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.

As Large Language Models (LLMs) become available in a wider range of domains and applications, evaluating the truthfulness of multilingual LLMs is an issue of increasing relevance. TruthfulQA (Lin et al., 2022) is one of few benchmarks designed to evaluate how models imitate widespread falsehoods. However, it is strongly English-centric and starting to become outdated. We present VeritasQA, a context- and time-independent truthfulness benchmark built with multilingual transferability in mind, and available in Spanish, Catalan, Galician and English. VeritasQA comprises a set of 353 questions and answers inspired by common misconceptions and falsehoods that are not tied to any particular country or recent event. We release VeritasQA under an open license and present the evaluation results of 15 models of various architectures and sizes.

The current best practice to measure the performance of base Large Language Models is to establish a multi-task benchmark that covers a range of capabilities of interest. Currently, however, such benchmarks are only available in a few high-resource languages. To address this situation, we present IberoBench, a multilingual, multi-task benchmark for Iberian languages (i.e., Basque, Catalan, Galician, European Spanish and European Portuguese) built on the LM Evaluation Harness framework. The benchmark consists of 62 tasks divided into 179 subtasks. We evaluate 33 existing LLMs on IberoBench on 0- and 5-shot settings. We also explore the issues we encounter when working with the Harness and our approach to solving them to ensure high-quality evaluation.

Trainable metrics for machine translation evaluation have been scoring the highest correlations with human judgements in the latest meta-evaluations, outperforming traditional lexical overlap metrics such as BLEU, which is still widely used despite its well-known shortcomings. In this work we look at COMET, a prominent neural evaluation system proposed in 2020, to analyze the extent of its language support restrictions, and to investigate strategies to extend this support to new, under-resourced languages. Our case study focuses on English-Maltese and Spanish-Basque. We run a crowd-based evaluation campaign to collect direct assessments and use the annotated dataset to evaluate COMET-22, further fine-tune it, and to train COMET models from scratch for the two language pairs. Our analysis suggests that COMET’s performance can be improved with fine-tuning, and that COMET can be highly susceptible to the distribution of scores in the training data, which especially impacts low-resource scenarios.

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.