This work represents significant progress in bridging the gap between advanced AI methods and the critical need for transparency and interpretability in psychiatric diagnosis and prediction.

Extensive experiments explore deep learning fusion strategies, including feature-level, model, and task fusion, with evaluation across two datasets covering five mental health conditions (attention deficit hyperactivity disorder, anxiety, bipolar disorder, depression and psychological stress).

We present SMHD-GER, a carefully constructed German dataset, alongside benchmark models incorporating (psycho-)linguistic features and BERT-German. This resource facilitates advanced research into digital biomarkers derived from verbal behavior.

We present our system for the SMMH23 Shared Task on binary classification of English Reddit posts self-reporting social anxiety disorder. We contrast hybrid and ensemble models employing domain-adapted transformers and BiLSTM neural networks, achieving 89.31% F1 on the validation set and 83.76% F1 on the test set.

Our work advances text simplification (TS) by enhancing explainable complexity prediction through psycho-linguistic features and pre-trained language models, while also extending a state-of-the-art Seq2Seq TS model to enable explicit control of ten attributes. Experiment results demonstrate improved performance in predicting complexity and significant enhancements in model performance across different settings.

We present our contribution to the TSAR-2022 Shared Task on Lexical Simplification, which enhances the LSBert system with a RoBERTa transformer model for candidate selection and introduces a new feature weighting scheme for substitution ranking. Achieving a 5.9% accuracy boost, our system secures the second position among 33 ranked solutions.

Recent efforts to predict emotions from text encounter challenges in real-world applications due to poor generalizability across domains. We introduce a novel approach combining transformer models with Bidirectional Long Short-Term Memory networks trained on psycholinguistic features. The proposed hybrid models exhibit improved out-of-domain robustness compared to standard transformer-based approaches and competitive performance on in-domain data.

We aim to uncover how individuals adapt their language use to different communication environments via quantitative analysis of linguistic complexity measures across writing tasks.

We detect six mental disorders leveraging BLSTM networks trained on psycholinguistic features for explainable detection, and combining them with Transformers to enhance prediction accuracy, while uncovering nuanced linguistic markers for specific conditions.

Our submission to the SMM4H22 focuses on detecting self-reported chronic stress on Twitter. Combining a pre-trained transformer model (RoBERTa) with a Bidirectional Long Short-Term Memory (BiLSTM) network trained on diverse psycholinguistic features, we address the class imbalance issue by augmenting the training dataset with another dataset used for stress classification in social media.

We introduce SPADE, the first dataset featuring continuous samples of argumentative speech labeled with the Big Five personality traits and enriched with socio-demographic data. Leveraging 436 engineered features and transformer models, our benchmark models aim to facilitate research in automatic personality detection, with feature ablation experiments shedding light on the predictive power of different types of features for individual personality traits.

We introduce two key advancements in personality prediction from verbal behavior: a comprehensive set of psycholinguistic features and hybrid models combining BERT with BLSTM networks trained on these features within-text distributions. Our models, evaluated on benchmark datasets for Big Five traits and MBTI types, outperform existing approaches, with ablation experiments assessing feature impact.

We investigate the use of NLP and ML to predict gaze patterns in naturalistic reading, with a focus on the impact of text characteristics on transformer-based language models (BERT and GPT-2). Through experiments, we determine how feature selection and model architecture influence eye-tracking prediction, while employing SP-LIME to assess the relative importance of different feature groups.

We predict TED talk-style affective ratings using a crowdsourced dataset of argumentative speech we created, elicited through debating prompts. Employing fine-tuning on a pre-trained model, we combine fluency features from automatic speech recognition with linguistic features. We employ SP-LIME to assess feature importance.

We address the critical need for benchmark datasets to combat COVID-19-related disinformation, particularly in the German language context. Introducing FANG-COVID, our dataset comprises real and fake German news articles, complemented by Twitter propagation data. Additionally, we propose an explainable model for fake news detection, comparing its performance to black-box models and conducting feature ablation to evaluate human-interpretable features

Alzheimers disease (AD) is a rapidly increasing neurodegenerative disease. Automated detection methods are crucial due to its increasing prevalence and high cost. In our work, linguistic features and pre-trained models are used to achieve 83.1% accuracy in detecting AD, with an ensemble model showing robust performance.

We evaluate the impacts of state-of-the-art Automatic Speech Recognition (ASR) systems on downstream text analytics systems in challenging scenarios, particularly focusing on spontaneously produced speech by second language learners. We assess 30 measures that capture the complexity, diversity, and sophistication of language production, revealing distinct effects of ASR on specific complexity measures while also accounting for variations in task types.

We predict second language learner proficiency using machine learning. Introducing complexity contours from a sliding window method, we employ recurrent neural network (RNN) classifiers to capture sequential information. Results show RNN classifiers trained on complexity contours outperform traditional methods, and sensitivity-based pruning highlights feature importance, validating CEFR levels.

Effective communication is paramount to personal, academic and professional success as it fosters social relationships and facilitates knowledge sharing. Prior research has addressed its verbal and nonverbal aspects, including auditory measures and nonverbal cues. Our work extends this research by employing a comprehensive set of psycholinguistic features in a multi-label classification task to predict affective ratings of online viewers across 14 categories.

