Grammatical error detection involves diverse error types, with real-word errors being particularly challenging because incorrect words remain lexically valid but contextually inappropriate. Transformer-based models achieve strong performance but typically require repeated task-specific fine-tuning, which is computationally expensive and impractical for large-scale or evolving document processing scenarios. We address this limitation by formulating cross–error-type transfer as a representation-level domain adaptation problem. Our analysis shows that fine-tuning on different error types induces systematic distributional shifts in embedding space. Instead of updating model parameters, we propose aligning target embedding distributions to a source domain, enabling zero-shot transfer while keeping the encoder fixed. We evaluate two lightweight alignment methods, PCA-based alignment and CORAL, on three linguistically complex German real-word error types (case, verb, and capitalization). Both approaches consistently outperform zero-shot baselines and recover a substantial portion of the performance gap to full target-domain fine-tuning, while reducing adaptation time and computational cost by more than an order of magnitude. The results demonstrate an efficient and practical strategy for adapting grammatical error detection systems to new error types in resource-constrained, real-world NLP applications.

Over the past years, extensive research has led to significant advancements in tools for the automatic detection and correction of errors in documents. Despite this progress, several challenges remain unresolved. In particular, the identification of real-word errors – errors involving words that are grammatically valid but contextually inappropriate within a given sentence – continues to pose a considerable difficulty. Addressing such errors requires models with a sophisticated understanding of linguistic context. Transformer-based language models are particularly well-suited for this task due to their contextual modeling capabilities. To further enhance their performance, we propose a boosting-based training approach in conjunction with a synthetically generated data set created via pattern-based noise injection. We evaluate this method across three transformer-based architectures, viz. mBERT, LLaMA 3, and Mistral 7B. Our experimental results show that the boosting-based strategy consistently improves real-word error detection across all models. A subsequent in-depth error analysis reveals limitations in the synthetic training data, prompting the development of a targeted fine-tuning procedure designed to address these shortcomings and further optimize model performance. A comparison with prompt-based inference using a large language model demonstrates that specialized, fine-tuned models yield more reliable performance for this task. Finally, an evaluation under realistic class imbalance highlights practical trade-offs between ranking quality and threshold-based detection, particularly for rare error types.

The present paper introduces a multilingual data set of erroneous and correct text sentences. The novel data set marks a significant advancement from an existing corpus by incorporating additional samples and refining its overall structure. The primary purpose of this data set is to support the research and development of automated error detection systems, especially in the multilingual setting where high-quality data sets are scarce. A distinctive feature of our data set is that it incorporates only incorrect sentences and their corresponding correct versions. These sentences are sourced from a variety of texts written by native speakers from different industries, such as pharmaceuticals, banking, insurance, retail, communications, and more. Each sentence in the data set has been annotated by professional proofreaders. The paper includes a comprehensive error analysis, where we classify and scrutinize the different types of errors within the data set. By categorizing and analysing the errors in the data set, we aim to identify patterns and common issues. Additionally, we conduct a thorough experimental evaluation using a well-established language model. Our analysis assesses the classification accuracy measured over all errors and the accuracy of each specific error type. Interestingly, our results show that while some error types can be detected with an accuracy exceeding 80

The increasing complexity and scale of technical document corpora present challenges for consistency verification, particularly in politically sensitive or high-stakes contexts. This paper proposes an iterative approach that integrates long-context large language models (LLMs), human expertise, and hybrid clustering mechanisms to address these challenges. The approach focuses on two types of inconsistencies: real inconsistencies, such as contradictory statements or omissions, and fabricated inconsistencies, which are plausible yet artificially introduced. This paper uses the Swiss National Cooperative for the Disposal of Radioactive Waste (Nagra) and its corpus of up to 300 technical documents as a case study. Experimental results suggest that targeted structuring of document contexts improves recall in inconsistency detection. The findings highlight the potential of combining structured human input with LLM-based reasoning for improving document integrity and trustworthiness. Future work will focus on refining the approach, including automated clustering strategies and optimization of prompt engineering.

