Swabha Swayamdipta

Alisa Liu, Zhaofeng Wu, Julian Michael et al. · allen-ai, berkeley

Ambiguity is an intrinsic feature of natural language. Managing ambiguity is a key part of human language understanding, allowing us to anticipate misunderstanding as communicators and revise our interpretations as listeners. As language models (LMs) are increasingly employed as dialogue interfaces and writing aids, handling ambiguous language is critical to their success. We characterize ambiguity in a sentence by its effect on entailment relations with another sentence, and collect AmbiEnt, a linguist-annotated benchmark of 1,645 examples with diverse kinds of ambiguity. We design a suite of tests based on AmbiEnt, presenting the first evaluation of pretrained LMs to recognize ambiguity and disentangle possible meanings. We find that the task remains extremely challenging, including for GPT-4, whose generated disambiguations are considered correct only 32% of the time in human evaluation, compared to 90% for disambiguations in our dataset. Finally, to illustrate the value of ambiguity-sensitive tools, we show that a multilabel NLI model can flag political claims in the wild that are misleading due to ambiguity. We encourage the field to rediscover the importance of ambiguity for NLP.

Xuhui Zhou, Hao Zhu, Akhila Yerukola et al. · allen-ai, cmu

Warning: This paper contains content that may be offensive or upsetting. Understanding the harms and offensiveness of statements requires reasoning about the social and situational context in which statements are made. For example, the utterance "your English is very good" may implicitly signal an insult when uttered by a white man to a non-white colleague, but uttered by an ESL teacher to their student would be interpreted as a genuine compliment. Such contextual factors have been largely ignored by previous approaches to toxic language detection. We introduce COBRA frames, the first context-aware formalism for explaining the intents, reactions, and harms of offensive or biased statements grounded in their social and situational context. We create COBRACORPUS, a dataset of 33k potentially offensive statements paired with machine-generated contexts and free-text explanations of offensiveness, implied biases, speaker intents, and listener reactions. To study the contextual dynamics of offensiveness, we train models to generate COBRA explanations, with and without access to the context. We find that explanations by context-agnostic models are significantly worse than by context-aware ones, especially in situations where the context inverts the statement's offensiveness (29% accuracy drop). Our work highlights the importance and feasibility of contextualized NLP by modeling social factors.

Chandra Bhagavatula, Jena D. Hwang, Doug Downey et al. · allen-ai, uw

Commonsense capabilities of pre-trained language models dramatically improve with scale, leading many to believe that scale is the only winning recipe. But is it? Here, we investigate an alternative that a priori seems impossible: can smaller language models (e.g., GPT-2) win over models that are orders of magnitude larger and better (e.g., GPT-3), if powered with novel commonsense distillation algorithms? The key intellectual challenge is to design a learning algorithm that achieve a competitive level of commonsense acquisition, without relying on the benefits of scale. In particular, we study generative models of commonsense knowledge, focusing on the task of generating generics, statements of commonsense facts about everyday concepts, e.g., birds can fly. We introduce I2D2, a novel commonsense distillation framework that loosely follows the Symbolic Knowledge Distillation of West et al. but breaks the dependence on the extreme-scale teacher model with two innovations: (1) the novel adaptation of NeuroLogic Decoding to enhance the generation quality of the weak, off-the-shelf language models, and (2) self-imitation learning to iteratively learn from the model's own enhanced commonsense acquisition capabilities. Empirical results suggest that scale is not the only way, as novel algorithms can be a promising alternative. Moreover, our study leads to a new corpus of generics, Gen-A-tomic, that is the largest and highest quality available to date.

Krishna Pillutla, Lang Liu, John Thickstun et al. · uw

Generative artificial intelligence has made significant strides, producing text indistinguishable from human prose and remarkably photorealistic images. Automatically measuring how close the generated data distribution is to the target distribution is central to diagnosing existing models and developing better ones. We present MAUVE, a family of comparison measures between pairs of distributions such as those encountered in the generative modeling of text or images. These scores are statistical summaries of divergence frontiers capturing two types of errors in generative modeling. We explore three approaches to statistically estimate these scores: vector quantization, non-parametric estimation, and classifier-based estimation. We provide statistical bounds for the vector quantization approach. Empirically, we find that the proposed scores paired with a range of $f$-divergences and statistical estimation methods can quantify the gaps between the distributions of human-written text and those of modern neural language models by correlating with human judgments and identifying known properties of the generated texts. We demonstrate in the vision domain that MAUVE can identify known properties of generated images on par with or better than existing metrics. In conclusion, we present practical recommendations for using MAUVE effectively with language and image modalities.

Jiao Sun, Swabha Swayamdipta, Jonathan May et al. · cmu

Free-form rationales aim to aid model interpretability by supplying the background knowledge that can help understand model decisions. Crowdsourced rationales are provided for commonsense QA instances in popular datasets such as CoS-E and ECQA, but their utility remains under-investigated. We present human studies which show that ECQA rationales indeed provide additional background information to understand a decision, while over 88% of CoS-E rationales do not. Inspired by this finding, we ask: can the additional context provided by free-form rationales benefit models, similar to human users? We investigate the utility of rationales as an additional source of supervision, by varying the quantity and quality of rationales during training. After controlling for instances where rationales leak the correct answer while not providing additional background knowledge, we find that incorporating only 5% of rationales during training can boost model performance by 47.22% for CoS-E and 57.14% for ECQA during inference. Moreover, we also show that rationale quality matters: compared to crowdsourced rationales, T5-generated rationales provide not only weaker supervision to models, but are also not helpful for humans in aiding model interpretability.

Matthew Finlayson, Andreas Grivas, Xiang Ren et al.

Language model parameters are known to impose unique (to each model) geometric constraints on their logit outputs, which serves as a signature that identifies the model, but also leaks the model's final layer parameters when an API distributes logits. We investigate more restrictive APIs that expose token rankings (i.e., their ordering by probability, but not the probability values) and find that rankings also constitute a signature: every model has a unique set of feasible top-$k$ rankings for sufficiently large $k$. Furthermore, the ranking signature is the first known (polynomially) unforgeable signature, since finding a model with the same set of feasible rankings is NP-hard. On the security front, we find that token rankings are already sufficient to approximately steal the final layer of the model, similar to logits, though the approximation is too coarse to forge the signature, and can be effectively countered by restricting the API to top-$k$ tokens with sufficiently small $k$. Since the top-$k$ required to present the model signature is generally smaller than the $k$ required to prevent stealing, it is possible for an API to present an unforgeable signature without leaking model parameters.

