As large language models (LLMs) have increased in their capabilities, so doestheir potential for dual use. To reduce harmful outputs, produces and vendors ofLLMs have used reinforcement learning with human feedback (RLHF). In tandem,LLM vendors have been increasingly enabling fine-tuning of their most powerfulmodels. However, concurrent work has shown that fine-tuning can remove RLHFprotections. We may expect that the most powerful models currently available(GPT-4) are less susceptible to fine-tuning attacks. In this work, we show the contrary: fine-tuning allows attackers to remove RLHFprotections with as few as 340 examples and a 95% success rate. These trainingexamples can be automatically generated with weaker models. We further show thatremoving RLHF protections does not decrease usefulness on non-censored outputs,providing evidence that our fine-tuning strategy does not decrease usefulnessdespite using weaker models to generate training data. Our results show the needfor further research on protections on LLMs.
Large language models (LLMs) have shown promise for automatic summarization but the reasons behind their successes are poorly understood. By conducting a human evaluation on ten LLMs across different pretraining methods, prompts, and model scales, we make two important observations. First, we find instruction tuning, not model size, is the key to the LLM’s zero-shot summarization capability. Second, existing studies have been limited by low-quality references, leading to underestimates of human performance and lower few-shot and finetuning performance. To better evaluate LLMs, we perform human evaluation over high-quality summaries we collect from freelance writers. Despite major stylistic differences such as the amount of paraphrasing, we find that LLM summaries are judged to be on par with human written summaries.
While pretrained language models (PLMs) have greatly improved text generation, they have also been known to produce unfaithful or inappropriate content. In contrast, classic template-based systems provide strong guarantees of faithfulness at the cost of fluency. We propose TempLM, which achieves the best of both worlds by distilling a PLM into a template-based generator. On the E2E and SynthBio data-to-text datasets, we show that TempLM is more faithful than the original PLM and is more fluent than prior template systems. Notably, on an out-of-domain evaluation, TempLM reduces a finetuned BART model’s unfaithfulness rate from 83% to 0%. In a human study, we find that TempLM’s templates substantially improve upon human-written ones in BERTScore.
Large language models (LLMs) are subject to sociocultural and other biases previously identified using intrinsic evaluations. However, when and how these intrinsic biases in pre-trained LM representations propagate to downstream, fine-tuned NLP tasks like summarization is not well understood. In this work, we investigate one type of bias—name-nationality bias—and trace it from the pre-training stage to a downstream summarization task across multiple summarization modeling choices. We show that these biases manifest themselves as hallucinations in summarization, leading to factually incorrect summaries. We also find that this propagation of biases is algorithm-dependent: more abstractive models allow biases to propagate more directly to downstream tasks as hallucinated facts. Building on these observations, we further analyze how changes to the adaptation method and fine-tuning data set affect name nationality biases and show that while they can reduce the overall rate of hallucinations, they do not change the types of biases that do appear.
The increased deployment of LMs for real-world tasks involving knowledge and facts makes it important to understand model epistemology: what LMs think they know, and how their attitudes toward that knowledge are affected by language use in their inputs. Here, we study an aspect of model epistemology: how epistemic markers of certainty, uncertainty, or evidentiality like “I’m sure it’s”, “I think it’s”, or “Wikipedia says it’s” affect models, and whether they contribute to model failures. We develop a typology of epistemic markers and inject 50 markers into prompts for question answering. We find that LMs are highly sensitive to epistemic markers in prompts, with accuracies varying more than 80%. Surprisingly, we find that expressions of high certainty result in a 7% decrease in accuracy as compared to low certainty expressions; similarly, factive verbs hurt performance, while evidentials benefit performance. Our analysis of a popular pretraining dataset shows that these markers of uncertainty are associated with answers on question-answering websites, while markers of certainty are associated with questions. These associations may suggest that the behavior of LMs is based on mimicking observed language use, rather than truly reflecting epistemic uncertainty.
Task-oriented dialogue systems often assist users with personal or confidential matters. For this reason, the developers of such a system are generally prohibited from observing actual usage. So how can they know where the system is failing and needs more training data or new functionality? In this work, we study ways in which realistic user utterances can be generated synthetically, to help increase the linguistic and functional coverage of the system, without compromising the privacy of actual users. To this end, we propose a two-stage Differentially Private (DP) generation method which first generates latent semantic parses, and then generates utterances based on the parses. Our proposed approach improves MAUVE by 2.5X and parse tree function-type overlap by 1.3X relative to current approaches for private synthetic data generation, improving both on fluency and semantic coverage. We further validate our approach on a realistic domain adaptation task of adding new functionality from private user data to a semantic parser, and show overall gains of 8.5% points on its accuracy with the new feature.
