JCDL '24: Proceedings of the 24th ACM/IEEE Joint Conference on Digital Libraries

The role of public libraries is evolving with the Smart City (SC) trend, but the impact of SC policies on libraries remains under-evaluated. Using the Chinese SC program as a testbed, we applied the Difference-in-Differences (DID) method to assess these policies' effects on infrastructure and services. Results show SC policies enhanced physical collections but negatively impacted space and seating, with no significant effect on services. We discuss the complexity underlying the interplay between digitization and public libraries' roles in the SC context, suggesting a need for further research and policy consideration.

Digital libraries that maintain extensive textual collections may want to further enrich their content for certain downstream applications, e.g., building knowledge graphs, semantic enrichment of documents, or implementing novel access paths. All of these applications require some text processing, either to identify relevant entities, extract semantic relationships between them, or to classify documents into some categories. However, implementing reliable, supervised workflows can become quite challenging for a digital library because suitable training data must be crafted, and reliable models must be trained. While many works focus on achieving the highest accuracy on some benchmarks, we tackle the problem from a digital library practitioner. In other words, we also consider tradeoffs between accuracy and application costs, dive into training data generation through distant supervision and large language models such as ChatGPT, LLama, and Olmo, and discuss how to design final pipelines. Therefore, we focus on relation extraction and text classification, using the showcase of eight biomedical benchmarks.

Peer review is a cornerstone of the academic editorial decision-making process, yet it faces significant challenges. Artificial intelligence can help address these challenges, but its use raises concerns about reliability and the potential for reproducing existing biases. In this research, we employ a formal argumentation-theoretic framework that allows for explicit analysis of arguments and their interrelations, combined with argument mining techniques to streamline the formalization of peer reviews, and resulting in a neuro-symbolic approach to dispute resolution. Our method involves identifying parties' arguments in peer reviews and representing them as abstract argumentation frameworks, which facilitate dispute resolution through logical inference. We annotate these frameworks within a corpus of scientific peer reviews, achieving a high Krippendorff's alpha of 0.81. Having the annotated corpus, we implement an argument mining pipeline that integrates BERT sentence embeddings with an LSTM model, classifying sentences into three categories: authors' arguments, reviewers' arguments, and non-arguments. We achieved an accuracy of 0.634 and an F1 score of 0.631, which are comparable to models trained on other datasets. However, our approach stands out by enabling the processing of the extracted argumentation with logical inference.

Performances of state-of-the-art keyphrase generation models improve with the size of the training dataset. But obtaining large amounts of keyphrase-labeled documents can be challenging and costly. Data augmentation methods allow to increase the training set size without additional cost. However, those techniques rely most of the time on external data or additional resources than can be as difficult to obtain as new annotated data. To tackle this issue, we present a self-compositional data augmentation method which creates additional training samples that keep domain coherence, without relying on any external data or resources. More specifically, we measure the relatedness of training documents based on their shared keyphrases, and combine similar documents to generate synthetic samples. Our results on multiple datasets spanning three different domains, demonstrate that our method consistently improves keyphrase generation. A qualitative analysis of the generated keyphrases for the Computer Science domain, confirms this improvement towards their representativity property.

Layout analysis of academic papers aims to identify various components within unstructured papers, benefiting researchers in quickly locating and extracting critical information. The effectiveness of this process depends heavily on the datasets and models used for training. However, existing datasets often have issues with annotation accuracy, granularity, scale, and acquisition cost. Current models treat each document image in isolation, ignoring the position information of a page within the entire paper. To address these challenges, we propose DLAgen, a method for rapidly, accurately, and cost-effectively generating fine-grained annotated paper datasets. DLAgen uses context-free grammar to generate textual content in LaTeX format, and incorporates visual content, such as images, tables, and formulas, from real papers, thus creating synthetic papers with accurate annotations. Concurrently, to leverage the high correlation between page numbers and components in academic papers and to make better use of textual information, we introduce MDT, a multimodal academic paper layout analysis model that utilizes page position information and correctly ordered text. Experiments show that MDT trained with data generated by DLAgen achieves higher accuracy in fine-grained layout analysis of real academic papers compared to existing state-of-the-art models. The mAP is improved from 85.13 to 88.61, which is a 4.09% enhancement, validating the effectiveness of our approach. Both the model and dataset will be released to the public.

There are lots of scientific articles are being published every year, it is increasingly challenging for researchers to maintain oversight and track scientific progress. Meanwhile, Large Language Models (LLMs) have revolutionized natural language processing tasks. This research focuses on generating summaries from research paper abstracts by utilizing LLMs and comprehensively evaluating the performance of the summarization. LLMs offer customizable outputs through Prompt Engineering by leveraging descriptive instructions including instructive examples and injection of context knowledge. We investigate the performance of various prompting techniques for various LLMs using both GPT-4 and human evaluation. For that purpose, we created a comprehensive benchmark dataset for scholarly summarization covering multiple scientific domains. We integrated our approach in the Open Research Knowledge Graph (ORKG) to enable quicker syn- thesis of research findings and trends across multiple studies, facilitating the dissemination of scientific knowledge to policymakers, practitioners, and the public.

Scientific knowledge entities encapsulate the core information and key elements within scientific literature, functioning as relatively independent and complete knowledge modules. To enhance the efficient organization, utilization, and analysis of knowledge, this study aims to automate the extraction of valuable scientific knowledge entities from general-domain scientific literature. However, the identification of scientific knowledge entities presents unique challenges compared to general entity recognition tasks, due to the lack of a unified annotation system and high-quality annotated corpus in general domains, as well as issues such as limited availability of entity resources, semantic conflicts and boundary ambiguity. To address these challenges, we have defined nine types of scientific knowledge entities and constructed an annotated corpus covering eight domains. This paper proposes a unified model, SCL-MRC-Diff, based on the Machine Reading Comprehension (MRC) framework, integrating a Semantic Contrastive Learning (SCL) module and a Boundary Perception module based on a semantic differential mechanism (Diff). The MRC framework provides prior knowledge for low-resource entities, while the SCL module introduces label prototypes and explicitly injects information into the context and target entities through contrastive learning to enhance semantic clarity. Additionally, the Diff module uses a semantic differential mechanism to learn span boundary information, thereby improving boundary constraints for the model. Comparative experiments demonstrate that our model achieves optimal performance among mainstream models, with a Macro F1 score 2.7% higher than the second-best, reaching 72.46%. Ablation studies further validate the effectiveness of our modules.

The rapid growth of scientific literature, particularly in Artificial Intelligence (AI), has resulted in the frequent introduction of new methodologies and evolving meanings for existing terms. However, in AI, the fast-paced development poses challenges for researchers in keeping up with recent methodologies (e.g., `R-CNN', `ELMo'), crucial for tasks such as idea generation, baseline selection, and peer review. Existing automated methods for extracting key concepts from research articles are limited by their static evaluation setups and lack of adaptability to newer terminology. Addressing these limitations, this paper proposes a factored approach leveraging category information (e.g., `NLP', `RL') to methodology extraction, which utilizes domain-specific partitioning either by input or label space. We conduct experiments under a zero- and few-shot setup and introduce a chronological evaluation framework. Our framewokrs learn incrementally using silver-standard data from previous predictions to adapt to recent data. Empirical results demonstrate that our proposed method outperforms prior approaches like SciREX in effectively identifying emerging methodology names in scientific literature. Experiments on a simulated setup using historical and incrementally updated data show that our approach improves future predictions on newer papers, outperforming baselines by up to 9.257% in the few-shot setting.

The method entities are the distillation of the knowledge within scientific papers. However, the contribution of different method entities to a study varies. Therefore, discerning the relation of a method and a paper has become a common concern in many application scenarios such as in impact evaluation of a paper from the method dimension, in method knowledge acquisition and retrieval and so on. Nevertheless, existing studies have often simplified this issue to a text classification task, modeling the semantics of entire method sentences rather than individual method entities. This approach falls short in addressing the complexity of a single sentence containing multiple method entities with inconsistent semantic relations. To overcome this limitation, this study formalizes the problem as an end-to-end named entity recognition task and introduces MPR, a dataset of the relation labels for the sequence labeling task. The dataset supports to identify method entities and their relations with the host paper in an integrated process. Several deep learning-based models are developed to solve the problem, with the SciBERT-based algorithm showing particularly excellent performance and thus establishing a strong baseline for the task. Part of the MPR dataset is publicly accessible at https://github.com/ChenCbnu/methodanno. The communities are welcome to utilize it as a foundation for exploring advanced performance and promising directions in fine-grained knowledge mining within the domain of scientific literature in the new technological era.

We address the extraction of mathematical statements and their proofs from scholarly PDF articles as a multimodal classification problem, utilizing text, font features, and bitmap image renderings of PDFs as distinct modalities. We propose a modular sequential multimodal machine learning approach specifically designed for extracting theorem-like environments and proofs. This is based on a cross-modal attention mechanism to generate multimodal paragraph embeddings, which are then fed into our novel multimodal sliding window transformer architecture to capture sequential information across paragraphs. Our approach demonstrates performance improvements obtained by transitioning from unimodality to multimodality, and finally by incorporating sequential modeling over paragraphs.

