2023-07-18: ACM / IEEE Joint Conference on Digital Libraries (JCDL' 23) Doctoral Consortium Trip Report

By -

The 2023 ACM/IEEE Joint Conference on Digital Libraries (JCDL' 23) was held in Santa Fe, New Mexico, USA, from June 26-30, 2023. The first day started with Doctoral Consortium (DC) with a few other workshops and tutorials. The JCDL DC is a pre-conference event that allows doctoral students to discuss their dissertation work with other doctoral students and an international panel of professional researchers and scientists. DC aims to help Ph.D. students with their theses and research plans by providing feedback and general advice in a constructive environment.

Source: https://2023.jcdl.org/

Due to flight problems, the co-chairs of the DC: Dr. George Buchanan and Dr. Dana McKay, could not attend JCDL this year. However, many thanks to the following session chairs and mentors for providing feedback during DC. 

This year eight students were accepted to present their thesis work at JCDL DC,  four of whom were from the Web Science and Digital Libraries Research Group (WSDL) at Old Dominion University. The session was hybrid, as two students presented their thesis work online. It was a full-day session from 9 am to 5 pm MDT. In this blog, I will cover the work of all accepted students. 

WSDL Presentations

Muntabir Hasan Choudhury

ETDSuite: An Library for Mining Electronic Theses and Dissertations

Muntabir introduced one of the understudied datasets called Electronic Theses and Dissertations (ETDs), which can be found in university library repositories or a centralized library system called ProQuest. Most university library repositories consist of ETDs, supported by a digital library application, which can do little to meet users' needs beyond simple searching and browsing. Digital Libraries (DLs) of ETDs lack computation models, tools, and services for discovering and accessing the knowledge found in ETDs. Moreover, ETDs have distinct features compared with conference proceedings and journals in many aspects. They are book-length documents, the topics may shift across chapters, exhibits significant contribution to the research work of a student, and have different metadata schema (e.g., department, discipline,  advisor) from regular scholarly papers. In addition, ETD contains domain knowledge, rich metadata, bibliographies, figures, tables, and discoveries. Although several state-of-the-art (SOTA) frameworks (e.g., GROBIOD, LayoutLMv2) are available for metadata extraction and document page-level classification, designed for journals and conference proceedings.
There is a lack of frameworks to extract information from ETDs, including ETD segmentation, metadata extraction, metadata quality improvement, and parsing reference strings. To address the research gap, Muntabir will develop an ETDSuite library containing novel methods that segment, extract, parse, and restructure raw ETD documents into structured JSON documents leveraging natural language processing (NLP) and computer vision (CV).
While addressing the research gap, Muntabir developed AutoMeta, a metadata extraction tool, and MetaEnhance, a metadata correction tool, which he contributed to JCDL in 2021 and 2023. He ended this presentation by proposing models to solve the problems of ETD segmentation, parsing citations in many styles from ETDs, enhancing the performance of MetaEnhance, and enhancing the extraction capability of AutoMeta.

Yasasi Abeysinghe

Can Reading Web-based Information Affect Human Working Memory Capacity: An Eye Tracking Study

In today's world, reading and consuming information from the web has become a significant part of many commercial, academic, occupational, and personal activities. Web-based information is often structured with a non-linear presentation of digital text and multimedia. The non-linear structure of a digital text can make considerable demands on the cognitive load of readers, affecting their reading comprehension. Reading comprehension is the reader's ability to extract or construct meaning from written text and integrate it with what the reader already knows. Yassasi introduced that human working memory plays an essential role in reading comprehension by holding relevant information in short-term memory and integrating it with information retrieved from long-term memory. Understanding how different presentations of digital documents can affect human working memory capacity (WMC) can help design or present web-based information.

The WMC of humans can be used to understand how the human working memory uses attention to hold and process new information. To assess an individual's WMC, complex span tasks (e.g., reading span task, operation span task) and n-back tasks are often used. However, they can only be measured by observing how they affect people with a secondary processing task. Thus, understanding WMC has been challenging as there currently needs to be a direct method of determining a person's WMC without distracting them from their tasks. The advancement of eye-tracking technology allows one to use eye movement measures for real-time monitoring of human subjects (eye movements) while performing a task and performing real-time implementation with a reduced computational cost. Hence, eye-tracking metrics can be used to measure human WMC without overloading users with additional processing tasks.

Yasasi will develop a modeling framework using machine learning algorithms to use eye-tracking measurements for indexing WMC. She will explore different human eye features, including oculomotor plant features, traditional positional gaze metrics, and advanced eye metrics, to find an objective relationship between participants' eye-tracking measurements and WMC measures.

Bhanuka Mahanama

Eye Tracking for Collaborative Digital Documents

Eye tracking data provides informative signals on human behavior during various collaborative activities on information interfaces. Traditional approaches in eye-tracking studies fail to capture collaborative behavior as they are often conducted in isolation in the form of single-user studies. Bhanuka proposed a multi-user eye-tracking system utilizing common off-the-shelf eye-trackers for aggregating eye-tracking data of participants connected across a network. Further, he outlined the implementation details of the proposed method, future developments, and his plans for evaluating the overall system. Bhanuka also demonstrated the preliminary results, which showed the potential utility of the system for understanding collaborative human behavior on collaborative information interfaces.


