As LLMs excel on standard reading comprehension benchmarks, attention is shifting toward evaluating their capacity for complex abstract reasoning and inference. Literature-based benchmarks, with their rich narrative and moral depth, provide a compelling framework for evaluating such deeper comprehension skills. Here, we present Morables, a human-verified benchmark built from fables and short stories drawn from historical literature. The main task is structured as multiple-choice questions targeting moral inference, with carefully crafted distractors that challenge models to go beyond shallow, extractive question answering. To further stress-test model robustness, we introduce adversarial variants designed to surface vulnerabilities such as data contamination. Our findings show that while larger models outperform smaller ones, they remain susceptible to adversarial manipulation and often rely on superficial patterns rather than true moral reasoning. Interestingly, reasoning-enhanced models fail to bridge this gap, suggesting that scale -- not reasoning ability -- is the primary driver of performance.

Morables: A Benchmark for Assessing Abstract Moral Reasoning in LLMs with Fables

Puns are a form of humorous wordplay that exploits polysemy and phonetic similarity. While LLMs have shown promise in detecting puns, we show in this paper that their understanding often remains shallow, lacking the nuanced grasp typical of human interpretation. By systematically analyzing and reformulating existing pun benchmarks, we demonstrate how subtle changes in puns are sufficient to mislead LLMs. Our contributions include comprehensive and nuanced pun detection benchmarks, human evaluation across recent LLMs, and an analysis of the robustness challenges these models face in processing puns.

Pun Unintended: LLMs and the Illusion of Humor Understanding

The detection of sensitive content in large datasets is crucial for ensuring that shared and analysed data is free from harmful material. However, current moderation tools, such as external APIs, suffer from limitations in customisation, accuracy across diverse sensitive categories, and privacy concerns. Additionally, existing datasets and open-source models focus predominantly on toxic language, leaving gaps in detecting other sensitive categories such as substance abuse or self-harm. In this paper, we put forward a unified dataset tailored for social media content moderation across six sensitive categories: conflictual language, profanity,sexually explicit material, drug-related content, self-harm, and spam. By collecting and annotating data with consistent retrieval strategies and guidelines, we address the shortcomings of previous focalised research. Our analysis demonstrates that fine-tuning large language models (LLMs) on this novel dataset yields significant improvements in detection performance compared to open off-the-shelf models such as LLaMA, and even proprietary OpenAI models, which underperform by 10-15% overall. This limitation is even more pronounced on popular moderation APIs, which cannot be easily tailored to specific sensitive content categories, among others.

Sensitive Content Classification in Social Media: A Holistic Resource and Evaluation

Social media offers the potential to provide detection of outbreaks or public health incidents faster than traditional reporting mechanisms. In this paper, we developed and tested a pipeline to produce alerts of influenza-like illness (ILI) using Twitter data. Data was collected from the Twitter API, querying keywords referring to ILI symptoms and geolocated to Wales. Tweets that described first-hand descriptions of symptoms (as opposed to non-personal descriptions) were classified using transformer-based language models specialised on social media (BERTweet and TimeLMs), which were trained on a manually labelled dataset matching the above criteria. After gathering this data, weekly tweet counts were applied to the regression-based Noufaily algorithm to identify exceedances throughout 2022. The algorithm was also applied to counts of ILI-related GP consultations for comparison. Exceedance detection applied to the classified tweet counts produced alerts starting four weeks earlier than by using GP consultation data. These results demonstrate the potential to facilitate advanced preparedness for unexpected increases in healthcare burdens.

Towards a Social Media-based Disease Surveillance System for Early Detection of Influenza-like Illnesses: A Twitter Case Study in Wales

The ability to compare by analogy, metaphorically or not, lies at the core of how humans understand the world and communicate. In this paper, we study the likelihood of metaphoric outputs, and the capability of a wide range of pretrained transformer-based language models to identify metaphors from other types of analogies, including anomalous ones. In particular, we are interested in discovering whether language models recognise metaphorical analogies equally well as other types of analogies, and whether the model size has an impact on this ability. The results show that there are relevant differences using perplexity as a proxy, with the larger models reducing the gap when it comes to analogical processing, and for distinguishing metaphors from incorrect analogies. This behaviour does not result in increased difficulties for larger generative models in identifying metaphors in comparison to other types of analogies from anomalous sentences in a zero-shot generation setting, when perplexity values of metaphoric and non-metaphoric analogies are similar.

