In humans, cognitive biases are systematic deviations from rationality in judgment that simplify complex decisions. They typically manifest as a consequence of learned behaviors or limitations on information processing capabilities. Recent work has shown that these biases can percolate through training data and ultimately be learned by language models. We examine different groups of models, factoring in model size and type (base or instructed) for four kinds of cognitive bias: primacy, recency, common token, and majority class bias. We evaluate the performance of each model for each type of bias in different settings using simple and complex variants of datasets. Our results show that some biases have much stronger effects than others, and that task complexity plays a part in eliciting stronger effects for some of these biases as measured by effect size. We show that some cognitive biases such as common token and majority class bias are not straightforward to evaluate, and that, contrary to some of the previous literature, some effects that have been previously classified as common token bias in the literature are actually due to primacy and recency bias.

We present the findings and results of our pseudonymisation system, which has been developed for a real-life use-case involving users and an informative chatbot in the context of the COVID-19 pandemic. Message exchanges between the two involve the former group providing information about themselves and their residential area, which could easily allow for their re-identification. We create a modular pipeline to detect PIIs and perform basic deidentification such that the data can be stored while mitigating any privacy concerns. The use-case presents several challenging aspects, the most difficult of which is the logistic challenge of not being able to directly view or access the data due to the very privacy issues we aim to resolve. Nevertheless, our system achieves a high recall of 0.99, correctly identifying almost all instances of personal data. However, this comes at the expense of precision, which only reaches 0.64. We describe the sensitive information identification in detail, explaining the design principles behind our decisions. We additionally highlight the particular challenges we’ve encountered.

We present and describe two language resources in this paper: CATalog 1.0, the largest text corpus in Catalan to date, and CURATE (Corpus Utility for RAting TExt), a modular, parallelizable pipeline used for processing and scoring documents based on text quality that we have optimised to run in High Performance Cluster (HPC) environments. In the coming sections we describe our data preprocessing pipeline at length; traditional pipelines usually implement a set of binary filters such that a given document is either in or out. In our experience with Catalan, in lower-resource settings it is more practical to instead assign a document a soft score to allow for more flexible decision-making. We describe how the document score is calculated and highlight its interpretability by showing that it is significantly correlated with human judgements as obtained from a comparative judgement experiment. We additionally describe the different subcorpora that make up CATalog 1.0.

Much work has gone into developing language models of increasing size, but only recently have we begun to examine them for pernicious behaviour that could lead to harming marginalised groups. Following Lin et al. (2022) in rooting our work in psychological research, we prompt two masked language models (MLMs) of different specialisations in English and Spanish with statements from a questionnaire developed to measure stigma to determine if they treat physical and mental illnesses equally. In both models we find a statistically significant difference in the treatment of physical and mental illnesses across most if not all latent constructs as measured by the questionnaire, and thus they are more likely to associate mental illnesses with stigma. We then examine their training data or data retrieved from the same domain using a computational implementation of the Stereotype Content Model (SCM) (Fiske et al., 2002; Fraser et al., 2021) to interpret the questionnaire results based on the SCM values as reflected in the data. We observe that model behaviour can largely be explained by the distribution of the mentions of illnesses according to their SCM values.

Training classification models on clinical speech is a time-saving and effective solution for many healthcare challenges, such as screening for Alzheimer’s Disease over the phone. One of the primary limiting factors of the success of artificial intelligence (AI) solutions is the amount of relevant data available. Clinical data is expensive to collect, not sufficient for large-scale machine learning or neural methods, and often not shareable between institutions due to data protection laws. With the increasing demand for AI in health systems, generating synthetic clinical data that maintains the nuance of underlying patient pathology is the next pressing task. Previous work has shown that automated evaluation of clinical speech tasks via automatic speech recognition (ASR) is comparable to manually annotated results in diagnostic scenarios even though ASR systems produce errors during the transcription process. In this work, we propose to generate synthetic clinical data by simulating ASR deletion errors on the transcript to produce additional data. We compare the synthetic data to the real data with traditional machine learning methods to test the feasibility of the proposed method. Using a dataset of 50 cognitively impaired and 50 control Dutch speakers, ten additional data points are synthetically generated for each subject, increasing the training size for 100 to 1000 training points. We find consistent and comparable performance of models trained on only synthetic data (AUC=0.77) to real data (AUC=0.77) in a variety of traditional machine learning scenarios. Additionally, linear models are not able to distinguish between real and synthetic data.

Effective management of dementia hinges on timely detection and precise diagnosis of the underlying cause of the syndrome at an early mild cognitive impairment (MCI) stage. Verbal fluency tasks are among the most often applied tests for early dementia detection due to their efficiency and ease of use. In these tasks, participants are asked to produce as many words as possible belonging to either a semantic category (SVF task) or a phonemic category (PVF task). Even though both SVF and PVF share neurocognitive function profiles, the PVF is typically believed to be less sensitive to measure MCI-related cognitive impairment and recent research on fine-grained automatic evaluation of VF tasks has mainly focused on the SVF. Contrary to this belief, we show that by applying state-of-the-art semantic and phonemic distance metrics in automatic analysis of PVF word productions, in-depth conclusions about production strategy of MCI patients are possible. Our results reveal a dissociation between semantically- and phonemically-guided search processes in the PVF. Specifically, we show that subjects with MCI rely less on semantic- and more on phonemic processes to guide their word production as compared to healthy controls (HC). We further show that semantic similarity-based features improve automatic MCI versus HC classification by 29% over previous approaches for the PVF. As such, these results point towards the yet underexplored utility of the PVF for in-depth assessment of cognition in MCI.

We describe the Saarland University submission to the shared task on Cross-Framework Meaning Representation Parsing (MRP) at the 2019 Conference on Computational Natural Language Learning (CoNLL).