Xin Wei

Exploring Semantic Information Extraction on Scholarly Documents Using Deep Learning Methods

The growth of scientific papers in the past decades and the large volume of patent documents call for effective semantic information extraction tools to automatically and accurately obtain the information needed. The information extracted can be key to understanding those documents more efficiently. However, the publicly available datasets in this field are limited in size. She explored different research scenarios to solve the specific data scarcity problem and proposed a solution for each research scenario. In the first research scenario, she annotated a dataset of 3000 sentences to train the model. It can automatically extract the information needed for annotating millions of patent figures with high time efficiency. Xin proposed extracting theory and model names from scientific papers in the second research scenario. However, annotation is challenging due to a lack of domain expertise. She proposed using distantly supervised learning to solve the problem, and this method can automatically construct training data instead of human annotation. In the third research scenario, she explored how to extract key findings and contributions in abstracts of scientific papers, namely, claim extraction. To solve the limited size problem, she proposed using supervised contrastive learning to improve the results.

Other Presentations

Rand Alchokr

The academic publishing landscape is changing so fast, making quality assessment and evaluation processes challenging with the advent of different methodologies. In modern data science, quality and impact methods must be non-discriminatory or fair for some attributes (e.g., prevalence or gender). Understanding and improving these methods has become a significant research area involving fundamental research that yields more profound insights into the validity and fairness of these methods. However, researchers regularly analyze the impact of these methods without concentrating on specific groups of researchers who may be more affected. Therefore, Rand proposed to work towards filling this gap by eliciting and systematically analyzing the pre and post-publication quality methods, their correlation, and their impact on underrepresented researchers (e.g., juniors) in the Computer Science domain. In particular, She will define concepts and guidelines on the best practices and derive integrated techniques to detect bias.

Stanislava Gardasevic

Stanislava started the presentation with an open end question, how do Ph.D. students discover the resources and relationships conducive to satisfaction and success in their degree programs? Her research seeks to make this information intuitively discoverable through knowledge graph technology, which captures and visualizes relationships between people based on their activities and relations to resources in a particular domain. The knowledge graph should facilitate access to information currently scattered across different web locations and the tacit knowledge that resides only within the community.
The interdisciplinary Ph.D. Program in Communication and Information Sciences (CIS) was taken for a case study because of the multitude of interdisciplinary collaborations and topics that can be represented as a rich network. Ph.D. students of this program are the target user population for the visualization and navigation prototypes currently being designed using information architecture and user experience research approaches. In this study, participatory design methodology plays the central role, and Ph.D. students of the CIS program are involved in all design stages through a series of workshops. This research contributes to foregrounding Ph.D. students in the information-retrieval loop that will allow a direct exchange of tacit knowledge and designing a novel multiplex graph visualization tool.

Satvik Chekuri

Knowledge Graphs (KGs) are a powerful tool for representing and organizing information in a structured and machine-readable format. They consist of nodes representing entities and edges representing relationships between those entities. By organizing information in this way, KGs enable machines to understand the meaning of the data they contain and to reason about it in a human-like manner. This makes them particularly well-suited for applications such as search engines, recommendation systems, and question-answering systems. Constructing KGs for long scholarly documents such as ETDs presents several challenges. These documents often contain complex language and concepts that can be difficult for machines to understand. Additionally, they may contain errors or inconsistencies that can affect the accuracy of the resulting KG. To address these challenges, Satvik proposed framing the KG construction problem as a series of regular NLP tasks such as entity extraction, relation extraction, entity linking, and entity disambiguation.

Sonia Pascua

Machine learning and information retrieval systems require efficient information representation, knowledge organization, and advanced predictive analytics to extract meaningful insights from extensive datasets and tackle intricate feature relationships. In this research, Ke's Discounted Least Information Theory of Entropy (DLITE) is explored and investigated for quantifying information in related processes. In particular, an innovative framework is proposed that expands upon traditional approaches, such as TF-IDF, BM25, Shannon entropy, and information gain, by integrating additional information-theoretic measures. This framework is designed to capture the statistical properties of terms and their interrelationships, which are crucial for quantifying fundamental information properties like uncertainty, relevance, and diversity. Grounded in axioms and mathematical principles that ensure consistency, coherence, generality, and computational efficiency, these methods exhibit strong potential for enhancing the performance of predictive analytics tasks, including classification. This study has conducted preliminary tests and outlined plans for further experiments using standard benchmark datasets from diverse domains, encompassing text and image classification. Ultimately, the framework employs information theory to pinpoint informative features and optimize machine learning models, offering a novel alternative for representing, organizing, and analyzing information and knowledge across various domains, such as natural language processing and text retrieval.

Conclusion

I thank all mentors and participants for providing valuable insights and guidance throughout this event. Moreover, attending the DC provided insights into improving some of my research techniques, especially a better representation of the data collection. I hope to apply what I learned to sharpen the focus of my research and, overall, to make an impactful contribution to the world.

-- Muntabir Choudhury (@TasinChoudhury)

Comments

Post a Comment