Trust and user experience play an important role in visualisation in digital humanities. Digital humanities visualisation is not only a research process, tool, technique and methodology, but it is also increasingly becoming an experiential and exploratory activity for the public. However, there are few research fully investigated applications of visualisation from a user experience perspective, especially in terms of overall user experience beyond usability, such as pragmatic and hedonic experiences. We conducted an interview and eye-tracking study with 30 participants using timelines and geographical maps as representatives of the digital humanities visualisations. The results show that casual users have richer hedonic than pragmatic experiences from their interactions with visualisations, and that timeline and geographical map visualisations bring a greater difference in immersion to the user experience. Users' experiences with different dimensions of data, narrative, visual, and visualisation all have an impact on trust, which further influences continuous adoption intentions.

This study aims to explore whether embodied experiences in virtual reality (VR) can promote the cultural education in digital humanities and heritages. Using a controlled experimental method, data on users' working memory, cognitive load, and emotional response were collected in three environments with different levels of embodied experience: VR, video, and printed materials. The results show that embodied experience has a strong positive effect on emotional responses; there is a certain cognitive resource competition between emotional cognition and working memory; and working memory processed information using both body-based and sociocultural-based cognitive frames. In the future, it is necessary to improve the narrative emotions during embodied experiences and to guide the allocation of limited cognitive resources to knowledge construction tasks through the integration of multimodal representations.

The development of digital humanities necessitates scholars to adopt more data-intensive methods and engage in multidisciplinary collaborations. Understanding their collaborative data behaviors becomes essential for providing more curated data, tailored tools, and a collaborative research environment. This study explores how interdisciplinary researchers collaborate on data activities by conducting focus group interviews with 19 digital humanities research groups. Through inductive coding, the study identified seven primary and supportive data activities and found that different collaborative modes are adopted in various data activities. The collaborative modes include humanities-driven, technically-driven, and balanced, depending on how team members naturally adjusted their responsibilities based on their expertise. These findings establish a preliminary framework for examining collaborative data behavior and interdisciplinary collaboration in digital humanities.

Recommender systems have been increasingly entering people's lives, predominantly to newsfeeds, entertainment, and product selection. An ideal recommendation system can help to get content in front of the right people, thereby satisfying users. Mental models, formed by users and used to enhance their understanding and predict the behavior of systems they interact with, can be based on different approaches. We deem it important to understand where the mental models come from because they have caused many false expectations regarding the recommender system's capability.

We conducted a qualitative study with a unique sample of video recommender system users. Eighteen typical and regular YouTube users, aged between 60 and 75, participated in this study. In-depth interviews were used for data collection, and inductive content analysis was used for data analysis. Our analysis of participants' responses demonstrates that Baby boomers' mental models about video recommender systems are 1) shaped by their experiences with mass media, 2) imagined by the usage experience of similar internet applications, and 3) improved through interaction with the current system. The results of this study provide valuable insights into the field. Identifying the process of forming a mental model is a necessary first step in informing the design to reduce cognitive load and create more inclusive, user-friendly interfaces.

Most traditional medical search engines rely on similarity-based retrieval between the user's query and document content, ignoring important contextual aspects, such as the quality of evidence and specific biomedical concepts. Addressing these limitations, we developed WisPerMed, a medical search engine that enhances search efficiency by incorporating Level of Evidence (LoE) and bio-concepts into both the ranking process and the user interface. This study compares the performance of WisPerMed to PubMed through a user study involving 131 medical experts. Our findings indicate that WisPerMed significantly reduces the number of queries needed and the time spent on searches, with large effect sizes compared to PubMed (0.822 and 0.551, respectively). The integration of LoE and bio-concepts not only improves the ranking of relevant and authoritative articles but also facilitates their identification through enhanced GUI elements. These results highlight the potential of WisPerMed to enhance search efficiency in clinical and research settings, ultimately supporting faster and more accurate access to medical knowledge.

In English literature, the 19th century witnessed a significant transition in styles, themes, and genres. Consequently, the novels from this period display remarkable diversity. This paper explores these variations by examining the evolution of term usage in 19th century English novels through the lens of information retrieval. By applying a query expansion-based approach to a decade-segmented collection of fiction from the British Library, we examine how related terms vary over time. Our analysis employs multiple standard metrics including Kendall's tau, Jaccard similarity, and Jensen-Shannon divergence to assess overlaps and shifts in expanded query term sets. Our results indicate a significant degree of divergence in the related terms across decades as selected by the query expansion technique, suggesting substantial linguistic and conceptual changes throughout the 19th century novels.

In academic writing, bibliography compilations is essential but time-consuming, often requiring repeated searches for references. Hence, an efficient tool for faster bibliography compilation is needed. Our work offers a solution to the challenges of managing large-scale bibliographic databases, introducing a new algorithm that improves both efficiency and sensitivity. Using two-vector semantic modelling, bibliographic entries and queries are embedded into the same vector space to select relevant references based on semantic similarity. Experimental results with 3.37 million entries show the method reduces the time needed to generate a manageable subset, streamlining scholarly writing. Our code and dataset are publicly available at https://github.com/cestwc/bibliography-pre-selection.

Using scientific papers as the data foundation, identifying the current global science center in the field of physics, and assessing the distance between each country and the global scientific center based on the analysis of its migration path are of significant importance for countries in formulating technology strategies and promoting technological advancements. Based on the physics paper data published in the Web of Science database from 2003 to 2022, this study extracts the countries of the first and corresponding authors using the fractional counting method, taking countries as units, and then counts the number of publications for each country. Subsequently, employing the Technological Diversity Index, Technical Specialization Index, and social network analysis, the study identifies the science center in physics. Furthermore, it accomplishes the recognition of the transfer paths of scientific centers by calculating the distances between each country and the science centers. We found that (1) The United States has consistently held the position of the global science center in the field of physics, but the scientific center is gradually exhibiting a trend toward decentralization. (2) The distances between countries such as Germany, the United Kingdom, France, Switzerland, Japan, China, and the United States are gradually decreasing, revealing a phenomenon of "one superpower and multiple strong contenders" in the international physics and technology capabilities. (3) Although the United States remains the global science center in the field of physics, the transfer path of the international physics center seems to show a trend of shifting from the United States to Europe and then from Europe to Asia.

The publication of a large number of non-citable items in journals may lead to an increase in the Journal Impact Factor (JIF). This study uses data from 56 multi-disciplinary journals published in 2019 and 2020, as well as citation information, to explore whether the publication of non-citable items affects JIF. The empirical results show that for multidisciplinary journals, the impact of non-citable items on JIF is not significant.

Citations are crucial in scholarly papers, aiding in the acknowledgment of previous research and enhancing accessibility to related works. Creating accurate citations takes time and requires expertise. To promote appropriate citations, previous studies have tackled the citation-worthy detection of reference tags, which detect the sentence that needs citations. However, scholarly papers also use URL citations, which are pivotal yet understudied. This paper introduces the novel task of citation-worthy detection for URL citations. This task aims to detect the locations within sentences where URL citations are needed. In experiments, we compared a transfer learning method using Named Entity Recognition (NER) with a simple token classification approach. The NER-based method performed better on citations after the noun and demonstrated better learning ability despite distribution gaps between training and test sets. These findings indicate that leveraging scientific domain knowledge through NER is a promising approach for accurate URL citations detection.

The "limitations" sections of scientific articles play a crucial role in highlighting the boundaries and shortcomings of research, thereby guiding future studies and improving research methods. Analyzing these limitations benefits researchers, reviewers, funding agencies, and the broader academic community. We introduce LimTopic, a strategy where Topic generation in Limitation sections in scientific articles with Large Language Models (LLMs). Here, each topic contains the title and `Topic Summary.' This study focuses on effectively extracting and understanding these limitations through topic modeling and text summarization, utilizing the capabilities of LLMs. We extracted limitations from research articles and applied an LLM-based topic modeling integrated with the BERtopic approach to generate a title for each topic and `Topic Sentences.' To enhance comprehension and accessibility, we employed LLM-based text summarization to create concise and generalizable summaries for each topic's Topic Sentences and produce a `Topic Summary.' Our experimentation involved prompt engineering, fine-tuning LLM and BERTopic, and integrating BERTopic with LLM to generate topics, titles, and a topic summary. We also experimented with various LLMs with BERTopic for topic modeling and various LLMs for text summarization tasks. Our results showed that the combination of BERTopic and GPT 4 performed the best in terms of silhouette and coherence scores in topic modeling, and the GPT4 summary outperformed other LLM tasks as a text summarizer. Our code and dataset are available at https://github.com/IbrahimAlAzhar/LimTopic/tree/master.

