Human annotation plays a core role in machine learning — annotations for supervised models, safety guardrails for generative models, and human feedback for reinforcement learning, to cite a few avenues. However, the fact that many of these human annotations are inherently subjective is often overlooked. Recent work has demonstrated that ignoring rater subjectivity (typically resulting in rater disagreement) is problematic within specific tasks and for specific subgroups. Generalizable methods to harness rater disagreement and thus understand the socio-cultural leanings of subjective tasks remain elusive. In this paper, we propose GRASP, a comprehensive disagreement analysis framework to measure group association in perspectives among different rater subgroups, and demonstrate its utility in assessing the extent of systematic disagreements in two datasets: (1) safety annotations of human-chatbot conversations, and (2) offensiveness annotations of social media posts, both annotated by diverse rater pools across different socio-demographic axes. Our framework (based on disagreement metrics) reveals specific rater groups that have significantly different perspectives than others on certain tasks, and helps identify demographic axes that are crucial to consider in specific task contexts.
How people interpret content is deeply influenced by their socio-cultural backgrounds and lived experiences. This is especially crucial when evaluating AI systems for safety, where accounting for such diversity in interpretations and potential impacts on human users will make them both more successful and inclusive. While recent work has demonstrated the importance of diversity in human ratings that underlie AI pipelines, effective and efficient ways to incorporate diverse perspectives in human data annotation pipelines is still largely elusive. In this paper, we discuss the primary challenges faced in incorporating diversity into model evaluations, and propose a practical diversity-aware annotation approach. Using an existing dataset with highly parallel safety annotations, we take as a test case a policy that prioritizes recall of safety issues, and demonstrate that our diversity-aware approach can efficiently obtain a higher recall of safety issues flagged by minoritized rater groups without hurting overall precision.
Human experiences are complex and subjective. This subjectivity is reflected in the way people label images for machine vision models. While annotation tasks are often assumed to deliver objective results, this assumption does not allow for the subjectivity of human experience. This paper examines the implications of subjective human judgments in the behavioral task of labeling images used to train machine vision models. We identify three primary sources of ambiguity: (1) depictions of labels in the images can be simply ambiguous, (2) raters’ backgrounds and experiences can influence their judgments and (3) the way the labeling task is defined can also influence raters’ judgments. By taking steps to address these sources of ambiguity, we can create more robust and reliable machine vision models.
State-of-the-art conversational AI exhibits a level of sophistication that promises to have profound impacts on many aspects of daily life, including how people seek information, create content, and find emotional support. It has also shown a propensity for bias, offensive language, and false information. Consequently, understanding and moderating safety risks posed by interacting with AI chatbots is a critical technical and social challenge. Safety annotation is an intrinsically subjective task, where many factors—often intersecting—determine why people may express different opinions on whether a conversation is safe. We apply Bayesian multilevel models to surface factors that best predict rater behavior to a dataset of 101,286 annotations of conversations between humans and an AI chatbot, stratified by rater gender, age, race/ethnicity, and education level. We show that intersectional effects involving these factors play significant roles in validating safety in conversational AI data. For example, race/ethnicity and gender show strong intersectional effects, particularly among South Asian and East Asian women. We also find that conversational degree of harm impacts raters of all race/ethnicity groups, but that Indigenous and South Asian raters are particularly sensitive. Finally, we discover that the effect of education is uniquely intersectional for Indigenous raters. Our results underscore the utility of multilevel frameworks for uncovering underrepresented social perspectives.
Among the problems with leaderboard culture in NLP has been the widespread lack of confidence estimation in reported results. In this work, we present a framework and simulator for estimating p-values for comparisons between the results of two systems, in order to understand the confidence that one is actually better (i.e. ranked higher) than the other. What has made this difficult in the past is that each system must itself be evaluated by comparison to a gold standard. We define a null hypothesis that each system’s metric scores are drawn from the same distribution, using variance found naturally (though rarely reported) in test set items and individual labels on an item (responses) to produce the metric distributions. We create a test set that evenly mixes the responses of the two systems under the assumption the null hypothesis is true. Exploring how to best estimate the true p-value from a single test set under different metrics, tests, and sampling methods, we find that the presence of response variance (from multiple raters or multiple model versions) has a profound impact on p-value estimates for model comparison, and that choice of metric and sampling method is critical to providing statistical guarantees on model comparisons.
