Diego Frassinelli

Anastasiia Sedova, Robert Litschko, Diego Frassinelli et al.

One of the major aspects contributing to the striking performance of large language models (LLMs) is the vast amount of factual knowledge accumulated during pre-training. Yet, many LLMs suffer from self-inconsistency, which raises doubts about their trustworthiness and reliability. This paper focuses on entity type ambiguity, analyzing the proficiency and consistency of state-of-the-art LLMs in applying factual knowledge when prompted with ambiguous entities. To do so, we propose an evaluation protocol that disentangles knowing from applying knowledge, and test state-of-the-art LLMs on 49 ambiguous entities. Our experiments reveal that LLMs struggle with choosing the correct entity reading, achieving an average accuracy of only 85%, and as low as 75% with underspecified prompts. The results also reveal systematic discrepancies in LLM behavior, showing that while the models may possess knowledge, they struggle to apply it consistently, exhibit biases toward preferred readings, and display self-inconsistencies. This highlights the need to address entity ambiguity in the future for more trustworthy LLMs.

Urban Knupleš, Diego Frassinelli, Sabine Schulte im Walde

Humans tend to strongly agree on ratings on a scale for extreme cases (e.g., a CAT is judged as very concrete), but judgements on mid-scale words exhibit more disagreement. Yet, collected rating norms are heavily exploited across disciplines. Our study focuses on concreteness ratings and (i) implements correlations and supervised classification to identify salient multi-modal characteristics of mid-scale words, and (ii) applies a hard clustering to identify patterns of systematic disagreement across raters. Our results suggest to either fine-tune or filter mid-scale target words before utilising them.

Jian Lan, Diego Frassinelli, Barbara Plank

Large vision-language models frequently struggle to accurately predict responses provided by multiple human annotators, particularly when those responses exhibit human uncertainty. In this study, we focus on the Visual Question Answering (VQA) task, and we comprehensively evaluate how well the state-of-the-art vision-language models correlate with the distribution of human responses. To do so, we categorize our samples based on their levels (low, medium, high) of human uncertainty in disagreement (HUD) and employ not only accuracy but also three new human-correlated metrics in VQA, to investigate the impact of HUD. To better align models with humans, we also verify the effect of common calibration and human calibration. Our results show that even BEiT3, currently the best model for this task, struggles to capture the multi-label distribution inherent in diverse human responses. Additionally, we observe that the commonly used accuracy-oriented calibration technique adversely affects BEiT3's ability to capture HUD, further widening the gap between model predictions and human distributions. In contrast, we show the benefits of calibrating models towards human distributions for VQA, better aligning model confidence with human uncertainty. Our findings highlight that for VQA, the consistent alignment between human responses and model predictions is understudied and should become the next crucial target of future studies.

Shijia Zhou, Saif M. Mohammad, Barbara Plank et al.

Internet memes represent a popular form of multimodal online communication and often use figurative elements to convey layered meaning through the combination of text and images. However, it remains largely unclear how multimodal large language models (MLLMs) combine and interpret visual and textual information to identify figurative meaning in memes. To address this gap, we evaluate eight state-of-the-art generative MLLMs across three datasets on their ability to detect and explain six types of figurative meaning. In addition, we conduct a human evaluation of the explanations generated by these MLLMs, assessing whether the provided reasoning supports the predicted label and whether it remains faithful to the original meme content. Our findings indicate that all models exhibit a strong bias to associate a meme with figurative meaning, even when no such meaning is present. Qualitative analysis further shows that correct predictions are not always accompanied by faithful explanations.

Robert Litschko, Barbara Plank, Diego Frassinelli

Automatic induction of high-quality dictionaries is essential for building lexical resources, yet low-resource languages and dialects pose several challenges: limited access to annotators, high degree of spelling variations, and poor performance of large language models (LLMs). We empirically show that statistical models (random forests) trained on string similarity features are surprisingly effective for inducing German dialect lexicons. They outperform LLMs, enable cross-dialect transfer, and offer a lightweight data-driven alternative. We evaluate our models intrinsically on bilingual lexicon induction (BLI) and extrinsically on dialect information retrieval (IR). On BLI, random forests outperform Mistral-123b while being more resource-lean. On dialect IR with BM25, using our dialect dictionaries for query expansion yields relative improvements of up to 28.9% in nDCG@10 and 50.7% in Recall@100. Motivated by the resource scarcity in dialects, we further investigate the extent to which models transfer across different German dialects, and their performance under varying amounts of training data.

Hyewon Jang, Diego Frassinelli

We tested the robustness of sarcasm detection models by examining their behavior when fine-tuned on four sarcasm datasets containing varying characteristics of sarcasm: label source (authors vs. third-party), domain (social media/online vs. offline conversations/dialogues), style (aggressive vs. humorous mocking). We tested their prediction performance on the same dataset (intra-dataset) and across different datasets (cross-dataset). For intra-dataset predictions, models consistently performed better when fine-tuned with third-party labels rather than with author labels. For cross-dataset predictions, most models failed to generalize well to the other datasets, implying that one type of dataset cannot represent all sorts of sarcasm with different styles and domains. Compared to the existing datasets, models fine-tuned on the new dataset we release in this work showed the highest generalizability to other datasets. With a manual inspection of the datasets and post-hoc analysis, we attributed the difficulty in generalization to the fact that sarcasm actually comes in different domains and styles. We argue that future sarcasm research should take the broad scope of sarcasm into account.

Elisa Leonardelli, Silvia Casola, Siyao Peng et al.