How Are Metaphors Processed by Language Models? The Case of Analogies

Social biases such as gender or racial biases have been reported in language models (LMs), including Masked Language Models (MLMs). Given that MLMs are continuously trained with increasing amounts of additional data collected over time, an important yet unanswered question is how the social biases encoded with MLMs vary over time. In particular, the number of social media users continues to grow at an exponential rate, and it is a valid concern for the MLMs trained specifically on social media data whether their social biases (if any) would also amplify over time. To empirically analyse this problem, we use a series of MLMs pretrained on chronologically ordered temporal snapshots of corpora. Our analysis reveals that, although social biases are present in all MLMs, most types of social bias remain relatively stable over time (with a few exceptions). To further understand the mechanisms that influence social biases in MLMs, we analyse the temporal corpora used to train the MLMs. Our findings show that some demographic groups, such as male, obtain higher preference over the other, such as female on the training corpora constantly.

Evaluating Short-Term Temporal Fluctuations of Social Biases in Social Media Data and Masked Language Models

In the dynamic realm of social media, diverse topics are discussed daily, transcending linguistic boundaries. However, the complexities of understanding and categorising this content across various languages remain an important challenge with traditional techniques like topic modelling often struggling to accommodate this multilingual diversity. In this paper, we introduce X-Topic, a multilingual dataset featuring content in four distinct languages (English, Spanish, Japanese, and Greek), crafted for the purpose of tweet topic classification. Our dataset includes a wide range of topics, tailored for social media content, making it a valuable resource for scientists and professionals working on cross-linguistic analysis, the development of robust multilingual models, and computational scientists studying online dialogue. Finally, we leverage X-Topic to perform a comprehensive cross-linguistic and multilingual analysis, and compare the capabilities of current general- and domain-specific language models.

Multilingual Topic Classification in X: Dataset and Analysis

Social media is an integral part of the daily life of an increasingly large number of people worldwide. Used for entertainment, communication and news updates, it constitutes a source of information that has been extensively used to study human behaviour. Unfortunately, the open nature of social media platforms along with the difficult task of supervising their content has led to a proliferation of misinformation posts. 
In this paper, we aim to identify the textual differences between the profiles of user that share misinformation from questionable sources and those that do not. Our goal is to better understand user behaviour in order to be better equipped to combat this issue. To this end, we identify Twitter (X) accounts of potential misinformation spreaders and apply transformer models specialised in social media to extract characteristics such as sentiment, emotion, topic and presence of hate speech. Our results indicate that, while there may be some differences between the behaviour of users that share misinformation and those that do not, there are no large differences when it comes to the type of content shared.

A Multi-Faceted NLP Analysis of Misinformation Spreaders in Twitter

In machine learning, temporal shifts occur when there are differences between training and test splits in terms of time. For streaming data such as news or social media, models are commonly trained on a fixed corpus from a certain period of time, and they can become obsolete due to the dynamism and evolving nature of online content. This paper focuses on temporal shifts in social media and, in particular, Twitter. We propose a unified evaluation scheme to assess the performance of language models (LMs) under temporal shift on standard social media tasks. LMs are tested on five diverse social media NLP tasks under different temporal settings, which revealed two important findings: (i) the decrease in performance under temporal shift is consistent across different models for entity-focused tasks such as named entity recognition or disambiguation, and hate speech detection, but not significant in the other tasks analysed (i.e., topic and sentiment classification); and (ii) continuous pre-training on the test period does not improve the temporal adaptability of LMs.

A Systematic Analysis on the Temporal Generalization of Language Models in Social Media

***Warning**: this paper contains content that may be offensive or upsetting.* Most hate speech datasets neglect the cultural diversity within a single language, resulting in a critical shortcoming in hate speech detection. To address this, we introduce **CREHate**, a **CR**oss-cultural **E**nglish **Hate** speech dataset. To construct CREHate, we follow a two-step procedure: 1) cultural post collection and 2) cross-cultural annotation. We sample posts from the SBIC dataset, which predominantly represents North America, and collect posts from four geographically diverse English-speaking countries (Australia, United Kingdom, Singapore, and South Africa) using culturally hateful keywords we retrieve from our survey. Annotations are collected from the four countries plus the United States to establish representative labels for each country. Our analysis highlights statistically significant disparities across countries in hate speech annotations. Only 56.2% of the posts in CREHate achieve consensus among all countries, with the highest pairwise label difference rate of 26%. Qualitative analysis shows that label disagreement occurs mostly due to different interpretations of sarcasm and the personal bias of annotators on divisive topics. Lastly, we evaluate large language models (LLMs) under a zero-shot setting and show that current LLMs tend to show higher accuracies on Anglosphere country labels in CREHate. Our dataset and codes are available at: https://github.com/nlee0212/CREHate