Hanjie Chen, Faeze Brahman, Xiang Ren et al.

Generating free-text rationales is a promising step towards explainable NLP, yet evaluating such rationales remains a challenge. Existing metrics have mostly focused on measuring the association between the rationale and a given label. We argue that an ideal metric should focus on the new information uniquely provided in the rationale that is otherwise not provided in the input or the label. We investigate this research problem from an information-theoretic perspective using conditional V-information (Hewitt et al., 2021). More concretely, we propose a metric called REV (Rationale Evaluation with conditional V-information), to quantify the amount of new, label-relevant information in a rationale beyond the information already available in the input or the label. Experiments across four benchmarks with reasoning tasks, including chain-of-thought, demonstrate the effectiveness of REV in evaluating rationale-label pairs, compared to existing metrics. We further demonstrate REV is consistent with human judgments on rationale evaluations and provides more sensitive measurements of new information in free-text rationales. When used alongside traditional performance metrics, REV provides deeper insights into models' reasoning and prediction processes.

Phillip Howard, Gadi Singer, Vasudev Lal et al.

While counterfactual data augmentation offers a promising step towards robust generalization in natural language processing, producing a set of counterfactuals that offer valuable inductive bias for models remains a challenge. Most existing approaches for producing counterfactuals, manual or automated, rely on small perturbations via minimal edits, resulting in simplistic changes. We introduce NeuroCounterfactuals, designed as loose counterfactuals, allowing for larger edits which result in naturalistic generations containing linguistic diversity, while still bearing similarity to the original document. Our novel generative approach bridges the benefits of constrained decoding, with those of language model adaptation for sentiment steering. Training data augmentation with our generations results in both in-domain and out-of-domain improvements for sentiment classification, outperforming even manually curated counterfactuals, under select settings. We further present detailed analyses to show the advantages of NeuroCounterfactuals over approaches involving simple, minimal edits.

Saleh Afroogh, Seyd Ishtiaque Ahmed, Petra Ahrweiler et al. · cmu

This study provides a cross-disciplinary examination of Explainable Artificial Intelligence (XAI) approaches-focusing on deep neural networks (DNNs) and large language models (LLMs)-and identifies empirical and conceptual limitations in current XAI. We discuss critical symptoms that stem from deeper root causes (i.e., two paradoxes, two conceptual confusions, and five false assumptions). These fundamental problems within the current XAI research field reveal three insights: experimentally, XAI exhibits significant flaws; conceptually, it is paradoxical; and pragmatically, further attempts to reform the paradoxical XAI might exacerbate its confusion-demanding fundamental shifts and new research directions. To move beyond XAI's limitations, we propose a four-pronged synthesized paradigm shift toward reliable and certified AI development. These four components include: verification-focused Interactive AI (IAI) to establish scientific community protocols for certifying AI system performance rather than attempting post-hoc explanations, AI Epistemology for rigorous scientific foundations, User-Sensible AI to create context-aware systems tailored to specific user communities, and Model-Centered Interpretability for faithful technical analysis-together offering comprehensive post-XAI research directions.

Matthew Finlayson, John Hewitt, Alexander Koller et al.

Despite their ubiquity in language generation, it remains unknown why truncation sampling heuristics like nucleus sampling are so effective. We provide a theoretical explanation for the effectiveness of the truncation sampling by proving that truncation methods that discard tokens below some probability threshold (the most common type of truncation) can guarantee that all sampled tokens have nonzero true probability. However, thresholds are a coarse heuristic, and necessarily discard some tokens with nonzero true probability as well. In pursuit of a more precise sampling strategy, we show that we can leverage a known source of model errors, the softmax bottleneck, to prove that certain tokens have nonzero true probability, without relying on a threshold. Based on our findings, we develop an experimental truncation strategy and the present pilot studies demonstrating the promise of this type of algorithm. Our evaluations show that our method outperforms its threshold-based counterparts under automatic and human evaluation metrics for low-entropy (i.e., close to greedy) open-ended text generation. Our theoretical findings and pilot experiments provide both insight into why truncation sampling works, and make progress toward more expressive sampling algorithms that better surface the generative capabilities of large language models.

Yoonsoo Nam, Adam Lehavi, Daniel Yang et al. · allen-ai

Video summarization remains a huge challenge in computer vision due to the size of the input videos to be summarized. We propose an efficient, language-only video summarizer that achieves competitive accuracy with high data efficiency. Using only textual captions obtained via a zero-shot approach, we train a language transformer model and forego image representations. This method allows us to perform filtration amongst the representative text vectors and condense the sequence. With our approach, we gain explainability with natural language that comes easily for human interpretation and textual summaries of the videos. An ablation study that focuses on modality and data compression shows that leveraging text modality only effectively reduces input data processing while retaining comparable results.

Atharva Kulkarni, Jacob Mitchell Springer, Arjun Subramonian et al. · cmu, meta-ai

Geometric properties of Transformer weights, particularly the unembedding matrix, have been widely useful in language model interpretability research. Yet, their utility for estimating downstream performance remains unclear. In this work, we systematically investigate the relationship between model performance and the unembedding matrix geometry, particularly its effective rank. Our experiments, involving a suite of 108 OLMo-style language models trained under controlled variation, reveal several key findings. While the best-performing models often exhibit a high effective rank, this trend is not universal across tasks and training setups. Contrary to prior work, we find that low effective rank does not cause late-stage performance degradation in small models, but instead co-occurs with it; we find adversarial cases where low-rank models do not exhibit saturation. Moreover, we show that effective rank is strongly influenced by pre-training hyperparameters, such as batch size and weight decay, which in-turn affect the model's performance. Lastly, extending our analysis to other geometric metrics and final-layer representation, we find that these metrics are largely aligned, but none can reliably predict downstream performance. Overall, our findings suggest that the model's geometry, as captured by existing metrics, primarily reflects training choices rather than performance.

Aryan Gulati, Xingjian Dong, Carlos Hurtado et al. · allen-ai

As language models become more general purpose, increased attention needs to be paid to detecting out-of-distribution (OOD) instances, i.e., those not belonging to any of the distributions seen during training. Existing methods for detecting OOD data are computationally complex and storage-intensive. We propose a novel soft clustering approach for OOD detection based on non-negative kernel regression. Our approach greatly reduces computational and space complexities (up to 11x improvement in inference time and 87% reduction in storage requirements) and outperforms existing approaches by up to 4 AUROC points on four different benchmarks. We also introduce an entropy-constrained version of our algorithm, which leads to further reductions in storage requirements (up to 97% lower than comparable approaches) while retaining competitive performance. Our soft clustering approach for OOD detection highlights its potential for detecting tail-end phenomena in extreme-scale data settings.

