Best Paper Nominierungen

Please help me!” Using large language models to improve titles of user-generated posts in online health communities

J.Chen

In online health communities, users can post questions to seek health-related advice from healthcare professionals. However, the titles they formulate often lack key information. Given that many people only scan titles, users may not get their questions answered. Large language models (LLM) offer a potential solution by generating titles that better align with the information needs of healthcare professionals. In this study, we fine-tuned an LLM using over 330.000 posts from the subreddit r/askdocs. Subsequently, we conducted a survey with 70 healthcare professionals to evaluate their preference between user- and LLM-generated titles. Our findings indicate that healthcare professionals perceive LLM-generated titles as better suited to the corresponding posts, more informative, and conveying a greater sense of urgency. With our work, we contribute to research on online health communities and large language models by demonstrating that LLMs can improve the titles of user-generated posts compared to those generated by users themselves.

Track: Thought Provoking Papers 1

Age Ain’t Just a Number: Exploring the Volume vs. Age Dilemma for Textual Data to Enhance Decision Making

L. Hägele, M. Klier, A. Obermeier, T. Widmann

The common belief that more data leads to better results often leads to all available data being used to derive the best possible decision. However, the age of data can strongly affect data-driven decision making. Consequently, the desire for larger data volume and at the same time contemporary data leads to the “volume vs. age” dilemma, which has not yet been sufficiently researched. In this work, we rigorously investigate the “volume vs. age” dilemma for textual data using four experiments with real-world data containing customer reviews from the Yelp platform. Contributing to theory and practice, we show that more data is not always better, as the effect of data age can outweigh the effect of data volume, resulting in overall poorer performance. Moreover, we demonstrate that different aspects within textual data can exhibit different temporal effects and that considering these effects when selecting training data can clearly outperform existing practices.

Track: Thought Provoking Papers 1

Evaluation of Outlier Detection Methods for Anomaly Detection in Journal Entries: A Use Case Analysis

T. Schreier, N. Gnoss, M. Tropmann-Frick, M. Schultz

Detecting anomalous journal entries in a company’s general ledger is essential for external auditors. An increasing trend employs outlier detection (OD) methods, especially machine learning methods, for anomaly detection in journal entry data. Recent research often lacks comparative analysis of OD methods. Thus, this study provides a comparative analysis of OD methods for journal entry anomaly detection using real-world accounting data. Additionally, in the context of domain-specific data preprocessing, we give special consideration to the amount, due to its importance for auditors. This yields three different dataset variants. We conduct our analysis based on three example accounts manually labeled by external auditors. Autoencoders, clustering-based local outlier factor (CBLOF), and histogram-based outlier score (HBOS) consistently outperform other methods across different accounts and dataset variants. With the provided results, this research enhances the understanding and applicability of OD methods for journal entry anomaly detection.

Track 6: Digital Finance

(When) Is a Summary Really Worth a Thousand Words? – How Textual and Visual Customer Review Summaries Affect Cognitive Load During Online Purchase Decisions

S. Erlebach

Online customer reviews are crucial to inform customers’ online purchase decisions. However, the complex nature and immense volume of customer reviews can easily overload customers’ cognitive resources, with significant negative consequences for customers and online shopping sites. In response, first online shopping sites provide summaries of textual review content (i.e., customer review summaries) alongside individual customer reviews. Building on cognitive load theory, this study examines how customer review summaries in textual and visual presentation formats affect customers’ cognitive load during online purchase decisions. With an incentive-compatible online experiment, we find that displaying customer review summaries reduces cognitive load. While textual and visual review summaries are equally effective, displaying both textual and visual review summaries reduces cognitive load the most. These findings contribute to literature on the design of customer review systems and provide relevant insights for practitioners on how to support customers’ online purchase decisions.

Track 12: Human Computer Interaction and Social Computing 1

Transparency of Algorithmic Control Systems and Worker Judgments

M. Kempf, F. Simić, A. Alizadeh, A. Benlian

The use of algorithms to guide worker behavior, referred to as algorithmic control (AC), is increasingly prevalent in organizations. Despite its potential operational benefits, prior research indicates that workers often struggle with the opaque nature of such systems. Our research aims to explore how workers perceive, judge, and react to AC systems when exposed to two distinct facets of algorithmic transparency (AT): input and transformation AT.
Through an experimental study with 121 participants, we provide empirical evidence that increased transparency about the algorithm’s transformation process significantly enhances workers’ perceived AT, which in turn positively impacts workers’ judgments and, ultimately, their continuance intention and acceptance of an AC system. In doing so, we provide practical recommendations for organizations to mitigate the adverse effects associated with algorithmic control.

Track 3: Changing Nature of Work 2

Bridging Fields of Practice: How Boundary Objects Enable Collaboration in Data Science Initiatives

N. Rahlmeier, K. Hopf

Data-intensive technologies draw high investments. Yet, data science projects are reported to suffer from poor collaboration, unrealistic expectations, and difficulties in realizing practical solutions between business and data science units. Moving beyond the currently prevalent approach to study data science practices, our study emphasizes the use of boundary objects between data science and collaborating fields. We interviewed collaborators from diverse fields in six organizational data science initiatives. Our inductive analysis of this rich data source uncovered six distinct mechanisms and six archetypes of boundary objects in data science projects. While archetypes that we label Alignment, Temporary, Collaboration, and Outcome are procedural and appear in selective stages of the data value creation process, the archetypes Infrastructure and Upskilling support projects along the value creation process. The archetypes and their mechanisms inform the management of data science initiatives, help to advance boundary object theory, and provide instruments to study data science initiatives.

Track 14: IT Strategy, Governance and Management 2