The widespread issue of fake news involves false or misleading information presented as legitimate news, undermining fact-based reporting and hindering accurate perceptions for political actors, authorities, media, and citizens. While existing language-based methods for detecting fake news often rely on opaque models with high accuracy, there is a growing need to adopt transparent (explainable) models, particularly in critical sectors like healthcare, finances, military, and news. This study employs interpretable features derived from multi-disciplinary language approaches to train bi-directional recurrent neural network classification models. Applied to benchmark datasets, our approach yields comparable results to top-performing black box models, highlighting its potential. We perform ablation experiments to assess the significance of human-interpretable features in distinguishing fake news from genuine news.

Research spanning two decades highlights statistical learning mechanisms aiding language processing in children and adults, though often utilizing simplified artificial languages. We investigate the sensitivity of both native and non-native adult speakers to authentic language statistics across various English registers, revealing their ability to adapt to multiple distributional statistics during online processing.

Recent studies show significant differences in first language (L1) acquisition across different language components and life stages, which has led to an investigation of their impact on second language (L2) acquisition. Previous research has focused mainly on L1-L2 reading comprehension, here we extend it to the domain of L1-L2 writing. The results indicate strong correlations between L1 and L2 complexity on various linguistic measures and emphasize the persistent influence of the L1 on L2 ability.

We present a novel methodology for investigating the developmental trajectory of writing skills in English and German schoolchildren across various grade levels, utilizing classification tasks. Using two benchmark corpora, we leverage "complexity contours," sequences of measurements capturing the evolution of linguistic complexity within written texts. Integral to our approach is the application of Recurrent Neural Network (RNN) classifiers, adept at processing the sequential information inherent in these complexity contours.

The purpose of our work is twofold: (1) to introduce, to our knowledge, the largest available resource of keystroke logging (KSL) data generated by Etherpad (https://etherpad. org/), an open-source, web-based collaborative real-time editor, that captures the dynamics of second language (L2) production and (2) to relate the behavioral data from KSL to indices of syntactic and lexical complexity of the texts produced obtained from a tool that implements a sliding window approach capturing the progression of complexity within a text.

We investigate the relationship between verbal working memory (vWM), second language (L2) experience, and L2 sentence comprehension in L2 learners. Results indicate positive correlations between vWM and L2 experience, with vWM significantly influencing L2 sentence comprehension across different sentence types, underscoring their interconnection.

We provide an overview of the experience-related, cognitive, and affective factors influencing individual differences in both native and non-native language attainment. We address recent topics in cognitive science research and introduce methodological paradigms and statistical techniques to model variability in language performance data and connect it with individual differences factors.

Evidence indicates a strong connection between individual differences (ID) in statistical learning ability (SL) and language performance in native speakers, with potential implications for second language (L2) learners. We address the interaction between SL, personality traits, and L2 experience, revealing a nuanced relationship where SLs impact on language comprehension is moderated by personality traits, highlighting the complexity of ID factors in L2 comprehension.

Learning analytics and educational data mining have garnered increased attention as valuable tools for understanding human learning processes. This paper presents an adaptive language learning system aimed at monitoring and enhancing academic vocabulary skills, resulting in comprehensive longitudinal data on individual vocabulary development trajectories. We pursue a dual objective: firstly, to investigate the pace and pattern of vocabulary growth trajectories, and secondly, to elucidate the influence of various individual differences factors on the variability observed in these growth trajectories.

We investigate whether intermediate-advanced L2 speakers of English develop sensitivity to the frequencies of multiword sequences (MWS), akin to native speakers. Using eye-tracking, we replicated MWS frequency effects found in native speakers, revealing faster processing of sentences containing MWS compared to control items, indicating similar language processing mechanisms between native and non-native speakers.

Recent research has focused primarily on demonstrating that both native and non-native speakers learn the statistical properties of multiword sequences (MWS), but there is also increasing emphasis on accounting for individual differences (IDs) in language processing. We investigate the relationship between individual variability in online processing of MWS and statistical learning (SL) ability using a within-subject design and a series of SL tasks in different modalities.

We present a novel method that evaluates text complexity locally as "complexity contours" rather than using global summary statistics. The approach proves effective in automatically categorizing texts from different text genres.

We investigate the potential for implicit learning of novel form-meaning mappings outside controlled laboratory settings, utilizing crowdsourcing experiments, revealing that awareness at both noticing and understanding levels facilitated learning outcomes.

We investigate the role of statistical multi-word sequences in native and second language learning and explore how individual differences in cognitive and affective factors influence the processing of these sequences.

We introduce a novel method for automatically assessing text complexity, employing a sliding-window technique to monitor complexity distribution within a text. This distribution is represented by "complexity contours," derived from a series of measurements for a chosen linguistic complexity metric. Our approach is realized in the form of CoCoGen – Complexity Contour Generator, a computational tool currently supporting 32 indices of linguistic complexity. This paper aims to (1) outline the design of our computational tool based on the sliding-window technique and (2) demonstrate its application in the domain of second language (L2) learning, particularly in L2 writing assessment.