An official website of the United States government
Recent advances in machine learning have led to increased deployment of black-box classifiers across a wide variety of applications. In many such situations there is a critical need to both reliably assess the performance of these pre-trained models and to perform this assessment in a label-efficient manner (given that labels may be scarce and costly to collect). In this paper, we introduce an active Bayesian approach for assessment of classifier performance to satisfy the desiderata of both reliability and label-efficiency. We begin by developing inference strategies to quantify uncertainty for common assessment metrics such as accuracy, misclassification cost, and calibration error. We then propose a general framework for active Bayesian assessment using inferred uncertainty to guide efficient selection of instances for labeling, enabling better performance assessment with fewer labels. We demonstrate significant gains from our proposed active Bayesian approach via a series of systematic empirical experiments assessing the performance of modern neural classifiers (e.g., ResNet and BERT) on several standard image and text classification datasets.
more »
« less
One Rating to Rule Them All?: Evidence of Multidimensionality in Human Assessment of Topic Labeling Quality
Two general approaches are common for evaluating automatically generated labels in topic modeling: direct human assessment; or performance metrics that can be calculated without, but still correlate with, human assessment. However, both approaches implicitly assume that the quality of a topic label is single-dimensional. In contrast, this paper provides evidence that human assessments about the quality of topic labels consist of multiple latent dimensions. This evidence comes from human assessments of four simple labeling techniques. For each label, study participants responded to several items asking them to assess each label according to a variety of different criteria. Exploratory factor analysis shows that these human assessments of labeling quality have a two-factor latent structure. Subsequent analysis demonstrates that this multi-item, two-factor assessment can reveal nuances that would be missed using either a single-item human assessment of perceived label quality or established performance metrics. The paper concludes by sug- gesting future directions for the development of human-centered approaches to evaluating NLP and ML systems more broadly.
more »
« less
- Award ID(s):
- 1757787
- PAR ID:
- 10386523
- Date Published:
- Journal Name:
- Proceedings of the 31st ACM International Conference on Information and Knowledge Management
- Page Range / eLocation ID:
- 768 to 779
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Activity recognition is central to many motion analysis applications ranging from health assessment to gaming. However, the need for obtaining sufficiently large amounts of labeled data has limited the development of personalized activity recognition models. Semi-supervised learning has traditionally been a promising approach in many application domains to alleviate reliance on large amounts of labeled data by learning the label information from a small set of seed labels. Nonetheless, existing approaches perform poorly in highly dynamic settings, such as wearable systems, because some algorithms rely on predefined hyper-parameters or distribution models that needs to be tuned for each user or context. To address these challenges, we introduce LabelForest 1, a novel non-parametric semi-supervised learning framework for activity recognition. LabelForest has two algorithms at its core: (1) a spanning forest algorithm for sample selection and label inference; and (2) a silhouette-based filtering method to finalize label augmentation for machine learning model training. Our thorough analysis on three human activity datasets demonstrate that LabelForest achieves a labeling accuracy of 90.1% in presence of a skewed label distribution in the seed data. Compared to self-training and other sequential learning algorithms, LabelForest achieves up to 56.9% and 175.3% improvement in the accuracy on balanced and unbalanced seed data, respectively.more » « less
-
Abstract As use of artificial intelligence (AI) has increased, concerns about AI bias and discrimination have been growing. This paper discusses an application called PyrEval in which natural language processing (NLP) was used to automate assessment and provide feedback on middle school science writing without linguistic discrimination. Linguistic discrimination in this study was operationalized as unfair assessment of scientific essays based on writing features that are not considered normative such as subject‐verb disagreement. Such unfair assessment is especially problematic when the purpose of assessment is not assessing English writing but rather assessing the content of scientific explanations. PyrEval was implemented in middle school science classrooms. Students explained their roller coaster design by stating relationships among such science concepts as potential energy, kinetic energy and law of conservation of energy. Initial and revised versions of scientific essays written by 307 eighth‐grade students were analyzed. Our manual and NLP assessment comparison analysis showed that PyrEval did not penalize student essays that contained non‐normative writing features. Repeated measures ANOVAs and GLMM analysis results revealed that essay quality significantly improved from initial to revised essays after receiving the NLP feedback, regardless of non‐normative writing features. Findings and implications are discussed. Practitioner notesWhat is already known about this topicAdvancement in AI has created a variety of opportunities in education, including automated assessment, but AI is not bias‐free.Automated writing assessment designed to improve students' scientific explanations has been studied.While limited, some studies reported biased performance of automated writing assessment tools, but without looking into actual linguistic features about which the tools may have discriminated.What this paper addsThis study conducted an actual examination of non‐normative linguistic features in essays written by middle school students to uncover how our NLP tool called PyrEval worked to assess them.PyrEval did not penalize essays containing non‐normative linguistic features.Regardless of non‐normative linguistic features, students' essay quality scores significantly improved from initial to revised essays after receiving feedback from PyrEval. Essay quality improvement was observed regardless of students' prior knowledge, school district and teacher variables.Implications for practice and/or policyThis paper inspires practitioners to attend to linguistic discrimination (re)produced by AI.This paper offers possibilities of using PyrEval as a reflection tool, to which human assessors compare their assessment and discover implicit bias against non‐normative linguistic features.PyrEval is available for use ongithub.com/psunlpgroup/PyrEvalv2.more » « less
-
A major bottleneck preventing the extension of deep learning systems to new domains is the prohibitive cost of acquiring sufficient training labels. Alternatives such as weak supervision, active learning, and fine-tuning of pretrained models reduce this burden but require substantial human input to select a highly informative subset of instances or to curate labeling functions. REGAL (Rule-Enhanced Generative Active Learning) is an improved framework for weakly supervised text classification that performs active learning over labeling functions rather than individual instances. REGAL interactively creates high-quality labeling patterns from raw text, enabling a single annotator to accurately label an entire dataset after initialization with three keywords for each class. Experiments demonstrate that REGAL extracts up to 3 times as many high-accuracy labeling functions from text as current state-of-the-art methods for interactive weak supervision, enabling REGAL to dramatically reduce the annotation burden of writing labeling functions for weak supervision. Statistical analysis reveals REGAL performs equal or significantly better than interactive weak supervision for five of six commonly used natural language processing (NLP) baseline datasets.more » « less
-
null (Ed.)Topic models are typically evaluated with respect to the global topic distributions that they generate, using metrics such as coherence, but without regard to local (token-level) topic assignments. Token-level assignments are important for downstream tasks such as classification. Even recent models, which aim to improve the quality of these token-level topic assignments, have been evaluated only with respect to global metrics. We propose a task designed to elicit human judgments of token-level topic assignments. We use a variety of topic model types and parameters and discover that global metrics agree poorly with human assignments. Since human evaluation is expensive we propose a variety of automated metrics to evaluate topic models at a local level. Finally, we correlate our proposed metrics with human judgments from the task on several datasets. We show that an evaluation based on the percent of topic switches correlates most strongly with human judgment of local topic quality. We suggest that this new metric, which we call consistency, be adopted alongside global metrics such as topic coherence when evaluating new topic models.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1145/3511808.3557410
