Crowdsourcing is a common approach to rapidly annotate large volumes of data in machine learning applications. Typically, crowd workers are compensated with a flat rate based on an estimated completion time to meet a target hourly wage. Unfortunately, prior work has shown that variability in completion times among crowd workers led to overpayment by 168 in another. However, by setting a time limit for task completion, it is possible to manage the risk of overpaying or underpaying while still facilitating flat rate payments. In this paper, we present an analysis of the impact of a time limit on crowd worker performance and satisfaction. We conducted a human study with a maximum view time for a crowdsourced image classification task. We find that the impact on overall crowd worker performance diminishes as view time increases. Despite some images being challenging under time limits, a consensus algorithm remains effective at preserving data quality and filters images needing more time. Additionally, crowd workers' consistent performance throughout the time-limited task indicates sustained effort, and their psychometric questionnaire scores show they prefer shorter limits. Based on our findings, we recommend implementing task time limits as a practical approach to making compensation more equitable and predictable.

RVL-CDIP is a popular dataset for benchmarking document classifiers. However, the dataset contains ample amounts of label errors as well as non-trivial amounts of test-train overlap, both of which may impact model performance metrics. In this paper, we address these two problems by (1) finding and fixing label errors, and (2) detecting and addressing test-train overlap. We produce several variations of RVL-CDIP with label error and test-train overlap fixes, and benchmark document classification performance on these new RVL-CDIP variations. Our rigorous analysis of RVL-CDIP finds that the corpus contains 12% label error and approximately 35

Crowdsourcing is a vital tool for rapid data annotation, yet flat-rate compensation often results in significant pay inequity due to worker speed variability. This paper investigates using task time limits to stabilize pay rates while maintaining data quality. Through a human study on an image classification task, we found that worker performance diminishes only slightly as view time decreases, and consensus algorithms remain effective at filtering complex images to preserve overall accuracy. Quantitatively, participants maintained consistent effort throughout the study and showed a psychometric trend favoring shorter time limits. These findings suggest that implementing task time limits is a practical approach to achieving more equitable compensation, mitigating the risks of overpayment and underpayment by creating a more predictable hourly rate. Our code and data are available at https://github.com/gordon-lim/sdogs-10h.

Rapid document classification is critical in several time-sensitive applications like digital forensics and large-scale media classification. Traditional approaches that rely on heavy-duty deep learning models fall short due to high inference times over vast input datasets and computational resources associated with analyzing whole documents. In this paper, we present a method using lightweight supervised learning models, combined with a TF-IDF feature extraction-based tokenization method, to accurately and efficiently classify documents based solely on their file names, which substantially reduces inference time. Experiments on two datasets introduced in this paper show that our file name classifiers correctly predict more than 90% of in-scope documents with 99.63% and 96.57% accuracy while being 442x faster than more complex models such as DiT. Our results demonstrate that incorporating lightweight file name classification as a front-end to document analysis pipelines can efficiently process vast document datasets in critical scenarios, enabling fast and more reliable document classification.

RVL-CDIP and Tobacco3482 are commonly used document classification benchmarks, but recent work on explainability has revealed that ID codes stamped on the documents in these datasets may be used by machine learning models to learn shortcuts on the classification task. In this paper, we present an in-depth investigation into the influence and impact of these ID codes on model performance. We annotate ID codes in documents from RVL-CDIP and Tobacco3482 and find that shallow learning models can achieve classification accuracy scores of roughly 40% on RVL-CDIP and 60% on Tobacco3482 using only features derived from the ID codes. We also find that a state-of-the-art document classifier sees a performance drop of 11 accuracy points on RVL-CDIP when ID codes are removed from the data. Finally, we train an ID code detection model in order to remove ID codes from RVL-CDIP and Tobacco3482 and make this data publicly available.

In deep learning (DL) systems, label noise in training datasets often degrades model performance, as models may learn incorrect patterns from mislabeled data. The area of Learning with Noisy Labels (LNL) has introduced methods to effectively train DL models in the presence of noisily-labeled datasets. Traditionally, these methods are tested using synthetic label noise, where ground truth labels are randomly (and automatically) flipped. However, recent findings highlight that models perform substantially worse under human label noise than synthetic label noise, indicating a need for more realistic test scenarios that reflect noise introduced due to imperfect human labeling. This underscores the need for generating realistic noisy labels that simulate human label noise, enabling rigorous testing of deep neural networks without the need to collect new human-labeled datasets. To address this gap, we present Cluster-Based Noise (CBN), a method for generating feature-dependent noise that simulates human-like label noise. Using insights from our case study of label memorization in the CIFAR-10N dataset, we design CBN to create more realistic tests for evaluating LNL methods. Our experiments demonstrate that current LNL methods perform worse when tested using CBN, highlighting its use as a rigorous approach to testing neural networks. Next, we propose Soft Neighbor Label Sampling (SNLS), a method designed to handle CBN, demonstrating its improvement over existing techniques in tackling this more challenging type of noise.

