Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.
Pre-trained masked language models successfully perform few-shot learning by formulating downstream tasks as text infilling. However, as a strong alternative in full-shot settings, discriminative pre-trained models like ELECTRA do not fit into the paradigm. In this work, we adapt prompt-based few-shot learning to ELECTRA and show that it outperforms masked language models in a wide range of tasks. ELECTRA is pre-trained to distinguish if a token is generated or original. We naturally extend that to prompt-based few-shot learning by training to score the originality of the target options without introducing new parameters. Our method can be easily adapted to tasks involving multi-token predictions without extra computation overhead. Analysis shows that ELECTRA learns distributions that align better with downstream tasks.
Dense retrieval uses a contrastive learning framework to learn dense representations of queries and contexts. Trained encoders are directly used for each test query, but they often fail to accurately represent out-of-domain queries. In this paper, we introduce a framework that refines instance-level query representations at test time, with only the signals coming from the intermediate retrieval results. We optimize the query representation based on the retrieval result similar to pseudo relevance feedback (PRF) in information retrieval. Specifically, we adopt a cross-encoder labeler to provide pseudo labels over the retrieval result and iteratively refine the query representation with a gradient descent method, treating each test query as a single data point to train on. Our theoretical analysis reveals that our framework can be viewed as a generalization of the classical Rocchio's algorithm for PRF, which leads us to propose interesting variants of our method. We show that our test-time query refinement strategy improves the performance of phrase retrieval (+8.1% Acc@1) and passage retrieval (+3.7% Acc@20) for open-domain QA with large improvements on out-of-domain queries.
Recent work has improved language models remarkably by equipping them with a non-parametric memory component. However, most existing approaches only introduce memories at testing time, or represent them using a separately trained encoder -- resulting in sub-optimal training of the language model. In this work, we present TRIME, a novel yet simple training approach designed for training language models with memory augmentation. Our approach uses a training objective that directly takes in-batch examples as accessible memory. We also present new methods for memory construction and data batching, which are used for adapting to different sets of memories -- local, long-term, and external memory -- at testing time. We evaluate our approach on multiple language modeling and machine translation benchmarks. We find that simply replacing the vanilla language modeling objective by ours greatly reduces the perplexity, without modifying the model architecture or incorporating extra context (e.g., 18.70 $\to$ 17.76 on WikiText-103). We further augment language models with long-range contexts and external knowledge and demonstrate significant gains over previous memory-augmented approaches.
Deductive reasoning (drawing conclusions from assumptions) is a challenging problem in NLP. In this work, we focus on proof generation: given a hypothesis and a set of supporting facts in natural language, the model generates a proof tree indicating how to deduce the hypothesis from supporting facts. Instead of generating the entire proof in one shot, prior work has demonstrated the promise of stepwise generation but achieved limited success on real-world data. Existing stepwise methods struggle to generate proof steps that are both valid and relevant. In this paper, we present a novel stepwise method NLProofS (Natural Language Proof Search), which learns to generate relevant steps conditioning on the hypothesis. At the core of our approach, we train an independent verifier to check the validity of proof steps. Instead of generating steps greedily, we search for proofs maximizing a global proof score judged by the verifier. NLProofS achieves state-of-the-art performance on EntailmentBank and RuleTaker. For example, it improves the percentage of correctly predicted proofs from 20.9% to 33.3% in the distractor setting of EntailmentBank. This is the first time stepwise methods have led to better generation of challenging human-authored proofs.
Federated learning allows distributed users to collaboratively train a model while keeping each user's data private. Recently, a growing body of work has demonstrated that an eavesdropping attacker can effectively recover image data from gradients transmitted during federated learning. However, little progress has been made in recovering text data. In this paper, we present a novel attack method FILM for federated learning of language models -- for the first time, we show the feasibility of recovering text from large batch sizes of up to 128 sentences. Different from image-recovery methods which are optimized to match gradients, we take a distinct approach that first identifies a set of words from gradients and then directly reconstructs sentences based on beam search and a prior-based reordering strategy. The key insight of our attack is to leverage either prior knowledge in pre-trained language models or memorization during training. Despite its simplicity, we demonstrate that FILM can work well with several large-scale datasets -- it can extract single sentences with high fidelity even for large batch sizes and recover multiple sentences from the batch successfully if the attack is applied iteratively. We hope our results can motivate future work in developing stronger attacks as well as new defense methods for training language models in federated learning. Our code is publicly available at https://github.com/Princeton-SysML/FILM.
A growing line of work has investigated the development of neural NLP models that can produce rationales--subsets of input that can explain their model predictions. In this paper, we ask whether such rationale models can also provide robustness to adversarial attacks in addition to their interpretable nature. Since these models need to first generate rationales ("rationalizer") before making predictions ("predictor"), they have the potential to ignore noise or adversarially added text by simply masking it out of the generated rationale. To this end, we systematically generate various types of 'AddText' attacks for both token and sentence-level rationalization tasks, and perform an extensive empirical evaluation of state-of-the-art rationale models across five different tasks. Our experiments reveal that the rationale models show the promise to improve robustness, while they struggle in certain scenarios--when the rationalizer is sensitive to positional bias or lexical choices of attack text. Further, leveraging human rationale as supervision does not always translate to better performance. Our study is a first step towards exploring the interplay between interpretability and robustness in the rationalize-then-predict framework.
The growing size of neural language models has led to increased attention in model compression. The two predominant approaches are pruning, which gradually removes weights from a pre-trained model, and distillation, which trains a smaller compact model to match a larger one. Pruning methods can significantly reduce the model size but hardly achieve large speedups as distillation. However, distillation methods require large amounts of unlabeled data and are expensive to train. In this work, we propose a task-specific structured pruning method CoFi (Coarse- and Fine-grained Pruning), which delivers highly parallelizable subnetworks and matches the distillation methods in both accuracy and latency, without resorting to any unlabeled data. Our key insight is to jointly prune coarse-grained (e.g., layers) and fine-grained (e.g., heads and hidden units) modules, which controls the pruning decision of each parameter with masks of different granularity. We also devise a layerwise distillation strategy to transfer knowledge from unpruned to pruned models during optimization. Our experiments on GLUE and SQuAD datasets show that CoFi yields models with over 10x speedups with a small accuracy drop, showing its effectiveness and efficiency compared to previous pruning and distillation approaches.