Loading...
Improving Reasoning in Question Answering Systems Using Deep Learning
Rahimi, Zahra | 2024
300
Viewed
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 56942 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Sameti, Hossein
- Abstract:
- Nowadays Artificial Intelligence systems are ubiquitous. One of the important applications is textual question-answering systems, which provide a means of information retrieval in a user-friendly manner. Reasoning is an inseparable part of human daily life, and people use reasoning to judge and find rational and correct answers to questions. To get the desired output from question-answering systems, these systems must be equipped with reasoning. This research focuses on improving question answering by considering Commonsense Reasoning. The two most important weaknesses of the existing question-answering systems are the questions being in the form of multiple-choice, which is far from a real-world setting, and the other using knowledge graphs as the source of knowledge. In this research, the answers to the questions are extracted from a corpus containing sentences in natural language instead of the traditional knowledge graph. Also, the questions of the dataset do not provide answer options, and the answers to the questions are presented as a sorted list of concepts after retrieving the relevant facts. The proposed system’s architecture is inspired by how humankind thinks, and it is implemented by using neural and symbolic artificial intelligence methods. To evaluate the performance of the proposed system, in addition to the automatic evaluation, a human evaluation has also been done, in such a way that the logically correct answers of the system to the questions not included in the HIT automatic evaluation are counted in the human evaluation. In this work, we have improved reasoning and the overall performance of the question-answer system by improving information retrieval and taking advantage of the power of graph neural networks. We have achieved Hit@50 of 80.77%, 76.46% and 76.88%, and Recall@50 of 47.30%, 41.88% and 33.50% on ARC, QASC and OBQA datasets, which has an improvement of 10%, 40% and 69% for Hit@50, and 35%, 27% and 69% for Recall@50 over the best existing baseline
- Keywords:
- Natural Language Processing ; Deep Learning ; Information Retrieval ; Graph Neural Network ; Textual Question Answering ; Commonsense Reasoning
-
محتواي کتاب
- view
- مقدمه
- تعریف مساله
- اهمیت موضوع
- اهداف پژوهش
- ساختار پایاننامه
- مفاهیم اولیه
- مقدمه
- گره موجودیت یا مفهوم
- یال یا رابطه
- سهگانه
- گراف دانش
- دانش درکعام
- حقیقت
- استدلال دانشدرکعام
- استدلال پراگماتیک
- استدلال شهودی
- پرسش و پاسخ انتهاباز
- پرسش و پاسخ مکالمهای
- پرسش و پاسخ چندگامی
- درک خواندنی ماشین
- آزمون کلوز
- نتیجهگیری
- پژوهشهای پیشین
- مقدمه
- مدلهای مبتنی بر مدل زبانی
- مدلهای متن به متن
- مدلهای مبتنی بر گرافدانش
- مدلهای مبتنی بر بازیابی عصبی روی متن
- سایر مدلهای عصبی-نمادین مبتنی بر متن
- مدلهای مورد استفاده در سامانه پیشنهادی
- مدل بارت
- مدل MP-Net
- مدل تبدیلگر گراف ناهمگون(HGT)
- مدلهای لبه دانش روی وظیفه OpenCSR
- مدل Dr.Fact
- مدل CoSe-Co (CommonSense Contextualizer)
- مجموعهدادگان مورد استفاده
- منبع دانش
- معیارهای ارزیابی
- جمعبندی
- سامانهپیشنهادی
- مقدمه
- معماری سامانه
- جمعآوری دادگان آموزش مدل بارت
- جزییات آموزش مدل بارت
- جزییات آموزش مدل تبدیلگر جمله MPNet
- جزییات آموزش مدل تبدیلگر گراف ناهمگون
- نتایج
- تحلیل نتایج
- جمعبندی
- نتیجهگیری و پیشنهادهای پژوهشی
- نتیجهگیری
- پیشنهادهایی برای پژوهشهای آینده