Exploring Cross-Cultural Differences in English Hate Speech Annotations: From Dataset Construction to Analysis

Relations such as "is influenced by", "is known for" or "is a competitor of" are inherently graded: we can rank entity pairs based on how well they satisfy these relations, but it is hard to draw a line between those pairs that satisfy them and those that do not. Such graded relations play a central role in many applications, yet they are typically not covered by existing Knowledge Graphs. In this paper, we consider the possibility of using Large Language Models (LLMs) to fill this gap. To this end, we introduce a new benchmark, in which entity pairs have to be ranked according to how much they satisfy a given graded relation. The task is formulated as a few-shot ranking problem, where models only have access to a description of the relation and five prototypical instances. We use the proposed benchmark to evaluate state-of-the-art relation embedding strategies as well as several publicly available LLMs and closed conversational models such as GPT-4. We find that smaller language models struggle to outperform a naive baseline. Overall, the best results are obtained with the 11B parameter Flan-T5 model and the 13B parameter OPT model, where further increasing the model size does not seem to be beneficial. For all models, a clear gap with human performance remains.

A RelEntLess Benchmark for Modelling Graded Relations between Named Entities

Various types of social biases have been reported with pretrained Masked Language Models (MLMs) in prior work. However, multiple underlying factors are associated with an MLM such as its model size, size of the training data, training objectives, the domain from which pretraining data is sampled, tokenization, and languages present in the pretrained corpora, to name a few. It remains unclear as to which of those factors influence social biases that are learned by MLMs. To study the relationship between model factors and the social biases learned by an MLM, as well as the downstream task performance of the model, we conduct a comprehensive study over 39 pretrained MLMs covering different model sizes, training objectives, tokenization methods, training data domains and languages. Our results shed light on important factors often neglected in prior literature, such as tokenization or model objectives.

A Predictive Factor Analysis of Social Biases and Task-Performance in Pretrained Masked Language Models

Metaphor identification aims at understanding whether a given expression is used figuratively in context. However, in this paper we show how existing metaphor identification datasets can be gamed by fully ignoring the potential metaphorical expression or the context in which it occurs. We test this hypothesis in a variety of datasets and settings, and show that metaphor identification systems based on language models without complete information can be competitive with those using the full context. This is due to the construction procedures to build such datasets, which introduce unwanted biases for positive and negative classes. Finally, we test the same hypothesis on datasets that are carefully sampled from natural corpora and where this bias is not present, making these datasets more challenging and reliable.

Construction Artifacts in Metaphor Identification Datasets | VIDEO

Despite its relevance, the maturity of NLP for social media pales in comparison with general-purpose models, metrics and benchmarks. This fragmented landscape makes it hard for the community to know, for instance, given a task, which is the best performing model and how it compares with others. To alleviate this issue, we introduce a unified benchmark for NLP evaluation in social media, SuperTweetEval, which includes a heterogeneous set of tasks and datasets combined, adapted and constructed from scratch. We benchmarked the performance of a wide range of models on SuperTweetEval and our results suggest that, despite the recent advances in language modelling, social media remains challenging.

SuperTweetEval: A Challenging, Unified and Heterogeneous Benchmark for Social Media NLP Research

This paper presents the Visual Word Sense Disambiguation (Visual-WSD) task. The objective of Visual-WSD is to identify among a set of ten images the one that corresponds to the intended meaning of a given ambiguous word which is accompanied with minimal context. The task provides datasets for three different languages: English, Italian, and Farsi. We received a total of 96 different submissions. Out of these, 40 systems outperformed a strong zero-shot CLIP based baseline (Radford et al., 2021). Participating systems proposed different zero- and few-shot approaches, often involving generative models and data augmentation. More information can be found on the task’s website: https://raganato.github.io/vwsd/.

SemEval-2023 Task 1: Visual Word Sense Disambiguation

The automatic detection of hate speech online is an active research area in NLP. Most of the studies to date are based on social media datasets that contribute to the creation of hate speech detection models trained on them. However, data creation processes contain their own biases, and models inherently learn from these dataset-specific biases. In this paper, we perform a large-scale cross-dataset comparison where we fine-tune language models on different hate speech detection datasets. This analysis shows how some datasets are more generalizable than others when used as training data. Crucially, our experiments show how combining hate speech detection datasets can contribute to the development of robust hate speech detection models. This robustness holds even when controlling by data size and compared with the best individual datasets. 


Jose Camacho-Collados

35

31

SHORT BIO

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES