Ratan Sebastian
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Ratan Sebastian has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Anett Hoppe
TIB – Leibniz Information Centre for Science and Technology
5 shared publications - Ralph Ewerth
TIB – Leibniz Information Centre for Science and Technology
2 shared publications - Andreas Nehring
Leibniz University Hannover
1 shared publication - Anett Hoppe1 shared publication
- Christoph Rippe
Goethe University Frankfurt, University Library, Frankfurt, Germany
1 shared publication - Marc Allassonnière-Tang
CNRS/MNHN/University Paris City
1 shared publication
Latest Publications
- Open Access
- 13 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-5
Real-Time, Context-Specific Educational Metadata Enrichment of Learning Resources on the Web
Web resources rarely carry the educational metadata that teachers use to judge whether a resource fits their lesson, and dedicated educational repositories that do carry it cover only a fraction of what teachers find through web search. Closing this gap in real time, at the point of search, requires classifiers that are fast enough to run per query and that target a vocabulary specific to the teacher’s curriculum and grade structure, for which labelled training data is scarce. We investigate whether Large Language Model (LLM)-generated labels and paraphrases can train small, fast classifiers for this task. Across nine LRMI-aligned educational metadata properties for STEM education in Lower Saxony, Germany, matching model complexity to property difficulty reaches micro F1 > 0.8 for every property at inference latencies under 15ms, with models compact enough for client-side, browser-based deployment. We report this as preliminary, ongoing work: evaluation for six properties currently relies on LLM-generated labels validated against a 50-page human-annotated sample, and train/test splits do not yet separate resources by website, which we flag as open validity questions for the digital libraries community to weigh in on.
- 13 October 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-5
Open AccessReal-Time, Context-Specific Educational Metadata Enrichment of Learning Resources on the Web
Web resources rarely carry the educational metadata that teachers use to judge whether a resource fits their lesson, and dedicated educational repositories that do carry it cover only a fraction of what teachers find through web search. Closing this gap in real time, at the point of search, requires classifiers that are fast enough to run per query and that target a vocabulary specific to the teacher’s curriculum and grade structure, for which labelled training data is scarce. We investigate whether Large Language Model (LLM)-generated labels and paraphrases can train small, fast classifiers for this task. Across nine LRMI-aligned educational metadata properties for STEM education in Lower Saxony, Germany, matching model complexity to property difficulty reaches micro F1 > 0.8 for every property at inference latencies under 15ms, with models compact enough for client-side, browser-based deployment. We report this as preliminary, ongoing work: evaluation for six properties currently relies on LLM-generated labels validated against a 50-page human-annotated sample, and train/test splits do not yet separate resources by website, which we flag as open validity questions for the digital libraries community to weigh in on.
- Open Access
- 25 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 70-74
Examining Gender and Racial Bias in Vision-language Models within Colonial Contexts
Vision-language models (VLMs) are increasingly used for image retrieval in cultural heritage collections. Yet these models are trained and evaluated on modern web-scraped data which might cause unobserved biases when applied to historical imagery. While prior work has documented biases related to gender and skin tone in VLMs using contemporary datasets, no evaluation datasets exist for historical archives. In this paper, we address this gap by introducing a new dataset for evaluating demographic bias in VLMs on photograph collections from colonial contexts. Our dataset comprises annotated photographs for demographic queries spanning gender and race. Using this dataset, we examine bias in multiple VLMs through retrieval performance across queries, measuring disparities using established information retrieval metrics. Experimental results show that gender biases documented in contemporary datasets persist on historical imagery, with all tested models retrieving certain demographic groups substantially better than others. Notably, models specifically trained for debiasing fail to mitigate bias on historical images and can introduce new disparities across skin tones. These findings underscore the need for domain-specific debiasing approaches and caution against deploying off-the-shelf VLMs in archival contexts without proper evaluation.
- 25 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 70-74
Open AccessExamining Gender and Racial Bias in Vision-language Models within Colonial Contexts
Vision-language models (VLMs) are increasingly used for image retrieval in cultural heritage collections. Yet these models are trained and evaluated on modern web-scraped data which might cause unobserved biases when applied to historical imagery. While prior work has documented biases related to gender and skin tone in VLMs using contemporary datasets, no evaluation datasets exist for historical archives. In this paper, we address this gap by introducing a new dataset for evaluating demographic bias in VLMs on photograph collections from colonial contexts. Our dataset comprises annotated photographs for demographic queries spanning gender and race. Using this dataset, we examine bias in multiple VLMs through retrieval performance across queries, measuring disparities using established information retrieval metrics. Experimental results show that gender biases documented in contemporary datasets persist on historical imagery, with all tested models retrieving certain demographic groups substantially better than others. Notably, models specifically trained for debiasing fail to mitigate bias on historical images and can introduce new disparities across skin tones. These findings underscore the need for domain-specific debiasing approaches and caution against deploying off-the-shelf VLMs in archival contexts without proper evaluation.