Sayan Ghosh, Tejas Srinivasan, Swabha Swayamdipta · allen-ai

Human evaluation of generated language through pairwise preference judgments is pervasive. However, under common scenarios, such as when generations from a model pair are very similar, or when stochastic decoding results in large variations in generations, it results in inconsistent preference ratings. We address these challenges by introducing a meta-evaluation measure, separability, which estimates how suitable a test instance is for pairwise preference evaluation. For a candidate test instance, separability samples multiple generations from a pair of models, and measures how distinguishable the two sets of generations are. Our experiments show that instances with high separability values yield more consistent preference ratings from both human- and auto-raters. Further, the distribution of separability allows insights into which test benchmarks are more valuable for comparing models. Finally, we incorporate separability into ELO ratings, accounting for how suitable each test instance might be for reliably ranking LLMs. Overall, separability has implications for consistent, efficient and robust preference evaluation of LLMs with both human- and auto-raters.

Jaspreet Ranjit, Ke Zhou, Swabha Swayamdipta et al. · allen-ai

Prior work has mapped which workplace tasks are exposed to AI, but less is known about whether workers perceive these tasks as meaningful or as busywork. We examined: (1) which dimensions of meaningful work do workers associate with tasks exposed to AI; and (2) how do the traits of existing AI systems compare to the traits workers want. We surveyed workers and developers on a representative sample of 171 tasks and use language models (LMs) to scale ratings to 10,131 computer-assisted tasks across all U.S. occupations. Worryingly, we find that tasks that workers associate with a sense of agency or happiness may be disproportionately exposed to AI. We also document design gaps: developers report emphasizing politeness, strictness, and imagination in system design; by contrast, workers prefer systems that are straightforward, tolerant, and practical. To address these gaps, we call for AI whose design explicitly focuses on meaningful work and worker needs, proposing a five-part research agenda.

Yue Huang, Chujie Gao, Siyuan Wu et al.

Generative Foundation Models (GenFMs) have emerged as transformative tools. However, their widespread adoption raises critical concerns regarding trustworthiness across dimensions. This paper presents a comprehensive framework to address these challenges through three key contributions. First, we systematically review global AI governance laws and policies from governments and regulatory bodies, as well as industry practices and standards. Based on this analysis, we propose a set of guiding principles for GenFMs, developed through extensive multidisciplinary collaboration that integrates technical, ethical, legal, and societal perspectives. Second, we introduce TrustGen, the first dynamic benchmarking platform designed to evaluate trustworthiness across multiple dimensions and model types, including text-to-image, large language, and vision-language models. TrustGen leverages modular components--metadata curation, test case generation, and contextual variation--to enable adaptive and iterative assessments, overcoming the limitations of static evaluation methods. Using TrustGen, we reveal significant progress in trustworthiness while identifying persistent challenges. Finally, we provide an in-depth discussion of the challenges and future directions for trustworthy GenFMs, which reveals the complex, evolving nature of trustworthiness, highlighting the nuanced trade-offs between utility and trustworthiness, and consideration for various downstream applications, identifying persistent challenges and providing a strategic roadmap for future research. This work establishes a holistic framework for advancing trustworthiness in GenAI, paving the way for safer and more responsible integration of GenFMs into critical applications. To facilitate advancement in the community, we release the toolkit for dynamic evaluation.

Urja Khurana, Eric Nalisnick, Antske Fokkens et al.

Subjective tasks in NLP have been mostly relegated to objective standards, where the gold label is decided by taking the majority vote. This obfuscates annotator disagreement and the inherent uncertainty of the label. We argue that subjectivity should factor into model decisions and play a direct role via calibration under a selective prediction setting. Specifically, instead of calibrating confidence purely from the model's perspective, we calibrate models for subjective tasks based on crowd worker agreement. Our method, Crowd-Calibrator, models the distance between the distribution of crowd worker labels and the model's own distribution over labels to inform whether the model should abstain from a decision. On two highly subjective tasks, hate speech detection and natural language inference, our experiments show Crowd-Calibrator either outperforms or achieves competitive performance with existing selective prediction baselines. Our findings highlight the value of bringing human decision-making into model predictions.

Guy Kaplan, Zorik Gekhman, Zhen Zhu et al.

Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t. knowledge acquired during pre-training. In this work, we explore whether SFT-induced hallucinations can be mitigated using established tools from the continual learning literature, since they arise as a by-product of knowledge degradation during training. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. We also show that, in settings where new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups, can preserve task performance while reducing hallucinations. Lastly, we investigate the mechanism behind SFT-induced hallucinations through three hypotheses: capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, and that self-distillation succeeds by mitigating this interference.

Kritee Kondapally, Claire J. Smerdon, Pooja C. Patel et al.

Language models (LMs) can exhibit systematic biases against speakers based on variations in their dialects, even in the absence of a dialect label, a behavior known as covert dialect bias. In this work, we quantify covert dialect bias in online discourse by evaluating how LMs associate stereotypical traits (derived from social psychology research on racial bias) with intent-equivalent tweets in Standard American English (SAE) and African-American Vernacular English (AAVE). While prior work shows that LMs associate more negative stereotypes with AAVE when evaluating tweets in isolation, we are surprised to find that this bias is significantly exacerbated when SAE / AAVE tweet pairs are compared side by side, a setting that more closely reflects high-impact decision making contexts in which models are used to rank candidates. The bias only worsens when dialect labels are explicitly specified. This is striking, given the extensive efforts from commercial developers to mitigate bias in their LMs. Encouragingly, we show that counterfactual fairness finetuning can mitigate covert dialect bias for some stereotypical traits, reducing average disparities when evaluating tweets in isolation, however, these improvements do not consistently hold across traits when evaluating SAE / AAVE tweets side by side. Our findings show that existing evaluation settings for covert dialect bias may underestimate its severity, specifically in contrastive settings. Additionally, overt dialect bias remains pronounced even after safety aligned finetuning, indicating that it remains an unresolved problem, and motivates the need for more robust evaluation and mitigation frameworks.

Harshita Diddee, Gregory Yauney, Swabha Swayamdipta et al.

