The move towards preserving judgement disagreements in NLP requires the identification of adequate evaluation metrics. We identify a set of key properties that such metrics should have, and assess the extent to which natural candidates for soft evaluation such as Cross Entropy satisfy such properties. We employ a theoretical framework, supported by a visual approach, by practical examples, and by the analysis of a real case scenario. Our results indicate that Cross Entropy can result in fairly paradoxical results in some cases, whereas other measures Manhattan distance and Euclidean distance exhibit a more intuitive behavior, at least for the case of binary classification.

Research on abusive language detection and content moderation is crucial to combat online harm. However, current limitations set by regulatory bodies and social media platforms can make it difficult to share collected data. We address this challenge by exploring the possibility to replace existing datasets in English for abusive language detection with synthetic data obtained by rewriting original texts with an instruction-based generative model.We show that such data can be effectively used to train a classifier whose performance is in line, and sometimes better, than a classifier trained on original data. Training with synthetic data also seems to improve robustness in a cross-dataset setting. A manual inspection of the generated data confirms that rewriting makes it impossible to retrieve the original texts online.

This paper presents the submissions of the DH-FBK team for the three tasks of Task 10 at SemEval 2023. The Explainable Detection of Online Sexism (EDOS) task aims at detecting sexism in English text in an accurate and explainable way, thanks to a fine-grained annotation that follows a three-level schema: sexist or not (Task A), category of sexism (Task B) and vector of sexism (Task C) exhibited. We use a multi-task learning approach in which models share representations from all three tasks, allowing for knowledge to be shared across them. Notably, with our approach a single model can solve all three tasks. In addition, motivated by the subjective nature of the task, we incorporate inter-annotator agreement information in our multi-task architecture. Although disaggregated annotations are not available, we artificially estimate them using a 5-classifier ensemble, and show that ensemble agreement can be a good approximation of crowd agreement. Our approach achieves competitive results, ranking 32nd out of 84, 24th out of 69 and 11th out of 63 for Tasks A, B and C respectively. We finally show that low inter-annotator agreement levels are associated with more challenging examples for models, making agreement information use ful for this kind of task.

NLP datasets annotated with human judgments are rife with disagreements between the judges. This is especially true for tasks depending on subjective judgments such as sentiment analysis or offensive language detection. Particularly in these latter cases, the NLP community has come to realize that the common approach of reconciling’ these different subjective interpretations risks misrepresenting the evidence. Many NLP researchers have therefore concluded that rather than eliminating disagreements from annotated corpora, we should preserve themindeed, some argue that corpora should aim to preserve all interpretations produced by annotators. But this approach to corpus creation for NLP has not yet been widely accepted. The objective of the Le-Wi-Di series of shared tasks is to promote this approach to developing NLP models by providing a unified framework for training and evaluating with such datasets. We report on the second such shared task, which differs from the first edition in three crucial respects: (i) it focuses entirely on NLP, instead of both NLP and computer vision tasks in its first edition; (ii) it focuses on subjective tasks, instead of covering different types of disagreements as training with aggregated labels for subjective NLP tasks is in effect a misrepresentation of the data; and (iii) for the evaluation, we concentrated on soft approaches to evaluation. This second edition of Le-Wi-Di attracted a wide array of partici- pants resulting in 13 shared task submission papers.

In this work, we investigate olfactory perception shifts, analysing how the description of the smells emitted by specific sources has changed over time. We first create a benchmark of selected smell sources, relying upon existing historical studies related to olfaction. We also collect an English text corpus by retrieving large collections of documents from freely available resources, spanning from 1500 to 2000 and covering different domains. We label such corpus using a system for olfactory information extraction inspired by frame semantics, where the semantic roles around the smell sources in the benchmark are marked. We then analyse how the roles describing Qualities of smell sources change over time and how they can contribute to characterise perception shifts, also in comparison with more standard statistical approaches.