- 15 July 2025
- Published by Springer Nature
- p. 236-250
Supporting Teachers Through AI-Augmented Web Search for Educational Resource Discovery
- 15 July 2025
- Published by Springer Nature
- p. 236-250
Supporting Teachers Through AI-Augmented Web Search for Educational Resource Discovery
- Open Access
Validating LLM-Generated Relevance Labels for Educational Resource Search
Manual relevance judgements in Information Retrieval are costly and require expertise, driving interest in using Large Language Models (LLMs) for automatic assessment. While LLMs have shown promise in general web search scenarios, their effectiveness for evaluating domain-specific search results, such as educational resources, remains unexplored. To investigate different ways of including domain-specific criteria in LLM prompts for relevance judgement, we collected and released a dataset of 401 human relevance judgements from a user study involving teaching professionals performing search tasks related to lesson planning. We compared three approaches to structuring these prompts: a simple two-aspect evaluation baseline from prior work on using LLMs as relevance judges, a comprehensive 12-dimensional rubric derived from educational literature, and criteria directly informed by the study participants. Using domain-specific frameworks, LLMs achieved strong agreement with human judgements (Cohen's $κ$ up to 0.650), significantly outperforming the baseline approach. The participant-derived framework proved particularly robust, with GPT-3.5 achieving $κ$ scores of 0.639 and 0.613 for 10-dimension and 5-dimension versions respectively. System-level evaluation showed that LLM judgements reliably identified top-performing retrieval approaches (RBO scores 0.71-0.76) while maintaining reasonable discrimination between systems (RBO 0.52-0.56). These findings suggest that LLMs can effectively evaluate educational resources when prompted with domain-specific criteria, though performance varies with framework complexity and input structure.Open AccessValidating LLM-Generated Relevance Labels for Educational Resource Search
Manual relevance judgements in Information Retrieval are costly and require expertise, driving interest in using Large Language Models (LLMs) for automatic assessment. While LLMs have shown promise in general web search scenarios, their effectiveness for evaluating domain-specific search results, such as educational resources, remains unexplored. To investigate different ways of including domain-specific criteria in LLM prompts for relevance judgement, we collected and released a dataset of 401 human relevance judgements from a user study involving teaching professionals performing search tasks related to lesson planning. We compared three approaches to structuring these prompts: a simple two-aspect evaluation baseline from prior work on using LLMs as relevance judges, a comprehensive 12-dimensional rubric derived from educational literature, and criteria directly informed by the study participants. Using domain-specific frameworks, LLMs achieved strong agreement with human judgements (Cohen's $κ$ up to 0.650), significantly outperforming the baseline approach. The participant-derived framework proved particularly robust, with GPT-3.5 achieving $κ$ scores of 0.639 and 0.613 for 10-dimension and 5-dimension versions respectively. System-level evaluation showed that LLM judgements reliably identified top-performing retrieval approaches (RBO scores 0.71-0.76) while maintaining reasonable discrimination between systems (RBO 0.52-0.56). These findings suggest that LLMs can effectively evaluate educational resources when prompted with domain-specific criteria, though performance varies with framework complexity and input structure. - 25 September 2024
- Published by Springer Nature
- p. 85-94
An Updated Analysis of Learning Resource Metadata Usage on the Web
- 25 September 2024
- Published by Springer Nature
- p. 85-94
An Updated Analysis of Learning Resource Metadata Usage on the Web
- Open Access
- 27 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 20-24
Adaptive Search Support for Teachers in Lesson Planning
Teachers report difficulties engaging in search in the context of lesson planning. Despite the availability of customised search solutions like Learning Object Repositories, most teachers report using general-purpose search engines like Google. Unlike students - to whom most research in educational search is directed - teachers are experts capable of using many types of resources for their tasks. In addition, since they teach multiple classes, they switch educational contexts frequently. These circumstances present unique challenges for user modelling. This thesis investigates how an adaptive browser-based search engine augmentation could support teachers in both exploratory and lookup search tasks. A preliminary user study was conducted to understand the actual search behaviour of teachers in their everyday environment and to investigate the inference of search tasks and context from search behaviour. Using this inferred information, personalized search aids could be delivered through the extension such as search engine result page (SERP) summarization for exploratory search and context-based search snippets. The relevance of existing search snippets was investigated and potential for improvement using teachers’ search context was identified. With the rest of the PhD, I plan to expand the user study to get more concrete results about teacher search behaviour and implement and test this adaptive browser extension to see how it affects the search experience for teachers.