Do language model benchmarks actually measure what practitioners intend them to ? High-level metadata is too coarse to convey the granular reality of benchmarks: a "poetry" benchmark may never test for haikus, while "instruction-following" benchmarks will often test for an arbitrary mix of skills. This opacity makes verifying alignment with practitioner goals a laborious process, risking an illusion of competence even when models fail on untested facets of user interests. We introduce BenchBrowser, a retriever that surfaces evaluation items relevant to natural language use cases over 20 benchmark suites. Validated by a human study confirming high retrieval precision, BenchBrowser generates evidence to help practitioners diagnose low content validity (narrow coverage of a capability's facets) and low convergent validity (lack of stable rankings when measuring the same capability). BenchBrowser, thus, helps quantify a critical gap between practitioner intent and what benchmarks actually test.

Risha Surana, Qinyuan Ye, Swabha Swayamdipta

Emergency responders managing hazardous material HAZMAT incidents face critical, time-sensitive decisions, manually navigating extensive chemical guidelines. We investigate whether today's language models can assist responders by rapidly and reliably understanding critical information, identifying hazards, and providing recommendations. We introduce the Chemical Emergency Response Evaluation Framework (ChEmREF), a new benchmark comprising questions on 1,035 HAZMAT chemicals from the Emergency Response Guidebook and the PubChem Database. ChEmREF is organized into three tasks: (1) translation of chemical representation between structured and unstructured forms (e.g., converting C2H6O to ethanol), (2) emergency response generation (e.g., recommending appropriate evacuation distances) and (3) domain knowledge question answering from chemical safety and certification exams. Our best evaluated models received an exact match of 68.0% on unstructured HAZMAT chemical representation translation, a LLM Judge score of 52.7% on incident response recommendations, and a multiple-choice accuracy of 63.9% on HAMZAT examinations. These findings suggest that while language models show potential to assist emergency responders in various tasks, they require careful human oversight due to their current limitations.

Graham Neubig, Chris Dyer, Yoav Goldberg et al.

We describe DyNet, a toolkit for implementing neural network models based on dynamic declaration of network structure. In the static declaration strategy that is used in toolkits like Theano, CNTK, and TensorFlow, the user first defines a computation graph (a symbolic representation of the computation), and then examples are fed into an engine that executes this computation and computes its derivatives. In DyNet's dynamic declaration strategy, computation graph construction is mostly transparent, being implicitly constructed by executing procedural code that computes the network outputs, and the user is free to use different network structures for each input. Dynamic declaration thus facilitates the implementation of more complicated network architectures, and DyNet is specifically designed to allow users to implement their models in a way that is idiomatic in their preferred programming language (C++ or Python). One challenge with dynamic declaration is that because the symbolic computation graph is defined anew for every training example, its construction must have low overhead. To achieve this, DyNet has an optimized C++ backend and lightweight graph representation. Experiments show that DyNet's speeds are faster than or comparable with static declaration toolkits, and significantly faster than Chainer, another dynamic declaration toolkit. DyNet is released open-source under the Apache 2.0 license and available at http://github.com/clab/dynet.

Lincan Li, Jiaqi Li, Catherine Chen et al.

In recent years, large language models (LLMs) have been widely adopted in political science tasks such as election prediction, sentiment analysis, policy impact assessment, and misinformation detection. Meanwhile, the need to systematically understand how LLMs can further revolutionize the field also becomes urgent. In this work, we--a multidisciplinary team of researchers spanning computer science and political science--present the first principled framework termed Political-LLM to advance the comprehensive understanding of integrating LLMs into computational political science. Specifically, we first introduce a fundamental taxonomy classifying the existing explorations into two perspectives: political science and computational methodologies. In particular, from the political science perspective, we highlight the role of LLMs in automating predictive and generative tasks, simulating behavior dynamics, and improving causal inference through tools like counterfactual generation; from a computational perspective, we introduce advancements in data preparation, fine-tuning, and evaluation methods for LLMs that are tailored to political contexts. We identify key challenges and future directions, emphasizing the development of domain-specific datasets, addressing issues of bias and fairness, incorporating human expertise, and redefining evaluation criteria to align with the unique requirements of computational political science. Political-LLM seeks to serve as a guidebook for researchers to foster an informed, ethical, and impactful use of Artificial Intelligence in political science. Our online resource is available at: http://political-llm.org/.

Xinyue Cui, Johnny Tian-Zheng Wei, Swabha Swayamdipta et al. · allen-ai

Data watermarking in language models injects traceable signals, such as specific token sequences or stylistic patterns, into copyrighted text, allowing copyright holders to track and verify training data ownership. Previous data watermarking techniques primarily focus on effective memorization during pretraining, while overlooking challenges that arise in other stages of the LLM lifecycle, such as the risk of watermark filtering during data preprocessing and verification difficulties due to API-only access. To address these challenges, we propose a novel data watermarking approach that injects plausible yet fictitious knowledge into training data using generated passages describing a fictitious entity and its associated attributes. Our watermarks are designed to be memorized by the LLM through seamlessly integrating in its training data, making them harder to detect lexically during preprocessing. We demonstrate that our watermarks can be effectively memorized by LLMs, and that increasing our watermarks' density, length, and diversity of attributes strengthens their memorization. We further show that our watermarks remain effective after continual pretraining and supervised finetuning. Finally, we show that our data watermarks can be evaluated even under API-only access via question answering.

Atharva Kulkarni, Yuan Zhang, Joel Ruben Antony Moniz et al. · cmu

Hallucinations pose a significant obstacle to the reliability and widespread adoption of language models, yet their accurate measurement remains a persistent challenge. While many task- and domain-specific metrics have been proposed to assess faithfulness and factuality concerns, the robustness and generalization of these metrics are still untested. In this paper, we conduct a large-scale empirical evaluation of 6 diverse sets of hallucination detection metrics across 4 datasets, 37 language models from 5 families, and 5 decoding methods. Our extensive investigation reveals concerning gaps in current hallucination evaluation: metrics often fail to align with human judgments, take an overtly myopic view of the problem, and show inconsistent gains with parameter scaling. Encouragingly, LLM-based evaluation, particularly with GPT-4, yields the best overall results, and mode-seeking decoding methods seem to reduce hallucinations, especially in knowledge-grounded settings. These findings underscore the need for more robust metrics to understand and quantify hallucinations, and better strategies to mitigate them.

Murtaza Nazir, Matthew Finlayson, John X. Morris et al.

