Sinclair Schneider
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Sinclair Schneider has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Florian Steuber
Bundeswehr University Munich Bavaria Germany
6 shared publications - Gabi Dreo Rodosek
University of the Bundeswehr Munich, Research Institute CODE, Munich, Germany
5 shared publications - Joao A. G. Schneider
Bundeswehr University Munich Bavaria Germany
2 shared publications - João A. G. Schneider
University of the Bundeswehr Munich Germany
2 shared publications - Mirco Schoenfeld
University of Bayreuth, Bayreuth, Germany
1 shared publication
Latest Publications
- Open Access
- 9 January 2026
- Published by Springer Nature in Social Network Analysis and Mining
- Vol. 16 (1)
Guess where this post is from: a hierarchical approach to predict tweet locations
The effective use of social media data for applications such as urban and regional studies or public health monitoring often requires reliable location information. Yet, in the case of Twitter, only a vanishingly small 2% of tweets are provided with post-related geo-annotations. This paper addresses this shortcoming by presenting a classification framework that infers tweet locations with high precision, relying solely on textual and contextual information provided within the post itself. The approach has worldwide coverage and adapts its spatial granularity based on post volume within the respective area, ranging from country-level predictions in less active regions to state-level predictions in high-activity zones. We perform extensive evaluations to examine the impact of various factors on the algorithm’s performance with respect to different regions of the world. Furthermore, our analyses suggest that discarding a small percentage of tweets containing few hints on location significantly increases overall model quality.
- 9 January 2026
- Published by Springer Nature in Social Network Analysis and Mining
- Vol. 16 (1)
Open AccessGuess where this post is from: a hierarchical approach to predict tweet locations
The effective use of social media data for applications such as urban and regional studies or public health monitoring often requires reliable location information. Yet, in the case of Twitter, only a vanishingly small 2% of tweets are provided with post-related geo-annotations. This paper addresses this shortcoming by presenting a classification framework that infers tweet locations with high precision, relying solely on textual and contextual information provided within the post itself. The approach has worldwide coverage and adapts its spatial granularity based on post volume within the respective area, ranging from country-level predictions in less active regions to state-level predictions in high-activity zones. We perform extensive evaluations to examine the impact of various factors on the algorithm’s performance with respect to different regions of the world. Furthermore, our analyses suggest that discarding a small percentage of tweets containing few hints on location significantly increases overall model quality.
- Open Access
- 10 March 2025
- Published by Cambridge University Press (CUP) in Data & Policy
- Vol. 7
Detection avoidance techniques for large language models
The increasing popularity of large language models has not only led to widespread use but has also brought various risks, including the potential for systematically spreading fake news. Consequently, the development of classification systems such as DetectGPT has become vital. These detectors are vulnerable to evasion techniques, as demonstrated in an experimental series: Systematic changes of the generative models’ temperature proofed shallow learning—detectors to be the least reliable (Experiment 1). Fine-tuning the generative model via reinforcement learning circumvented BERT-based—detectors (Experiment 2). Finally, rephrasing led to a >90% evasion of zero-shot—detectors like DetectGPT, although texts stayed highly similar to the original (Experiment 3). A comparison with existing work highlights the better performance of the presented methods. Possible implications for society and further research are discussed.
- 10 March 2025
- Published by Cambridge University Press (CUP) in Data & Policy
- Vol. 7
Open AccessDetection avoidance techniques for large language models
The increasing popularity of large language models has not only led to widespread use but has also brought various risks, including the potential for systematically spreading fake news. Consequently, the development of classification systems such as DetectGPT has become vital. These detectors are vulnerable to evasion techniques, as demonstrated in an experimental series: Systematic changes of the generative models’ temperature proofed shallow learning—detectors to be the least reliable (Experiment 1). Fine-tuning the generative model via reinforcement learning circumvented BERT-based—detectors (Experiment 2). Finally, rephrasing led to a >90% evasion of zero-shot—detectors like DetectGPT, although texts stayed highly similar to the original (Experiment 3). A comparison with existing work highlights the better performance of the presented methods. Possible implications for society and further research are discussed.
- Open Access
- 1 January 2024
- Published by HICSS Conference Office
How good can machine generated texts be identified and can language models be trained to avoid identification?
With the rise of generative pre-trained transformer models such as GPT-3, GPT-NeoX, or OPT, distinguishing human-generated texts from machine-generated ones has become important. We refined five separate language models to generate synthetic tweets, uncovering that shallow learning classification algorithms, like Naive Bayes, achieve detection accuracy between 0.6 and 0.8. Shallow learning classifiers differ from human-based detection, especially when using higher temperature values during text generation, resulting in a lower detection rate. Humans prioritize linguistic acceptability, which tends to be higher at lower temperature values. In contrast, transformer-based classifiers have an accuracy of 0.9 and above. We found that using a reinforcement learning approach to refine our generative models can successfully evade BERT-based classifiers with a detection accuracy of 0.15 or less.