- 27 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 20-24
Open AccessAdaptive Search Support for Teachers in Lesson Planning
Teachers report difficulties engaging in search in the context of lesson planning. Despite the availability of customised search solutions like Learning Object Repositories, most teachers report using general-purpose search engines like Google. Unlike students - to whom most research in educational search is directed - teachers are experts capable of using many types of resources for their tasks. In addition, since they teach multiple classes, they switch educational contexts frequently. These circumstances present unique challenges for user modelling. This thesis investigates how an adaptive browser-based search engine augmentation could support teachers in both exploratory and lookup search tasks. A preliminary user study was conducted to understand the actual search behaviour of teachers in their everyday environment and to investigate the inference of search tasks and context from search behaviour. Using this inferred information, personalized search aids could be delivered through the extension such as search engine result page (SERP) summarization for exploratory search and context-based search snippets. The relevance of existing search snippets was investigated and potential for improvement using teachers’ search context was identified. With the rest of the PhD, I plan to expand the user study to get more concrete results about teacher search behaviour and implement and test this adaptive browser extension to see how it affects the search experience for teachers.
- 26 June 2023
- Published by Springer Nature
- p. 515-527
Dropout Prediction in a Web Environment Based on Universal Design for Learning
- 26 June 2023
- Published by Springer Nature
- p. 515-527
Dropout Prediction in a Web Environment Based on Universal Design for Learning
- 30 March 2021
- Published by Oxford University Press (OUP) in Digital Scholarship in the Humanities
- Vol. 36 (4) , 886-904
Topic modelling on archive documents from the 1970s: global policies on refugees
This study conducts a historical analysis of global policies on refugees within typewritten and digitally born documents (c. 55,000 pages) from international and national archives. The data originate from the 1970s and are stored in archives from the UK and US governments, plus the United Nations High Commissioner for Refugees (UNHCR). The overarching theme is to analyse the involvement of the UK, the USA, and the UNHCR in different refugee cases that occurred during the 1970s. To do so, we (1) identify the main topics in each document; (2) investigate the transmission of topics horizontally (between organizations) and vertically (through time); and (3) suggest targeted areas of the document set for further close reading by historians. Standard Optical Character Recognition and object detection are used to extract information from documents and categorize them. Then, natural language processing (NLP) methods like topic modelling and clustering are used to identify topics and the relationships between them across time. The results identify several main themes covered by different organizations and how the focus of each organization changes diachronically. Besides its academic contribution, this study also demonstrates how, through the use of existing techniques with limited customization, digital technologies in the hands of the historian can augment and complement qualitative methods in bringing to light the themes and trends demonstrated in large bodies of historical documents.
- 30 March 2021
- Published by Oxford University Press (OUP) in Digital Scholarship in the Humanities
- Vol. 36 (4) , 886-904
Topic modelling on archive documents from the 1970s: global policies on refugees
This study conducts a historical analysis of global policies on refugees within typewritten and digitally born documents (c. 55,000 pages) from international and national archives. The data originate from the 1970s and are stored in archives from the UK and US governments, plus the United Nations High Commissioner for Refugees (UNHCR). The overarching theme is to analyse the involvement of the UK, the USA, and the UNHCR in different refugee cases that occurred during the 1970s. To do so, we (1) identify the main topics in each document; (2) investigate the transmission of topics horizontally (between organizations) and vertically (through time); and (3) suggest targeted areas of the document set for further close reading by historians. Standard Optical Character Recognition and object detection are used to extract information from documents and categorize them. Then, natural language processing (NLP) methods like topic modelling and clustering are used to identify topics and the relationships between them across time. The results identify several main themes covered by different organizations and how the focus of each organization changes diachronically. Besides its academic contribution, this study also demonstrates how, through the use of existing techniques with limited customization, digital technologies in the hands of the historian can augment and complement qualitative methods in bringing to light the themes and trends demonstrated in large bodies of historical documents.