries.
• Further research can build upon this hybrid approach to bridge the gap in deep learning model performance for complex question types.