Recent work has identified noisy and misannotated data as a core cause of hallucinations and unfaithful outputs in Natural Language Generation (NLG) tasks. Consequently, identifying and removing these examples is a key open challenge in creating reliable NLG systems. In this work, we introduce a framework to identify and remove low-quality training instances that lead to undesirable outputs, such as faithfulness errors in text summarization. We show that existing approaches for error tracing, such as gradient-based influence measures, do not perform reliably for detecting faithfulness errors in NLG datasets. We overcome the drawbacks of existing error tracing methods through a new, contrast-based estimate that compares undesired generations to human-corrected outputs. Our proposed method can achieve a mean average precision of 0.93 at detecting known data errors across synthetic tasks with known ground truth, substantially outperforming existing approaches. Using this approach and re-training models on cleaned data leads to a 70% reduction in entity hallucinations on the NYT dataset and a 55% reduction in semantic errors on the E2E dataset.
Given a language model (LM), maximum probability is a poor decoding objective for open-ended generation, because it produces short and repetitive text. On the other hand, sampling can often produce incoherent text that drifts from the original topics. We propose contrastive decoding (CD), a reliable decoding approach that optimizes a contrastive objective subject to a plausibility constraint. The contrastive objective returns the difference between the likelihood under a large LM (called the expert, e.g. OPT-13B) and a small LM (called the amateur, e.g. OPT-125M), and the constraint ensures that the outputs are plausible. CD is inspired by the fact that the failures of larger LMs (e.g., repetition, inco- herence) are even more prevalent in smaller LMs, and that this difference signals which texts should be preferred. CD requires zero additional training, and produces higher quality text than decoding from the larger LM alone. It also works across model scales (OPT-13B and GPT2-1.5B) and significantly outperforms four strong decoding algorithms (e.g., nucleus, top-k) in automatic and human evaluations across wikipedia, news and story domains.
Model-based, reference-free evaluation metricshave been proposed as a fast and cost-effectiveapproach to evaluate Natural Language Generation(NLG) systems. Despite promising recentresults, we find evidence that reference-freeevaluation metrics of summarization and dialoggeneration may be relying on spuriouscorrelations with measures such as word overlap,perplexity, and length. We further observethat for text summarization, these metrics havehigh error rates when ranking current state-ofthe-art abstractive summarization systems. Wedemonstrate that these errors can be mitigatedby explicitly designing evaluation metrics toavoid spurious features in reference-free evaluation.
Meta-learning promises few-shot learners that can adapt to new distributions by repurposing knowledge acquired from previous training. However, we believe meta-learning has not yet succeeded in NLP due to the lack of a well-defined task distribution, leading to attempts that treat datasets as tasks. Such an ad hoc task distribution causes problems of quantity and quality. Since there’s only a handful of datasets for any NLP problem, meta-learners tend to overfit their adaptation mechanism and, since NLP datasets are highly heterogeneous, many learning episodes have poor transfer between their support and query sets, which discourages the meta-learner from adapting. To alleviate these issues, we propose DReCA (Decomposing datasets into Reasoning Categories), a simple method for discovering and using latent reasoning categories in a dataset, to form additional high quality tasks. DReCA works by splitting examples into label groups, embedding them with a finetuned BERT model and then clustering each group into reasoning categories. Across four few-shot NLI problems, we demonstrate that using DReCA improves the accuracy of meta-learners by 1.5-4%
We study how masking and predicting tokens in an unsupervised fashion can give rise to linguistic structures and downstream performance gains. Recent theories have suggested that pretrained language models acquire useful inductive biases through masks that implicitly act as cloze reductions for downstream tasks. While appealing, we show that the success of the random masking strategy used in practice cannot be explained by such cloze-like masks alone. We construct cloze-like masks using task-specific lexicons for three different classification datasets and show that the majority of pretrained performance gains come from generic masks that are not associated with the lexicon. To explain the empirical success of these generic masks, we demonstrate a correspondence between the Masked Language Model (MLM) objective and existing methods for learning statistical dependencies in graphical models. Using this, we derive a method for extracting these learned statistical dependencies in MLMs and show that these dependencies encode useful inductive biases in the form of syntactic structures. In an unsupervised parsing evaluation, simply forming a minimum spanning tree on the implied statistical dependence structure outperforms a classic method for unsupervised parsing (58.74 vs. 55.91 UUAS).
We introduce GEM, a living benchmark for natural language Generation (NLG), its Evaluation, and Metrics. Measuring progress in NLG relies on a constantly evolving ecosystem of automated metrics, datasets, and human evaluation standards. Due to this moving target, new models often still evaluate on divergent anglo-centric corpora with well-established, but flawed, metrics. This disconnect makes it challenging to identify the limitations of current models and opportunities for progress. Addressing this limitation, GEM provides an environment in which models can easily be applied to a wide set of tasks and in which evaluation strategies can be tested. Regular updates to the benchmark will help NLG research become more multilingual and evolve the challenge alongside models. This paper serves as the description of the data for the 2021 shared task at the associated GEM Workshop.