Language model inversion seeks to recover hidden prompts using only language model outputs. This capability has implications for security and accountability in language model deployments, such as leaking private information from an API-protected language model's system message. We propose a new method -- prompt inversion from logprob sequences (PILS) -- that recovers hidden prompts by gleaning clues from the model's next-token probabilities over the course of multiple generation steps. Our method is enabled by a key insight: The vector-valued outputs of a language model occupy a low-dimensional subspace. This enables us to losslessly compress the full next-token probability distribution over multiple generation steps using a linear map, allowing more output information to be used for inversion. Our approach yields massive gains over previous state-of-the-art methods for recovering hidden prompts, achieving 2--3.5 times higher exact recovery rates across test sets, in one case increasing the recovery rate from 17% to 60%. Our method also exhibits surprisingly good generalization behavior; for instance, an inverter trained on 16 generations steps gets 5--27 points higher prompt recovery when we increase the number of steps to 32 at test time. Furthermore, we demonstrate strong performance of our method on the more challenging task of recovering hidden system messages. We also analyze the role of verbatim repetition in prompt recovery and propose a new method for cross-family model transfer for logit-based inverters. Our findings show that next-token probabilities are a considerably more vulnerable attack surface for inversion attacks than previously known.

Brihi Joshi, Keyu He, Sahana Ramnath et al.

Language models today are widely used in education, yet their ability to tailor responses for learners with varied informational needs and knowledge backgrounds remains under-explored. To this end, we introduce ELI-Why, a benchmark of 13.4K "Why" questions to evaluate the pedagogical capabilities of language models. We then conduct two extensive human studies to assess the utility of language model-generated explanatory answers (explanations) on our benchmark, tailored to three distinct educational grades: elementary, high-school and graduate school. In our first study, human raters assume the role of an "educator" to assess model explanations' fit to different educational grades. We find that GPT-4-generated explanations match their intended educational background only 50% of the time, compared to 79% for lay human-curated explanations. In our second study, human raters assume the role of a learner to assess if an explanation fits their own informational needs. Across all educational backgrounds, users deemed GPT-4-generated explanations 20% less suited on average to their informational needs, when compared to explanations curated by lay people. Additionally, automated evaluation metrics reveal that explanations generated across different language model families for different informational needs remain indistinguishable in their grade-level, limiting their pedagogical effectiveness.

Joseph Liu, Yoonsoo Nam, Xinyue Cui et al. · allen-ai

Despite the successes of language models, their evaluation remains a daunting challenge for new and existing tasks. We consider the task of text simplification, commonly used to improve information accessibility, where evaluation faces two major challenges. First, the data in existing benchmarks might not reflect the capabilities of current language models on the task, often containing disfluent, incoherent, or simplistic examples. Second, existing human ratings associated with the benchmarks often contain a high degree of disagreement, resulting in inconsistent ratings; nevertheless, existing metrics still have to show higher correlations with these imperfect ratings. As a result, evaluation for the task is not reliable and does not reflect expected trends (e.g., more powerful models being assigned higher scores). We address these challenges for the task of text simplification through three contributions. First, we introduce SynthSimpliEval, a synthetic benchmark for text simplification featuring simplified sentences generated by models of varying sizes. Through a pilot study, we show that human ratings on our benchmark exhibit high inter-annotator agreement and reflect the expected trend: larger models produce higher-quality simplifications. Second, we show that auto-evaluation with a panel of LLM judges (LLMs-as-a-jury) often suffices to obtain consistent ratings for the evaluation of text simplification. Third, we demonstrate that existing learnable metrics for text simplification benefit from training on our LLMs-as-a-jury-rated synthetic data, closing the gap with pure LLMs-as-a-jury for evaluation. Overall, through our case study on text simplification, we show that a reliable evaluation requires higher quality test data, which could be obtained through synthetic data and LLMs-as-a-jury ratings.

Brihi Joshi, Xiang Ren, Swabha Swayamdipta et al.

Language models prompted with a user description or persona are being used to predict the user's preferences and opinions. However, existing approaches to building personas mostly rely on a user's demographic attributes and/or prior judgments, but not on any underlying reasoning behind a user's judgments. We introduce PB&J (Psychology of Behavior and Judgments), a framework that improves LM personas by incorporating potential rationales for why the user could have made a certain judgment. Our rationales are generated by a language model to explicitly reason about a user's behavior on the basis of their experiences, personality traits, or beliefs. Our method employs psychological scaffolds: structured frameworks such as the Big 5 Personality Traits or Primal World Beliefs to help ground the generated rationales in existing theories. Experiments on public opinion and movie preference prediction tasks demonstrate that language model personas augmented with PB&J rationales consistently outperform personas conditioned only on user demographics and / or judgments, including those that use a model's default chain-of-thought, which is not grounded in psychological theories. Additionally, our PB&J personas perform competitively with those using human-written rationales, suggesting the potential of synthetic rationales guided by existing theories.

Matthew Finlayson, Xiang Ren, Swabha Swayamdipta

The ubiquity of closed-weight language models with public-facing APIs has generated interest in forensic methods, both for extracting hidden model details (e.g., parameters) and for identifying models by their outputs. One successful approach to these goals has been to exploit the geometric constraints imposed by the language model architecture and parameters. In this work, we show that a lesser-known geometric constraint--namely, that language model outputs lie on the surface of a high-dimensional ellipse--functions as a signature for the model and can be used to identify the source model of a given output. This ellipse signature has unique properties that distinguish it from existing model-output association methods like language model fingerprints. In particular, the signature is hard to forge: without direct access to model parameters, it is practically infeasible to produce log-probabilities (logprobs) on the ellipse. Secondly, the signature is naturally occurring, since all language models have these elliptical constraints. Thirdly, the signature is self-contained, in that it is detectable without access to the model inputs or the full weights. Finally, the signature is compact and redundant, as it is independently detectable in each logprob output from the model. We evaluate a novel technique for extracting the ellipse from small models and discuss the practical hurdles that make it infeasible for production-scale models. Finally, we use ellipse signatures to propose a protocol for language model output verification, analogous to cryptographic symmetric-key message authentication systems.

