Hate speech detection is key to online content moderation, but current models struggle to generalise beyond their training data.This has been linked to dataset biases and the use of sentence-level labels, which fail to teach models the underlying structure of hate speech.In this work, we show that even when models are trained with more fine-grained, spanlevel annotations (e.g., "artists" is labeled as target and "are parasites" as dehumanising comparison), they struggle to disentangle the meaning of these labels from the surrounding context.As a result, combinations of expressions that deviate from those seen during training remain particularly difficult for models to detect.We investigate whether training on a dataset where expressions occur with equal frequency across all contexts can improve generalisation.To this end, we create Unseen-PLEAD (U-PLEAD), a dataset of ∼364,000 synthetic posts, along with a novel compositional generalisation benchmark of ∼8,000 posts.Training on a combination of U-PLEAD and real data improves compositional generalisation while achieving state-of-the-art performance on the human-sourced PLEAD.

Abstract To proactively offer social media users a safe online experience, there is a need for systems that can detect harmful posts and promptly alert platform moderators. In order to guarantee the enforcement of a consistent policy, moderators are provided with detailed guidelines. In contrast, most state-of-the-art models learn what abuse is from labeled examples and as a result base their predictions on spurious cues, such as the presence of group identifiers, which can be unreliable. In this work we introduce the concept of policy-aware abuse detection, abandoning the unrealistic expectation that systems can reliably learn which phenomena constitute abuse from inspecting the data alone. We propose a machine-friendly representation of the policy that moderators wish to enforce, by breaking it down into a collection of intents and slots. We collect and annotate a dataset of 3,535 English posts with such slots, and show how architectures for intent classification and slot filling can be used for abuse detection, while providing a rationale for model decisions.1

Thanks to the wealth of high-quality annotated images available in popular repositories such as ImageNet, multimodal language-vision research is in full bloom.However, events, feelings and many other kinds of concepts which can be visually grounded are not well represented in current datasets.Nevertheless, we would expect a wide-coverage language understanding system to be able to classify images depicting RECESS and REMORSE, not just CATS, DOGS and BRIDGES.We fill this gap by presenting BabelPic, a hand-labeled dataset built by cleaning the image-synset association found within the BabelNet Lexical Knowledge Base (LKB).BabelPic explicitly targets nonconcrete concepts, thus providing refreshing new data for the community.We also show that pre-trained language-vision systems can be used to further expand the resource by exploiting natural language knowledge available in the LKB.BabelPic is available for download at http://babelpic.org.

The problem of grounding language in vision is increasingly attracting scholarly efforts. As of now, however, most of the approaches have been limited to word embeddings, which are not capable of handling polysemous words. This is mainly due to the limited coverage of the available semantically-annotated datasets, hence forcing research to rely on alternative technologies (i.e., image search engines). To address this issue, we introduce EViLBERT, an approach which is able to perform image classification over an open set of concepts, both concrete and non-concrete. Our approach is based on the recently introduced Vision-Language Pretraining (VLP) model, and builds upon a manually-annotated dataset of concept-image pairs. We use our technique to clean up the image-to-concept mapping that is provided within a multilingual knowledge base, resulting in over 258,000 images associated with 42,500 concepts. We show that our VLP-based model can be used to create multimodal sense embeddings starting from our automatically-created dataset. In turn, we also show that these multimodal embeddings improve the performance of a Word Sense Disambiguation architecture over a strong unimodal baseline. We release code, dataset and embeddings at http://babelpic.org.