In response to the growing complexity and volume of scientific literature, this paper introduces the LLMs4Synthesis framework, designed to enhance the capabilities of Large Language Models (LLMs) to synthesize the key insights from scientific texts as high-quality and concise summaries. This framework addresses the need for rapid, coherent, and contextually rich integration of key scientific insights, leveraging both open-source and proprietary LLMs. It also examines the effectiveness of LLMs in evaluating the integrity and reliability of these syntheses, alleviating inadequacies in current quantitative metrics. The contributions of this study are a novel methodology for synthesizing key scientific insights, definition of new synthesis types, and establishing nine detailed quality criteria for evaluating syntheses. The implementation fits LLMs with reinforcement learning to optimize for synthesis quality by alignment with our established quality criteria. The LLMs4Synthesis framework and its components are available, promising to improve the generation and evaluation of scientific research synthesis.

Recent advances in deep learning have led to the development of well-known AI language models, such as ChatGPT. Such models have gained widespread attention across various domains, including scientific research. In this context, discussions about the use of these models for writing and reviewing publications have started. Within this paper, we discuss the implications of integrating AI language models into the scientific writing process and provide a comprehensive overview of existing research on this topic. Therefore, we searched, describe, summarize, and organize existing research following systematic literature-review guidelines using the digital library Scopus. Since peer review is a crucial part of scientific research, we also focus on exploring the consequent impact of emerging models on the peer-reviewing process. Existing studies show that AI language models are used significantly in scientific writing. However, this usage requires guidelines and control to overcome potential challenges and problems.

Acknowledgments are the most overlooked part of a scientific publication that is often taken for granted. The acknowledgment section of dissertations is historically understudied and is considered as "Cinderella" of academic writing. Examining the acknowledgment sections of dissertations is crucial for understanding the cultural aspects embedded within academic practices and their impact on wider societal values and norms. In this data paper, we introduce an ongoing project involving a manually coded dataset consisting of 4603 acknowledgment sentences derived from 3737 dissertations sourced from the institutional repository of the University of Illinois Urbana-Champaign. A team of 12 coders, divided into 6 groups, used a tailor-made streamlit web-based tool developed specifically for qualitative coding, utilizing 17 support, 4 sentiment, and 2 non-support tags. This dataset will be an important asset for researchers in natural language processing, computational social science, and information science for various downstream analyses, including machine learning and named entity recognition, among others.

Scientific academies have played a crucial role in shaping modern epistemological frameworks. Today, many academies have evolved to expand their roles, acting as key intermediaries between science and the public, while fostering academic exchange and international collaboration. However, despite their prominent influence, there is a lack of comprehensive resources that capture the full scope of their identities, functions, and societal contributions.

This paper addresses this gap by presenting a detailed dataset of scientific academies, which integrates and organizes information from diverse sources, significantly enhancing both coverage and quality compared to existing datasets. A web mining approach is employed to further characterize the organizational behaviors and activities of these academies. The resulting profiling provides valuable insights into their geographical distribution and individual characteristics. This dataset serves as a foundational resource for future research into scientific academies and their broader impact.

High-quality biomedical datasets are essential for medical research and disease treatment innovation. The NIH-funded Bridge2AI project strives to facilitate such innovations by uniting top-tier, diverse teams to curate datasets designed for AI-driven biomedical research. In this study, we examined 1,699 dataset papers from the Nucleic Acids Research (NAR) database issues and the Bridge2AI Talent Knowledge Graph. By treating each paper's authors as a team, we explored the relationship between team attributes (team power and fairness) and dataset paper quality, measured by scientific impact (Relative Citation Ratio percentile) and clinical translation power (APT, likelihood of citation by clinical trials and guidelines). Utilizing the SHAP explainable AI framework, we identified correlations between team attributes and the success of dataset papers in both citation impact and clinical translation. Key findings reveal that (1) PI (Principal Investigator) leadership and team academic prowess are strong predictors of dataset success; (2) team size and career age are positively correlated with scientific impact but show inverse patterns for clinical translation; and (3) higher female representation correlates with greater dataset success. Although our results are correlational, they offer valuable insights into forming high-performing data generation teams. Future research should incorporate causal frameworks to deepen understanding of these relationships.

Digital libraries should organize background materials on any foreign text, and customize it for different users (e.g., expert vs. novice vs. language learner). Editors have written such background materials, called "commentaries", for millenia. We review types of information currently in commentaries, and also how print publication traditions shape gathering and presentation of information. We then review how digital libraries can make these commentaries more accessible with more thorough coverage and flexible presentation transcending the conventional limits of print.

The development of information technology has altered people's reading habits, which is going to leads VR reading to an emerging form in the near future. This study explores the reading outcomes in a virtual reality environment from the perspective of personal immersion. After analyzing the experimental data of thirty college students, we found the results showed in a VR reading environment, immersion had no significant impact on their reading experience and reading comprehension. However, readers comprehended materials with low cognitive load significantly better than those with high cognitive load. After the experiment, brief interviews were conducted with the participants. This study will provide insights for the development and promotion of virtual reading content.

Recent advancements in artificial intelligence (AI) have led to transformative impacts across various sectors, including the library and information science (LIS) domain. The incorporation of AI into traditional and digital library services facilitates the automation of routine tasks such as circulation and cataloging, while simultaneously enhancing patrons' experiences through improved book recommendations and informational chatbots. This perspective paper reviews the literature to identify the current perceptions of AI in libraries, applications of AI in public and academic libraries, future research directions in the field, and the potential challenges of adopting AI in libraries. Through a systematic search and investigation, we detail the purpose, methodologies, and findings of existing literature on AI applications in libraries. This paper documents three main areas of artificial intelligence with potential for integration into library services: recommendation systems, information and resource retrieval, and optical character recognition.

Standing at the forefront of knowledge dissemination, digital libraries curate vast collections of scientific literature. However, these scholarly writings are often laden with jargon and tailored for domain experts rather than the general public. As librarians, we strive to offer services to a diverse audience, including those with lower reading levels. To extend our services beyond mere access, we propose fine-tuning a language model to rewrite scholarly abstracts into more comprehensible versions, thereby making scholarly literature more accessible when requested. We began by introducing a corpus specifically designed for training models to simplify scholarly abstracts. This corpus consists of over three thousand pairs of abstracts and significance statements from diverse disciplines. We then fine-tuned four language models using this corpus. The outputs from the models were subsequently examined both quantitatively for accessibility and semantic coherence, and qualitatively for language quality, faithfulness, and completeness. Our findings show that the resulting models can improve readability by over three grade levels compared to the original abstracts, while maintaining fidelity to the original content. Although commercial state-of-the-art models still hold an edge, our models are much more compact, can be deployed locally in an affordable manner, and alleviate the privacy concerns associated with using commercial models. We envision this work as a step toward more accessible libraries, improving our services for young readers and those without advanced degrees.

This paper proposes a digital library approach to preserve traditional dance as a form of living cultural heritage. It explores using technology to capture knowledge and principles, enabling future generations to dynamically engage with and evolve these traditions. By democratizing access to cultural knowledge, digital technology challenges conservative ideologies that centralize cultural evolution. Using the Thai traditional dance principle Mae Bot Yai as a case study, we present Open Dance Lab, a web-based platform designed to preserve and innovate Thai traditional dance. The platform features a digital archive of 59 Mae Bot Yai poses as interactive 3D models with expert annotations, incorporates Pichet Klunchun's deconstruction of Thai dance into six core elements, and includes an AI-powered system for generating new dance sequences based on traditional principles. This research demonstrates how digital technologies can safeguard and transmit intangible cultural heritage while facilitating its evolution in the digital age.

The explosive growth of web archives presents a significant challenge: manually curating specialized document collections from this vast data. Existing approaches rely on supervised techniques, but recent advancements in Large Language Models (LLMs) offer new possibilities for automating collection creation. Large Language Models (LLMs) are demonstrating impressive performance on various tasks even without fine-tuning. This paper investigates the effectiveness of prompt design in achieving results comparable to fine-tuned models. We explore different prompting techniques for collecting specialized documents from web archives like UNT.edu, Michigan.gov, and Texas.gov. We then analyze the performance of LLMs under various prompt configurations. Our findings highlight the significant impact of incorporating task descriptions within prompts. Additionally, including the document type as justification for the search scope leads to demonstrably better results. This research suggests that well-crafted prompts can unlock the potential of LLMs for specialized tasks, potentially reducing reliance on resource-intensive fine-tuning. This research paves the way for automating specialized collection creation using LLMs and prompt engineering.

Digital Libraries Federation (FBC), the largest Polish aggregator in the cultural heritage domain, has recently been redesigned to enable full-text and melodic-content search functionalities aimed at improving the discovery and content-based retrieval of over 45k sheet music objects aggregated in FBC. Enabling music content retrieval required implementing three key changes to the FBC operations workflow: i) adapting its Solr search engine to process MEI documents; ii) adding AI-based Optical Music Recognition framework that converts the results to the MEI format to be processed by Solr; and iii) creating a brand new user interface for entering melodic queries and search results presentation.