Gregory Yauney, Shahzaib Saqib Warraich, Swabha Swayamdipta · allen-ai

Micro-benchmarking offers a solution to the often prohibitive time and cost of language model development: evaluate on a very small subset of existing benchmarks. Can these micro-benchmarks, however, rank models as consistently as the full benchmarks they replace? And can they rank models more consistently than selecting a random subset of data points? In many scenarios, we find that the answer is no. We introduce a meta-evaluation measure for micro-benchmarking which investigates how well a micro-benchmark can rank two models as a function of their performance difference on the full benchmark. This approach can determine which model pairs can be ranked correctly by a micro-benchmark, allowing for a finer-grained analysis of the trade-off between micro-benchmark size and reliability. Prior work has suggested selecting as few as 10 examples; we find that no micro-benchmarking method can consistently rank model pairs 3.5 points of accuracy apart on MMLU-Pro or 4 points apart on BIG-bench Hard. In order to consistently rank model pairs with relatively similar performances, we show that often as many as 250 examples must be selected, at which point random sampling is competitive with existing micro-benchmarking methods. When comparing only 8B instruction-tuned models on MMLU-Pro micro-benchmarks with 25 examples, we find that more than half of pairwise comparisons are not likely to be preserved. Our work provides actionable guidance for both micro-benchmark users and developers in navigating the trade-off between evaluation efficiency and reliability.

Sayan Ghosh, Shahzaib Saqib Warraich, Dhruv Tarsadiya et al. · allen-ai

Language models can be sampled multiple times to access the distribution underlying their responses, but existing methods cannot efficiently synthesize rich epistemic signals across different long-form responses. We introduce Consensus Graphs (ConGrs), a flexible DAG-based data structure that represents shared information, as well as semantic variation in a set of sampled LM responses to the same prompt. We construct ConGrs using a light-weight lexical sequence alignment algorithm from bioinformatics, supplemented by the targeted usage of a secondary LM judge. Further, we design task-dependent decoding methods to synthesize a single, final response from our ConGr data structure. Our experiments show that synthesizing responses from ConGrs improves factual precision on two biography generation tasks by up to 31% over an average response and reduces reliance on LM judges by more than 80% compared to other methods. We also use ConGrs for three refusal-based tasks requiring abstention on unanswerable queries and find that abstention rate is increased by up to 56%. We apply our approach to the MATH and AIME reasoning tasks and find an improvement over self-verification and majority vote baselines by up to 6 points of accuracy. We show that ConGrs provide a flexible method for capturing variation in LM responses and using the epistemic signals provided by response variation to synthesize more effective responses.

Keyu He, Tejas Srinivasan, Brihi Joshi et al.

When people query Vision-Language Models (VLMs) but cannot see the accompanying visual context (e.g. for blind and low-vision users), augmenting VLM predictions with natural language explanations can signal which model predictions are reliable. However, prior work has found that explanations can easily convince users that inaccurate VLM predictions are correct. To remedy undesirable overreliance on VLM predictions, we propose evaluating two complementary qualities of VLM-generated explanations via two quality scoring functions. We propose Visual Fidelity, which captures how faithful an explanation is to the visual context, and Contrastiveness, which captures how well the explanation identifies visual details that distinguish the model's prediction from plausible alternatives. On the A-OKVQA and VizWiz tasks, these quality scoring functions are better calibrated with model correctness than existing explanation qualities. We conduct a user study in which participants have to decide whether a VLM prediction is accurate without viewing its visual context. We observe that showing our quality scores alongside VLM explanations improves participants' accuracy at predicting VLM correctness by 11.1%, including a 15.4% reduction in the rate of falsely believing incorrect predictions. These findings highlight the utility of explanation quality scores in fostering appropriate reliance on VLM predictions.

Ryan Wang, Matthew Finlayson, Luca Soldaini et al. · allen-ai

Language model developers typically filter out high-risk content -- such as toxic or copyrighted text -- from their pre-training data to prevent models from generating similar outputs. However, removing such data altogether limits models' ability to recognize and appropriately respond to harmful or sensitive content. In this paper, we introduce Selective Loss to Understand but Not Generate (SLUNG), a pre-training paradigm through which models learn to understand high-risk data without learning to generate it. Instead of uniformly applying the next-token prediction loss, SLUNG selectively avoids incentivizing the generation of high-risk tokens while ensuring they remain within the model's context window. As the model learns to predict low-risk tokens that follow high-risk ones, it is forced to understand the high-risk content. Through our experiments, we show that SLUNG consistently improves models' understanding of high-risk data (e.g., ability to recognize toxic content) without increasing its generation (e.g., toxicity of model responses). Overall, our SLUNG paradigm enables models to benefit from high-risk text that would otherwise be filtered out.

Jaspreet Ranjit, Brihi Joshi, Rebecca Dorn et al.

Warning: Contents of this paper may be upsetting. Public attitudes towards key societal issues, expressed on online media, are of immense value in policy and reform efforts, yet challenging to understand at scale. We study one such social issue: homelessness in the U.S., by leveraging the remarkable capabilities of large language models to assist social work experts in analyzing millions of posts from Twitter. We introduce a framing typology: Online Attitudes Towards Homelessness (OATH) Frames: nine hierarchical frames capturing critiques, responses and perceptions. We release annotations with varying degrees of assistance from language models, with immense benefits in scaling: 6.5x speedup in annotation time while only incurring a 3 point F1 reduction in performance with respect to the domain experts. Our experiments demonstrate the value of modeling OATH-Frames over existing sentiment and toxicity classifiers. Our large-scale analysis with predicted OATH-Frames on 2.4M posts on homelessness reveal key trends in attitudes across states, time periods and vulnerable populations, enabling new insights on the issue. Our work provides a general framework to understand nuanced public attitudes at scale, on issues beyond homelessness.

Xinyue Cui, Swabha Swayamdipta

Despite the remarkable generative capabilities of language models in producing naturalistic language, their effectiveness on explicit manipulation and generation of linguistic structures remain understudied. In this paper, we investigate the task of generating new sentences preserving a given semantic structure, following the FrameNet formalism. We propose a framework to produce novel frame-semantically annotated sentences following an overgenerate-and-filter approach. Our results show that conditioning on rich, explicit semantic information tends to produce generations with high human acceptance, under both prompting and finetuning. Our generated frame-semantic structured annotations are effective at training data augmentation for frame-semantic role labeling in low-resource settings; however, we do not see benefits under higher resource settings. Our study concludes that while generating high-quality, semantically rich data might be within reach, the downstream utility of such generations remains to be seen, highlighting the outstanding challenges with automating linguistic annotation tasks.

