Real-Time Visual Feedback to Guide Benchmark Creation:
A Human-and-Metric-in-the-Loop Workflow

Researchers invest significant effort to create benchmarks in machine learning, including ImageNet Deng et al. (2009), SQUAD Rajpurkar et al. (2016), and SNLI Bowman et al. (2015), as well as to develop models that solve them. Can we rely on these benchmarks? A growing body of recent research Schwartz et al. (2017); Poliak et al. (2018); Kaushik and Lipton (2018) is revealing that models exploit spurious bias/artifacts– unintended correlations between input and output Torralba and Efros (2011) (e.g. the word ‘not’ is associated with the label ‘contradiction’ in Natural Language Inference (NLI) Gururangan et al. (2018))– instead of the actual underlying features, to solve many popular benchmarks. Models, therefore, fail to generalize, and experience drastic performance drops when testing with out-of-distribution (OOD) data or adversarial examples  Bras et al. (2020); Mishra et al. (2020a); McCoy et al. (2019); Zhang et al. (2019); Larson et al. (2019b); Sakaguchi et al. (2019); Hendrycks and Gimpel (2016). This begs the question: Shouldn’t ML researchers consequently focus on creating ‘better’ datasets rather than developing increasingly complex models on bias-laden benchmarks?

Deletion of samples based on bias baseline reports– hypothesis-only baseline in NLI Dua et al. (2019))– and mitigation approaches such as AFLite Sakaguchi et al. (2019) (adversarial filtering which deletes targeted data subsets),  Clark et al. (2019); Kaushik et al. (2019), have the following limitations: (i) data deletion/augmentation and residual learning do not justify the original investment in data creation, and (ii) crowdworkers are not provided adequate feedback to learn what constitutes high-quality data– and so have additional overhead due to the manual effort involved in sample creation/validation. Furthermore, Parmar et al. (2022a) show that biased samples are created even when crowdworkers are provided with an initial set of annotation instructions. One potential solution to these problems is continuous, in situ feedback about artifacts while benchmark data is being created. To our knowledge, there are no approaches that provide real-time artifact identification, feedback, and reconciliation opportunities to data creators, nor guide them on data quality.

Contributions: (i) We propose VAIDA (Visual Analytics for Interactively Discouraging Artifacts), a novel system for benchmark creation that provides continuous visual feedback to data creators in real-time. VAIDA supports artifact identification and resolution, implicitly educating crowdworkers and analysts on data quality. (ii) We evaluate VAIDA empirically through expert review and a user study to understand the cognitive workload it imposes. The results indicate that VAIDA decreases mental demand, temporal demand, effort, and frustration of crowdworkers (31.1%) and analysts (14.3%); it increases their performance by 34.6% and 30.8% respectively, and educates crowdworkers on how to create high-quality samples. Overall, we see a 45.8% decrease in the presence of artifacts in created samples. (iii) Even though our main goal is to reduce artifacts in samples, we observe that samples created in our user study are adversarial across language models with performance decreases of 31.3% (BERT), 22.5% (RoBERTa), and 14.98% (GPT-3 fewshot).

This work sits at the intersection of two primary areas: (1) visual analysis of data quality (higher presence of artifacts indicates lower quality), and (2) development of a novel data collection pipeline.²²2Detailed related work is in the Supplemental Material.

In this section, we describe VAIDA’s important backend processes.

DQI: DQI can be expressed as a quality metric that examines different sources of artifacts in text, by scoring samples along 7 different components. We use DQI in order to demonstrate VAIDA’s ability to cover multiple facets of artifact creation, although VAIDA is metric agnostic. VAIDA uses an intuitive traffic signal color coding (high$>>$moderate$>>$low) to indicate levels of artifacts (i.e., quality) in samples. Hyperparameters for color mapping with respect to DQI component values depend on (i) the application type, and (ii) characteristics of pre-existing samples present in the corpus at the time of new sample creation.⁴⁴4 See Supplemental Material: Evaluation, for details across all DQI components, hyperparameter tuning, and analyses. For instance, when recreating SNLI Bowman et al. (2015) with VAIDA, we tune hyperparameters separating the boundary between red, yellow, and green flags on 0.01% of the SNLI training dataset manually in a supervised manner Mishra et al. (2020b).

AutoFix: We propose AutoFix as a module to help crowdworkers avoid creating bad samples by recommending changes to a sample to improve its quality. The AutoFix algorithm is explained in Figure 1. Given a premise, hypothesis, and DQI values for the hypothesis, AutoFix sequentially masks each word in the hypothesis and ranks words based on their influence on model output, i.e., their importance. The hypothesis word of highest importance is replaced, to achieve at least moderate quality. DQI hence controls the amount and aspect of changes made by AutoFix. By incrementally changing the sample, users can understand how and why their sample is being modified and how DQI values are affected.