Accurately identifying Science and Technology(S&T) priorities of various countries is important for understanding their S&T planning and future arrangements. We propose a framework for identifying S&T Priorities. It employs a comprehensive approach, integrating domain knowledge, feature definition, automatic classification, and text refinement. This framework enables the stepwise filtering and accurate identification of S&T priorities within news texts related to S&T policies and strategic plans. We have preliminarily verified the effectiveness of this framework through manual checks of identification results by librarians. In addition, the processing procedures and techniques involved in each step of the framework exhibit strong universality, offering good applicability and generalization when used for other types of text analysis. However, it should be noted that there is room for improvement in terms of threshold determination and model performance.

This paper draws on Activity Theory to explore the structure and process of scientific data practices in research context. Firstly, scientific data practices involve six elements, namely subject (researchers), object (scientific data), community, tool, rule, and division of labor. Through the interaction of these elements, scientific research tasks are completed and results (data outcomes) are generated. Secondly, each adjacent combination of three elements of data practices forms four subsystems, that is to say, production system, exchange system, distribution system, and consumption system. Thirdly, data practices are hierarchically divided into activities, actions, and operations, corresponding to motives, goals, and conditions, respectively. Different elements affect different stages of data practices in the activity theory subsystem. This study, by analyzing data behaviors and practices from an activity perspective, offers a new theoretical framework. It provides a foundation for future research on the mechanisms, motivations, and pathways involved in data behaviors and practices.

With the growing demand for aligning research with society, researchers and policymakers have been unequivocal in increasing research use in the policy-making process in recent years. Integrating research into policy formulation not only results in more effective policies but also allows researchers to extend the reach and influence of their work to a broader and more diverse audience. In this study, we focused on one of the protected and disadvantaged groups in society, youth, whose perspectives are frequently overlooked in research studies and policymaking, even though they comprise a significant portion of the US population. We built a model that can predict whether a research article will get more citations after the first citation or not using the textual features of the research articles. Experimenting with 11 classification algorithms and analyzing the results, we identified the most important textual features and the best model to drive researchers to generate research with a higher impact on policy. We found that avoiding using complex words in research articles increases the likelihood of getting more policy citations. Moreover, semantic relevance between the research article and the policy document does not inherently guarantee increased attention from policymakers.

Short video platforms are increasingly critical to information dissemination, especially during emergencies, where high dissemination-force videos rapidly shape online public opinion and exert significant social impact. Predicting the dissemination-force of emergency short videos can enhance public opinion management and improve decision-making foresight. This study aims to develop a predictive system for the dissemination-force of emergency short videos using machine learning and to identify key factors influencing dissemination-force through feature contribution analysis. Consequently, the study quantified dissemination force through metrics such as likes, comments, and retweets, and developed a multidimensional feature system encompassing user features, title features, audio & video (AV) features, AV & title features, and time features. Machine learning algorithms were applied for prediction, with XGBoost identified as the optimal model, achieving 96.39% accuracy, and recall and F1 scores exceeding 95%. Feature contribution analysis highlighted the total_favorited metric as the most significant predictor, followed by posting date and dubbing, with user features being the most important dimension in prediction outcomes. This research enriches the theoretical understanding of short video dissemination-force and offers practical models for effective public opinion management, thereby enhancing control, accuracy, and predictability during emergencies.

The detection of hate speech on social networks presents significant challenges due to its increasingly subtle nature. The advent of large language models (LLMs) has revolutionized text understanding and generation, presenting novel avenues for hate speech detection. This study evaluates the performance of the ChatGLM model in detecting hate speech through a small sample prompt method. Our analysis uncovers three key limitations when employing the LLMs to directly generate answers: inconsistency in output format, the illusion of comprehensiveness inherent in LLMs, and the inability to respond due to security concerns. To mitigate these limitations, we investigate four LLMs: LLaMA, Llama-2, Llama-3, and ChatGLM-3. These models are equipped with Multi-Layer Perception (MLP) and Low-Rank Adaptation (LoRA), which specifically tailored for hate speech detection. Extensive comparisons with baseline models are conducted across three hate speech datasets with six classification tasks. Our findings demonstrate that our improved LLMs can surpass traditional methods in detecting hate speech, highlighting their potential for further improvement and refinement in addressing this critical societal issue.

In the Internet age, crowdsourced data from social media is important for Alien Invasive Plants. To obtain high-quality crowdsourced data from social media, this study introduces an innovative framework for analyzing the multimodal crowdsourced data by adopting Large Language Models. Taking Canada Goldenrod as a case study, the experimental results demonstrate that the method is efficient and effective at analyzing multimodal crowdsourced data. The research contributes to the construction of biodiversity digital platforms.

This study investigates the several nuanced rationales for countering the rise of political bias. We evaluate the performance of the Llama-3 (70B) language model on the Media Bias Identification Benchmark (MBIB), based on a novel prompting technique that incorporates subtle reasons for identifying political leaning. Our findings underscore the challenges of detecting political bias and highlight the potential of transfer learning methods to enhance future models. Through our framework, we achieve a comparable performance with the supervised and fully fine-tuned ConvBERT model, which is the state-of-the-art model, performing best among other baseline models for the political bias task on MBIB. By demonstrating the effectiveness of our approach, we contribute to the development of more robust tools for mitigating the spread of misinformation and polarization. Our codes and dataset are made publicly available in github¹.

Large language models (LLMs) have gained tremendous popularity in domains such as ecommerce, finance, healthcare, and education. Fine-tuning is a common approach to customize an LLM on a domain-specific dataset for a desired downstream task. In this paper, we present a valuable dataset for fine-tuning LLMs developed for the Spanish language to perform a variety of tasks such as classification, masked language modeling, clustering, and others. Our dataset is a collection of handwritten notary records from the seventeenth century obtained from the National Archives of Argentina. This collection contains a combination of original images and transcribed text (and metadata) of 160+ pages that were handwritten by two notaries, namely, Estenban Agreda de Vergara and Nicolas de Valdivia y Brisuela nearly 400 years ago. Our transcription is accurate as it was prepared by experts in 17^th-century Spanish. Through empirical evaluation, we demonstrate that our dataset can be used to fine-tune Spanish LLMs for tasks such as classification and masked language modeling, and can outperform pretrained Spanish models and ChatGPT-3.5/ChatGPT-4o. Our dataset will be an invaluable resource for historical text analysis in the era of LLMs and is available via GitHub at https://github.com/raopr/SpanishNotaryCollection.

Scientific domains encompass many concepts that require a concise term definition to enable a common understanding among researchers, in particular for interdisciplinary fields. In digital libraries, information access and sharing is often facilitated by terminology databases. However, building up such resources is expensive to produce manually and requires expert knowledge. Automatically generating definitions for scientific terms has become a hot research topic recently that can reduce the manual burden. However, current methods heavily rely on large language models (LLMs) that store factual knowledge in their parameters, so that knowledge cannot be easily updated for emerging scientific terms. Furthermore, a major shortcoming of these models is that they are prone to hallucination and their output is difficult to control. To bridge these gaps, we propose to address the task of definition generation through guided abstractive summarization, incorporating key information from external resources. At test time, we augment the model with retrieved abstracts from Scopus and use automatically extracted topics and keywords as guidance, both essential for definition generation. To this aim, our approach takes into account two relevant sub-tasks in the process, a) predicting the topic class and b) generating hypernym candidates for the term. Our proposed pipelined approach for automatic guided definition generation achieves significant performance improvement over the standard baselines as well as relevant prior works on this problem. We use BLEU, ROUGE and BERTScore to automatically evaluate the quality of the systems on our benchmark and carry out a human evaluation to assess fluency, relevancy, coherence and factuality of the output. Our experiments show that LLMs can provide fluent and coherent definitions, and are often on par with human created definitions. Yet, there is still room for improvement on identifying relevant content and improving factual correctness.

Nowadays, the accessibility and long-term preservation of historical records are significantly impacted by the sharp increase in the digitization of these archives. This shift creates new opportunities for researchers and students in multiple disciplines to broaden their knowledge or conduct multidisciplinary research. However, given the vast amount of data that needs to be analyzed, using this knowledge is not easy. Different natural language processing tasks such as named entity recognition, entity linking, and article separation have been developed to make this accessibility easier for the public by extracting information and structuring data. However, historical newspaper article aggregation is still unexplored. In this work, we demonstrate the potential of the retrieval-augmented generation framework that integrates large language models (LLMs), a semantic retrieval module, and knowledge bases to create a system capable of aggregating historical newspaper articles. In addition, we propose a set of metrics that permit evaluating these generative systems without requiring any ground truth. The results of our proposed RAG pipeline are promising at this early stage of the system. They show that semantic retrieval with the help of reranking and additional information (NER) reduces the impact of OCR errors and query misspellings.

