The Board Members of EETN, during their meeting on Monday 11 July 2022, approved the recommendation of the three-member committee, which consisted of Professor Grigoris Antoniou of the University of Huddersfield, Professor Giorgos Stamou of the NTUA and Assistant Professor Christos Tzamos of the University Wisconsin, for the best dissertation award in the area of Artificial Intelligence for the years 2020 to 2021.
10 PhDs from Greek Universities, who met the specifications of the announcement, responded to the invitation of EETN. The committee, found the level of the submitted dissertations satisfactorily high and decided to award the prize for the best AI-dissertation jointly to doctors Maria Tzelepi and Ilias Chalkidis, while an honorable mention was awarded to doctor Michalis Moudantonakis. The awards were announced at the 12th Conference on Artificial Intelligence (SETN-22), which took place in Corfu on September 7 – 9, 2022.
Below you can find information about the awarded PhDs and their dissertations.
Dr. Maria Tzelepi completed her PhD in 2021, at the School of Informatics of the Aristotle University of Thessaloniki, under the supervision of Professor Anastasios Tefas, entitled “Deep Learning Techniques in Digital Media“.
Abstract: Recent advances in deep learning (DL) provided significant performance increase on various digital media analysis tasks, such as image classification and retrieval. However, despite their effectiveness, DL models suffer from high complexity. This constitutes a major impediment on applying these models on devices with restricted computational power. In this Ph.D thesis, we deal with three different digital media analysis problems, that is content based image retrieval, image classification, and video captioning, utilizing DL techniques. The principal goals of this thesis can be summarized in developing deep representation learning methods oriented to the specific digital media analysis tasks, and in developing lightweight DL methods that allow for deploying them on devices with restricted computational power. To this end, firstly a deep representation learning method for producing efficient retrieval oriented representations was proposed. Subsequently, the proposed method was properly adapted in order to learn more efficient representations considering both the retrieval performance, and the memory requirements and the retrieval speed. Next, lightweight DL models capable of operating even in real-time for high resolution input, on devices with limited computation power were proposed, for addressing generic problems of image classification. In addition, various regularization techniques based on the concept of multitask learning were proposed (e.g., graph embedding based regularization, regularization based on the criterion of quadratic mutual information), improving the generalization ability of the proposed lightweight models. Furthermore, two online self distillation methods were proposed, allowing for training efficient lightweight models in generic classification problems. Finally, a video captioning method was proposed. The proposed method was capable of capturing different kinds of information, producing improved performance in the video captioning task.
Dr. Ilias Chalkidis completed his PhD thesis in 2021, at the Department of Informatics of the Athens University of Economics and Business, under the supervision of Professor Ionas Androutsopoulos, entitled “Deep Neural Networks for Mining Information from Legal Texts“.
