To address the global issue of online hate, hate speech detection (HSD) systems are typically developed on datasets from the United States, thereby failing to generalize to English dialects from the Majority World. Furthermore, HSD models are often evaluated on non-representative samples, raising concerns about overestimating model performance in real-world settings. In this work, we introduce NaijaHate, the first dataset annotated for HSD which contains a representative sample of Nigerian tweets. We demonstrate that HSD evaluated on biased datasets traditionally used in the literature consistently overestimates real-world performance by at least two-fold. We then propose NaijaXLM-T, a pretrained model tailored to the Nigerian Twitter context, and establish the key role played by domain-adaptive pretraining and finetuning in maximizing HSD performance. Finally, owing to the modest performance of HSD systems in real-world conditions, we find that content moderators would need to review about ten thousand Nigerian tweets flagged as hateful daily to moderate 60% of all hateful content, highlighting the challenges of moderating hate speech at scale as social media usage continues to grow globally. Taken together, these results pave the way towards robust HSD systems and a better protection of social media users from hateful content in low-resource settings.

Adapting general-purpose language models has proven to be effective in tackling downstream tasks within specific domains. In this paper, we address the task of extracting entities from the economics literature on impact evaluation. To this end, we release EconBERTa, a large language model pretrained on scientific publications in economics, and ECON-IE, a new expert-annotated dataset of economics abstracts for Named Entity Recognition (NER). We find that EconBERTa reaches state-of-the-art performance on our downstream NER task. Additionally, we extensively analyze the model’s generalization capacities, finding that most errors correspond to detecting only a subspan of an entity or failure to extrapolate to longer sequences. This limitation is primarily due to an inability to detect part-of-speech sequences unseen during training, and this effect diminishes when the number of unique instances in the training set increases. Examining the generalization abilities of domain-specific language models paves the way towards improving the robustness of NER models for causal knowledge extraction.

A central quest of probing is to uncover how pre-trained models encode a linguistic property within their representations. An encoding, however, might be spurious—i.e., the model might not rely on it when making predictions. In this paper, we try to find an encoding that the model actually uses, introducing a usage-based probing setup. We first choose a behavioral task which cannot be solved without using the linguistic property. Then, we attempt to remove the property by intervening on the model’s representations. We contend that, if an encoding is used by the model, its removal should harm the performance on the chosen behavioral task. As a case study, we focus on how BERT encodes grammatical number, and on how it uses this encoding to solve the number agreement task. Experimentally, we find that BERT relies on a linear encoding of grammatical number to produce the correct behavioral output. We also find that BERT uses a separate encoding of grammatical number for nouns and verbs. Finally, we identify in which layers information about grammatical number is transferred from a noun to its head verb.

Word order, an essential property of natural languages, is injected in Transformer-based neural language models using position encoding. However, recent experiments have shown that explicit position encoding is not always useful, since some models without such feature managed to achieve state-of-the art performance on some tasks. To understand better this phenomenon, we examine the effect of removing position encodings on the pre-training objective itself (i.e., masked language modelling), to test whether models can reconstruct position information from co-occurrences alone. We do so by controlling the amount of masked tokens in the input sentence, as a proxy to affect the importance of position information for the task. We find that the necessity of position information increases with the amount of masking, and that masked language models without position encodings are not able to reconstruct this information on the task. These findings point towards a direct relationship between the amount of masking and the ability of Transformers to capture order-sensitive aspects of language using position encoding.

Although transformer-based Neural Language Models demonstrate impressive performance on a variety of tasks, their generalization abilities are not well understood. They have been shown to perform strongly on subject-verb number agreement in a wide array of settings, suggesting that they learned to track syntactic dependencies during their training even without explicit supervision. In this paper, we examine the extent to which BERT is able to perform lexically-independent subject-verb number agreement (NA) on targeted syntactic templates. To do so, we disrupt the lexical patterns found in naturally occurring stimuli for each targeted structure in a novel fine-grained analysis of BERT’s behavior. Our results on nonce sentences suggest that the model generalizes well for simple templates, but fails to perform lexically-independent syntactic generalization when as little as one attractor is present.

Both humans and neural language models are able to perform subject verb number agreement (SVA). In principle, semantics shouldn’t interfere with this task, which only requires syntactic knowledge. In this work we test whether meaning interferes with this type of agreement in English in syntactic structures of various complexities. To do so, we generate both semantically well-formed and nonsensical items. We compare the performance of BERT-base to that of humans, obtained with a psycholinguistic online crowdsourcing experiment. We find that BERT and humans are both sensitive to our semantic manipulation: They fail more often when presented with nonsensical items, especially when their syntactic structure features an attractor (a noun phrase between the subject and the verb that has not the same number as the subject). We also find that the effect of meaningfulness on SVA errors is stronger for BERT than for humans, showing higher lexical sensitivity of the former on this task.