Large language models (LLMs) have recently revolutionized human language understanding and generation. Ontology is considered one of the primary cornerstones for representing knowledge in a more meaningful way on the semantic web. It's significant to explore whether LLMs know and understand such ontological knowledge. In this paper, we report an experiment to investigate the performance of LLMs in the task of classifying cultural heritage domain terms to upper-level ontology. We first probed the understanding and memorization of CIDOC CRM ontological knowledge by LLMs. Then, we further leverage LLMs to classify domain terms into the structure of CRM, and compare the match type with experts. Our initial findings indicate that LLMs demonstrate a certain level of awareness and comprehension of CIDOC CRM ontological knowledge. LLMs have shown potential as valuable assistants in enhancing ontology engineering and knowledge-intensive tasks.

Emotional information in narrative texts is an important resource for deepening text interpretation and promoting literature utilization, but there is still a lack of dedicated information organization schemes. Based on the systematic analysis of related theories, this study constructed the narrative text emotion ontology (NTEO), which contains three core components of event, appraisal, and emotion, as well as three related components of emotion category, emotion expression, and action tendency, for structuring and organizing the emotion information. Meanwhile, six relations of temporal, cause, circumstance, regulation, change, and trigger were summarized for embedding emotional units into narrative sequences. Furthermore, taking Tsen Shui-fang's Diary, an archive related to the Nanjing Massacre in China, as an example, we explored the innovative applications of fine-grained emotion classification, fine-grained emotion visualization, and emotion attribution analysis supported by this emotion ontology.

In the cultural heritage and digital humanities fields, digitization of Arabic script documents presents unique challenges due to the cursive nature and homographic characteristics of the script. This study evaluates the efficacy of Optical Character Recognition (OCR) post-processing with Large Language Models, specifically GPT-4 and GPT-4 Turbo. The proposed experiments, though in a preliminary form, span various configurations, including different model temperatures and approaches from zero-shot to few-shot learning. Results indicate that despite the high expectations of AI capabilities, the post-processing performance of GPT-4 models remains sub-optimal for Arabic scripts. This paper discusses potential reasons for these limitations and suggests directions for future research, emphasizing the need for models trained to deal with the complexities of Arabic orthography and semantics.

Understanding the evolution of science from billions of publications has always been a challenge, with the key to solving it lying in paper source tracing (PST). Few-shot learners based on large language models (LLMs) offer a promising solution, yet they face two major challenges. First, tracing source papers requires a profound understanding of scientific texts and complex research logic reasoning, where existing chain-of-thought (CoT) methods struggle to fully grasp this complexity. Second, PST involves linking a target paper with multiple references, and when these texts are used for few-shot learning, the information-rich content of academic texts complicates the learning of reasoning patterns from contextual examples. To address these challenges, this paper presents a novel prompting framework, Decomposition and Branch Reasoning Prompting (DBRP). The core of our approach is the Decomposition-of-Thought (DoT) strategy, which decomposes texts into several components for easier analysis. Additionally, we introduce Branch Reasoning Demonstrations (BRD), a new few-shot learning method that enhances DoT. This approach enables LLMs to efficiently focus on crucial information and learn reasoning patterns. Our experimental results using the advanced language model (GPT-3.5) on the PST-benchmark show that: (1) the DBRP method significantly outperforms the state-of-the-art CoT method by up to 72.8% in mAP and 37% in NDCG; and (2) BRD surpasses standard in-context learning methods, achieving substantial improvements in DoT zero-shot performance. The DBRP method effectively extends the typical CoT paradigm, enabling highly accurate and interpretable automated paper tracing, even under constraints of annotated data. This advancement lays a solid foundation for precise analysis of knowledge trajectories and research lineage.

Across many humanities disciplines, researchers analyze how word frequencies change over time. The Google Books Ngram Corpus provides this data for roughly 6% of all printed books. However, current tools and methods for analyzing this massive corpus are limited, hindering researchers' ability to address all their questions effectively. We propose a flexible workflow adaptable to existing research problems from literature. This approach enables more comprehensive and efficient corpus analysis. It builds on the observation that many current methods involve two distinct, computationally unbalanced steps: subsampling (expensive) and analysis (typically less so). By separating these steps into a preprocessing stage (long-running) and an interactive analysis stage, our workflow gives way to efficiency when working with that corpus. We demonstrate this by replicating existing studies using our proposed workflow.

Citations play a unique role in scientific discourse and are crucial for assessing the impact of scientific findings. The continuous development of deep learning has advanced significantly the task of citation intent classification, and the fine-tuning of large pretrained models based on transformers has greatly benefited this research field. However, a potentially overlooked issue is that the rich information among citations is not adequately represented by the task. Additionally, the representational information obtained solely from text modalities does not fully capture citation intent. In light of this, we propose a Dynamic Heterogeneous Graph Attention Fusion Network (DHGAN) for citation intent classification in scientific publications. It captures complex relationships between different entities in the publications through a heterogeneous graph attention mechanism. Simultaneously, it dynamically interacts contextual information related to citations with the relationships between the entities to induce associative representations among citations. Experiments on publicly available datasets demonstrate that DHGAN has shown improved performance over previous baselines.

Archives of libraries contain many materials, which have not yet been made available to the public. The prioritization of which content to provide and especially how to design effective access paths depend on potential users' needs. As a case study we interviewed researchers working on topics related to one German philosopher to map out their information interaction workflow. Additionally, we deeply analyze study participants' requirements for a digital library system. Moreover, we discuss how existing methods may meet their requirements and which implications these methods may have in a practical digital library setting.

Emotions play an essential role in the human information interaction. This paper examines the impact of emotional tasks on information seeking behavior in both passive information encountering and active information searching. Study 1 recruited 31 participants and conducted in-depth interviews to investigate their impressive emotional information browsing experiences in daily life. Study 2 recruited 36 participants to conduct a quasiexperiment study, who were required to finish 3 positive tasks and 3 negative tasks. The findings indicate that in the case of information encountering, both positive and negative information task may expand, limit, terminate, and avoid information seeking, depending on their motivations. In purposive active information searching, participants are more active in query interactions when searching for positive information, and browsing deeper when searching for negative information. These findings could help information system providers better understand the impact of emotions on users' information seeking behavior and provide better emotional responses and supports.

Measuring the novelty of scientific papers is one of the pivotal concerns in the realm of scientometrics. From the combinatorial perspective, previous research has typically considered only the cooccurrence of knowledge units to quantify the novelty of scientific papers, which may not entirely capture the comprehensive novelty of scientific papers. To fill this gap, this study proposes a comprehensive novelty measurement that integrates three types of relationships, including network, semantic, and hierarchical relationships, to depict MeSH distance. The results show that only approximately 0.1% of PLoS One papers have a novelty score above 0.5. Moreover, our proposed novelty indicator shows a significant correlation with the Uzzi et al. (2013)'s indicator.

This study explores the automatic classification of citizens' appeals from the "12345" service hotline in a local Chinese government. We propose an innovative approach that combines knowledge graph-driven adaptive classification, a multi-level analysis integrating large language models (LLMs) with word embeddings, and a weighted voting mechanism. Our findings reveal that the top five categories of citizens' appeals are "Housing", "Daily Life", "Urban Management", "Transportation", and "Employment". The proposed method achieves a precision of 0.83 and f1 is 0.78.

In this poster, we present (interim) results achieved in the course of the design, development and evaluation of a prototype chatbot application. The focus is on the preparation of training data for machine learning and the integration of the chatbot application with established information services and practices at the institution.

Predicting user personality from online posts is a critical endeavor in many social science fields. Previous studies have confirmed that large language models (LLMs) often struggle with processing lengthy contexts. To this end, we propose a divide-and-conquer prompting strategy. Initially, the original text is divided into multiple chunks. The LLM is then prompted to estimate the probability of each chunk corresponding to various personality traits, grounded in established definitions. Finally, these estimates are aggregated using a voting mechanism to yield a final personality assessment. Extensive experiments show that our strategy effectively unlocks the potential of LLMs, surpassing direct inference results from GPT-4 on two benchmarks.

Algorithms are a cornerstone of the artificial intelligence (AI) era and are pivotal in advancing various disciplines. Comparing the use of algorithms in Chinese and international research can provide methodological references for Chinese researchers to keep pace with international trends and achieve innovations. Focusing on the field of library and information science (LIS), we automatically extracted algorithm entities from academic papers published in both Chinese and international journals. We then compared the frequency and time of mentioning specific algorithms used in these two parts of studies. Our findings reveal that while most algorithms used in Chinese and international LIS research overlap, machine learning algorithms are more prevalent in global studies. Although Chinese LIS studies show lower overall frequency and a time lag in algorithm use, this gap is narrowing over time. Our study underscores the differences between Chinese LIS research and international trends from an algorithmic perspective, providing a reference for future development and alignment with global standards.

