Recent advances in large pre-trained language models have greatly improved the performance on a broad set of NLP tasks. However, adapting an existing model to new tasks often requires (repeated) re-training over enormous labeled data that is prohibitively expensive to obtain. Moreover, models learned on new tasks may gradually "forget" about the knowledge learned from earlier tasks (i.e., catastrophic forgetting). In this paper, we study the challenge of lifelong learning to few-shot learn over a sequence of diverse NLP tasks, through continuously fine-tuning a language model. We investigate the model's ability of few-shot generalization to new tasks while retaining its performance on the previously learned tasks. We explore existing continual learning methods in solving this problem and propose a continual meta-learning approach which learns to generate adapter weights from a few examples while regularizing changes of the weights to mitigate catastrophic forgetting. We demonstrate our approach preserves model performance over training tasks and leads to positive knowledge transfer when the future tasks are learned.
Augmenting pre-trained language models with knowledge graphs (KGs) has achieved success on various commonsense reasoning tasks. Although some works have attempted to explain the behavior of such KG-augmented models by indicating which KG inputs are salient (i.e., important for the model's prediction), it is not always clear how these explanations should be used to make the model better. In this paper, we explore whether KG explanations can be used as supervision for teaching these KG-augmented models how to filter out unhelpful KG information. To this end, we propose SalKG, a simple framework for learning from KG explanations of both coarse (Is the KG salient?) and fine (Which parts of the KG are salient?) granularity. Given the explanations generated from a task's training set, SalKG trains KG-augmented models to solve the task by focusing on KG information highlighted by the explanations as salient. Across two popular commonsense QA benchmarks and three KG-augmented models, we find that SalKG's training process can consistently improve model performance.
We study the power of cross-attention in the Transformer architecture within the context of machine translation. In transfer learning experiments, where we fine-tune a translation model on a dataset with one new language, we find that, apart from the new language's embeddings, only the cross-attention parameters need to be fine-tuned to obtain competitive BLEU performance. We provide insights into why this is the case and further find that limiting fine-tuning in this manner yields cross-lingually aligned type embeddings. The implications of this finding include a mitigation of catastrophic forgetting in the network and the potential for zero-shot translation.
Recently, pre-trained language models (PLMs) have dominated conditional text generation tasks. Given the impressive performance and prevalence of the PLMs, it is seemingly natural to assume that they could figure out what to attend to in the input and what to include in the output via seq2seq learning without more guidance than the training input/output pairs. However, a rigorous study regarding the above assumption is still lacking. In this paper, we present a systematic analysis of conditional generation to study whether current PLMs are good enough for preserving important concepts in the input and to what extent explicitly guiding generation with lexical constraints is beneficial. We conduct extensive analytical experiments on a range of conditional generation tasks and try to answer in what scenarios guiding generation with lexical constraints works well and why. We then propose a framework for automatic constraint extraction, denoising, and enforcement that is shown to perform comparably or better than unconstrained generation. We hope that our findings could serve as a reference when determining whether it is appropriate and worthwhile to use explicit constraints for a specific task or dataset.\footnote{Our code is available at \url{https://github.com/morningmoni/LCGen-eval}.}
Warning: this paper contains content that may be offensive or upsetting. Numerous natural language processing models have tried injecting commonsense by using the ConceptNet knowledge base to improve performance on different tasks. ConceptNet, however, is mostly crowdsourced from humans and may reflect human biases such as "lawyers are dishonest." It is important that these biases are not conflated with the notion of commonsense. We study this missing yet important problem by first defining and quantifying biases in ConceptNet as two types of representational harms: overgeneralization of polarized perceptions and representation disparity. We find that ConceptNet contains severe biases and disparities across four demographic categories. In addition, we analyze two downstream models that use ConceptNet as a source for commonsense knowledge and find the existence of biases in those models as well. We further propose a filtered-based bias-mitigation approach and examine its effectiveness. We show that our mitigation approach can reduce the issues in both resource and models but leads to a performance drop, leaving room for future work to build fairer and stronger commonsense models.