Matthew Finlayson, Xiang Ren, Swabha Swayamdipta

Large language model (LLM) providers often hide the architectural details and parameters of their proprietary models by restricting public access to a limited API. In this work we show that, with only a conservative assumption about the model architecture, it is possible to learn a surprisingly large amount of non-public information about an API-protected LLM from a relatively small number of API queries (e.g., costing under $1000 USD for OpenAI's gpt-3.5-turbo). Our findings are centered on one key observation: most modern LLMs suffer from a softmax bottleneck, which restricts the model outputs to a linear subspace of the full output space. We exploit this fact to unlock several capabilities, including (but not limited to) obtaining cheap full-vocabulary outputs, auditing for specific types of model updates, identifying the source LLM given a single full LLM output, and even efficiently discovering the LLM's hidden size. Our empirical investigations show the effectiveness of our methods, which allow us to estimate the embedding size of OpenAI's gpt-3.5-turbo to be about 4096. Lastly, we discuss ways that LLM providers can guard against these attacks, as well as how these capabilities can be viewed as a feature (rather than a bug) by allowing for greater transparency and accountability.

Phillip Howard, Junlin Wang, Vasudev Lal et al.

Comparative knowledge (e.g., steel is stronger and heavier than styrofoam) is an essential component of our world knowledge, yet understudied in prior literature. In this paper, we harvest the dramatic improvements in knowledge capabilities of language models into a large-scale comparative knowledge base. While the ease of acquisition of such comparative knowledge is much higher from extreme-scale models like GPT-4, compared to their considerably smaller and weaker counterparts such as GPT-2, not even the most powerful models are exempt from making errors. We thus ask: to what extent are models at different scales able to generate valid and diverse comparative knowledge? We introduce NeuroComparatives, a novel framework for comparative knowledge distillation overgenerated from language models such as GPT-variants and LLaMA, followed by stringent filtering of the generated knowledge. Our framework acquires comparative knowledge between everyday objects, producing a corpus of up to 8.8M comparisons over 1.74M entity pairs - 10X larger and 30% more diverse than existing resources. Moreover, human evaluations show that NeuroComparatives outperform existing resources in terms of validity (up to 32% absolute improvement). Our acquired NeuroComparatives leads to performance improvements on five downstream tasks. We find that neuro-symbolic manipulation of smaller models offers complementary benefits to the currently dominant practice of prompting extreme-scale language models for knowledge distillation.

Alisa Liu, Swabha Swayamdipta, Noah A. Smith et al.

A recurring challenge of crowdsourcing NLP datasets at scale is that human writers often rely on repetitive patterns when crafting examples, leading to a lack of linguistic diversity. We introduce a novel approach for dataset creation based on worker and AI collaboration, which brings together the generative strength of language models and the evaluative strength of humans. Starting with an existing dataset, MultiNLI for natural language inference (NLI), our approach uses dataset cartography to automatically identify examples that demonstrate challenging reasoning patterns, and instructs GPT-3 to compose new examples with similar patterns. Machine generated examples are then automatically filtered, and finally revised and labeled by human crowdworkers. The resulting dataset, WANLI, consists of 107,885 NLI examples and presents unique empirical strengths over existing NLI datasets. Remarkably, training a model on WANLI improves performance on eight out-of-domain test sets we consider, including by 11% on HANS and 9% on Adversarial NLI, compared to training on the 4x larger MultiNLI. Moreover, it continues to be more effective than MultiNLI augmented with other NLI datasets. Our results demonstrate the promise of leveraging natural language generation techniques and re-imagining the role of humans in the dataset creation process.

Sarah Wiegreffe, Jack Hessel, Swabha Swayamdipta et al.

Large language models are increasingly capable of generating fluent-appearing text with relatively little task-specific supervision. But can these models accurately explain classification decisions? We consider the task of generating free-text explanations using human-written examples in a few-shot manner. We find that (1) authoring higher quality prompts results in higher quality generations; and (2) surprisingly, in a head-to-head comparison, crowdworkers often prefer explanations generated by GPT-3 to crowdsourced explanations in existing datasets. Our human studies also show, however, that while models often produce factual, grammatical, and sufficient explanations, they have room to improve along axes such as providing novel information and supporting the label. We create a pipeline that combines GPT-3 with a supervised filter that incorporates binary acceptability judgments from humans in the loop. Despite the intrinsic subjectivity of acceptability judgments, we demonstrate that acceptability is partially correlated with various fine-grained attributes of explanations. Our approach is able to consistently filter GPT-3-generated explanations deemed acceptable by humans.

Maarten Sap, Swabha Swayamdipta, Laura Vianna et al.

The perceived toxicity of language can vary based on someone's identity and beliefs, but this variation is often ignored when collecting toxic language datasets, resulting in dataset and model biases. We seek to understand the who, why, and what behind biases in toxicity annotations. In two online studies with demographically and politically diverse participants, we investigate the effect of annotator identities (who) and beliefs (why), drawing from social psychology research about hate speech, free speech, racist beliefs, political leaning, and more. We disentangle what is annotated as toxic by considering posts with three characteristics: anti-Black language, African American English (AAE) dialect, and vulgarity. Our results show strong associations between annotator identity and beliefs and their ratings of toxicity. Notably, more conservative annotators and those who scored highly on our scale for racist beliefs were less likely to rate anti-Black language as toxic, but more likely to rate AAE as toxic. We additionally present a case study illustrating how a popular toxicity detection system's ratings inherently reflect only specific beliefs and perspectives. Our findings call for contextualizing toxicity labels in social variables, which raises immense implications for toxic language annotation and detection.

Kawin Ethayarajh, Yejin Choi, Swabha Swayamdipta

Estimating the difficulty of a dataset typically involves comparing state-of-the-art models to humans; the bigger the performance gap, the harder the dataset is said to be. However, this comparison provides little understanding of how difficult each instance in a given distribution is, or what attributes make the dataset difficult for a given model. To address these questions, we frame dataset difficulty -- w.r.t. a model $\mathcal{V}$ -- as the lack of $\mathcal{V}$-$\textit{usable information}$ (Xu et al., 2019), where a lower value indicates a more difficult dataset for $\mathcal{V}$. We further introduce $\textit{pointwise $\mathcal{V}$-information}$ (PVI) for measuring the difficulty of individual instances w.r.t. a given distribution. While standard evaluation metrics typically only compare different models for the same dataset, $\mathcal{V}$-$\textit{usable information}$ and PVI also permit the converse: for a given model $\mathcal{V}$, we can compare different datasets, as well as different instances/slices of the same dataset. Furthermore, our framework allows for the interpretability of different input attributes via transformations of the input, which we use to discover annotation artefacts in widely-used NLP benchmarks.