The use of knowledge graph technology empowers the integration of book resource association, and provides multi-dimensional intelligent knowledge services, which is of great significance to stimulate the vitality of digital resources and release the value of books [1]. This paper designed a structured framework of book knowledge graph and knowledge service application and explored the new modes of knowledge organization, knowledge management and knowledge service of book resources in-depth. The book knowledge graph and knowledge service system constructed in this study realizes the functions of intelligent retrieval, knowledge question and answer, and map editing of digital resources related to books, which can provide new ideas for the subsequent optimization of the book knowledge management system and enhancement of the book knowledge service level.

AI technology development has a growing impact on daily life and industry creation, and how to governance AI become a hot topic. This study screened out 66 US and 75 EU policies, using the content analysis method, network analysis and graphing methods. This research revealed their AI governance state from the perspective of policy tools and stakeholders' collaboration.

Knowledge of Chinese traditional opera is primarily preserved in informal forms, such as ancient texts, cultural relics, and oral traditions. To appreciate the unique charm of traditional opera, we focus on Yu Opera as a case study and develop the Yu Opera Knowledge Resource (YOKR) Ontology. Our approach involves collecting data from specialized literature, semantically segmenting it, and using the Bert4torch model for entity and relationship extraction. This paper explores the connections among historical figures, roles, plays, music, costumes, and other elements, aiming to enrich knowledge in Yu Opera and provide effective knowledge management services.

Data is now utilized in advanced research and scholarship across various fields. Utilizing data in the field of literary research represents a burgeoning area within the digital humanities. World Literature holds a distinct cultural significance in Korea. Traditional cataloging, such as MARC, present challenges in gathering, analyzing, and comprehending various facets. Despite the recent attention given to various data-related topics, techniques, and solutions, the efforts toward data aggregation and dissemination have been disorganized and occasionally inconsistent. This highlights the significance of standards-based data models and metadata. This study has selected 1950~60, part of the published period of the World Literature Collection. The data model has been developed based on BIBFRAME with metadata elements drawn from MODS. The current research has developed a preliminary data model and enhanced metadata for the World Literature Data Collection. The findings of the study have two implications: first, it could contribute to the creation of the data model for organizing book series in a digital library, and second, it could provide valuable insights to literature researchers working in their area.

In recent decades, the landscape of scholarly data has witnessed unprecedented transformations. The quantity of scholarly data has gained remarkable growth, and many techniques have emerged that enable effective analytics and the processing of scholarly big data. These changes give rise to Linked Scholarly Databases (LSD), a promising method that can effectively improve the quality of scholarly databases by establishing dense linkages among databases and performing large-scale metadata comparison and restoration through the linkages. LSD helps facilitate an open, interconnected ecosystem of scholarly databases that break through the knowledge boundaries of a single database, thus it has the potential to improve the quality of scholarly databases on a large scale. In this paper, we briefly describe the architecture of LSD and the feasibility of this approach in improving scholarly databases.

By integrating digital content with the physical world, augmented reality (AR) creates a seamless blend of digital information and real environments. Simultaneously, a tangible user interface (TUI) enables users to interact with digital information through physical objects, promoting more natural and effective content understanding and manipulation. In this preliminary study, we develop an interface, "Poetic Moon," that combines AR and TUI for human-information interaction within the context of Chinese poetry in digital libraries. Our results indicate this approach offers immersive and intuitive experiences, enhancing users' engagement and understanding of the presented content. This method has the potential for digital libraries, offering innovative ways to explore and interact with content in literary and other subjects.

Crowdfunding have gained popularity as a platform for individuals to harness the power of social solidarity to raise public funds for a range of objectives, such as for community projects, social good, medical expenses, etc. While there are many crowdfunding campaigns with altruistic goals, not all campaigns go viral and many more do not even meet their fundraising targets. This paper analyses the various factors that contribute to successful campaigns and develops five models for predicting fundraising success based on various combination of campaign features. We experiment on a dataset of 18,473 crowdfunding campaigns and discuss our main findings in terms of how different factors affect campaign success.

Social media platforms have become integral to daily life and serve as powerful channels for influencing public opinion, spanning applications from viral marketing to disinformation campaigns. While these platforms amplify the reach of marketing efforts, they also present significant risks when used for spreading misinformation via campaigns. Thus, it is crucial to understand such information campaigns by identifying the source, the different discussion topics within this campaign and how they change over time. Towards addressing this problem, we develop and present an interactive system for information campaign source tracing and analysis. This includes a demonstration and visualization of the main components of a information campaign, including the source of the campaign via explicit and implicit links, the discussion topics/clusters, their content, and how these evolve over different time periods.

This poster presents the preliminary findings of ongoing doctoral research examining the information behavior of scholars of African digital humanities (ADH). The study focuses on ethical and moral considerations in developing digital visualization tools such as GIS maps and 3D visuals showcasing information on sensitive topics such as slavery. This research employs a qualitative analysis of data collected through in-depth, semi-structured interviews aiming to 1) identify the challenges faced by scholars of African origin/descent in developing or using such visualization tools and 2) record their recommendations to address such challenges. Analyses of the first five interviews reveal prominent challenges faced by participants, including lack of technical skills, usability issues, content-related complexity, and the representation of primary source material. To overcome these challenges, the participants recommend collaboration, inclusive research designs, and involvement of community members. The five pilot interviews have allowed adjustments to the interview guide for an additional 10--13 prospective participants. This work aims to produce a guide for digital humanities scholars and professional web developers on designing efficient websites that showcase the results of historical research in an ethical and morally considerate manner.

This study explores the differences in user behavior between ChatGPT-assisted and traditional information retrieval modes during academic research. By setting up five academic information retrieval tasks with varying cognitive levels---remembering, understanding, analyzing, evaluating, and creating---and recruiting 20 first-year master's students, the participants were randomly assigned to complete the tasks under one of the retrieval modes. The results indicate that in each task type, the ChatGPT-assisted mode reduced the number of clicks on the search engine results page, the number of URLs visited, the duration of URL browsing, and the total retrieval time, suggesting that ChatGPT assistance can enhance retrieval efficiency. Further analysis revealed that the assistance effect of ChatGPT was more significant in remembering, understanding, and creating tasks. This study supports the necessity and feasibility of further exploring ChatGPT's application in the field of information retrieval.

Novelty is an important criterion for evaluating academic papers, and reliable measurement of novelty has also attracted attention from the research community. Recent studies used word embeddings to quantify the distance between all references and assess the novelty of the focal paper. However, the effectiveness of such methods remains unclear due to the absence of the "gold standard" for novelty evaluation. Furthermore, the availability of the peerreview score provide us with a new opportunity to address this research gap, as the scores provided by experts serve as a relatively objective measure. Validating different word embedding models can offer diverse perspectives for novelty assessment. This study proposes employing novel scores provided by reviewers alongside various word embedding models for the purpose of measuring novelty based on word vector representations. The findings suggest that there is a limited correlation between the novelty computed by different word embeddings and scores from actual reviewers. Moreover, various word embedding models demonstrate distinct performance across different novelty scores.

Statistical reports are indispensable sources of information containing crucial data presented through various document objects such as text, tables, figures, and equations. Extracting these objects accurately from PDFs is essential for digital services like Q&A chatbots, which rely on precise statistical information. This study proposes a methodological approach involving three main steps: converting PDFs to images, employing an object detection model trained on annotated document objects, and subsequently extracting pertinent information. The process leverages AI-assisted annotation systems for initial extraction, followed by manual adjustments to ensure accuracy.

Patent infringement reflects the conflict of interests between innovation entities. Although scenarios have been constructed to indicate how such information can be used, little attention has been paid to its organization. This study adopted a knowledge graph approach to organize patent infringement information before pointing out its potential use. It highlights the value of such information in competitive environment analysis and calls for deeper mining to assist entities in future innovation.

The purpose of this paper is to investigate the factors influencing user adoption behavior in the context of conflicting health information. This helps to understand users' adoption behaviors in complex information environments. This paper combines DEMATEL with ISM modeling. We first identified the key causal factors including information type, user basic characteristics, and experience accumulation. The key result factors consist of psychological emotion, perceived trade-offs, and source credibility. And, user adoption behavior in the context of conflicting health information is a manifestation of the interplay among surface-level, mid-level, and deep-level factors.

Understanding public concerns and attitudes toward sensitive AIGC applications is essential for establishing ethical guidelines that minimize risks. This study introduces a cross-cultural framework for extracting topics and stances from public comments, focusing on AI resurrection as a case study. The framework identifies key ethical concerns (dignity, mental health, and emotional manipulation) and emphasizes the need for culturally sensitive guidelines. Our case study demonstrates its effectiveness.

To comprehensively explore the driving forces behind users' participation in value co-creation of smart library services, this study draws on theories of value co-creation, behavioral dynamics, and system dynamics. User participation motivations are analyzed from the aspects of primary motivation, internal motivation and external motivation, and the system dynamics model is constructed. Then the software VensimPLE is adopted to conduct simulation analysis, and identify key motivations and the underlying mechanisms of user participation in the value co-creation of library smart services.