In patients undergoing systemic treatment for cancer, symptom tracking via electronic patient-reported outcomes (ePROs) has been used to optimize communication and monitoring, and facilitate the early detection of adverse effects and to compare the side effects of similar drugs. We aimed to examine whether the patterns in electronic patient-reported outcomes, without any additional clinician data input, are predictive of the underlying cancer type and reflect tumor- and treatment-associated symptom clusters (SCs). The data were derived from a total of 226 patients who self-reported on the presence and severity (according to the Common Terminology Criteria for Adverse Events (CTCAEs)) of more than 90 available symptoms via the mediduxTM app (versions 2.0 and 3.2, developed by mobile Health AG based in Zurich, Switzerland). Among these, 172 had breast cancer as the primary tumor, 19 had lung, 16 had gut, 12 had blood-lymph, and 7 had prostate cancer. For this secondary analysis, a subgroup of 25 patients with breast cancer were randomly selected to reduce the risk of overfitting. The symptoms were aggregated by counting the days on which a particular symptom was reported, resulting in a symptom vector for each patient. A logistic regression model was trained to predict the type of the respective tumor from the symptom vectors, and the symptoms with coefficients above (0.1) were graphically displayed. The machine learning model was not able to recognize any of the patients with prostate and blood-lymph cancer, likely as these cancer types were barely represented in the dataset. The Area Under the Curve (AUC) values for the three remaining cancer types were breast cancer: 0.74 (95% CI [0.624, 0.848]); gut cancer: 0.78 (95% CI [0.659, 0.893]); and lung cancer: 0.63 (95% CI [0.495, 0.771]). Despite the small datasets, for the breast and gut cancers, the respective models demonstrated a fair predictive performance (AUC > 0.7). The generalization of the findings are limited especially due to the heterogeneity of the dataset. This line of research could be especially interesting to monitor individual treatment trajectories. Deviations in the electronic patient-reported symptoms from the treatment-associated symptom patterns could dynamically indicate treatment non-adherence or lower treatment efficacy, without clinician input or additional costs. Similar analyses on larger patient cohorts are needed to validate these preliminary findings and to identify specific and robust treatment profiles.

We propose conversational agents as a means to simulate expert interviews, integrated into a simulational learning environment: ChEdventure. Designing and developing conversational agents using the existing tools and frameworks requires technical knowledge and a considerable learning curve. Recently, LLMs are being leveraged for their adaptability to different domains and their ability to perform various tasks in a natural, human-like conversational style. In this work, we explore if LLMs can help educators easily create conversational agents for their individual teaching goals. We propose a generalized template-based approach using LLMs that can instantiate conversational agents as an integrable component of teaching and learning activities. We evaluate our approach using prototypes generated from this template and identify guidelines to improve the experience of educators.

This position paper presents a novel approach of semantic verification in Large Language Model-based Retrieval Augmented Generation (LLM-RAG) systems, focusing on the critical need for factually accurate information dissemination during public debates, especially prior to plebiscites e.g. in direct democracies, particularly in the context of Switzerland. Recognizing the unique challenges posed by the current generation of Large Language Models (LLMs) in maintaining factual integrity, this research proposes an innovative solution that integrates retrieval mechanisms with enhanced semantic verification processes. The paper outlines a comprehensive methodology following a Design Science Research approach, which includes defining user personas, designing conversational interfaces, and iteratively developing a hybrid dialogue system. Central to this system is a robust semantic verification framework that leverages a knowledge graph for fact-checking and validation, ensuring the correctness and consistency of information generated by LLMs. The paper discusses the significance of this research in the context of Swiss direct democracy, where informed decision-making is pivotal. By improving the accuracy and reliability of information provided to the public, the proposed system aims to support the democratic process, enabling citizens to make well-informed decisions on complex issues. The research contributes to advancing the field of natural language processing and information retrieval, demonstrating the potential of AI and LLMs in enhancing civic engagement and democratic participation.

This study delves into the application of large language models (LLMs), such as ChatGPT-4, for the automated evaluation of student essays, with a focus on a case study conducted at the Swiss Institute of Business Administration. It explores the effectiveness of LLMs in assessing German-language student transfer assignments, and contrasts their performance with traditional evaluations by human lecturers. The primary findings highlight the challenges faced by LLMs in terms of accurately grading complex texts according to predefined categories and providing detailed feedback. This research illuminates the gap between the capabilities of LLMs and the nuanced requirements of student essay evaluation. The conclusion emphasizes the necessity for ongoing research and development in the area of LLM technology to improve the accuracy, reliability, and consistency of automated essay assessments in educational contexts.

