Neural language models have achieved human level performance across several NLP datasets. However, recent studies have shown that these models are not truly learning the desired task; rather, their high performance is attributed to overfitting using spurious biases, which suggests that the capabilities of AI systems have been over-estimated. We introduce a generic formula for Data Quality Index (DQI) to help dataset creators create datasets free of such unwanted biases. We evaluate this formula using a recently proposed approach for adversarial filtering, AFLite. We propose a new data creation paradigm using DQI to create higher quality data. The data creation paradigm consists of several data visualizations to help data creators (i) understand the quality of data and (ii) visualize the impact of the created data instance on the overall quality. It also has a couple of automation methods to (i) assist data creators and (ii) make the model more robust to adversarial attacks. We use DQI along with these automation methods to renovate biased examples in SNLI. We show that models trained on the renovated SNLI dataset generalize better to out of distribution tasks. Renovation results in reduced model performance, exposing a large gap with respect to human performance. DQI systematically helps in creating harder benchmarks using active learning. Our work takes the process of dynamic dataset creation forward, wherein datasets evolve together with the evolving state of the art, therefore serving as a means of benchmarking the true progress of AI.
Honeyfile deployment is a useful breach detection method in cyber deception that can also inform defenders about the intent and interests of intruders and malicious insiders. A key property of a honeyfile, enticement, is the extent to which the file can attract an intruder to interact with it. We introduce a novel metric, Topic Semantic Matching (TSM), which uses topic modelling to represent files in the repository and semantic matching in an embedding vector space to compare honeyfile text and topic words robustly. We also present a honeyfile corpus created with different Natural Language Pro…
While advances in pre-training have led to dramatic improvements in few-shot learning of NLP tasks, there is limited understanding of what drives successful few-shot adaptation in datasets. In particular, given a new dataset and a pre-trained model, what properties of the dataset make it \emph{few-shot learnable} and are these properties independent of the specific adaptation techniques used? We consider an extensive set of recent few-shot learning methods, and show that their performance across a large number of datasets is highly correlated, showing that few-shot hardness may be intrinsic to…
Amazon AWS AI（亚马逊AWS人工智能）；University of Waterloo（滑铁卢大学）
Behavior of deep neural networks can be inconsistent between different versions. Regressions during model update are a common cause of concern that often over-weigh the benefits in accuracy or efficiency gain. This work focuses on quantifying, reducing and analyzing regression errors in the NLP model updates. Using negative flip rate as regression measure, we show that regression has a prevalent presence across tasks in the GLUE benchmark. We formulate the regression-free model updates into a constrained optimization problem, and further reduce it into a relaxed form which can be approximately…
Queen Mary University of London（伦敦玛丽女王大学）；UC Berkeley（加州大学伯克利分校）；Jožef Stefan Institute（约瑟夫·斯特凡研究所）；Nanning Central Sub-branch of the People’s Bank of China（中国人民银行南宁中心支行）
In text-to-SQL tasks -- as in much of NLP -- compositional generalization is a major challenge: neural networks struggle with compositional generalization where training and test distributions differ. However, most recent attempts to improve this are based on word-level synthetic data or specific dataset splits to generate compositional biases. In this work, we propose a clause-level compositional example generation method. We first split the sentences in the Spider text-to-SQL dataset into sub-sentences, annotating each sub-sentence with its corresponding SQL clause, resulting in a new datase…
AtlasNLP: A Country-Aware Atlas of Dataset Representation in NLP
University of Michigan（密歇根大学）；Addis Ababa University（亚的斯亚贝巴大学）；San Jose State University（圣何塞州立大学）；Skyline High School（斯凯莱恩高中）；University of Pennsylvania（宾夕法尼亚大学）；Carnegie Mellon University（卡内基梅隆大学）