Neural networks are prone to learning spurious correlations from biased datasets, and are thus vulnerable when making inferences in a new target domain. Prior work reveals spurious patterns via post-hoc model explanations which compute the importance of input features, and further eliminates the unintended model behaviors by regularizing importance scores with human knowledge. However, such regularization technique lacks flexibility and coverage, since only importance scores towards a pre-defined list of features are adjusted, while more complex human knowledge such as feature interaction and pattern generalization can hardly be incorporated. In this work, we propose to refine a learned model by collecting human-provided compositional explanations on the models' failure cases. By describing generalizable rules about spurious patterns in the explanation, more training examples can be matched and regularized, tackling the challenge of regularization coverage. We additionally introduce a regularization term for feature interaction to support more complex human rationale in refining the model. We demonstrate the effectiveness of the proposed approach on two text classification tasks by showing improved performance in target domain after refinement.
Large neural networks are impractical to deploy on mobile devices due to their heavy computational cost and slow inference. Knowledge distillation (KD) is a technique to reduce the model size while retaining performance by transferring knowledge from a large "teacher" model to a smaller "student" model. However, KD on multimodal datasets such as vision-language datasets is relatively unexplored and digesting such multimodal information is challenging since different modalities present different types of information. In this paper, we propose modality-specific distillation (MSD) to effectively transfer knowledge from a teacher on multimodal datasets. Existing KD approaches can be applied to multimodal setup, but a student doesn't have access to modality-specific predictions. Our idea aims at mimicking a teacher's modality-specific predictions by introducing an auxiliary loss term for each modality. Because each modality has different importance for predictions, we also propose weighting approaches for the auxiliary losses; a meta-learning approach to learn the optimal weights on these loss terms. In our experiments, we demonstrate the effectiveness of our MSD and the weighting scheme and show that it achieves better performance than KD.
Pre-trained text-to-text transformers achieve impressive performance across a wide range of NLP tasks, and they naturally support zero-shot learning (ZSL) by using the task description as prompt in the input. However, this approach has potential limitations, as it learns from input-output pairs at instance level, instead of learning to solve tasks at task level. Alternatively, applying existing ZSL methods to text-to-text transformers is non-trivial due to their text generation objective and huge size. To address these issues, we introduce Hypter, a framework that improves zero-shot transferability by training a hypernetwork to generate task-specific adapters from task descriptions. This formulation enables learning at task level, and greatly reduces the number of parameters by using light-weight adapters. Experiments on two datasets demonstrate Hypter improves upon fine-tuning baselines.
A riddle is a mystifying, puzzling question about everyday concepts. For example, the riddle "I have five fingers but I am not alive. What am I?" asks about the concept of a glove. Solving riddles is a challenging cognitive process for humans, in that it requires complex commonsense reasoning abilities and an understanding of figurative language. However, there are currently no commonsense reasoning datasets that test these abilities. We propose RiddleSense, a novel multiple-choice question answering challenge for benchmarking higher-order commonsense reasoning models, which is the first large dataset for riddle-style commonsense question answering, where the distractors are crowdsourced from human annotators. We systematically evaluate a wide range of reasoning models over it and point out that there is a large gap between the best-supervised model and human performance -- pointing to interesting future research for higher-order commonsense reasoning and computational creativity.
Closed-book question-answering (QA) is a challenging task that requires a model to directly answer questions without access to external knowledge. It has been shown that directly fine-tuning pre-trained language models with (question, answer) examples yields surprisingly competitive performance, which is further improved upon through adding an intermediate pre-training stage between general pre-training and fine-tuning. Prior work used a heuristic during this intermediate stage, whereby named entities and dates are masked, and the model is trained to recover these tokens. In this paper, we aim to learn the optimal masking strategy for the intermediate pre-training stage. We first train our masking policy to extract spans that are likely to be tested, using supervision from the downstream task itself, then deploy the learned policy during intermediate pre-training. Thus, our policy packs task-relevant knowledge into the parameters of a language model. Our approach is particularly effective on TriviaQA, outperforming strong heuristics when used to pre-train BART.