BackgroundThe use of smartphone apps in cancer patients undergoing systemic treatment can promote the early detection of symptoms and therapy side effects and may be supported by machine learning (ML) for timely adaptation of therapies and reduction of adverse events and unplanned admissions.ObjectiveWe aimed to create an Early Warning System (EWS) to predict situations where supportive interventions become necessary to prevent unplanned visits. For this, dynamically collected standardized electronic patient reported outcome (ePRO) data were analyzed in context with the patient's individual journey. Information on well-being, vital parameters, medication, and free text were also considered for establishing a hybrid ML model. The goal was to integrate both the strengths of ML in sifting through large amounts of data and the long-standing experience of human experts. Given the limitations of highly imbalanced datasets (where only very few adverse events are present) and the limitations of humans in overseeing all possible cause of such events, we hypothesize that it should be possible to combine both in order to partially overcome these limitations.MethodsThe prediction of unplanned visits was achieved by employing a white-box ML algorithm (i.e., rule learner), which learned rules from patient data (i.e., ePROs, vital parameters, free text) that were captured via a medical device smartphone app. Those rules indicated situations where patients experienced unplanned visits and, hence, were captured as alert triggers in the EWS. Each rule was evaluated based on a cost matrix, where false negatives (FNs) have higher costs than false positives (FPs, i.e., false alarms). Rules were then ranked according to the costs and priority was given to the least expensive ones. Finally, the rules with higher priority were reviewed by two oncological experts for plausibility check and for extending them with additional conditions. This hybrid approach comprised the application of a sensitive ML algorithm producing several potentially unreliable, but fully human-interpretable and -modifiable rules, which could then be adjusted by human experts.ResultsFrom a cohort of 214 patients and more than 16’000 available data entries, the machine-learned rule set achieved a recall of 19% on the entire dataset and a precision of 5%. We compared this performance to a set of conditions that a human expert had defined to predict adverse events. This “human baseline” did not discover any of the adverse events recorded in our dataset, i.e., it came with a recall and precision of 0%. Despite more plentiful results were expected by our machine learning approach, the involved medical experts a) had understood and were able to make sense of the rules and b) felt capable to suggest modification to the rules, some of which could potentially increase their precision. Suggested modifications of rules included e.g., adding or tightening certain conditions to make them less sensitive or changing the rule consequences: sometimes further monitoring the situation, applying certain test (such as a CRP test) or applying some simple pain-relieving measures was deemed sufficient, making a costly consultation with the physician unnecessary. We can thus conclude that it is possible to apply machine learning as an inspirational tool that can help human experts to formulate rules for an EWS. While humans seem to lack the ability to define such rules without such support, they are capable of modifying the rules to increase their precision and generalizability.ConclusionsLearning rules from dynamic ePRO datasets may be used to assist human experts in establishing an early warning system for cancer patients in outpatient settings.

Choosing courses at the beginning of each semester is a complex decision that affects students' future careers and academic performance, especially when given the freedom to choose. Among other factors, the expected grade at the end of the semester and/or the expected ability to successfully complete a course plays an important role in course selection. This paper introduces a prototype for predicting student performance using state-of-the-art natural language processing techniques. The prototype, designed to assist students in course selection, uses historical course enrollment data and current course descriptions to predict possible grades and warn students of possible negative performance. A large language model, BERT, was used to analyse text and create course description embeddings. For this purpose, descriptions of courses a student has attended were considered and formed the basis for the student knowledge profile. In addition, student performance profiles are created by examining grades from the historical enrolment data. This two-pronged analysis is used to identify patterns that lead to negative study results. Although the idea of creating knowledge profiles based on course descriptions is promising, the evaluation showed room for improvement in terms of accuracy and recall.

In this paper, we present the early results of a hybrid intelligent approach that consists of an interpretable rule-based machine learning model for the prediction of unplanned visits of cancer patients. The approach is contextualized within the area of personalized medicine and will contribute to the development of an early-warning system (EWS) whose goal is to support cancer patients to cope with their daily symptoms remotely, by avoiding as much as possible physician visits. The interpretability of rules makes it possible to involve medical experts in the learning process who can accept, reject or modify rules, e.g. by adding conditions to increase their precision. The results appear to be promising as the discovered rules provide value in the identification of critical situations of patients. Experts also suggested the modification of rules for recommending not only visits to a physician, but also other (less costly) actions, such as increasing the dosage of pain killers – an extension to the EWS that would not have been possible without our hybrid approach. Overall, our first experiments showed how a new form of “dialogue” between the experts and the machine learning algorithm started to emerge.

The present paper introduces a novel benchmark data set for automatic error detection as well as error correction in text documents based on language models or other techniques. The data set contains a large number of sentences from various domains annotated with various types of errors (orthographic, grammatical, punctuation, and typography errors). The paper presents the method used to collect and annotate the documents, provides statistical analyses of the data set’s properties and evaluates two preliminary baseline models for automatic error detection on a specific benchmark task. The results show, on the one hand, the effectiveness of the proposed data set for the evaluation of automatic error detection systems. On the other hand, these initial analyses also reveal that the data set contains challenging cases that are difficult to detect. Finally, the paper discusses potential applications of the proposed data set in the development and research of error detection and error correction systems.