Annotators’ disagreement in linguistic data has been recently the focus of multiple initiatives aimed at raising awareness on issues related to ‘majority voting’ when aggregating diverging annotations. Disagreement can indeed reflect different aspects of linguistic annotation, from annotators’ subjectivity to sloppiness or lack of enough context to interpret a text. In this work we first propose a taxonomy of possible reasons leading to annotators’ disagreement in subjective tasks. Then, we manually label part of a Twitter dataset for offensive language detection in English following this taxonomy, identifying how the different categories are distributed. Finally we run a set of experiments aimed at assessing the impact of the different types of disagreement on classification performance. In particular, we investigate how accurately tweets belonging to different categories of disagreement can be classified as offensive or not, and how injecting data with different types of disagreement in the training set affects performance. We also perform offensive language detection as a multi-task framework, using disagreement classification as an auxiliary task.

Corpus-based studies on acceptability judgements have always stimulated the interest of researchers, both in theoretical and computational fields. Some approaches focused on spontaneous judgements collected through different types of tasks, others on data annotated through crowd-sourcing platforms, still others relied on expert annotated data available from the literature. The release of CoLA corpus, a large-scale corpus of sentences extracted from linguistic handbooks as examples of acceptable/non acceptable phenomena in English, has revived interest in the reliability of judgements of linguistic experts vs. non-experts. Several issues are still open. In this work, we contribute to this debate by presenting a 3D video game that was used to collect acceptability judgments on Italian sentences. We analyse the resulting annotations in terms of agreement among players and by comparing them with experts’ acceptability judgments. We also discuss different game settings to assess their impact on participants’ motivation and engagement. The final dataset containing 1,062 sentences, which were selected based on majority voting, is released for future research and comparisons.

The subtle and typically unconscious use of patronizing and condescending language (PCL) in large-audience media outlets undesirably feeds stereotypes and strengthens power-knowledge relationships, perpetuating discrimination towards vulnerable communities. Due to its subjective and subtle nature, PCL detection is an open and challenging problem, both for computational methods and human annotators. In this paper we describe the systems submitted by the DH-FBK team to SemEval-2022 Task 4, aiming at detecting PCL towards vulnerable communities in English media texts. Motivated by the subjectivity of human interpretation, we propose to leverage annotators’ uncertainty and disagreement to better capture the shades of PCL in a multi-task, multi-view learning framework. Our approach achieves competitive results, largely outperforming baselines and ranking on the top-left side of the leaderboard on both PCL identification and classification. Noticeably, our approach does not rely on any external data or model ensemble, making it a viable and attractive solution for real-world use.

The development of automated approaches to linguistic acceptability has been greatly fostered by the availability of the English CoLA corpus, which has also been included in the widely used GLUE benchmark. However, this kind of research for languages other than English, as well as the analysis of cross-lingual approaches, has been hindered by the lack of resources with a comparable size in other languages. We have therefore developed the ItaCoLA corpus, containing almost 10,000 sentences with acceptability judgments, which has been created following the same approach and the same steps as the English one. In this paper we describe the corpus creation, we detail its content, and we present the first experiments on this new resource. We compare in-domain and out-of-domain classification, and perform a specific evaluation of nine linguistic phenomena. We also present the first cross-lingual experiments, aimed at assessing whether multilingual transformer-based approaches can benefit from using sentences in two languages during fine-tuning.

Since state-of-the-art approaches to offensive language detection rely on supervised learning, it is crucial to quickly adapt them to the continuously evolving scenario of social media. While several approaches have been proposed to tackle the problem from an algorithmic perspective, so to reduce the need for annotated data, less attention has been paid to the quality of these data. Following a trend that has emerged recently, we focus on the level of agreement among annotators while selecting data to create offensive language datasets, a task involving a high level of subjectivity. Our study comprises the creation of three novel datasets of English tweets covering different topics and having five crowd-sourced judgments each. We also present an extensive set of experiments showing that selecting training and test data according to different levels of annotators’ agreement has a strong effect on classifiers performance and robustness. Our findings are further validated in cross-domain experiments and studied using a popular benchmark dataset. We show that such hard cases, where low agreement is present, are not necessarily due to poor-quality annotation and we advocate for a higher presence of ambiguous cases in future datasets, in order to train more robust systems and better account for the different points of view expressed online.