Many researchers have reached the conclusion that AI models should be trained to be aware of the possibility of variation and disagreement in human judgments, and evaluated as per their ability to recognize such variation. The LEWIDI series of shared tasks on Learning With Disagreements was established to promote this approach to training and evaluating AI models, by making suitable datasets more accessible and by developing evaluation methods. The third edition of the task builds on this goal by extending the LEWIDI benchmark to four datasets spanning paraphrase identification, irony detection, sarcasm detection, and natural language inference, with labeling schemes that include not only categorical judgments as in previous editions, but ordinal judgments as well. Another novelty is that we adopt two complementary paradigms to evaluate disagreement-aware systems: the soft-label approach, in which models predict population-level distributions of judgments, and the perspectivist approach, in which models predict the interpretations of individual annotators. Crucially, we moved beyond standard metrics such as cross-entropy, and tested new evaluation metrics for the two paradigms. The task attracted diverse participation, and the results provide insights into the strengths and limitations of methods to modeling variation. Together, these contributions strengthen LEWIDI as a framework and provide new resources, benchmarks, and findings to support the development of disagreement-aware technologies.

Andreas Säuberli, Diego Frassinelli, Barbara Plank

Knowing how test takers answer items in educational assessments is essential for test development, to evaluate item quality, and to improve test validity. However, this process usually requires extensive pilot studies with human participants. If large language models (LLMs) exhibit human-like response behavior to test items, this could open up the possibility of using them as pilot participants to accelerate test development. In this paper, we evaluate the human-likeness or psychometric plausibility of responses from 18 instruction-tuned LLMs with two publicly available datasets of multiple-choice test items across three subjects: reading, U.S. history, and economics. Our methodology builds on two theoretical frameworks from psychometrics which are commonly used in educational assessment, classical test theory and item response theory. The results show that while larger models are excessively confident, their response distributions can be more human-like when calibrated with temperature scaling. In addition, we find that LLMs tend to correlate better with humans in reading comprehension items compared to other subjects. However, the correlations are not very strong overall, indicating that LLMs should not be used for piloting educational assessments in a zero-shot setting.

Tarun Tater, Sabine Schulte im Walde, Diego Frassinelli · ibm-research

The visual representation of a concept varies significantly depending on its meaning and the context where it occurs; this poses multiple challenges both for vision and multimodal models. Our study focuses on concreteness, a well-researched lexical-semantic variable, using it as a case study to examine the variability in visual representations. We rely on images associated with approximately 1,000 abstract and concrete concepts extracted from two different datasets: Bing and YFCC. Our goals are: (i) evaluate whether visual diversity in the depiction of concepts can reliably distinguish between concrete and abstract concepts; (ii) analyze the variability of visual features across multiple images of the same concept through a nearest neighbor analysis; and (iii) identify challenging factors contributing to this variability by categorizing and annotating images. Our findings indicate that for classifying images of abstract versus concrete concepts, a combination of basic visual features such as color and texture is more effective than features extracted by more complex models like Vision Transformer (ViT). However, ViTs show better performances in the nearest neighbor analysis, emphasizing the need for a careful selection of visual features when analyzing conceptual variables through modalities other than text.

Andreas Säuberli, Darja Jepifanova, Diego Frassinelli et al.

The way our eyes move while reading can tell us about the cognitive effort required to process the text. In the present study, we use this fact to generate texts with controllable reading ease. Our method employs a model that predicts human gaze patterns to steer language model outputs towards eliciting certain reading behaviors. We evaluate the approach in an eye-tracking experiment with native and non-native speakers of English. The results demonstrate that the method is effective at making the generated texts easier or harder to read, measured both in terms of reading times and perceived difficulty of the texts. A statistical analysis reveals that the changes in reading behavior are mostly due to features that affect lexical processing. Possible applications of our approach include text simplification for information accessibility and generation of personalized educational material for language learning.

Robert Litschko, Verena Blaschke, Diana Burkhardt et al.

Dialects exhibit a substantial degree of variation due to the lack of a standard orthography. At the same time, the ability of Large Language Models (LLMs) to process dialects remains largely understudied. To address this gap, we use Bavarian as a case study and investigate the lexical dialect understanding capability of LLMs by examining how well they recognize and translate dialectal terms across different parts-of-speech. To this end, we introduce DiaLemma, a novel annotation framework for creating dialect variation dictionaries from monolingual data only, and use it to compile a ground truth dataset consisting of 100K human-annotated German-Bavarian word pairs. We evaluate how well nine state-of-the-art LLMs can judge Bavarian terms as dialect translations, inflected variants, or unrelated forms of a given German lemma. Our results show that LLMs perform best on nouns and lexically similar word pairs, and struggle most in distinguishing between direct translations and inflected variants. Interestingly, providing additional context in the form of example usages improves the translation performance, but reduces their ability to recognize dialect variants. This study highlights the limitations of LLMs in dealing with orthographic dialect variation and emphasizes the need for future work on adapting LLMs to dialects.

Ekta Sood, Simon Tannert, Diego Frassinelli et al.

While neural networks with attention mechanisms have achieved superior performance on many natural language processing tasks, it remains unclear to which extent learned attention resembles human visual attention. In this paper, we propose a new method that leverages eye-tracking data to investigate the relationship between human visual attention and neural attention in machine reading comprehension. To this end, we introduce a novel 23 participant eye tracking dataset - MQA-RC, in which participants read movie plots and answered pre-defined questions. We compare state of the art networks based on long short-term memory (LSTM), convolutional neural models (CNN) and XLNet Transformer architectures. We find that higher similarity to human attention and performance significantly correlates to the LSTM and CNN models. However, we show this relationship does not hold true for the XLNet models -- despite the fact that the XLNet performs best on this challenging task. Our results suggest that different architectures seem to learn rather different neural attention strategies and similarity of neural to human attention does not guarantee best performance.