- 1 January 2024
- Published by HICSS Conference Office
Open AccessHow good can machine generated texts be identified and can language models be trained to avoid identification?
With the rise of generative pre-trained transformer models such as GPT-3, GPT-NeoX, or OPT, distinguishing human-generated texts from machine-generated ones has become important. We refined five separate language models to generate synthetic tweets, uncovering that shallow learning classification algorithms, like Naive Bayes, achieve detection accuracy between 0.6 and 0.8. Shallow learning classifiers differ from human-based detection, especially when using higher temperature values during text generation, resulting in a lower detection rate. Humans prioritize linguistic acceptability, which tends to be higher at lower temperature values. In contrast, transformer-based classifiers have an accuracy of 0.9 and above. We found that using a reinforcement learning approach to refine our generative models can successfully evade BERT-based classifiers with a detection accuracy of 0.15 or less.
- Open Access
- 6 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 300-304
Real-Time Anomaly Detection and Popularity Prediction for Emerging Events on Twitter
Due to their high volume and data recency, communications from social media platforms have become an excellent source for monitoring information diffusion. The insights leveraged are invaluable for social media analysts in the areas of event analysis and emergency management. Existing work ranges from the initial detection of incidents over information enrichment to determining an incident's relevance and life span. Until now, individual parts of this process have been considered separately, but never in combination. In this work, we address this crucial need and present an approach for detecting the onset and context of emerging events and predicting their popularity two weeks after emergence on Twitter in real time. Our contribution is threefold. We first present an online learning anomaly detection method refined with temporal clustering to identify abnormal conversational volumes of keywords. Second, we reconstruct potentially underlying events causing the anomaly through the enrichment of contextual and temporal information. Third, we assess an event's relevance and life span by predicting the resonance corresponding tweets receive shortly after their publication.
- 6 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 300-304
Open AccessReal-Time Anomaly Detection and Popularity Prediction for Emerging Events on Twitter
Due to their high volume and data recency, communications from social media platforms have become an excellent source for monitoring information diffusion. The insights leveraged are invaluable for social media analysts in the areas of event analysis and emergency management. Existing work ranges from the initial detection of incidents over information enrichment to determining an incident's relevance and life span. Until now, individual parts of this process have been considered separately, but never in combination. In this work, we address this crucial need and present an approach for detecting the onset and context of emerging events and predicting their popularity two weeks after emergence on Twitter in real time. Our contribution is threefold. We first present an online learning anomaly detection method refined with temporal clustering to identify abnormal conversational volumes of keywords. Second, we reconstruct potentially underlying events causing the anomaly through the enrichment of contextual and temporal information. Third, we assess an event's relevance and life span by predicting the resonance corresponding tweets receive shortly after their publication.
- Open Access
How well can machine-generated texts be identified and can language models be trained to avoid identification?
With the rise of generative pre-trained transformer models such as GPT-3, GPT-NeoX, or OPT, distinguishing human-generated texts from machine-generated ones has become important. We refined five separate language models to generate synthetic tweets, uncovering that shallow learning classification algorithms, like Naive Bayes, achieve detection accuracy between 0.6 and 0.8. Shallow learning classifiers differ from human-based detection, especially when using higher temperature values during text generation, resulting in a lower detection rate. Humans prioritize linguistic acceptability, which tends to be higher at lower temperature values. In contrast, transformer-based classifiers have an accuracy of 0.9 and above. We found that using a reinforcement learning approach to refine our generative models can successfully evade BERT-based classifiers with a detection accuracy of 0.15 or less.
Open AccessHow well can machine-generated texts be identified and can language models be trained to avoid identification?
With the rise of generative pre-trained transformer models such as GPT-3, GPT-NeoX, or OPT, distinguishing human-generated texts from machine-generated ones has become important. We refined five separate language models to generate synthetic tweets, uncovering that shallow learning classification algorithms, like Naive Bayes, achieve detection accuracy between 0.6 and 0.8. Shallow learning classifiers differ from human-based detection, especially when using higher temperature values during text generation, resulting in a lower detection rate. Humans prioritize linguistic acceptability, which tends to be higher at lower temperature values. In contrast, transformer-based classifiers have an accuracy of 0.9 and above. We found that using a reinforcement learning approach to refine our generative models can successfully evade BERT-based classifiers with a detection accuracy of 0.15 or less.
- 1 February 2022
- Published by Elsevier in Big Data Research
- Vol. 27
Embedding Semantic Anchors to Guide Topic Models on Short Text Corpora
- 1 February 2022
- Published by Elsevier in Big Data Research
- Vol. 27
Embedding Semantic Anchors to Guide Topic Models on Short Text Corpora