The IIT-CDIP document collection is the source of several widely used and publicly accessible document understanding datasets. In this paper, manual inspection of 5 datasets derived from IIT-CDIP uncovers the presence of thousands of instances of sensitive personal data, including US Social Security Numbers (SSNs), birth places and dates, and home addresses of individuals. The presence of such sensitive personal data in commonly-used and publicly available datasets is startling and has ethical and potentially legal implications; we believe such sensitive data ought to be removed from the internet. Thus, in this paper, we develop a modular data de-identification pipeline that replaces sensitive data with synthetic, but realistic, data. Via experiments, we demonstrate that this de-identification method preserves the utility of the de-identified documents so that they can continue be used in various document understanding applications. We will release redacted versions of these datasets publicly.

Tobacco3482 is a widely used document classification benchmark dataset. However, our manual inspection of the entire dataset uncovers widespread ontological issues, especially large amounts of annotation label problems in the dataset. We establish data label guidelines and find that 11.7% of the dataset is improperly annotated and should either have an unknown label or a corrected label, and 16.7% of samples in the dataset have multiple valid labels. We then analyze the mistakes of a top-performing model and find that 35% of the model's mistakes can be directly attributed to these label issues, highlighting the inherent problems with using a noisily labeled dataset as a benchmark. Supplementary material, including dataset annotations and code, is available at https://github.com/gordon-lim/tobacco3482-mistakes/.

Intent classifiers must be able to distinguish when a user's utterance does not belong to any supported intent to avoid producing incorrect and unrelated system responses. Although out-of-scope (OOS) detection for intent classifiers has been studied, previous work has not yet studied changes in classifier performance against hard-negative out-of-scope utterances (i.e., inputs that share common features with in-scope data, but are actually out-of-scope). We present an automated technique to generate hard-negative OOS data using ChatGPT. We use our technique to build five new hard-negative OOS datasets, and evaluate each against three benchmark intent classifiers. We show that classifiers struggle to correctly identify hard-negative OOS utterances more than general OOS utterances. Finally, we show that incorporating hard-negative OOS data for training improves model robustness when detecting hard-negative OOS data and general OOS data. Our technique, datasets, and evaluation address an important void in the field, offering a straightforward and inexpensive way to collect hard-negative OOS data and improve intent classifiers' robustness.

The RVL-CDIP benchmark is widely used for measuring performance on the task of document classification. Despite its widespread use, we reveal several undesirable characteristics of the RVL-CDIP benchmark. These include (1) substantial amounts of label noise, which we estimate to be 8.1% (ranging between 1.6% to 16.9% per document category); (2) presence of many ambiguous or multi-label documents; (3) a large overlap between test and train splits, which can inflate model performance metrics; and (4) presence of sensitive personally-identifiable information like US Social Security numbers (SSNs). We argue that there is a risk in using RVL-CDIP for benchmarking document classifiers, as its limited scope, presence of errors (state-of-the-art models now achieve accuracy error rates that are within our estimated label error rate), and lack of diversity make it less than ideal for benchmarking. We further advocate for the creation of a new document classification benchmark, and provide recommendations for what characteristics such a resource should include.

Document denoising and binarization are fundamental problems in the document processing space, but current datasets are often too small and lack sufficient complexity to effectively train and benchmark modern data-driven machine learning models. To fill this gap, we introduce ShabbyPages, a new document image dataset designed for training and benchmarking document denoisers and binarizers. ShabbyPages contains over 6,000 clean "born digital" images with synthetically-noised counterparts ("shabby pages") that were augmented using the Augraphy document augmentation tool to appear as if they have been printed and faxed, photocopied, or otherwise altered through physical processes. In this paper, we discuss the creation process of ShabbyPages and demonstrate the utility of ShabbyPages by training convolutional denoisers which remove real noise features with a high degree of human-perceptible fidelity, establishing baseline performance for a new ShabbyPages benchmark.

This paper introduces Augraphy, a Python library for constructing data augmentation pipelines which produce distortions commonly seen in real-world document image datasets. Augraphy stands apart from other data augmentation tools by providing many different strategies to produce augmented versions of clean document images that appear as if they have been altered by standard office operations, such as printing, scanning, and faxing through old or dirty machines, degradation of ink over time, and handwritten markings. This paper discusses the Augraphy tool, and shows how it can be used both as a data augmentation tool for producing diverse training data for tasks such as document denoising, and also for generating challenging test data to evaluate model robustness on document image modeling tasks.

The RVL-CDIP benchmark is widely used for measuring performance on the task of document classification. Despite its widespread use, we reveal several undesirable characteristics of the RVL-CDIP benchmark. These include (1) substantial amounts of label noise, which we estimate to be 8.1% (ranging between 1.6% to 16.9% per document category); (2) presence of many ambiguous or multi-label documents; (3) a large overlap between test and train splits, which can inflate model performance metrics; and (4) presence of sensitive personally-identifiable information like US Social Security numbers (SSNs). We argue that there is a risk in using RVL-CDIP for benchmarking document classifiers, as its limited scope, presence of errors (state-of-the-art models now achieve accuracy error rates that are within our estimated label error rate), and lack of diversity make it less than ideal for benchmarking. We further advocate for the creation of a new document classification benchmark, and provide recommendations for what characteristics such a resource should include.