Adversarially testing large language models (LLMs) is crucial for their safe and responsible deployment in practice. We introduce an AI-assisted approach for automated generation of adversarial evaluation datasets to test the safety of LLM generations on new downstream applications. We call it AART AI-assisted Red-Teaming - an automated alternative to current manual red-teaming efforts. AART offers a data generation and augmentation pipeline of reusable and customizable recipes that reduce significantly human effort and enable integration of adversarial testing earlier in new product development. AART generates evaluation datasets with high diversity of content characteristics critical for effective adversarial testing (e.g. sensitive and harmful concepts, specific to a wide range of cultural and geographic regions and application scenarios). The data generation is steered by AI-assisted recipes to define, scope and prioritize diversity within a new application context. This feeds into a structured LLM-generation process that scales up evaluation priorities. This provides transparency of developers evaluation intentions and enables quick adaptation to new use cases and newly discovered model weaknesses. Compared to some of the state-of-the-art tools AART shows promising results in terms of concept coverage and data quality.
Large neural models have brought a new challenge to natural language generation (NLG): It has become imperative to ensure the safety and reliability of the output of models that generate freely. To this end, we present an evaluation framework, Attributable to Identified Sources (AIS), stipulating that NLG output pertaining to the external world is to be verified against an independent, provided source. We define AIS and a two-stage annotation pipeline for allowing annotators to evaluate model output according to annotation guidelines. We successfully validate this approach on generation datasets spanning three tasks (two conversational QA datasets, a summarization dataset, and a table-to-text dataset). We provide full annotation guidelines in the appendices and publicly release the annotated data at https://github.com/google-research-datasets/AIS.
Annotator disagreement is often dismissed as noise or the result of poor annotation process quality. Others have argued that it can be meaningful. But lacking a rigorous statistical foundation, the analysis of disagreement patterns can resemble a high-tech form of tea-leaf-reading. We contribute a framework for analyzing the variation of per-item annotator response distributions to data for humans-in-the-loop machine learning. We provide visualizations for, and use the framework to analyze the variance in, a crowdsourced dataset of hard-to-classify examples from the OpenImages archive.
When collecting annotations and labeled data from humans, a standard practice is to use inter-rater reliability (IRR) as a measure of data goodness (Hallgren, 2012). Metrics such as Krippendorff’s alpha or Cohen’s kappa are typically required to be above a threshold of 0.6 (Landis and Koch, 1977). These absolute thresholds are unreasonable for crowdsourced data from annotators with high cultural and training variances, especially on subjective topics. We present a new alternative to interpreting IRR that is more empirical and contextualized. It is based upon benchmarking IRR against baseline measures in a replication, one of which is a novel cross-replication reliability (xRR) measure based on Cohen’s (1960) kappa. We call this approach the xRR framework. We opensource a replication dataset of 4 million human judgements of facial expressions and analyze it with the proposed framework. We argue this framework can be used to measure the quality of crowdsourced datasets.
We present a resource for the task of FrameNet semantic frame disambiguation of over 5,000 word-sentence pairs from the Wikipedia corpus. The annotations were collected using a novel crowdsourcing approach with multiple workers per sentence to capture inter-annotator disagreement. In contrast to the typical approach of attributing the best single frame to each word, we provide a list of frames with disagreement-based scores that express the confidence with which each frame applies to the word. This is based on the idea that inter-annotator disagreement is at least partly caused by ambiguity that is inherent to the text and frames. We have found many examples where the semantics of individual frames overlap sufficiently to make them acceptable alternatives for interpreting a sentence. We have argued that ignoring this ambiguity creates an overly arbitrary target for training and evaluating natural language processing systems - if humans cannot agree, why would we expect the correct answer from a machine to be any different? To process this data we also utilized an expanded lemma-set provided by the Framester system, which merges FN with WordNet to enhance coverage. Our dataset includes annotations of 1,000 sentence-word pairs whose lemmas are not part of FN. Finally we present metrics for evaluating frame disambiguation systems that account for ambiguity.