Pretrained language models, especially masked language models (MLMs) have seen success across many NLP tasks. However, there is ample evidence that they use the cultural biases that are undoubtedly present in the corpora they are trained on, implicitly creating harm with biased representations. To measure some forms of social bias in language models against protected demographic groups in the US, we introduce the Crowdsourced Stereotype Pairs benchmark (CrowS-Pairs). CrowS-Pairs has 1508 examples that cover stereotypes dealing with nine types of bias, like race, religion, and age. In CrowS-Pai…
GlobalBench: A Benchmark for Global Progress in Natural Language Processing
Carnegie Mellon University（卡内基梅隆大学）；Harvard University（哈佛大学）；Shanghai Jiaotong University（上海交通大学）；George Mason University（乔治梅森大学）；The Hong Kong University of Science and Technology（香港科技大学）；MBZUAI（穆罕默德·本·扎耶德人工智能大学）；University of Washington（华盛顿大学）
We propose a benchmark to measure whether a language model is truthful in generating answers to questions. The benchmark comprises 817 questions that span 38 categories, including health, law, finance and politics. We crafted questions that some humans would answer falsely due to a false belief or misconception. To perform well, models must avoid generating false answers learned from imitating human texts. We tested GPT-3, GPT-Neo/J, GPT-2 and a T5-based model. The best model was truthful on 58% of questions, while human performance was 94%. Models generated many false answers that mimic popul…
The Chinese University of Hong Kong（香港中文大学）；Amazon AWS AI Labs（亚马逊AWS人工智能实验室）；University of Cambridge（剑桥大学）
Recent advance in deep learning has led to the rapid adoption of machine learning-based NLP models in a wide range of applications. Despite the continuous gain in accuracy, backward compatibility is also an important aspect for industrial applications, yet it received little research attention. Backward compatibility requires that the new model does not regress on cases that were correctly handled by its predecessor. This work studies model update regression in structured prediction tasks. We choose syntactic dependency parsing and conversational semantic parsing as representative examples of …
INRIA（法国国家信息与自动化研究所）；The Systems Biology Institute（系统生物学研究所）
This paper evaluates the difference between human pathway curation and current NLP systems. We propose graph analysis methods for quantifying the gap between human curated pathway maps and the output of state-of-the-art automatic NLP systems. Evaluation is performed on the popular mTOR pathway. Based on analyzing where current systems perform well and where they fail, we identify possible avenues for progress.
University of Colorado Colorado Springs（科罗拉多大学科罗拉多斯普林斯分校）
Artificial Neural networks are mathematical models at their core. This truismpresents some fundamental difficulty when networks are tasked with Natural Language Processing. A key problem lies in measuring the similarity or distance among vectors in NLP embedding space, since the mathematical concept of distance does not always agree with the linguistic concept. We suggest that the best way to measure linguistic distance among vectors is by employing the Language Model (LM) that created them. We introduce Language Model Distance (LMD) for measuring accuracy of vector transformations based on th…
Facebook AI Research（脸书人工智能研究院）；University of Southern California（南加州大学）
Real-world natural language processing (NLP) models need to be continually updated to fix the prediction errors in out-of-distribution (OOD) data streams while overcoming catastrophic forgetting. However, existing continual learning (CL) problem setups cannot cover such a realistic and complex scenario. In response to this, we propose a new CL problem formulation dubbed continual model refinement (CMR). Compared to prior CL settings, CMR is more practical and introduces unique challenges (boundary-agnostic and non-stationary distribution shift, diverse mixtures of multiple OOD data clusters, e…
As language models grow in popularity, it becomes increasingly important to clearly measure all possible markers of demographic identity in order to avoid perpetuating existing societal harms. Many datasets for measuring bias currently exist, but they are restricted in their coverage of demographic axes and are commonly used with preset bias tests that presuppose which types of biases models can exhibit. In this work, we present a new, more inclusive bias measurement dataset, HolisticBias, which includes nearly 600 descriptor terms across 13 different demographic axes. HolisticBias was assembl…
Bar-Ilan University（巴伊兰大学）；Allen Institute for Artificial Intelligence（艾伦人工智能研究所）；LMU Munich（慕尼黑大学）；Technion（以色列理工学院）；Carnegie Mellon University（卡内基梅隆大学）；Saarland University（萨尔大学）
