Aristotle University Of Thessaloniki (AUTH)
Machine learning and natural language processing techniques for question answering
Abstract
dc:descriptionNowadays, advanced services for extracting quick and precise information are required. Although search engines provide a flexible way for users to search for multimedia data, most of the time users must spend more time creating complex or multiple queries to get the information they demand. Question Answering may be able to solve this problem. The purpose of this thesis is to present research directions and novel ideas for dealing with the question-answering problem. A question-answering system provides a human-readable answer to a question posed by a user in natural language. Many approaches to dealing with the problem have been proposed; however, there is still room for improvement and new insights. Our research interests are in machine learning and natural language processing techniques, which can provide advanced methods and approaches for developing trustworthy question-answering systems when combined.In particular, we used traditional machine learning techniques to define a rich set of features for factoid question-answering. In addition, we investigated deep learning techniques for developing more advanced question-answering systems for yes/no questions. Finally, we tested and experimented with a variety of biomedical tools in order to propose solutions to biomedical question-answering. We proposed a conventional QA system for factoid question-answering that can extract a list of candidate answers by utilizing several biomedical resources. To create a comprehensive feature set, we combined hand-crafted features with word embeddings. Initially, we developed a linear function that did not involve any trainable weights to rank the candidate answers. Later, we incorporated a machine learning model that consisted of a group of supervised machine learning models. This ensemble model selected the most suitable candidate answer based on the given question.In a well-known dataset for biomedical question-answering, the proposed system achieved state-of-the-art performance. The majority of this dissertation's content is concerned with yes/no question answering, and we proposed several approaches for dealing with the problem and providing new insights to the community. In particular, we emphasized the role of sentiment in yes/no question-answering and how it can be used to improve the results of deep learning models. We also demonstrated how to create artificial tasks and datasets for fine-tuning a pre-trained language model in order to improve the performance of a yes/no question-answering model. Finally, we used the multitask learning technique to show that when a model is trained on both this task and the yes/no question answering task, the model's accuracy is higher than when the technique is not used. Our experience with the problem of question-answering led us to develop an end-to-end question-answering system for the biomedical domain. The system returned a set of biomedical articles from the PubMed biomedical resource using information retrieval techniques. The articles were then processed further, and a set of passages most relevant to the given question was returned. A method searched the passages for the most relevant answer to the user. The system recognized factoid questions and provided appropriate answers.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2022
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Dimitriadis, Dimitrios
- Δημητριάδης, Δημήτριος
Subjects
dc:subject × 14- Μηχανική μάθηση
- Επεξεργασία φυσικής γλώσσας
- Απάντηση ερώτησης από κείμενα
- Απάντηση ερώτησης με συλλογισμό
- Machine learning
- Natural language processing
- Question answering from texts
- Deductive question answering
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53517