Distant supervision is a popular method for performing relation extraction from text that is known to produce noisy labels. Most progress in relation extraction and classification has been made with crowdsourced corrections to distant-supervised labels, and there is evidence that indicates still more would be better. In this paper, we explore the problem of propagating human annotation signals gathered for open-domain relation classification through the CrowdTruth methodology for crowdsourcing, that captures ambiguity in annotations by measuring inter-annotator disagreement. Our approach propagates annotations to sentences that are similar in a low dimensional embedding space, expanding the number of labels by two orders of magnitude. Our experiments show significant improvement in a sentence-level multi-class relation classifier.
In this paper, automatic homophone- and homograph detection are suggested as new useful features for humor recognition systems. The system combines style-features from previous studies on humor recognition in short text with ambiguity-based features. The performance of two potentially useful homograph detection methods is evaluated using crowdsourced annotations as ground truth. Adding homophones and homographs as features to the classifier results in a small but significant improvement over the style-features alone. For the task of humor recognition, recall appears to be a more important quality measure than precision. Although the system was designed for humor recognition in oneliners, it also performs well at the classification of longer humorous texts.
This paper reports on a reimplementation of a system on detecting implicit positive meaning from negated statements. In the original regression experiment, different positive interpretations per negation are scored according to their likelihood. We convert the scores to classes and report our results on both the regression and classification tasks. We show that a baseline taking the mean score or most frequent class is hard to beat because of class imbalance in the dataset. Our error analysis indicates that an approach that takes the information structure into account (i.e. which information is new or contrastive) may be promising, which requires looking beyond the syntactic and semantic characteristics of negated statements.
This paper presents a framework and methodology for the annotation of perspectives in text. In the last decade, different aspects of linguistic encoding of perspectives have been targeted as separated phenomena through different annotation initiatives. We propose an annotation scheme that integrates these different phenomena. We use a multilayered annotation approach, splitting the annotation of different aspects of perspectives into small subsequent subtasks in order to reduce the complexity of the task and to better monitor interactions between layers. Currently, we have included four layers of perspective annotation: events, attribution, factuality and opinion. The annotations are integrated in a formal model called GRaSP, which provides the means to represent instances (e.g. events, entities) and propositions in the (real or assumed) world in relation to their mentions in text. Then, the relation between the source and target of a perspective is characterized by means of perspective annotations. This enables us to place alternative perspectives on the same entity, event or proposition next to each other.
This paper presents a collection of annotations (tags or keywords) for a set of 2,133 environmental sounds taken from the Freesound database (www.freesound.org). The annotations are acquired through an open-ended crowd-labeling task, in which participants were asked to provide keywords for each of three sounds. The main goal of this study is to find out (i) whether it is feasible to collect keywords for a large collection of sounds through crowdsourcing, and (ii) how people talk about sounds, and what information they can infer from hearing a sound in isolation. Our main finding is that it is not only feasible to perform crowd-labeling for a large collection of sounds, it is also very useful to highlight different aspects of the sounds that authors may fail to mention. Our data is freely available, and can be used to ground semantic models, improve search in audio databases, and to study the language of sound.
The increasing streams of information pose challenges to both humans and machines. On the one hand, humans need to identify relevant information and consume only the information that lies at their interests. On the other hand, machines need to understand the information that is published in online data streams and generate concise and meaningful overviews. We consider events as prime factors to query for information and generate meaningful context. The focus of this paper is to acquire empirical insights for identifying salience features in tweets and news about a target event, i.e., the event of “whaling”. We first derive a methodology to identify such features by building up a knowledge space of the event enriched with relevant phrases, sentiments and ranked by their novelty. We applied this methodology on tweets and we have performed preliminary work towards adapting it to news articles. Our results show that crowdsourcing text relevance, sentiments and novelty (1) can be a main step in identifying salient information, and (2) provides a deeper and more precise understanding of the data at hand compared to state-of-the-art approaches.