Text detoxification has the potential to mitigate the harms of toxicity by rephrasing text to remove offensive meaning, but subtle toxicity remains challenging to tackle. We introduce MaRCo, a detoxification algorithm that combines controllable generation and text rewriting methods using a Product of Experts with autoencoder language models (LMs). MaRCo uses likelihoods under a non-toxic LM (expert) and a toxic LM (anti-expert) to find candidate words to mask and potentially replace. We evaluate our method on several subtle toxicity and microaggressions datasets, and show that it not only outperforms baselines on automatic metrics, but MaRCo's rewrites are preferred 2.1 $\times$ more in human evaluation. Its applicability to instances of subtle toxicity is especially promising, demonstrating a path forward for addressing increasingly elusive online hate.
translated by 谷歌翻译
我们提出了两种小型无监督方法,用于消除文本中的毒性。我们的第一个方法结合了最近的两个想法:(1)使用小型条件语言模型的生成过程的指导和(2)使用释义模型进行风格传输。我们使用良好的令人措辞的令人愉快的释放器,由风格培训的语言模型引导,以保持文本内容并消除毒性。我们的第二种方法使用BERT用他们的非攻击性同义词取代毒性单词。我们通过使BERT替换具有可变数量的单词的屏蔽令牌来使该方法更灵活。最后,我们介绍了毒性去除任务的风格转移模型的第一个大规模比较研究。我们将模型与许多用于样式传输的方法进行比较。使用无监督的样式传输指标的组合以可参考方式评估该模型。两种方法都建议产生新的SOTA结果。
translated by 谷歌翻译
Large pretrained language models generate fluent text but are notoriously hard to controllably sample from. In this work, we study constrained sampling from such language models: generating text that satisfies user-defined constraints, while maintaining fluency and the model's performance in a downstream task. We propose MuCoLa -- a sampling procedure that combines the log-likelihood of the language model with arbitrary (differentiable) constraints in a single energy function, and then generates samples in a non-autoregressive manner. Specifically, it initializes the entire output sequence with noise and follows a Markov chain defined by Langevin Dynamics using the gradients of the energy function. We evaluate MuCoLa on text generation with soft and hard constraints as well as their combinations obtaining significant improvements over competitive baselines for toxicity avoidance, sentiment control, and keyword-guided generation.
translated by 谷歌翻译
有毒语言检测系统通常会错误地将包含少数群体群体提及的毒性的错误标记文本,因为这些群体通常是在线仇恨的目标。这种对虚假相关性的过度依赖也导致系统在检测隐式有毒语言方面挣扎。为了帮助缓解这些问题,我们创建了Toxigen,这是一个新的大规模和机器生成的数据集,该数据集是274K有毒和良性陈述,约有13个少数群体。我们开发了一个基于示范的提示框架和一种对抗性分类器的解码方法,以使用大量预处理的语言模型生成微妙的有毒和良性文本。以这种方式控制机器的生成使毒素可以比以前的人写文本的资源更大的规模和大约人口组覆盖隐式有毒文本。我们对毒素的一个充满挑战的子集进行人体评估,发现注释者难以区分机器生成的文本和人类写的语言。我们还发现,94.5%的有毒例子被人类注释者标记为仇恨言论。我们使用三个公开可用的数据集,我们表明,对我们的数据进行毒性分类器的填充可以大大提高其在人体编写数据上的性能。我们还证明,毒素可用于抵抗机器生成的毒性,因为鉴定在我们的评估子集中大大改善了分类器。我们的代码和数据可以在https://github.com/microsoft/toxigen上找到。
translated by 谷歌翻译
Text classifiers have promising applications in high-stake tasks such as resume screening and content moderation. These classifiers must be fair and avoid discriminatory decisions by being invariant to perturbations of sensitive attributes such as gender or ethnicity. However, there is a gap between human intuition about these perturbations and the formal similarity specifications capturing them. While existing research has started to address this gap, current methods are based on hardcoded word replacements, resulting in specifications with limited expressivity or ones that fail to fully align with human intuition (e.g., in cases of asymmetric counterfactuals). This work proposes novel methods for bridging this gap by discovering expressive and intuitive individual fairness specifications. We show how to leverage unsupervised style transfer and GPT-3's zero-shot capabilities to automatically generate expressive candidate pairs of semantically similar sentences that differ along sensitive attributes. We then validate the generated pairs via an extensive crowdsourcing study, which confirms that a lot of these pairs align with human intuition about fairness in the context of toxicity classification. Finally, we show how limited amounts of human feedback can be leveraged to learn a similarity specification that can be used to train downstream fairness-aware models.
