We present UltraSparseBERT, a BERT variant that uses 0.3% of its neurons during inference while performing on par with similar BERT models. UltraSparseBERT selectively engages just 12 out of 4095 neurons for each layer inference. This is achieved by reorganizing feedforward networks into fast feedforward networks (FFFs).To showcase but one benefit of high sparsity, we provide an Intel MKL implementation achieving 78x speedup over the optimized feedforward baseline on CPUs, and an OpenAI Triton implementation performing forward passes 4.1x faster than the corresponding native GPU implementation. The training and benchmarking code is enclosed.

Recent work has demonstrated that pre-trained language models (PLMs) are zero-shot learners. However, most existing zero-shot methods involve heavy human engineering or complicated self-training pipelines, hindering their application to new situations. In this work, we show that zero-shot text classification can be improved simply by clustering texts in the embedding spaces of PLMs. Specifically, we fit the unlabeled texts with a Bayesian Gaussian Mixture Model after initializing cluster positions and shapes using class names. Despite its simplicity, this approach achieves superior or comparable performance on both topic and sentiment classification datasets and outperforms prior works significantly on unbalanced datasets. We further explore the applicability of our clustering approach by evaluating it on 14 datasets with more diverse topics, text lengths, and numbers of classes. Our approach achieves an average of 20% absolute improvement over prompt-based zero-shot learning. Finally, we compare different PLM embedding spaces and find that texts are well-clustered by topics even if the PLM is not explicitly pre-trained to generate meaningful sentence embeddings. This work indicates that PLM embeddings can categorize texts without task-specific fine-tuning, thus providing a new way to analyze and utilize their knowledge and zero-shot learning ability.

In this paper, we present an approach to improve the robustness of BERT language models against word substitution-based adversarial attacks by leveraging adversarial perturbations for self-supervised contrastive learning. We create a word-level adversarial attack generating hard positives on-the-fly as adversarial examples during contrastive learning. In contrast to previous works, our method improves model robustness without using any labeled data. Experimental results show that our method improves robustness of BERT against four different word substitution-based adversarial attacks, and combining our method with adversarial training gives higher robustness than adversarial training alone. As our method improves the robustness of BERT purely with unlabeled data, it opens up the possibility of using large text datasets to train robust language models against word substitution-based adversarial attacks.

Temporal knowledge graphs store the dynamics of entities and relations during a time period. However, typical temporal knowledge graphs often suffer from incomplete dynamics with missing facts in real-world scenarios. Hence, modeling temporal knowledge graphs to complete the missing facts is important. In this paper, we tackle the temporal knowledge graph completion task by proposing TempCaps, which is a Capsule network-based embedding model for Temporal knowledge graph completion. TempCaps models temporal knowledge graphs by introducing a novel dynamic routing aggregator inspired by Capsule Networks. Specifically, TempCaps builds entity embeddings by dynamically routing retrieved temporal relation and neighbor information. Experimental results demonstrate that TempCaps reaches state-of-the-art performance for temporal knowledge graph completion. Additional analysis also shows that TempCaps is efficient.

Different studies of the embedding space of transformer models suggest that the distribution of contextual representations is highly anisotropic - the embeddings are distributed in a narrow cone. Meanwhile, static word representations (e.g., Word2Vec or GloVe) have been shown to benefit from isotropic spaces. Therefore, previous work has developed methods to calibrate the embedding space of transformers in order to ensure isotropy. However, a recent study (Cai et al. 2021) shows that the embedding space of transformers is locally isotropic, which suggests that these models are already capable of exploiting the expressive capacity of their embedding space. In this work, we conduct an empirical evaluation of state-of-the-art methods for isotropy calibration on transformers and find that they do not provide consistent improvements across models and tasks. These results support the thesis that, given the local isotropy, transformers do not benefit from additional isotropy calibration.

Automatic ICD coding is the task of assigning codes from the International Classification of Diseases (ICD) to medical notes. These codes describe the state of the patient and have multiple applications, e.g., computer-assisted diagnosis or epidemiological studies. ICD coding is a challenging task due to the complexity and length of medical notes. Unlike the general trend in language processing, no transformer model has been reported to reach high performance on this task. Here, we investigate in detail ICD coding using PubMedBERT, a state-of-the-art transformer model for biomedical language understanding. We find that the difficulty of fine-tuning the model on long pieces of text is the main limitation for BERT-based models on ICD coding. We run extensive experiments and show that despite the gap with current state-of-the-art, pretrained transformers can reach competitive performance using relatively small portions of text. We point at better methods to aggregate information from long texts as the main need for improving BERT-based ICD coding.

We present Knowledge Enhanced Multimodal BART (KM-BART), which is a Transformer-based sequence-to-sequence model capable of reasoning about commonsense knowledge from multimodal inputs of images and texts. We adapt the generative BART architecture (Lewis et al., 2020) to a multimodal model with visual and textual inputs. We further develop novel pretraining tasks to improve the model performance on the Visual Commonsense Generation (VCG) task. In particular, our pretraining task of Knowledge-based Commonsense Generation (KCG) boosts model performance on the VCG task by leveraging commonsense knowledge from a large language model pretrained on external commonsense knowledge graphs. To the best of our knowledge, we are the first to propose a dedicated task for improving model performance on the VCG task. Experimental results show that our model reaches state-of-the-art performance on the VCG task (Park et al., 2020) by applying these novel pretraining tasks.

We take a deep look into the behaviour of self-attention heads in the transformer architecture. In light of recent work discouraging the use of attention distributions for explaining a model’s behaviour, we show that attention distributions can nevertheless provide insights into the local behaviour of attention heads. This way, we propose a distinction between local patterns revealed by attention and global patterns that refer back to the input, and analyze BERT from both angles. We use gradient attribution to analyze how the output of an attention head depends on the input tokens, effectively extending the local attention-based analysis to account for the mixing of information throughout the transformer layers. We find that there is a significant mismatch between attention and attribution distributions, caused by the mixing of context inside the model. We quantify this discrepancy and observe that interestingly, there are some patterns that persist across all layers despite the mixing.

Large pre-trained language models have repeatedly shown their ability to produce fluent text. Yet even when starting from a prompt, generation can continue in many plausible directions. Current decoding methods with the goal of controlling generation, e.g., to ensure specific words are included, either require additional models or fine-tuning, or work poorly when the task at hand is semantically unconstrained, e.g., story generation. In this work, we present a plug-and-play decoding method for controlled language generation that is so simple and intuitive, it can be described in a single sentence: given a topic or keyword, we add a shift to the probability distribution over our vocabulary towards semantically similar words. We show how annealing this distribution can be used to impose hard constraints on language generation, something no other plug-and-play method is currently able to do with SOTA language generators. Despite the simplicity of this approach, we see it works incredibly well in practice: decoding from GPT-2 leads to diverse and fluent sentences while guaranteeing the appearance of given guide words. We perform two user studies, revealing that (1) our method outperforms competing methods in human evaluations; and (2) forcing the guide words to appear in the generated text has no impact on the fluency of the generated text.

Generating adversarial examples for natural language is hard, as natural language consists of discrete symbols, and examples are often of variable lengths. In this paper, we propose a geometry-inspired attack for generating natural language adversarial examples. Our attack generates adversarial examples by iteratively approximating the decision boundary of Deep Neural Networks (DNNs). Experiments on two datasets with two different models show that our attack fools natural language models with high success rates, while only replacing a few words. Human evaluation shows that adversarial examples generated by our attack are hard for humans to recognize. Further experiments show that adversarial training can improve model robustness against our attack.