Abstract: Legal text processing (Ashley, 2017) is a growing research area where Natural Language Processing (NLP) techniques are applied in the legal domain. There are several applications such as legal text segmentation (Mencia, 2009; Hasan et al., 2008), legal topic classification (Mencia and Fürnkranzand, 2007; Nallapati and Manning, 2008), legal judgment prediction and analysis (Wang et al., 2012; Aletras et al., 2016), legal information extraction (Kiyavitskaya et al., 2008; Dozier et al., 2010; Asooja et al., 2015), and legal question answering (Kim et al., 2015b, 2016b). These applications and relevant NLP techniques arise from three main sub-domains, i.e, legislation, court cases, and legal agreements (contracts). In all three sub-domains, documents are much longer than in most other modern NLP applications. They also have different characteristics concerning the use of language, the writing style, and their structuring, compared to non-legal text. Given the rapid growth of deep learning technologies (Goodfellow et al., 2016; Goldberg, 2017), the goal of this thesis is to explore and advance deep learning methods for legal tasks, such as contract element and obligation extraction, legal judgment prediction, legal topic classification, and information retrieval, that have already been discussed in the literature (but not in the context of deep learning) or that were first addressed during the work of this thesis. In this direction, we aim to answer two main research questions: First and foremost on the adaptability of neural methods that have been proposed for related NLP tasks in other domains and how they are affected by legal language, writing, and structure; and second on providing explanations of neural models’ decisions (predictions). Considering the first research question we find and highlight several cases, where either legal language affects a model’s performance or suitable modeling is needed to imitate the document structure. To this end, we pre-train and use in-domain word representations and neural language models, while we also propose new methods with state-of-the-art performance. With respect to model explainability, we initially experiment with saliency (attention) heat-maps and highlight their limitations as a means for the explanation of the model’s decisions, especially in the most challenging task of legal judgment prediction, where it is most important. To overcome these limitations we further study rationale extraction techniques as a prominent methodology towards model explainability.In lack of publicly available annotated datasets in order to experiment with deep learning methods, we curate and publish five datasets for various legal tasks (contract element extraction, legal topic classification, legal judgment prediction and rationale extraction, and legal information retrieval), while we also publish legal word embeddings and a legal pre-trained language model to assist legal text processing research and development. We consider our work, a first, fundamental, step among other recent efforts, towards improving legal natural language understanding using state-of-the-art deep learning techniques, which further promotes the adaptation of new technologies and sheds light on the emerging field of legal text processing.
Dr. Michalis Moutantonakis completed his PhD thesis in 2020, at the Computer Science Department of the University of Crete, under the supervision of Professor Yiannis Tzitzikas, entitled “Services for Connecting and Integrating Big Number of Linked Datasets“.
Abstract: Linked Data is a method for publishing structured data that facilitates their sharing, linking, searching and re-use. A big number of such datasets (or sources), has already been published and their number and size keeps increasing. Although the main objective of Linked Data is linking and integration, this target has not yet been satisfactorily achieved.Even seemingly simple tasks, such as finding all the available information for an entity is challenging, since this presupposes knowing the contents of all datasets and performing cross-dataset identity reasoning, i.e., computing the symmetric and transitive closure of the equivalence relationships that exist among entities and schemas. Another big challenge is Dataset Discovery, since current approaches exploit only the metadata of datasets,without taking into consideration their contents.In this dissertation, we analyze the research work done in the area of Linked Data Integration, by giving emphasis on methods that can be used at large scale. Specifically, we factorize the integration process according to various dimensions, for better understanding the overall problem and for identifying the open challenges. Then, we propose indexes and algorithms for tackling the above challenges, i.e., methods for performing cross-dataset identity reasoning, for finding all the available information for an entity, methods for offering content-based Dataset Discovery, and others. Due to the large number and volume of datasets, we propose techniques that include incremental and parallelized algorithms. We show that content-based Dataset Discovery is reduced to solving optimization problems, and we propose techniques for solving them in an efficient way. The aforementioned indexes and algorithms have been implemented in a suite of services that we have developed, called LODsyndesis, which offers all these services in real time. Furthermore, we present an extensive connectivity analysis for a big subset of LOD cloud datasets. In particular, we introduce measurements (concerning connectivity and efficiency) for 2 billion triples, 412 million URIs and 44 million equivalence relationships derived from 400 datasets, by using from 1 to 96 machines for indexing the datasets. Just indicatively, by using the proposed indexes and algorithms, with 96 machines it takes less than 10 minutes to compute the closure of 44 million equivalence relationships, and 81minutes for indexing 2 billion triples. Furthermore, the dedicated indexes, along with the proposed incremental algorithms, enable the computation of connectivity metrics for 1million subsets of datasets in 1 second (three orders of magnitude faster than conventional methods), while the provided services offer responses in a few seconds. These services enable the implementation of other high level services, such as services for Data Enrichment which can be exploited for Machine-Learning tasks, and techniques for Knowledge Graph Embeddings, and we show that this enrichment improves the prediction of machine-learning problems.