There exists a wide variety of efficiency methods for natural language processing (NLP) tasks, such as pruning, distillation, dynamic inference, quantization, etc. From a different perspective, we can consider an efficiency method as an operator applied on a model. Naturally, we may construct a pipeline of operators, i.e., to apply multiple efficiency methods on the model sequentially. In this paper, we study the plausibility of this idea, and more importantly, the commutativity and cumulativeness of efficiency operators. We make two interesting observations from our experiments: (1) The operators are commutative—the order of efficiency methods within the pipeline has little impact on the final results; (2) The operators are also cumulative—the final results of combining several efficiency methods can be estimated by combining the results of individual methods. These observations deepen our understanding of efficiency operators and provide useful guidelines for building them in real-world applications.

Layout detection is an essential step for accurately extracting structured contents from historical documents. The intricate and varied layouts present in these document images make it expensive to label the numerous layout regions that can be densely arranged on each page. Current active learning methods typically rank and label samples at the image level, where the annotation budget is not optimally spent due to the overexposure of common objects per image. Inspired by recent progress in semi-supervised learning and self-training, we propose OLALA, an Object-Level Active Learning framework for efficient document layout Annotation. OLALA aims to optimize the annotation process by selectively annotating only the most ambiguous regions within an image, while using automatically generated labels for the rest. Central to OLALA is a perturbation-based scoring function that determines which objects require manual annotation. Extensive experiments show that OLALA can significantly boost model performance and improve annotation efficiency, facilitating the extraction of masses of structured text for downstream NLP applications.

In selective prediction, a classifier is allowed to abstain from making predictions on low-confidence examples. Though this setting is interesting and important, selective prediction has rarely been examined in natural language processing (NLP) tasks. To fill this void in the literature, we study in this paper selective prediction for NLP, comparing different models and confidence estimators. We further propose a simple error regularization trick that improves confidence estimation without substantially increasing the computation budget. We show that recent pre-trained transformer models simultaneously improve both model accuracy and confidence estimation effectiveness. We also find that our proposed regularization improves confidence estimation and can be applied to other relevant scenarios, such as using classifier cascades for accuracy–efficiency trade-offs. Source code for this paper can be found at https://github.com/castorini/transformers-selective.

We address the problem of enhancing model robustness through regularization. Specifically, we focus on methods that regularize the model posterior difference between clean and noisy inputs. Theoretically, we provide a connection of two recent methods, Jacobian Regularization and Virtual Adversarial Training, under this framework. Additionally, we generalize the posterior differential regularization to the family of f-divergences and characterize the overall framework in terms of the Jacobian matrix. Empirically, we compare those regularizations and standard BERT training on a diverse set of tasks to provide a comprehensive profile of their effect on model generalization. For both fully supervised and semi-supervised settings, we show that regularizing the posterior difference with f-divergence can result in well-improved model robustness. In particular, with a proper f-divergence, a BERT-base model can achieve comparable generalization as its BERT-large counterpart for in-domain, adversarial and domain shift scenarios, indicating the great potential of the proposed framework for enhancing NLP model robustness.

The slow speed of BERT has motivated much research on accelerating its inference, and the early exiting idea has been proposed to make trade-offs between model quality and efficiency. This paper aims to address two weaknesses of previous work: (1) existing fine-tuning strategies for early exiting models fail to take full advantage of BERT; (2) methods to make exiting decisions are limited to classification tasks. We propose a more advanced fine-tuning strategy and a learning-to-exit module that extends early exiting to tasks other than classification. Experiments demonstrate improved early exiting for BERT, with better trade-offs obtained by the proposed fine-tuning strategy, successful application to regression tasks, and the possibility to combine it with other acceleration methods. Source code can be found at https://github.com/castorini/berxit.

Pre-trained language models such as BERT have shown their effectiveness in various tasks. Despite their power, they are known to be computationally intensive, which hinders real-world applications. In this paper, we introduce early exiting BERT for document ranking. With a slight modification, BERT becomes a model with multiple output paths, and each inference sample can exit early from these paths. In this way, computation can be effectively allocated among samples, and overall system latency is significantly reduced while the original quality is maintained. Our experiments on two document ranking datasets demonstrate up to 2.5x inference speedup with minimal quality degradation. The source code of our implementation can be found at https://github.com/castorini/earlyexiting-monobert.

Large-scale pre-trained language models such as BERT have brought significant improvements to NLP applications. However, they are also notorious for being slow in inference, which makes them difficult to deploy in real-time applications. We propose a simple but effective method, DeeBERT, to accelerate BERT inference. Our approach allows samples to exit earlier without passing through the entire model. Experiments show that DeeBERT is able to save up to ~40% inference time with minimal degradation in model quality. Further analyses show different behaviors in the BERT transformer layers and also reveal their redundancy. Our work provides new ideas to efficiently apply deep transformer-based models to downstream tasks. Code is available at https://github.com/castorini/DeeBERT.

In natural language processing, a recently popular line of work explores how to best report the experimental results of neural networks. One exemplar publication, titled “Show Your Work: Improved Reporting of Experimental Results” (Dodge et al., 2019), advocates for reporting the expected validation effectiveness of the best-tuned model, with respect to the computational budget. In the present work, we critically examine this paper. As far as statistical generalizability is concerned, we find unspoken pitfalls and caveats with this approach. We analytically show that their estimator is biased and uses error-prone assumptions. We find that the estimator favors negative errors and yields poor bootstrapped confidence intervals. We derive an unbiased alternative and bolster our claims with empirical evidence from statistical simulation. Our codebase is at https://github.com/castorini/meanmax.

Memory neurons of long short-term memory (LSTM) networks encode and process information in powerful yet mysterious ways. While there has been work to analyze their behavior in carrying low-level information such as linguistic properties, how they directly contribute to label prediction remains unclear. We find inspiration from biologists and study the affinity between individual neurons and labels, propose a novel metric to quantify the sensitivity of neurons to each label, and conduct experiments to show the validity of our proposed metric. We discover that some neurons are trained to specialize on a subset of labels, and while dropping an arbitrary neuron has little effect on the overall accuracy of the model, dropping label-specialized neurons predictably and significantly degrades prediction accuracy on the associated label. We further examine the consistency of neuron-label affinity across different models. These observations provide insight into the inner mechanisms of LSTMs.