The spreading of fake news or misinformation on social media is a serious threat to modern societies, making more and more people susceptible to being unfairly influenced in their decision-making, be it in elections or other democratic processes. We contribute to the body of work in the area of fake news detection by studying cross-platform, multivariate spreading patterns of fake news on Covid-19-related topics – where existing studies have focused strongly on single platforms and/or on single metrics or indicators. Our findings show that there are several attributes that are specific to the cross-platform spreading process that become important predictors of fake news: there is e.g. a clear tendency that fake news travels faster from one platform to the other than real news. Meanwhile, although we have compiled a cross-platform corpus of fake and real news that future research may build on, data availability remains a challenge for future work.

Although machine learning algorithms outperform humans in many predictive tasks, their quality depends much on the availability of sufficient and representative training data. On the other hand, humans are capable of making predictions based on “spontaneous” transfers of knowledge from other domains or situations in cases where no directly relevant experiences exist. This can be seen very well in the task of predicting lead times in goods transport, where sudden disruptions or shortages may occur that are not reflected in historical data, but known to a well-informed human. If the variation can be anticipated and more accurate lead times estimated, proactive measures can be taken to decrease the impact. Therefore, we describe three novel approaches for delivery time predictions, combining a machine learning model with human input. The proposed logic covers two phases, learning based on actual delivery data and capturing human knowledge to cover exceptional situations not reflected in historical data. The proposed models and the resulting estimates were evaluated using deliveries from a retail company. It was found that the pure machine learning model delivers better results than a combination of humans and machines. On the one hand, this is caused by the complexity of incorporating human knowledge into the algorithm in a suitable way. On the other hand, it is also due to the tendency of humans to over-generalise the impact of certain events. Thus, although the pure machine learning model delivers superior estimation accuracy than the human-machine combination, our systematic qualitative analysis of the results presents insights for future development in this area.

The flexible data models of knowledge graphs (KGs) are powerful tools for handling large and dynamic data sets and are increasingly used for the tasks of data processing and storage. Although a KG may contain rich data and powerful connections, it is upon the searchers to explore these graphs and make sense out of them. The objective of this research paper is to investigate if and how KG exploration can be improved from a user’s point of view, to enhance the discovery of information. A qualitative user study should deliver insights on how different users interact with a KG, at what point they struggle and missed potential discoveries. Recognizing and understanding the intentions of the users is necessary to create solutions that support them best in their particular situation. Based on the findings, new features and improvements are suggested, developed and added to a prototypical KG exploration application, to be finally tested with regard to their impact on user exploration and acceptance. Based on the collected data we could identify the best guidance mechanisms that improve KG exploration the most.

Preschool children, when diagnosed with Autism Spectrum Disorder (ASD), often ex- perience a long and painful journey on their way to self-advocacy. Access to standard of care is poor, with long waiting times and the feeling of stigmatization in many social set- tings. Early interventions in ASD have been found to deliver promising results, but have a high cost for all stakeholders. Some recent studies have suggested that digital biomarkers (e.g., eye gaze), tracked using affordable wearable devices such as smartphones or tablets, could play a role in identifying children with special needs. In this paper, we discuss the possibility of supporting neurodiverse children with technologies based on digital biomark- ers which can help to a) monitor the performance of children diagnosed with ASD and b) predict those who would benefit most from early interventions. We describe an ongoing feasibility study that uses the “DREAM dataset”, stemming from a clinical study with 61 pre-school children diagnosed with ASD, to identify digital biomarkers informative for the child’s progression on tasks such as imitation of gestures. We describe our vision of a tool that will use these prediction models and that ASD pre-schoolers could use to train certain social skills at home. Our discussion includes the settings in which this usage could be embedded.

Besides the typical applications of recommender systems in B2C scenarios such as movie or shopping platforms, there is a rising interest in transforming the human-driven advice provided, e.g., in consultancy via the use of recommender systems. We explore the special characteristics of such knowledge-based B2B services and propose a process that allows incorporating recommender systems into them. We suggest and compare several recommender techniques that allow incorporating the necessary contextual knowledge (e.g., company demographics). These techniques are evaluated in isolation on a test set of business intelligence consultancy cases. We then identify the respective strengths of the different techniques and propose a new hybridisation strategy to combine these strengths. Our results show that the hybridisation leads to substantial performance improvement over the individual methods.