To explore the influencing factors of online infant and toddler health information search behavior among Chinese new generation parents, we recruited 32 participants and conducted semi-structured interviews on their online infant and toddler health information search behavior. We analyzed the qualitative data by three rounds of coding, and proposed a model to identify the influencing factors. The findings show that six main categories of factors can trigger infant and toddler health information search behavior from the user-, information- and environment- related perspective.

Governments have implemented online protection policies for minors to protect their safety, but no one knows whether these policies are effective. This paper used questionnaires and interviews to study minors and parents in China. It explored the implementation of online protection policies for minors in urban and rural areas and factors influencing policy effectiveness. This study can provide Chinese evidence and experience for effectively implementing relevant policies in various countries.

This paper is part of the international and interdisciplinary ROSETTA project, exploring multilingualism in Online Public Access Catalogs (OPAC). Despite authors creating culturally relevant works in their local languages, metadata and vocabularies in OPACs are predominantly in English. We investigate Natural Language Processing approaches to measure multilingual representation and indexing bias in transnational literary works. As case study, we collected all multilingual bibliographic records of one transnational literary work: "The Adventure of Huckleberry Finn" from World-Cat and we analyzed the indexing terms used as subjects to describe the different translations of this work. Our experiment reports on 634 bibliographic records in 18 distinct languages and propose to measure the multilingual representation and the indexing bias for this novel.

The paper proposes a new space-efficient representation of citation datasets obtained via a two-stage process consisting in (1) smart re-ordering and re-encoding nodes and edges, and (2) passing the output of stage one through the LZMA compression algorithm. While it does not allow for performing operations on the graph without its prior decompression, it achieves superior compression compared to both general-purpose and web graph compressors.

Trolling has become an integral part of social media interactions that focus on social and ideological posts. As influencers' posts set the social and ideological agenda for their followers, our study aims to examine the impact of influencers' message frames on followers' trolling reactions. Using the goal (Gain, Loss) and cultural frames (Individualist, Collectivists) we analyzed 3,050 followers' comments on 160 influencers' posts on the Sina Weibo platform. We found that social media influencers adopted the Loss-Individualist frame in most of their posts, and when they adopted the Gain-Individualist frame and the Loss-Collectivist frame, their posts attracted more followers' trolling reactions, while posts under the Gain-Collectivist frame were less susceptible to trolling than the three other frames. We conclude by articulating the theoretical implications of our findings and providing suggestions for future research directions.

This study explores how librarians' anxiety towards AI influences their willingness to adopt AI technologies. Using a grounded theory approach and in-depth interviews, the research identifies key anxiety factors such as information privacy, technology dependence, learning challenges, and job security. These anxieties are shaped by contextual elements like the innovation environment, emotional contagion, and perceived AI risks. The findings show that AI-related anxieties directly impact librarians' trust in AI, affecting their continuous adoption.

Current journal impact indicators are predominantly field-normalized to address citation variations across fields. As fields intersect, the topics within each field diversify, making it necessary to develop topic-level normalized indicators that capture journals' micro-level impact. Expanding on our initial Journal Normalized Impact (JNI), this study introduces the JNI2 series, a set of indicators that utilize advanced semantic topic modeling and Z-score normalization to precisely measure journal impact. Empirical analysis validated the effectiveness of the JNI2 series in assessing both the overall impact and topic-specific impacts of journals. The JNI2 series optimizes prior indicators in authenticity, precision, interpretability, and suitability for specialized journals.

This study examines the formation and characteristics of non-English scientific communities within the global publication landscape. While English dominates as the lingua franca in scientific communication, publications in other languages often remain underrepresented, potentially forming distinct clusters of knowledge. Using network analysis, we analyze how scientific publications outside the English language form communities. Thus far, results indicate that multilingual research communities typically form either as entirely non-English groups or with an equal split between English and non-English papers. Additionally, as the demand for community cliqueness increases, non-English publications tend to cluster into more tightly-knit, insular communities. The code for this analysis is available in our GitHub repository [6].

Cultural heritage images are imbued with profound historical and cultural significance. However, the prevailing image annotation methods for this domain are often plagued by an inverse relationship between the semantic depth of the annotations and the efficiency. This study proposes a human-machine image collaborative annotation model for cultural heritage that integrates the advantages of deep learning and crowdsourcing. Dunhuang mural was took as the typical cultural heritage objects, our method is demonstrated through experimental recognition and annotation tasks targeting the knowledge of the "Buddha's face" within these mural images.

Digital maker activities can enhance learners' agency, cultivating their digital and multiliteracy skills, and improving their creativity. The products from digital maker activities can potentially enrich the content of digital libraries and be shared with everyone with network access. Virtual reality (VR) content, owing to its affordance in providing audiences with immersive experiences, is getting popular in education and information services. The integration of digital maker activity and VR, namely VR content creation, can potentially provide benefits from the two, yet can have a high learning curve for the creators. In this study, we leverage a low-tech VR content creation approach, to teach primary school students to create VR stories in the context of an environment conservation project. Preliminary results from 85 primary school students demonstrate the effectiveness of this approach in improving their digital literacy, which provides a viable model for outreach activities in digital libraries and related institutions.

The extent to which translated Japanese manga are held in public libraries worldwide can serve as an indicator of their international reception. This study utilized a list of translated Japanese manga compiled during our previous surveys to investigate the availability of these works in public libraries across the United Kingdom.

This panel focuses on a very important yet lack-of-discussed topic, namely causal inference in science of science. We plan to invite several researchers to present their causal inference-related works and inspire fruitful discussions with JCDL audiences. The objective of this panel is to bridge the scientometrics community with causal inference communities and to showcase how future science of science researchers and those from other related fields establish their research in a causal inference way instead of purely at a correlation level.

This panel attempts to initiate a dialog and collaboration between digital humanities workers and ACM digital library communities. The panel will focus on the discussion around tools, methods, and stages of developing digital libraries that can enable marginalized communities to create, curate, and steward their collections through community-based digital archives and libraries.

The promises and perils about scientific team diversity are still debated in the scholarly literature. In this paper, we emphasize the moderating role of team structure, which sheds light on the inconclusive findings surrounding team diversity but has been overlooked in the literature. Understanding how hierarchical structures shape diversity is not only important to maximize the benefits of diverse teams, but also for enhancing the representation and impact of underrepresented voices. By applying the Explainable AI method SHAP to 1,876,905 computer science teams from the DBLP dataset, we find that team hierarchy plays a critical role in predicting team performance. We also investigate how team hierarchy interacts with team diversity along three dimensions: authors' gender, race, and sector. Interestingly, we observe that flat team structures positively correlate with performance in diverse teams than in homogeneous ones across all examined diversity facets. We propose that flat team structures foster conditions for diverse teams to flourish, enabling minority groups to assume significant roles and wield influential power. Our study contributes valuable insights into team diversity within the scientific community, emphasizing the significance of meaningful inclusion beyond mere numerical representation.

The abundance of scientists' memory resources has opened up unprecedented research opportunities for digital humanities research. However, these narrative resources are multimodal, and multimedia, and have diverse formats, posing challenges for coherent storytelling. To address this, the research develops a method that combines ontologies, event evolutionary graphs (EEGs), and Natural Language Processing (NLP) techniques for information extraction and correlation. The extracted data will be integrated into an interactive Digital Storytelling (DST) platform. This work enriches both historical understanding and public engagement with scientists' lives and achievements. The results of this study will contribute to digital humanities and archival studies by offering new methods for representing and narrating scientists' memory resources.

This paper explores the unique challenges of artwork image captioning, a task that demands deep understanding of historical, cultural, and stylistic elements often absent in traditional image captioning. We conducted preliminary experiments using a Meshed Memory Transformer on the Iconclass AI Test Set, which revealed significant improvements in standard metrics but highlighted critical limitations in current datasets and evaluation methods. To address these issues, we propose a novel approach integrating knowledge graphs with large language models. This approach involves creating a specialized art ontology and knowledge graph, and developing new evaluation metrics specifically designed for artwork captioning. While not yet implemented, this proposed method aims to generate more comprehensive, contextually rich, and accurate captions for artwork images. Our research lays the groundwork for future advancements in artwork image captioning, potentially enhancing the accessibility and educational value of digital art collections.

Identifying and Generating research papers' limitations in scientific articles will enhance the transparency and rigor of scientific research. In this work, we aim to automatically generate Limitations sections in scientific articles based on other key sections, including Abstract, Introduction, Methodology, Related Work, Experiment, and Conclusion. To identify the most relevant content, we apply a cosine similarity-based approach to select important sections for input. For generating limitations, we experiment with several large language models (LLMs), including BART, T5, Pegasus, GPT-3.5, GPT-4, and Gemini. Our findings show that GPT-3.5 with Retrieval-Augmented Generation (RAG) outperforms other models in accurately generating limitations. Building on this, we plan to incorporate a graph-based model using a graph neural network (GNN) that leverages both citation networks and thematic similarity to enhance the generation of limitations. This graph model will be integrated with the LLM and RAG system. Additionally, we intend to create a strong ground truth by incorporating both OpenReview and Limitations sections in the training data. For evaluation, we will apply PointWise similarity with interpretable metrics, utilizing the LLM as a judge approach to assess outcomes, supplemented by a human-in-the-loop approach to further refine results.