TextFooler: From an analyst’s perspective, the quality of a submitted sample might be “too low” because (i) the crowdworker might not employ AutoFix appropriately, or (ii) there is a narrow acceptability range due to the criticality of the application domain, such as in BioNLP Lee et al. (2020); Parmar et al. (2022b). We therefore implement TextFooler Jin et al. (2019) to adversarially transform low-quality samples (instead of discarding them), to improve benchmark robustness, and ensure that crowdsourcing effort is not wasted. We initially use AFLite Bras et al. (2020), to bin samples into good (retained samples) and bad (filtered samples) splits. Using TextFooler, we adversarially transform bad split data to flip the label; we revert back to the original label and identify sample artifacts using DQI (see Figure 5).

VAIDA provides customized interfaces for both crowdworkers and analysts, as shown in Figures 2,  3 respectively.⁵⁵5See Supplemental Material: Interface Design for interface intuitions and detailed descriptions, with full-resolution images. We describe VAIDA’s workflow via a case study for sample creation (crowdworker) and review (analyst) in Figures 4,5.

The crowdworker interface provides instructions (A) to navigate through the panels and how to interpret feedback for created samples. Communication links for FAQs, and error reporting are also provided (F). Crowdworkers can then review the instructions for data creation (B)– here, we use the same instructions provided in the original SNLI crowdsourcing interface (b1).

1. Sample Creation: The premise field auto-populates with a caption from the Flickr30 corpus (b2). The crowdworker must create three hypotheses (for the entailment, neutral, contradiction labels) at a time, though they are reviewed individually.

2. DQI Evaluation Feedback: On clicking the review button, DQI feedback is shown in (C), for each component. Each colored circle (c1) indicates the level of artifacts present corresponding to each DQI component for the created sample; hovering displays a tooltip that explains and highlights (c2) the artifacts in the created sample, pertaining to the category of bias covered by that component. The overall sample quality (c3) is calculated, by averaging over the artifact percentiles of all 7 DQI components for the sample. By comparing sample quality with that of pre-existing samples, the probability that the sample will be accepted/rejected is shown (c4). The user can choose to either revise or submit the sample.

3a. Sample Revision: Manual Revision or AutoFix can be done. We illustrate the improvement of sample quality in this step in Figure 4.

3b. Sample Submission: After review (and potentially iterative DQI evaluations/sample revisions), the sample can be submitted for benchmark inclusion. On submission, a sample enters a pending state (d1) until review by the analyst.

While crowdworkers work within a single tightly-coordinated interface to create, submit, and review samples, analysts can navigate between a set of nine screens (Figure 3) to review samples in detail to make ‘accept’, ‘reject’, and ‘modification’ decisions, and to assess overall benchmark quality. Analysts review samples in batches of size 50.⁶⁶6DQI components 1, 2, 3, 6, and 7, gauge artifact presence relative to pre-existing samples. We observe given at least 50 pre-existing samples, DQI component values change by less than 5% for the overall dataset if another 50 samples are accepted.

4. Analyst Review: Review can consist of several different operations by the analyst.

(i) Direct Acceptance– The home page (UI) for the analyst provides a view similar to the crowdworker interface, and allows the analyst to review the work of a single crowdworker. If the data quality is deemed to be sufficiently high by just viewing the DQI color flags and quality percentile, the analyst can directly accept the sample into the corpus from this screen, as shown in Figure 5.

(ii) Visual Analysis and Modification– VAIDA provides several visualizations (C1-C7) to support detailed analysis and review of submitted samples, and to assess artifact presence (i.e., quality) in the overall benchmark. Each visualization allows the analyst to simulate how adding one or more submitted samples affects the benchmark’s quality.

For example, in Figure 5, we show the visualizations for analysis of DQI component 4 (C4), which deals with artifacts caused due to word similarity within a sample. For each sample in the corpus, the word similarities between all possible pairs of words are averaged; the samples are then hierarchically displayed as a treemap based on their DQI color mapping and average word similarity. We note that the color scale followed is bilinear, as while artifacts must be eliminated, there still needs to be a sufficient inductive bias for the sample to be solvable. The size of a rectangle indicates the distance of each sample from the average word similarity across all samples. The new sample is highlighted in the treemap with a black outline. In the example shown, we see that the dataset’s C4 value decreases slightly (0.807) when the new sample is added, though the flag color remains red.

Further examination of the new sample can be done, to establish why the sample has low effect on the dataset’s C4; the user clicks on the sample rectangle in the treemap, and is taken to a heat map view. The heatmap shows the similarity values between every word pair (from the premise/hypothesis/both) in a tooltip on hover, with the values mapped to the same color scale used for the treemap . For instance, in Figure 5, the words ‘man’ and ‘champagne are repeated verbatim, and ‘overjoyed’, ‘smiling’, ‘celebrating’, ‘pop’, and ‘shoot’ also have similarity of [0.6–1]. This indicates that several words in the sample are too closely related and constitute artifacts.