Large amounts of training data are one of the major reasons for the high performance of state-of-the-art NLP models. But what exactly in the training data causes a model to make a certain prediction? We seek to answer this question by providing a language for describing how training data influences predictions, through a causal framework. Importantly, our framework bypasses the need to retrain expensive models and allows us to estimate causal effects based on observational data alone. Addressing the problem of extracting factual knowledge from pretrained language models (PLMs), we focus on sim…
The University of Sheffield（谢菲尔德大学）；Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；Beijing Academy of Artificial Intelligence（北京人工智能研究院）；University of Michigan Ann Arbor（密歇根大学安娜堡分校）
Fairness has become a trending topic in natural language processing (NLP), which addresses biases targeting certain social groups such as genders and religions. However, regional bias in language models (LMs), a long-standing global discrimination problem, still remains unexplored. This paper bridges the gap by analysing the regional bias learned by the pre-trained language models that are broadly used in NLP tasks. In addition to verifying the existence of regional bias in LMs, we find that the biases on regional groups can be strongly influenced by the geographical clustering of the groups. …
German Research Center for Artificial Intelligence (DFKI)（德国人工智能研究中心（DFKI））
Recently, state-of-the-art NLP models gained an increasing syntactic and semantic understanding of language, and explanation methods are crucial to understand their decisions. Occlusion is a well established method that provides explanations on discrete language data, e.g. by removing a language unit from an input and measuring the impact on a model's decision. We argue that current occlusion-based methods often produce invalid or syntactically incorrect language data, neglecting the improved abilities of recent NLP models. Furthermore, gradient-based explanation methods disregard the discrete…
Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case …
Institute of Artificial Intelligence（人工智能研究所）；Center for Medical Statistics, Informatics, and Intelligent Systems（医学统计、信息学与智能系统中心）；Medical University of Vienna（维也纳医科大学）
Measuring the performance of natural language processing models is challenging. Traditionally used metrics, such as BLEU and ROUGE, originally devised for machine translation and summarization, have been shown to suffer from low correlation with human judgment and a lack of transferability to other tasks and languages. In the past 15 years, a wide range of alternative metrics have been proposed. However, it is unclear to what extent this has had an impact on NLP benchmarking efforts. Here we provide the first large-scale cross-sectional analysis of metrics used for measuring performance in nat…
In this paper, a method for measuring synchronic corpus (dis-)similarity put forward by Kilgarriff (2001) is adapted and extended to identify trends and correlated changes in diachronic text data, using the Corpus of Historical American English (Davies 2010a) and the Google Ngram Corpora (Michel et al. 2010a). This paper shows that this fully data-driven method, which extracts word types that have undergone the most pronounced change in frequency in a given period of time, is computationally very cheap and that it allows interpretations of diachronic trends that are both intuitively plausible …
Reproducibility in NLP: What Have We Learned from the Checklist?
Allen Institute for Artificial Intelligence（艾伦人工智能研究所）；Paul G. Allen School of Computer Science & Engineering, University of Washington（华盛顿大学保罗·G·艾伦计算机科学与工程学院）
University of Groningen（格罗宁根大学）；University of Malta（马耳他大学）
Contextualized word embeddings have been replacing standard embeddings as the representational knowledge source of choice in NLP systems. Since a variety of biases have previously been found in standard word embeddings, it is crucial to assess biases encoded in their replacements as well. Focusing on BERT (Devlin et al., 2018), we measure gender bias by studying associations between gender-denoting target words and names of professions in English and German, comparing the findings with real-world workforce statistics. We mitigate bias by fine-tuning BERT on the GAP corpus (Webster et al., 2018…
Multiple studies have probed representations emerging in neural networks trained for end-to-end NLP tasks and examined what word-level linguistic information may be encoded in the representations. In classical probing, a classifier is trained on the representations to extract the target linguistic information. However, there is a threat of the classifier simply memorizing the linguistic labels for individual words, instead of extracting the linguistic abstractions from the representations, thus reporting false positive results. While considerable efforts have been made to minimize the memoriza…
