Zafartavanaelmi, Hamid: Semantic Question Answering Over Knowledge Graphs: Pitfalls and Pearls. - Bonn, 2021. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc:
author = {{Hamid Zafartavanaelmi}},
title = {Semantic Question Answering Over Knowledge Graphs: Pitfalls and Pearls},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2021,
month = jun,

note = {Nowadays, the Web provides an infrastructure to share all kinds of information which are easily accessible to humans around the world. Furthermore, the amount of information is growing rapidly and requires computing machines to process, comprehend, and extract useful information tailored for the end-users. The Semantic Web and semantic technologies play a prominent role to enable knowledge representation and reasoning for these computational processes. Semantic technologies such as ontologies and knowledge graphs are being used in various application domains, including data governance, knowledge management, chatbots, biology, etc., which aim at providing proper infrastructure to analyze the knowledge and reasoning for the computers.
Semantic Question Answering systems are among the most desired platforms in recent years that facilitate access to information in knowledge graphs. They provide a natural language interface that permits the users to ask their questions posed in a natural language, without any understanding of the underlying technologies. We thus study question answering systems over knowledge graphs which aim to map an input question in natural language into a formal query, intending to retrieve a concise answer from the knowledge graph. This is a highly challenging task due to the intrinsic complexity of the natural language, such that the resulting query does not always accurately subserve the user intent, particularly, for more complex and less common questions.
In this thesis, we explore semantic question answering systems in a modular manner in order to discover the bottlenecks and mitigate the challenges in each part independently. Therefore, we focus on the individual modules and propose two innovative models: First, a reinforcement learning-based approach to parse the input question using distant labels, and second, an algorithm that generates the candidate formal queries based on a set of linked entities and relations. The latter additionally uses a neural network based model to rank the candidate queries by exploiting the structural similarity of the input question and the candidate queries. Through extensive empirical studies, we demonstrate that our proposed models perform well on three commonly used question answering datasets and increase the overall performance of the question answering system.
In addition, we design an interactive question answering system that solicits users for their feedback with the aim of guiding the system toward seizing the rectified semantic query, while striving to keep user satisfaction into account. Our oracle evaluation indicates that even a small number of user interactions can lead to a significant improvement in the performance of semantic question answering systems. Moreover, we conduct a user study to evaluate the performance of our system in interactive scenarios. We further devise a novel metric, called option gain, that is leveraged in the user interface and results in efficient and intuitive user interactions.
Moreover, we take the initial steps toward providing descriptive answers that enable the users to assess the correctness of the answer to their question. We present the first question answering dataset that includes the verbalization of the answers. This resource empowers researchers to train and evaluate a variety of models to generate answer verbalizations. Our experiments exhibit satisfactory results by natural language generation models that are trained on our proposed dataset.},

url = {}

Die folgenden Nutzungsbestimmungen sind mit dieser Ressource verbunden: