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.
The speaker-follower models have proven to be effective in vision-and-language navigation, where a speaker model is used to synthesize new instructions to augment the training data for a follower navigation model. However, in many of the previous methods, the generated instructions are not directly trained to optimize the performance of the follower. In this paper, we present \textsc{foam}, a \textsc{Fo}llower-\textsc{a}ware speaker \textsc{M}odel that is constantly updated given the follower feedback, so that the generated instructions can be more suitable to the current learning state of the follower. Specifically, we optimize the speaker using a bi-level optimization framework and obtain its training signals by evaluating the follower on labeled data. Experimental results on the Room-to-Room and Room-across-Room datasets demonstrate that our methods can outperform strong baseline models across settings. Analyses also reveal that our generated instructions are of higher quality than the baselines.
We propose a general and efficient framework to control auto-regressive generation models with NeurAlly-Decomposed Oracle (NADO). Given a pre-trained base language model and a sequence-level boolean oracle function, we propose to decompose the oracle function into token-level guidance to steer the base model in text generation. Specifically, the token-level guidance is approximated by a neural model trained with examples sampled from the base model, demanding no additional auxiliary labeled data. We present the closed-form optimal solution to incorporate the token-level guidance into the base model for controllable generation. We further provide a theoretical analysis of how the approximation quality of NADO affects the controllable generation results. Experiments conducted on two applications: (1) text generation with lexical constraints and (2) machine translation with formality control demonstrate that our framework efficiently guides the base model towards the given oracle while maintaining high generation quality.
Relational structure extraction covers a wide range of tasks and plays an important role in natural language processing. Recently, many approaches tend to design sophisticated graphical models to capture the complex relations between objects that are described in a sentence. In this work, we demonstrate that simple tagging models can surprisingly achieve competitive performances with a small trick -- priming. Tagging models with priming append information about the operated objects to the input sequence of pretrained language model. Making use of the contextualized nature of pretrained language model, the priming approach help the contextualized representation of the sentence better embed the information about the operated objects, hence, becomes more suitable for addressing relational structure extraction. We conduct extensive experiments on three different tasks that span ten datasets across five different languages, and show that our model is a general and effective model, despite its simplicity. We further carry out comprehensive analysis to understand our model and propose an efficient approximation to our method, which can perform almost the same performance but with faster inference speed.
Task-oriented dialogue systems aim to answer questions from users and provide immediate help. Therefore, how humans perceive their helpfulness is important. However, neither the human-perceived helpfulness of task-oriented dialogue systems nor its fairness implication has been studied yet. In this paper, we define a dialogue response as helpful if it is relevant & coherent, useful, and informative to a query and study computational measurements of helpfulness. Then, we propose utilizing the helpfulness level of different groups to gauge the fairness of a dialogue system. To study this, we collect human annotations for the helpfulness of dialogue responses and build a classifier that can automatically determine the helpfulness of a response. We design experiments under 3 information-seeking scenarios and collect instances for each from Wikipedia. With collected instances, we use carefully-constructed questions to query the state-of-the-art dialogue systems. Through analysis, we find that dialogue systems tend to be more helpful for highly-developed countries than less-developed countries, uncovering a fairness issue underlying these dialogue systems.
Numerous events occur worldwide and are documented in the news, social media, and various online platforms in raw text. Extracting useful and succinct information about these events is crucial to various downstream applications. Event Argument Extraction (EAE) deals with the task of extracting event-specific information from natural language text. In order to cater to new events and domains in a realistic low-data setting, there is a growing urgency for EAE models to be generalizable. Consequentially, there is a necessity for benchmarking setups to evaluate the generalizability of EAE models. But most existing benchmarking datasets like ACE and ERE have limited coverage in terms of events and cannot adequately evaluate the generalizability of EAE models. To alleviate this issue, we introduce a new dataset GENEVA covering a diverse range of 115 events and 187 argument roles. Using this dataset, we create four benchmarking test suites to assess the model's generalization capability from different perspectives. We benchmark various representative models on these test suites and compare their generalizability relatively. Finally, we propose a new model SCAD that outperforms the previous models and serves as a strong benchmark for these test suites.
The ability to infer pre- and postconditions of an action is vital for comprehending complex instructions, and is essential for applications such as autonomous instruction-guided agents and assistive AI that supports humans to perform physical tasks. In this work, we propose a task dubbed action condition inference, and collecting a high-quality, human annotated dataset of preconditions and postconditions of actions in instructional manuals. We propose a weakly supervised approach to automatically construct large-scale training instances from online instructional manuals, and curate a densely human-annotated and validated dataset to study how well the current NLP models can infer action-condition dependencies in the instruction texts. We design two types of models differ by whether contextualized and global information is leveraged, as well as various combinations of heuristics to construct the weak supervisions. Our experimental results show a >20% F1-score improvement with considering the entire instruction contexts and a >6% F1-score benefit with the proposed heuristics.
The vast majority of text transformation techniques in NLP are inherently limited in their ability to expand input space coverage due to an implicit constraint to preserve the original class label. In this work, we propose the notion of sibylvariance (SIB) to describe the broader set of transforms that relax the label-preserving constraint, knowably vary the expected class, and lead to significantly more diverse input distributions. We offer a unified framework to organize all data transformations, including two types of SIB: (1) Transmutations convert one discrete kind into another, (2) Mixture Mutations blend two or more classes together. To explore the role of sibylvariance within NLP, we implemented 41 text transformations, including several novel techniques like Concept2Sentence and SentMix. Sibylvariance also enables a unique form of adaptive training that generates new input mixtures for the most confused class pairs, challenging the learner to differentiate with greater nuance. Our experiments on six benchmark datasets strongly support the efficacy of sibylvariance for generalization performance, defect detection, and adversarial robustness.
Stories or narratives are comprised of a sequence of events. To compose interesting stories, professional writers often leverage a creative writing technique called flashback that inserts past events into current storylines as we commonly observe in novels and plays. However, it is challenging for machines to generate flashback as it requires a solid understanding of event temporal order (e.g. "feeling hungry" before "eat," not vice versa), and the creativity to arrange storylines so that earlier events do not always appear first in narrative order. Two major issues in existing systems that exacerbate the challenges: 1) temporal bias in pertaining and story datasets that leads to monotonic event temporal orders; 2) lack of explicit guidance that helps machines decide where to insert flashbacks. We propose to address these issues using structured storylines to encode events and their pair-wise temporal relations (before, after and vague) as temporal prompts that guide how stories should unfold temporally. We leverage a Plan-and-Write framework enhanced by reinforcement learning to generate storylines and stories end-to-end. Evaluation results show that the proposed method can generate more interesting stories with flashbacks while maintaining textual diversity, fluency, and temporal coherence.
In this paper, we propose a simple yet effective way to generate pun sentences that does not require any training on existing puns. Our approach is inspired by humor theories that ambiguity comes from the context rather than the pun word itself. Given a pair of definitions of a pun word, our model first produces a list of related concepts through a reverse dictionary. We then utilize one-shot GPT3 to generate context words and then generate puns incorporating context words from both concepts. Human evaluation shows that our method successfully generates pun 52\% of the time, outperforming well-crafted baselines and the state-of-the-art models by a large margin.