This research addresses the simply stacking status of multimedia convergence resources in digital cultural services. It focuses on the users' cognitive state associated with the use of multimedia combinations. The aim of the study is to construct a new theory of user multimedia fusion cognition in digital culture. The main research question is how to explain users' cognitive state during the utilization of multimedia fusion resources in DC. To answer that, a series of sub research questions have arisen. A mixed methodology of user experiment and algorithm optimization is designed. Following the research path, preliminarily, sub question 1 and sub question 2 have been solved. In the preliminary study, we designed an EEG-eye movement collaborative experiment to examine whether digital cultural multimedia fusion resources bring cognitive differences to users and attempted to use multimodal subjective and objective data collaborative modeling to predict user cognitive load. The findings showed that there were differences of users' cognition in digital cultural multimedia convergence and the multimodal subjective-objective synergistic prediction effect was the most optimal.

In recent years, cultural heritage institutions have experimented with opening up cultural heritage data and encouraging its reuse. While a growing body of digital humanities research focuses on the value of open data for cultural heritage and the development of data infrastructure, there is limited work exploring open data access and reuse in the cultural heritage domain at the user level. My PhD research aims to address this issue by investigating the practical process of cultural heritage open data access and reuse from the perspective of users. Specifically, I aim to: 1) identify the primary users of cultural heritage open data and their purpose, and 2) investigate the needs of individual data users and institutional data users, as well as how they locate, understand, and reuse cultural heritage open data. The findings will provide insights from a user perspective to facilitate the access and reuse of cultural heritage data, thereby unlocking its value and fostering sustainable development of cultural heritage data.

Wearable eye-tracking devices have rapidly evolved, providing researchers with innovative ways to capture and analyze multimodal interactions. In collaborative environments, individuals share visual attention and mental attention which affect the efficiency and success of the collaborative task. Synchronized eye-tracking studies can be employed during these collaborative tasks to understand Joint Visual Attention (JVA) and Joint Mental Attention (JMA). Here we aim to evaluate oculomotor plant features extracted from saccadic eye movements, traditional positional gaze metrics, and advanced eye metrics such as ambient/focal coefficient k, Index of Pupillary Activity (IPA), Low/High Index of Pupillary Activity (LHIPA), and Real-time Index of Pupillary Activity (RIPA) to study the JVA and JMA between two individuals. By leveraging modern wearable eye-tracking devices, we step beyond single-user eye-tracking studies to monitor and assess how individuals distribute attention, share information, and coordinate tasks in group settings. This research aims to explore the dynamic relationships between visual attention, mental attention, and teamwork while offering new insights to improve more synchronized collaboration in digital environments.

Complex scientific tables present unique challenges for information extraction due to their multi-level headers, merged cells, and domain-specific notations. Existing Table Structure Recognition (TSR) frameworks, often fall short when applied to these complex structures. How to perform UQ effectively and efficiently for table data extraction is a research question. To address these gaps, we propose an integrated pipeline that leverages artificial intelligence (AI) methods for mining complex scientific tables. Our approach combines TSR, Optical Character Recognition (OCR), and Large Language Models (LLMs) with uncertainty quantification techniques. We introduce the GenTSR benchmark for evaluating TSR methods across scientific domains and a modified Test-Time Augmentation (TTA-m) approach for uncertainty quantification. Additionally, we propose a novel benchmark for LLM-based table question-answering tasks using complex scientific tables. This comprehensive framework aims to enhance the accuracy and reliability of information extraction from scientific tables, facilitating more effective data analysis and interpretation in various research domains.

Digital library collections are valuable documents, which contain vast amounts of knowledge. The rise of artificial intelligence (AI) technologies, especially Large Language Models (LLM), has made an impact on various domains, and may also shift the focus of digital library services from storing and preserving information to utilizing and extracting knowledge from that information. We expect to add value created by LLMs to existing digital library methods so that users can better leverage the information stored in digital library collections. This research aims to improve access to digital library collections by providing users with classification labels and summaries to help easily find their works of interest leveraging large language models. Two expected major benefits are: (1) Enhancing metadata in the form of classification labels, which will help users discover and use digital library collections more easily. (2) Providing summaries of works stored in the digital library to aid researchers in finding works of interest without having to read the entire content, which will help remediate the information overload problem.

My doctoral research examines the information behavior of African scholars. I am reviewing their ethical and moral perspectives towards developing and using digital visualization tools such as GIS maps and 3D models on research websites, including digital archival collections of historical artifacts on the history of African slavery. These collections are housed and showcased through digital visualizations at various archives, courthouses, museums, and libraries. My research employs a qualitative study using in-depth, semi-structured interviews to 1) understand the meaning of ethical and moral research in digital spaces for African history, 2) investigate how and why African scholar uses visualization tools, 3) identify the challenges faced in developing or using such tools, and 4) record recommendations to overcome such challenges. I am inductively coding verbatim transcripts of the interviews to develop themes that address my research questions. My study design aims to reflect on the perspectives of marginalized communities with a sensitive background history of trauma that leads to the horror of racism and discrimination to date. Thus, this work applies more widely to digital libraries and archival studies that intend to develop digital tools sensitive to the ethical and moral implications of the information they contain.

The Parquet format is an open source, column-oriented data file format designed for efficient data storage and retrieval. To demonstrate the potential for Parquet files to be used in analyzing web archives, we will use the End of Term (EOT) Web Archive as a data testbed for this tutorial. This half-day in-person tutorial will introduce participants to columnar formats and how they can be used in the analysis and exploration of large web archive collections. Specifically, it will work with the Parquet file format to demonstrate the features and opportunities that the file format offers. The subject area used to demonstrate these features will be analyzing and exploring large web archives like the End of Term (EOT) Datasets. The tutorial will include both lecture components and live-coding examples of how these file formats can be used to solve challenges presented by large web archives.

Tables and figures are essential for conveying complex information visually and systematically in academic documents. Effective interpretation and extraction of information from academic tables and figures enhance the functionality of digital libraries, enabling more efficient data access and knowledge discovery. This half-day tutorial offers a comprehensive overview of the academic table and figure understanding process, covering detection, structural analysis, semantic interpretation, and practical applications. By discussing related benchmark datasets, recent techniques, and their pros and cons, this tutorial provides valuable insights for the JCDL community, supporting advancements in digital library research and development.

This half-day tutorial is designed for those who want an introduction to building a digital library using an open source software program. The tutorial will focus on the Greenstone digital library software. In particular, participants will work with the Greenstone Librarian Interface, a graphical user interface designed for developing and managing digital library collections. Attendees do not require programming expertise, however they should be familiar with HTML and the Web, and be aware of representation standards such as Unicode, Dublin Core and XML.

The Greenstone software has a pedigree of more than two decades, with over 1 million downloads from SourceForge. This tutorial centres on Greenstone 3, which is built on top of web standards such as XML Transforms to aid flexibility and extensibility. Emphasis in the tutorial is placed on recent capabilities introduced into Greenstone 3, such as the cloud-based version of the Greenstone Librarian Interface, and integration of the Google cloud-based Vision API into Greenstone3's ingest process for enhanced handling of images.

The increasingly mature artificial intelligence technologies, such as big data, deep learning, and large language models, provide technical support for research on automatic text understanding and bring development opportunities for innovative measurement of scientific communication, which has been a challenging and cutting-edge direction in informetrics. Inspired by the success of the 1st Workshop on Innovation Measurement for Scientific Communication (IMSC) in the Era of Big Data [3] at the ACM/IEEE Joint Conference on Digital Libraries (JCDL) 2023, we are excited to extend the second version at JCDL 2024. This workshop focuses on discussion and produces enlightening outcomes. We will engage broad audiences to share their ideas and pre-productions, enabling an interdisciplinary approach to exploring frontier areas. This workshop consists of a keynote, oral presentations, and poster sessions and will attract academic researchers, librarians, and decision-makers from governments and practical sectors.

This workshop tackles the challenges of managing and leveraging large-scale digital archives. As historical records are increasingly digitized, the vast volume of data poses accessibility and usability challenges. The workshop highlights artificial intelligence (AI) and machine learning (ML) applications in transforming archival processing, analysis, and retrieval, with the aim of advancing the field through new frameworks, tools, and best practices. Featuring keynotes, oral sessions, and posters, the workshop fosters interdisciplinary collaboration among experts in information science, computer science, and digital humanities. Participants will explore the latest AI/ML advancements in digital archives and contribute to the future of archival science.