translated by 谷歌翻译
文本样式传输是自然语言生成中的重要任务,旨在控制生成的文本中的某些属性,例如礼貌,情感,幽默和许多其他特性。它在自然语言处理领域拥有悠久的历史,最近由于深神经模型带来的有希望的性能而重大关注。在本文中,我们对神经文本转移的研究进行了系统调查,自2017年首次神经文本转移工作以来跨越100多个代表文章。我们讨论了任务制定,现有数据集和子任务,评估,以及丰富的方法在存在并行和非平行数据存在下。我们还提供关于这项任务未来发展的各种重要主题的讨论。我们的策据纸张列表在https://github.com/zhijing-jin/text_style_transfer_survey
translated by 谷歌翻译
大型语言模型会产生类似人类的文本,这些文本推动了越来越多的应用。但是,最近的文献以及越来越多的现实世界观察表明,这些模型可以产生有毒,有偏见,不真实或其他有害的语言。尽管正在进行评估语言模型危害的工作,但要远见卓识转换出可能出现的危害可能会引起严格的基准。为了促进这种翻译,我们概述了六种表征有害文本的方式,这些方法在设计新基准时值得明确考虑。然后,我们将这些特征用作镜头来识别现有基准中的趋势和差距。最后,我们将它们应用于视角API的案例研究,这是一种毒性分类器,被广泛用于HARS基准。我们的特征提供了一块桥梁,可以在远见和有效评估之间转化。
translated by 谷歌翻译
文本排毒是创建中性版本的有毒文本的样式转移任务。在本文中,我们使用文本编辑的概念来使用平行的俄罗斯文本语料库构建基于两步标记的排毒模型。通过此模型,我们在Russe排毒共享任务中所有模型中达到了最佳的样式转移精度,超过了较大的序列到序列模型。
translated by 谷歌翻译
在线仇恨言论已成为小时的需求。但是,由于几种地缘政治和文化原因,对此类活动的禁令是不可行的。为了减少问题的严重性,在本文中,我们介绍了一项新颖的任务,仇恨言语归一化,旨在削弱在线帖子表现出的仇恨强度。仇恨言语归一化的意图不是支持仇恨,而是为用户提供对非讨厌的垫脚石,同时为在线平台提供更多时间来监视用户行为的任何改进。为此,我们手动策划了平行语料库 - 仇恨文本及其标准化的同行(标准化文本较不憎恨,更良性)。我们介绍了NACL,这是一个简单而有效的仇恨言语归一化模型,该模型在三个阶段运行 - 首先,它测量了原始样本的仇恨强度;其次,它标识了其中的仇恨跨度;最后,它通过解释仇恨跨度来降低仇恨强度。我们进行了广泛的实验,以通过三向评估(内在,外部和人类研究)来衡量NaCl的功效。我们观察到,NaCl优于六个基准-NACL的强度预测得分为0.1365 RMSE,在SPAN识别中获得0.622 F1分数,而82.27 BLEU和80.05的差异和80.05的困惑为归一化​​文本生成。我们进一步显示了NACL在其他平台上的普遍性(Reddit,Facebook,GAB)。将NaCl的交互式原型放在一起进行用户研究。此外,该工具正在WIPRO AI的真实环境中部署,这是其在线平台上处理有害内容的任务的一部分。
translated by 谷歌翻译
在基于文本的分类器中测试公平性问题的一种常见方法是通过使用反事实来:如果更改输入中的敏感属性,则分类器输出是否会更改?现有的反事实生成方法通常依赖于单词列表或模板,产生不考虑语法,上下文或微妙敏感属性引用的简单反事实,并且可能会错过WordList创建者未考虑的问题。在本文中,我们介绍了一项为克服这些缺点而产生的反事实的任务,并证明了如何利用大型语言模型(LLM)来在此任务上取得进展。我们表明,这种基于LLM的方法可以产生现有方法无法实现的复杂反事实,从而比较了民事评论数据集中各种反事实生成方法的性能,并在评估毒性分类器时显示出它们的价值。
translated by 谷歌翻译
GPT-3显示了培训的大规模语言模型(LMS)的卓越情调学习能力,培训数十亿规模数据。在这里,我们解决了GPT-3纸张报告的一些剩余问题,例如非英语LM,不同大小模型的性能,以及最近引入的迅速优化对上下文学习的效果。为实现这一目标,我们介绍了HyperClova,一个韩国VPT-3的韩国变体训练在一个以韩国为中心的560b标准的令牌。通过我们的韩国特定标记化,HyperClova与我们的培训配置增强,显示了韩国各种下游任务的最先进的上下游零射击和几秒钟学习表演。此外,我们展示了基于及时的学习的性能优势,并演示如何集成到迅速的工程管道中。然后,我们讨论了通过引入Hyperclova Studio,互动提示工程界面向ML的非专家提供AI原型设计能力来实现No Code AI范例的可能性。最后,我们展示了我们具有三个成功的内部应用程序的方法的潜力。
translated by 谷歌翻译
公开可用的大型预磨语删除媒介(LMS)生成具有显着质量的文本,但仅从左右依次顺序地。因此,它们不会立即适用于打破单向假设的生成任务,例如释放或文本缺陷,需要特定于特定的监督。在本文中,我们呈现反射解码,这是一种新型无监督算法,其允许直接向非顺序任务应用单向LMS。我们的2步方法不需要监督甚至并行对象,只有两个离心的预磨损LMS相反的方向:向前和向后。首先,在上下文化步骤中,我们使用LMS生成过去和未来环境的集合,该上下文共同捕获输入(例如,索引源句)。其次,在反射步骤中,我们在这些“上下文集合”中的条件,生成与它们兼容的输出。综合经验结果表明,反思解码优于涉及释义和绑架文本缺陷的强烈无监督的基线,显着缩小无监督和监督方法之间的差距。反射解码超越了各种度量的多个监督基线,包括人为评估。
translated by 谷歌翻译
近年来,文本的风格特性吸引了计算语言学研究人员。具体来说,研究人员研究了文本样式转移(TST)任务,该任务旨在在保留其样式独立内容的同时改变文本的风格属性。在过去的几年中,已经开发了许多新颖的TST算法,而该行业利用这些算法来实现令人兴奋的TST应用程序。由于这种共生,TST研究领域迅速发展。本文旨在对有关文本样式转移的最新研究工作进行全面审查。更具体地说,我们创建了一种分类法来组织TST模型,并提供有关最新技术状况的全面摘要。我们回顾了针对TST任务的现有评估方法,并进行了大规模的可重复性研究,我们在两个公开可用的数据集上实验基准了19个最先进的TST TST算法。最后,我们扩展了当前趋势,并就TST领域的新开发发展提供了新的观点。
translated by 谷歌翻译
基于变压器的语言模型能够生成流利的文本,并在各种自然语言生成任务中有效地适应。但是,已证明在大型未标记的网络文本语料库中鉴定的语言模型已被证明会遭受堕落的有毒内容和社会偏见行为的损害,从而阻碍了他们的安全部署。提出了各种排毒方法来减轻语言模型的毒性;但是,这些方法是在包含与性别,种族或宗教相关的特定社会身份的提示条件下进行排毒语言模型的。在这项研究中,我们提出了增强氧化。一种基于强化学习的方法,用于降低语言模型中的毒性。我们应对语言模型中的安全性挑战,并提出了一种新的奖励模型,该模型能够检测有毒内容并减轻对毒性预测中社会身份的意外偏见。该实验表明,用于语言模型排毒的增强方法化方法优于自动评估指标中现有的排毒方法,这表明我们在语言模型排毒中的方法能力和对生成内容中社会认同的意外偏见的能力较小。
