The use of synthetic data for training models for a variety of NLP tasks is now widespread. However, previous work reports mixed results with regards to its effectiveness on highly subjective tasks such as hate speech detection. In this paper, we present an in-depth qualitative analysis of the potential and specific pitfalls of synthetic data for hate speech detection in English, with 3,500 manually annotated examples. We show that, across different models, synthetic data created through paraphrasing gold texts can improve out-of-distribution robustness from a computational standpoint. However, this comes at a cost: synthetic data fails to reliably reflect the characteristics of real-world data on a number of linguistic dimensions, it results in drastically different class distributions, and it heavily reduces the representation of both specific identity groups and intersectional hate.

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.

Exploring and understanding language data is a fundamental stage in all areas dealing with human language. It allows NLP practitioners to uncover quality concerns and harmful biases in data before training, and helps linguists and social scientists to gain insight into language use and human behavior. Yet, there is currently a lack of a unified, customizable tool to seamlessly inspect and visualize language variation and bias across multiple variables, language units, and diverse metrics that go beyond descriptive statistics. In this paper, we introduce Variationist, a highly-modular, extensible, and task-agnostic tool that fills this gap. Variationist handles at once a potentially unlimited combination of variable types and semantics across diversity and association metrics with regards to the language unit of choice, and orchestrates the creation of up to five-dimensional interactive charts for over 30 variable type-semantics combinations. Through our case studies on computational dialectology, human label variation, and text generation, we show how Variationist enables researchers from different disciplines to effortlessly answer specific research questions or unveil undesired associations in language data. A Python library, code, documentation, and tutorials are made publicly available to the research community.

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.

Generation-based data augmentation (DA) has been presented in several works as a way to improve offensive language detection. However, the effectiveness of generative DA has been shown only in limited scenarios, and the potential injection of biases when using generated data to classify offensive language has not been investigated. Our aim is that of analyzing the feasibility of generative data augmentation more in-depth with two main focuses. First, we investigate the robustness of models trained on generated data in a variety of data augmentation setups, both novel and already presented in previous work, and compare their performance on four widely-used English offensive language datasets that present inherent differences in terms of content and complexity. In addition to this, we analyze models using the HateCheck suite, a series of functional tests created to challenge hate speech detection systems. Second, we investigate potential lexical bias issues through a qualitative analysis on the generated data. We find that the potential positive impact of generative data augmentation on model performance is unreliable, and generative DA can also have unpredictable effects on lexical bias.

We introduce DiatopIt, the first corpus specifically focused on diatopic language variation in Italy for language varieties other than Standard Italian. DiatopIt comprises over 15K geolocated social media posts from Twitter over a period of two years, including regional Italian usage and content fully written in local language varieties or exhibiting code-switching with Standard Italian. We detail how we tackled key challenges in creating such a resource, including the absence of orthography standards for most local language varieties and the lack of reliable language identification tools. We assess the representativeness of DiatopIt across time and space, and show that the density of non-Standard Italian content across areas correlates with actual language use. We finally conduct computational experiments and find that modeling diatopic variation on highly multilingual areas such as Italy is a complex task even for recent language models.

In this paper we present our submission to sub-task A at SemEval 2020 Task 12: Multilingual Offensive Language Identification in Social Media (OffensEval2). For Danish, Turkish, Arabic and Greek, we develop an architecture based on transfer learning and relying on a two-channel BERT model, in which the English BERT and the multilingual one are combined after creating a machine-translated parallel corpus for each language in the task. For English, instead, we adopt a more standard, single-channel approach. We find that, in a multilingual scenario, with some languages having small training data, using parallel BERT models with machine translated data can give systems more stability, especially when dealing with noisy data. The fact that machine translation on social media data may not be perfect does not hurt the overall classification performance.