Loading...

Representation Based Multi-hop Question Answering

Hemmati, Nima | 2024

0 Viewed
  1. Type of Document: Ph.D. Dissertation
  2. Language: Farsi
  3. Document No: 57588 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Ghassem Sani, Gholamreza
  7. Abstract:
  8. The Question-Answering(QA) problem has long been a significant focus of researchers. Its connection with natural language understanding and knowledge retrieval makes it one of the most critical issues in Natural Language Processing (NLP). Given the inefficiency of simple question-answering methods, multi-hop question-answering (Multi-hop QA) across multiple documents has become one of the most attractive problems in recent years. In general, multi-hop question-answering is supposed to answer natural language questions that require extracting and combining information conained in several documents and performing reasoning about that information. The ability to answer questions and perform multi-step reasoning can significantly enhance the performance of natural language processing systems. Accordingly, this topic has been enriched in recent years by providing high-quality data, and introducing new models and methods. Graph neural networks have emerged as a promising tool for multi-hop question-answering tasks. However, these models ususlly face issues such as increased computational and model complexity, making them inefficient for real-world applications with limited resources. In this thesis, a graph-based approach called the Sparse Graph-based Multi-hop Question Answering system (SG-MQA) is proposed. This innovative approach is designed based on relational graph convolutional networks to reduce the complexity of the model and improve its performance. Additionally, leveraging the capabilities of Large Language Models (LLMs), the performance of the initial proposed model (SG-MQA+LLMs) has been enhanced. To evaluate the performance of the proposed models and the quality of the outputs, several experiments were conducted in this research. The effectiveness of the proposed approaches has been confirmed by the results of our experiments on two question-answering datasets, WikiHop and HotpotQA. The SG-MQA+LLMs model outperformed all existing advanced methods on WikiHop, achieving an accuracy of 81.8%. Furthermore, this model achieved a satisfactory performance on HotpotQA. Based on the F1 measure, this model has had the best performance among all previous methods. However, in terms of the exact match metric, its performance (although is still satisfactory)has been outperformed by some of the previous methods
  9. Keywords:
  10. Natural Language Processing ; Deep Learning ; Graph Convolutional Networks ; Attention Mechanism ; Large Language Model ; Multi-Hop Question Answering

 Digital Object List

 Bookmark

No TOC