translated by 谷歌翻译
真实世界的文本应用程序通常涉及组成广泛的文本控制操作,例如编辑文本W.R.T.属性,操纵关键字和结构,并生成所需属性的新文本。事先的工作通常会学习/芬太尼语言模型(LM)以执行操作的个人或特定子集。最近的研究以插件方式研究了合并操作,通常在复杂序列空间中以昂贵的搜索或优化进行了研究。本文提出了一种新的有效方法,用于在紧凑的文本潜在空间中进行可复合的文本操作。文本潜在矢量的低维度和不同性使我们能够基于给定的任意插入运算符(例如属性分类器)基于普通微分方程(ODE)开发有效的采样器。通过通过有效的适应性将预告片的LMS(例如GPT2)连接到潜在空间,然后我们将采样向量解码为所需的文本序列。灵活的方法允许使用来自不同域中的任何相关数据获取的各种控制操作员(情感,时态,形式,关键字等)。实验表明,在我们的方法中构成这些操作员可以生成或编辑高质量文本,从而在发电质量和效率方面显着改善了以前的方法。
translated by 谷歌翻译
随着人工智能系统变得越来越强大和普遍,人们对机器的道德或缺乏道德的关注变得越来越关注。然而,向机器讲授道德是一项艰巨的任务,因为道德仍然是人类中最激烈的争论问题之一,更不用说AI了。但是,部署到数百万用户的现有AI系统已经在做出充满道德影响的决策,这构成了一个看似不可能的挑战:教学机器的道德意义,而人类继续努力努力。为了探索这一挑战,我们介绍了Delphi,这是一个基于深层神经网络的实验框架,直接训练了描述性道德判断,例如,“帮助朋友”通常是不错的,而“帮助朋友传播假新闻”不是。经验结果提供了对机器伦理的承诺和局限性的新见解。面对新的道德情况,德尔菲(Delphi)表现出强大的概括能力,而现成的神经网络模型表现出明显差的判断,包括不公正的偏见,证实了对明确教学机器的道德意义的必要性。然而,德尔菲并不完美,表现出对普遍性偏见和不一致的敏感性。尽管如此,我们还是展示了不完美的Delphi的积极用例,包括在其他不完美的AI系统中将其用作组件模型。重要的是,我们根据著名的道德理论来解释Delphi的运营化,这使我们提出了重要的未来研究问题。
translated by 谷歌翻译
As text generated by large language models proliferates, it becomes vital to understand how humans engage with such text, and whether or not they are able to detect when the text they are reading did not originate with a human writer. Prior work on human detection of generated text focuses on the case where an entire passage is either human-written or machine-generated. In this paper, we study a more realistic setting where text begins as human-written and transitions to being generated by state-of-the-art neural language models. We show that, while annotators often struggle at this task, there is substantial variance in annotator skill and that given proper incentives, annotators can improve at this task over time. Furthermore, we conduct a detailed comparison study and analyze how a variety of variables (model size, decoding strategy, fine-tuning, prompt genre, etc.) affect human detection performance. Finally, we collect error annotations from our participants and use them to show that certain textual genres influence models to make different types of errors and that certain sentence-level features correlate highly with annotator selection. We release the RoFT dataset: a collection of over 21,000 human annotations paired with error classifications to encourage future work in human detection and evaluation of generated text.
translated by 谷歌翻译
This work aims to employ natural language generation (NLG) to rapidly generate items for English language learning applications: this requires both language models capable of generating fluent, high-quality English, and to control the output of the generation to match the requirements of the relevant items. We experiment with deep pretrained models for this task, developing novel methods for controlling items for factors relevant in language learning: diverse sentences for different proficiency levels and argument structure to test grammar. Human evaluation demonstrates high grammatically scores for all models (3.4 and above out of 4), and higher length (24%) and complexity (9%) over the baseline for the advanced proficiency model. Our results show that we can achieve strong performance while adding additional control to ensure diverse, tailored content for individual users.