What Are We Measuring in NLG? A Meta-Analysis of Evaluation Trends 2020-2025
Technische Universität Berlin（技术大学柏林）；BIFOLD – Berlin Institute for the Foundations of Learning and Data（柏林学习与数据基础研究所）；German Research Center for Artificial Intelligence (DFKI), Berlin（德国人工智能研究中心（DFKI），柏林）；ANITI；University of Potsdam（波茨坦大学）；CERTAIN
Many modern Artificial Intelligence (AI) systems make use of data embeddings, particularly in the domain of Natural Language Processing (NLP). These embeddings are learnt from data that has been gathered "from the wild" and have been found to contain unwanted biases. In this paper we make three contributions towards measuring, understanding and removing this problem. We present a rigorous way to measure some of these biases, based on the use of word lists created for social psychology applications; we observe how gender bias in occupations reflects actual gender bias in the same occupations in…
School of Electrical and Computer Engineering（电气与计算机工程学院）；College of Engineering（工程学院）；University of Tehran（德黑兰大学）
Human languages are full of metaphorical expressions. Metaphors help people understand the world by connecting new concepts and domains to more familiar ones. Large pre-trained language models (PLMs) are therefore assumed to encode metaphorical knowledge useful for NLP systems. In this paper, we investigate this hypothesis for PLMs, by probing metaphoricity information in their encodings, and by measuring the cross-lingual and cross-dataset generalization of this information. We present studies in multiple metaphor detection datasets and in four languages (i.e., English, Spanish, Russian, and …
The recent success of distributed word representations has led to an increased interest in analyzing the properties of their spatial distribution. Several studies have suggested that contextualized word embedding models do not isotropically project tokens into vector space. However, current methods designed to measure isotropy, such as average random cosine similarity and the partition score, have not been thoroughly analyzed and are not appropriate for measuring isotropy. We propose IsoScore: a novel tool that quantifies the degree to which a point cloud uniformly utilizes the ambient vector …
People Make Better Edits: Measuring the Efficacy of LLM-Generated Counterfactually Augmented Data for Harmful Language Detection
RWTH Aachen University（亚琛工业大学）；University of Konstanz（康斯坦茨大学）；GESIS - Leibniz Institute for Social Sciences（GESIS-莱布尼茨社会科学研究所）；Sapienza University of Rome（罗马智慧大学）；University of Copenhagen（哥本哈根大学）
Measuring and Improving Attentiveness to Partial Inputs with Counterfactuals
Allen Institute for AI（艾伦人工智能研究所）；University of Washington（华盛顿大学）；University of Illinois Urbana-Champaign（伊利诺伊大学厄巴纳-香槟分校）；University of Utah（犹他大学）；University of California, Irvine（加利福尼亚大学欧文分校）
School of Computer Science（计算机科学学院）；Engineering, Sun Yat-sen University（工程，中山大学）；School of Software Engineering（软件工程学院）；Sun Yat-sen University（中山大学）；Zhuhai Key Laboratory of Trusted Large Language Models（珠海可信大语言模型实验室）；Tencent AI Lab（腾讯AI实验室）；National University of Singapore（新加坡国立大学）；Shenzhen University（深圳大学）；Shanghai Jiao Tong University（上海交通大学）
Most NLP datasets are not annotated with protected attributes such as gender, making it difficult to measure classification bias using standard measures of fairness (e.g., equal opportunity). However, manually annotating a large dataset with a protected attribute is slow and expensive. Instead of annotating all the examples, can we annotate a subset of them and use that sample to estimate the bias? While it is possible to do so, the smaller this annotated sample is, the less certain we are that the estimate is close to the true bias. In this work, we propose using Bernstein bounds to represent…
McBE: A Multi-task Chinese Bias Evaluation Benchmark for Large Language Models
College of Computer Science, Inner Mongolia University（内蒙古大学计算机科学学院）；National & Local Joint Engineering Research Center of Intelligent Information Processing Technology for Mongolian（蒙古语智能信息处理技术国家与地方联合工程研究中心）；Inner Mongolia Key Laboratory of Multilingual Artificial Intelligence Technology（内蒙古多语种人工智能技术重点实验室）
Exploring the Use of Large Language Models for Reference-Free Text Quality Evaluation: An Empirical Study
Harbin Institute of Technology, Shenzhen（哈尔滨工业大学（深圳））；Peng Cheng Laboratory（鹏城实验室）；Guangdong Provincial Key Laboratory of Novel Security Intelligence Technologies（广东省新型安全智能技术重点实验室）