Each such visualization is therefore individually tailored to represent a specific DQI component of interest, based on the linguistic features examined for artifact creation in that component. We further elaborate on the design intuitions and analysis conducted with all the component visualizations available to the analyst in the supplementary.

Post analysis, if the analyst feels that the submitted sample requires only a minor change (for instance, reshuffling or the addition of a single word) to warrant acceptance, then they can invoke the TextFooler module to transform the sample adversarially, and then accept, thereby ensuring minimal data loss. In the case shown, TextFooler improves the sample’s C4 with most of the heatmap varying from 0.3-0.7. Due to this, the analyst decides to accept the updated sample.

5. Analyst Feedback: Once the analyst has reached a decision, the crowdworker sees updates (Figure 5) in (D) to the reviewed sample count (d2) – increases to 16– and the pie chart that indicates the distribution of actions taken by the analyst over all samples submitted by the crowdworker (d3). Additionally, in (E), the line chart (e1), which contains the history of previously submitted samples is updated. The x-axis denotes the sample number and the y-axis denotes the quality percentile (c3), of the corresponding sample. On click, the corresponding sample is loaded, along with its DQI component values and feedback to the crowdworker. The crowdworker can also use the box plot (e2), to view their current rank, and choose to view the sample history of another crowd worker.

We evaluate VAIDA’s efficacy at providing real-time feedback to educate crowdworkers during benchmark creation using expert review and a user study. We also evaluate model performance (BERT Devlin et al. (2018), RoBERTa Liu et al. (2019), GPT-3 (fewshot) Brown et al. (2020)) on data created with VAIDA during the user study.

We propose VAIDA, a paradigm to address benchmark artifacts, by integrating human-in-the-loop sensemaking with continuous visual feedback. VAIDA uses several visualization interfaces to analyze quality considerations (based on artifact levels) at multiple granularities. While we do not explicitly address computational quality control or fairness consideration (though some aspects can be targeted by currently integrated metrics), since VAIDA is extensible to the incorporation of customized backend-metrics, we believe our paradigm can support multi-faceted benchmark evaluation.

In our usability evaluation, we see that users report greater satisfaction, and lower difficulty with their work and system experience; this implies possible higher crowdworker retention and engagement. Additionally, in our study, we see that users effectively identify and avoid artifact patterns during sample creation. Based on our study results, we believe that a minimum of 30 annotators would be needed for large-scale data creation to ensure timely feedback (i.e., sample decision provided within 24 hours) to crowdworkers. Based on our secondary study, we believe analysts will exhibit increased performance and maintain satisfaction ratings in full-scale creation, as they will become well-versed in the nuances of bias for the data-creation task they are evaluating, as well as the visualizations being used. This, however, is contingent on restricting the visualization views to display only the  200–300 samples with closest artifact levels to the sample being evaluated.

Overall, samples created with VAIDA are found to not only of higher quality than achieved with conventional crowdsourcing, but are also seen to be adversarial across transformer models. This is also maintained across multiple task types– we additionally create StoryCLOZE Schwartz et al. (2017) samples with VAIDA⁹. This was done independent of the study described in the paper, with 4 crowdworkers creating samples. However, for this, several interface features had to be changed, so we focus on reporting only NLI results in the main paper. VAIDA hence demonstrates a novel, dynamic approach for building benchmarks and mitigating artifacts, and serves as a starting point for the next generation of benchmarks in machine learning.

In future work, we intend to integrate VAIDA with an actual crowdsourcing framework, and run a full-scale data creation study to create a high-quality benchmark. Expanding to such a setup will require additional back-end engineering, to ensure that (i) timely and accurate feedback continues to be provided in real-time to crowdworkers, (ii) analysts are available on hand to process samples in a timely manner. This is out of scope for the current paper (e.g., it would require a significant budget), but we see our current work as a stepping stone in this direction. Additionally, studying the problem and designing the visualizations and real-time feedback mechanisms are essential steps before moving to large-scale evaluation; the novel affordances and designs are a necessary first step, and we believe they will be impactful to the NLP community.

Crowdworker retention and engagement in this full-scale setting also need to be evaluated, in order to better contextualize the learning curve associated with system usage and handling artifact creation, given an increasingly higher number of pre-existing system samples. Comparing this setup directly with the effect of in-depth user training Roit et al. (2019) on artifact creation and review, prior to crowdsourcing, would also further help analyze and quantify if/how user strategy and performance changes during VAIDA usage.

Design modifications when creating different types of datasets will mainly require the redesigning of sample input fields, corresponding to the application and the type of metric used for artifact evaluation. However, in full-scale dataset creation, the visualization views for the analyst corresponding to different artifact types will have to be restricted to the $\sim$300 samples closest artifact levels, to the given sample being created, in order to facilitate scalable processing for analysts. Additionally, since visualization familiarity is required for the analyst to effectively review samples, the analysts may choose to streamline analysis by only using a subset of the provided visualization types in their version of the system, corresponding to the application domain.