In conversation, uptake happens when a speaker builds on the contribution of their interlocutor by, for example, acknowledging, repeating or reformulating what they have said. In education, teachers’ uptake of student contributions has been linked to higher student achievement. Yet measuring and improving teachers’ uptake at scale is challenging, as existing methods require expensive annotation by experts. We propose a framework for computationally measuring uptake, by (1) releasing a dataset of student-teacher exchanges extracted from US math classroom transcripts annotated for uptake by experts; (2) formalizing uptake as pointwise Jensen-Shannon Divergence (pJSD), estimated via next utterance classification; (3) conducting a linguistically-motivated comparison of different unsupervised measures and (4) correlating these measures with educational outcomes. We find that although repetition captures a significant part of uptake, pJSD outperforms repetition-based baselines, as it is capable of identifying a wider range of uptake phenomena like question answering and reformulation. We apply our uptake measure to three different educational datasets with outcome indicators. Unlike baseline measures, pJSD correlates significantly with instruction quality in all three, providing evidence for its generalizability and for its potential to serve as an automated professional development tool for teachers.
Neural language models are usually trained to match the distributional properties of large-scale corpora by minimizing the log loss. While straightforward to optimize, this approach forces the model to reproduce all variations in the dataset, including noisy and invalid references (e.g., misannotations and hallucinated facts). Even a small fraction of noisy data can degrade the performance of log loss. As an alternative, prior work has shown that minimizing the distinguishability of generated samples is a principled and robust loss that can handle invalid references. However, distinguishability has not been used in practice due to challenges in optimization and estimation. We propose loss truncation: a simple and scalable procedure which adaptively removes high log loss examples as a way to optimize for distinguishability. Empirically, we demonstrate that loss truncation outperforms existing baselines on distinguishability on a summarization task. Furthermore, we show that samples generated by the loss truncation model have factual accuracy ratings that exceed those of baselines and match human references.
How can we measure whether a natural language generation system produces both high quality and diverse outputs? Human evaluation captures quality but not diversity, as it does not catch models that simply plagiarize from the training set. On the other hand, statistical evaluation (i.e., perplexity) captures diversity but not quality, as models that occasionally emit low quality samples would be insufficiently penalized. In this paper, we propose a unified framework which evaluates both diversity and quality, based on the optimal error rate of predicting whether a sentence is human- or machine-generated. We demonstrate that this error rate can be efficiently estimated by combining human and statistical evaluation, using an evaluation metric which we call HUSE. On summarization and chit-chat dialogue, we show that (i) HUSE detects diversity defects which fool pure human evaluation and that (ii) techniques such as annealing for improving quality actually decrease HUSE due to decreased diversity.
Language models are generally trained on data spanning a wide range of topics (e.g., news, reviews, fiction), but they might be applied to an a priori unknown target distribution (e.g., restaurant reviews). In this paper, we first show that training on text outside the test distribution can degrade test performance when using standard maximum likelihood (MLE) training. To remedy this without the knowledge of the test distribution, we propose an approach which trains a model that performs well over a wide range of potential test distributions. In particular, we derive a new distributionally robust optimization (DRO) procedure which minimizes the loss of the model over the worst-case mixture of topics with sufficient overlap with the training distribution. Our approach, called topic conditional value at risk (topic CVaR), obtains a 5.5 point perplexity reduction over MLE when the language models are trained on a mixture of Yelp reviews and news and tested only on reviews.
We propose a new generative language model for sentences that first samples a prototype sentence from the training corpus and then edits it into a new sentence. Compared to traditional language models that generate from scratch either left-to-right or by first sampling a latent sentence vector, our prototype-then-edit model improves perplexity on language modeling and generates higher quality outputs according to human evaluation. Furthermore, the model gives rise to a latent edit vector that captures interpretable semantics such as sentence similarity and sentence-level analogies.
Continuous word representations have been remarkably useful across NLP tasks but remain poorly understood. We ground word embeddings in semantic spaces studied in the cognitive-psychometric literature, taking these spaces as the primary objects to recover. To this end, we relate log co-occurrences of words in large corpora to semantic similarity assessments and show that co-occurrences are indeed consistent with an Euclidean semantic space hypothesis. Framing word embedding as metric recovery of a semantic space unifies existing word embedding algorithms, ties them to manifold learning, and demonstrates that existing algorithms are consistent metric recovery methods given co-occurrence counts from random walks. Furthermore, we propose a simple, principled, direct metric recovery algorithm that performs on par with the state-of-the-art word embedding and manifold learning methods. Finally, we complement recent focus on analogies by constructing two new inductive reasoning datasets—series completion and classification—and demonstrate that word embeddings can be used to solve them as well.