The ability of a document classifier to handle inputs that are drawn from a distribution different from the training distribution is crucial for robust deployment and generalizability. The RVL-CDIP corpus is the de facto standard benchmark for document classification, yet to our knowledge all studies that use this corpus do not include evaluation on out-of-distribution documents. In this paper, we curate and release a new out-of-distribution benchmark for evaluating out-of-distribution performance for document classifiers. Our new out-of-distribution benchmark consists of two types of documents: those that are not part of any of the 16 in-domain RVL-CDIP categories (RVL-CDIP-O), and those that are one of the 16 in-domain categories yet are drawn from a distribution different from that of the original RVL-CDIP dataset (RVL-CDIP-N). While prior work on document classification for in-domain RVL-CDIP documents reports high accuracy scores, we find that these models exhibit accuracy drops of between roughly 15-30% on our new out-of-domain RVL-CDIP-N benchmark, and further struggle to distinguish between in-domain RVL-CDIP-N and out-of-domain RVL-CDIP-O inputs. Our new benchmark provides researchers with a valuable new resource for analyzing out-of-distribution performance on document classifiers. Our new out-of-distribution data can be found at https://github.com/gxlarson/rvl-cdip-ood.

Dialog systems must be capable of incorporating new skills via updates over time in order to reflect new use cases or deployment scenarios. Similarly, developers of such ML-driven systems need to be able to add new training data to an already-existing dataset to support these new skills. In intent classification systems, problems can arise if training data for a new skill's intent overlaps semantically with an already-existing intent. We call such cases collisions. This paper introduces the task of intent collision detection between multiple datasets for the purposes of growing a system's skillset. We introduce several methods for detecting collisions, and evaluate our methods on real datasets that exhibit collisions. To highlight the need for intent collision detection, we show that model performance suffers if new data is added in such a way that does not arbitrate colliding intents. Finally, we use collision detection to construct and benchmark a new dataset, Redwood, which is composed of 451 ntent categories from 13 original intent classification datasets, making it the largest publicly available intent classification benchmark.

Neural clone detection has attracted the attention of software engineering researchers and practitioners. However, most neural clone detection methods do not generalize beyond the scope of clones that appear in the training dataset. This results in poor model performance, especially in terms of model recall. In this paper, we present an Abstract Syntax Tree (AST) assisted approach for generalizable neural clone detection, or ASTRO, a framework for finding clones in codebases reflecting industry practices. We present three main components: (1) an AST-inspired representation for source code that leverages program structure and semantics, (2) a global graph representation that captures the context of an AST among a corpus of programs, and (3) a graph embedding for programs that, in combination with extant large-scale language models, improves state-of-the-art code clone detection. Our experimental results show that ASTRO improves state-of-the-art neural clone detection approaches in both recall and F-1 scores.

Interest in dialog systems has grown substantially in the past decade. By extension, so too has interest in developing and improving intent classification and slot-filling models, which are two components that are commonly used in task-oriented dialog systems. Moreover, good evaluation benchmarks are important in helping to compare and analyze systems that incorporate such models. Unfortunately, much of the literature in the field is limited to analysis of relatively few benchmark datasets. In an effort to promote more robust analyses of task-oriented dialog systems, we have conducted a survey of publicly available datasets for the tasks of intent classification and slot-filling. We catalog the important characteristics of each dataset, and offer discussion on the applicability, strengths, and weaknesses of each. Our goal is that this survey aids in increasing the accessibility of these datasets, which we hope will enable their use in future evaluations of intent classification and slot-filling models for task-oriented dialog systems.

To be robust enough for widespread adoption, document analysis systems involving machine learning models must be able to respond correctly to inputs that fall outside of the data distribution that was used to generate the data on which the models were trained. This paper explores the ability of text classifiers trained on standard document classification datasets to generalize to out-of-distribution documents at inference time. We take the Tobacco-3482 and RVL-CDIP datasets as a starting point and generate new out-of-distribution evaluation datasets in order to analyze the generalization performance of models trained on these standard datasets. We find that models trained on the smaller Tobacco-3482 dataset perform poorly on our new out-ofdistribution data, while text classification models trained on the larger RVL-CDIP exhibit smaller performance drops.

Open Information Extraction (OIE) systems seek to compress the factual propositions of a sentence into a series of n-ary tuples. These tuples are useful for downstream tasks in natural language processing like knowledge base creation, textual entailment, and natural language understanding. However, current OIE datasets are limited in both size and diversity. We introduce a new dataset by converting the QA-SRL 2.0 dataset to a large-scale OIE dataset LSOIE. Our LSOIE dataset is 20 times larger than the next largest human-annotated OIE dataset. We construct and evaluate several benchmark OIE models on LSOIE, providing baselines for future improvements on the task. Our LSOIE data, models, and code are made publicly available.