Ayush Pancholy, Miriam R. L. Petruck, Swabha Swayamdipta

While FrameNet is widely regarded as a rich resource of semantics in natural language processing, a major criticism concerns its lack of coverage and the relative paucity of its labeled data compared to other commonly used lexical resources such as PropBank and VerbNet. This paper reports on a pilot study to address these gaps. We propose a data augmentation approach, which uses existing frame-specific annotation to automatically annotate other lexical units of the same frame which are unannotated. Our rule-based approach defines the notion of a sister lexical unit and generates frame-specific augmented data for training. We present experiments on frame-semantic role labeling which demonstrate the importance of this data augmentation: we obtain a large improvement to prior results on frame identification and argument identification for FrameNet, utilizing both full-text and lexicographic annotations under FrameNet. Our findings on data augmentation highlight the value of automatic resource creation for improved models in frame-semantic parsing.

Alisa Liu, Maarten Sap, Ximing Lu et al.

Despite recent advances in natural language generation, it remains challenging to control attributes of generated text. We propose DExperts: Decoding-time Experts, a decoding-time method for controlled text generation that combines a pretrained language model with "expert" LMs and/or "anti-expert" LMs in a product of experts. Intuitively, under the ensemble, tokens only get high probability if they are considered likely by the experts, and unlikely by the anti-experts. We apply DExperts to language detoxification and sentiment-controlled generation, where we outperform existing controllable generation methods on both automatic and human evaluations. Moreover, because DExperts operates only on the output of the pretrained LM, it is effective with (anti-)experts of smaller size, including when operating on GPT-3. Our work highlights the promise of tuning small LMs on text with (un)desirable attributes for efficient decoding-time steering.

Alon Jacovi, Swabha Swayamdipta, Shauli Ravfogel et al.

Contrastive explanations clarify why an event occurred in contrast to another. They are more inherently intuitive to humans to both produce and comprehend. We propose a methodology to produce contrastive explanations for classification models by modifying the representation to disregard non-contrastive information, and modifying model behavior to only be based on contrastive reasoning. Our method is based on projecting model representation to a latent space that captures only the features that are useful (to the model) to differentiate two potential decisions. We demonstrate the value of contrastive explanations by analyzing two different scenarios, using both high-level abstract concept attribution and low-level input token/span attribution, on two widely used text classification tasks. Specifically, we produce explanations for answering: for which label, and against which alternative label, is some aspect of the input useful? And which aspects of the input are useful for and against particular decisions? Overall, our findings shed light on the ability of label-contrastive explanations to provide a more accurate and finer-grained interpretability of a model's decision.

Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers et al.

As major progress is made in open-ended text generation, measuring how close machine-generated text is to human language remains a critical open problem. We introduce MAUVE, a comparison measure for open-ended text generation, which directly compares the learnt distribution from a text generation model to the distribution of human-written text using divergence frontiers. MAUVE scales up to modern text generation models by computing information divergences in a quantized embedding space. Through an extensive empirical study on three open-ended generation tasks, we find that MAUVE identifies known properties of generated text, scales naturally with model size, and correlates with human judgments, with fewer restrictions than existing distributional evaluation metrics.

Xuhui Zhou, Maarten Sap, Swabha Swayamdipta et al.

Biased associations have been a challenge in the development of classifiers for detecting toxic language, hindering both fairness and accuracy. As potential solutions, we investigate recently introduced debiasing methods for text classification datasets and models, as applied to toxic language detection. Our focus is on lexical (e.g., swear words, slurs, identity mentions) and dialectal markers (specifically African American English). Our comprehensive experiments establish that existing methods are limited in their ability to prevent biased behavior in current toxicity detectors. We then propose an automatic, dialect-aware data correction method, as a proof-of-concept. Despite the use of synthetic labels, this method reduces dialectal associations with toxicity. Overall, our findings show that debiasing a model trained on biased toxic language data is not as effective as simply relabeling the data to remove existing biases.

Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie et al.

Large datasets have become commonplace in NLP research. However, the increased emphasis on data quantity has made it challenging to assess the quality of data. We introduce Data Maps---a model-based tool to characterize and diagnose datasets. We leverage a largely ignored source of information: the behavior of the model on individual instances during training (training dynamics) for building data maps. This yields two intuitive measures for each example---the model's confidence in the true class, and the variability of this confidence across epochs---obtained in a single run of training. Experiments across four datasets show that these model-dependent measures reveal three distinct regions in the data map, each with pronounced characteristics. First, our data maps show the presence of "ambiguous" regions with respect to the model, which contribute the most towards out-of-distribution generalization. Second, the most populous regions in the data are "easy to learn" for the model, and play an important role in model optimization. Finally, data maps uncover a region with instances that the model finds "hard to learn"; these often correspond to labeling errors. Our results indicate that a shift in focus from quantity to quality of data could lead to robust models and improved out-of-distribution generalization.

Yiben Yang, Chaitanya Malaviya, Jared Fernandez et al.

Recent advances in commonsense reasoning depend on large-scale human-annotated training data to achieve peak performance. However, manual curation of training examples is expensive and has been shown to introduce annotation artifacts that neural models can readily exploit and overfit on. We investigate G-DAUG^C, a novel generative data augmentation method that aims to achieve more accurate and robust learning in the low-resource setting. Our approach generates synthetic examples using pretrained language models, and selects the most informative and diverse set of examples for data augmentation. In experiments with multiple commonsense reasoning benchmarks, G-DAUG^C consistently outperforms existing data augmentation methods based on back-translation, and establishes a new state-of-the-art on WinoGrande, CODAH, and CommonsenseQA. Further, in addition to improvements in in-distribution accuracy, G-DAUG^C-augmented training also enhances out-of-distribution generalization, showing greater robustness against adversarial or perturbed examples. Our analysis demonstrates that G-DAUG^C produces a diverse set of fluent training examples, and that its selection and training approaches are important for performance. Our findings encourage future research toward generative data augmentation to enhance both in-distribution learning and out-of-distribution generalization.

Suchin Gururangan, Ana Marasović, Swabha Swayamdipta et al.

Language models pretrained on text from a wide variety of sources form the foundation of today's NLP. In light of the success of these broad-coverage models, we investigate whether it is still helpful to tailor a pretrained model to the domain of a target task. We present a study across four domains (biomedical and computer science publications, news, and reviews) and eight classification tasks, showing that a second phase of pretraining in-domain (domain-adaptive pretraining) leads to performance gains, under both high- and low-resource settings. Moreover, adapting to the task's unlabeled data (task-adaptive pretraining) improves performance even after domain-adaptive pretraining. Finally, we show that adapting to a task corpus augmented using simple data selection strategies is an effective alternative, especially when resources for domain-adaptive pretraining might be unavailable. Overall, we consistently find that multi-phase adaptive pretraining offers large gains in task performance.