translated by 谷歌翻译
Task agnostic generative pretraining (GPT) has recently proved promising for zero- and few-shot learning, gradually diverting attention from the expensive supervised learning paradigm. Although the community is accumulating knowledge as to capabilities of English-language autoregressive models such as GPT-3 adopting this generative approach, scholarship about these models remains acutely Anglocentric. Consequently, the community currently has serious gaps in its understanding of this class of models, their potential, and their societal impacts in diverse settings, linguistic traditions, and cultures. To alleviate this issue for Arabic, a collection of diverse languages and language varieties with more than $400$ million population, we introduce JASMINE, a suite of powerful Arabic autoregressive Transformer language models ranging in size between 300 million-13 billion parameters. We pretrain our new models with large amounts of diverse data (400GB of text) from different Arabic varieties and domains. We evaluate JASMINE extensively in both intrinsic and extrinsic settings, using a comprehensive benchmark for zero- and few-shot learning across a wide range of NLP tasks. We also carefully develop and release a novel benchmark for both automated and human evaluation of Arabic autoregressive models focused at investigating potential social biases, harms, and toxicity in these models. We aim to responsibly release our models with interested researchers, along with code for experimenting with them
translated by 谷歌翻译
Pretrained language models have demonstrated extraordinary capabilities in language generation. However, real-world tasks often require controlling the distribution of generated text in order to mitigate bias, promote fairness, and achieve personalization. Existing techniques for controlling the distribution of generated text only work with quantified distributions, which require pre-defined categories, proportions of the distribution, or an existing corpus following the desired distributions. However, many important distributions, such as personal preferences, are unquantified. In this work, we tackle the problem of generating text following arbitrary distributions (quantified and unquantified) by proposing Nano, a few-shot human-in-the-loop training algorithm that continuously learns from human feedback. Nano achieves state-of-the-art results on single topic/attribute as well as quantified distribution control compared to previous works. We also show that Nano is able to learn unquantified distributions, achieves personalization, and captures differences between different individuals' personal preferences with high sample efficiency.
translated by 谷歌翻译