skip to main content


Title: Motivated misreporting in crowdsourcing: Data quality concerns for scientific tasks
Crowdsourcing has become a popular means to solicit assistance for scientific research. From classifying images or texts to responding to surveys, tapping into the knowledge of crowds to complete complex tasks has become a common strategy in social and information sciences. Although the timeliness and cost-effectiveness of crowdsourcing may provide desirable advantages to researchers, the data it generates may be of lower quality for some scientific purposes. The quality control mechanisms, if any, offered by common crowdsourcing platforms may not provide robust measures of data quality. This study explores whether research task participants may engage in motivated misreporting whereby participants tend to cut corners to reduce their workload while performing various scientific tasks online. We conducted an experiment with three common crowdsourcing tasks: answering surveys, coding images, and classifying online social media content. The experiment recruited workers from three sources: a crowdsourcing platform (Amazon Mechanical Turk) and a commercial online survey panel. The analysis seeks to address the following two questions: (1) whether online panelists or crowd workers may engage in motivated misreporting differently and (2) whether the patterns of misreporting vary by different task types. The study focuses on the analysis of the experiment in answering surveys and offers quality assurance practice guideline of using crowdsourcing for social science research.  more » « less
Award ID(s):
1645307
NSF-PAR ID:
10080559
Author(s) / Creator(s):
; ; ;
Date Published:
Journal Name:
The 73th Annual Conference of the American Association for Public Opinion Research
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract: 100 words Jurors are increasingly exposed to scientific information in the courtroom. To determine whether providing jurors with gist information would assist in their ability to make well-informed decisions, the present experiment utilized a Fuzzy Trace Theory-inspired intervention and tested it against traditional legal safeguards (i.e., judge instructions) by varying the scientific quality of the evidence. The results indicate that jurors who viewed high quality evidence rated the scientific evidence significantly higher than those who viewed low quality evidence, but were unable to moderate the credibility of the expert witness and apply damages appropriately resulting in poor calibration. Summary: <1000 words Jurors and juries are increasingly exposed to scientific information in the courtroom and it remains unclear when they will base their decisions on a reasonable understanding of the relevant scientific information. Without such knowledge, the ability of jurors and juries to make well-informed decisions may be at risk, increasing chances of unjust outcomes (e.g., false convictions in criminal cases). Therefore, there is a critical need to understand conditions that affect jurors’ and juries’ sensitivity to the qualities of scientific information and to identify safeguards that can assist with scientific calibration in the courtroom. The current project addresses these issues with an ecologically valid experimental paradigm, making it possible to assess causal effects of evidence quality and safeguards as well as the role of a host of individual difference variables that may affect perceptions of testimony by scientific experts as well as liability in a civil case. Our main goal was to develop a simple, theoretically grounded tool to enable triers of fact (individual jurors) with a range of scientific reasoning abilities to appropriately weigh scientific evidence in court. We did so by testing a Fuzzy Trace Theory-inspired intervention in court, and testing it against traditional legal safeguards. Appropriate use of scientific evidence reflects good calibration – which we define as being influenced more by strong scientific information than by weak scientific information. Inappropriate use reflects poor calibration – defined as relative insensitivity to the strength of scientific information. Fuzzy Trace Theory (Reyna & Brainerd, 1995) predicts that techniques for improving calibration can come from presentation of easy-to-interpret, bottom-line “gist” of the information. Our central hypothesis was that laypeople’s appropriate use of scientific information would be moderated both by external situational conditions (e.g., quality of the scientific information itself, a decision aid designed to convey clearly the “gist” of the information) and individual differences among people (e.g., scientific reasoning skills, cognitive reflection tendencies, numeracy, need for cognition, attitudes toward and trust in science). Identifying factors that promote jurors’ appropriate understanding of and reliance on scientific information will contribute to general theories of reasoning based on scientific evidence, while also providing an evidence-based framework for improving the courts’ use of scientific information. All hypotheses were preregistered on the Open Science Framework. Method Participants completed six questionnaires (counterbalanced): Need for Cognition Scale (NCS; 18 items), Cognitive Reflection Test (CRT; 7 items), Abbreviated Numeracy Scale (ABS; 6 items), Scientific Reasoning Scale (SRS; 11 items), Trust in Science (TIS; 29 items), and Attitudes towards Science (ATS; 7 items). Participants then viewed a video depicting a civil trial in which the defendant sought damages from the plaintiff for injuries caused by a fall. The defendant (bar patron) alleged that the plaintiff (bartender) pushed him, causing him to fall and hit his head on the hard floor. Participants were informed at the outset that the defendant was liable; therefore, their task was to determine if the plaintiff should be compensated. Participants were randomly assigned to 1 of 6 experimental conditions: 2 (quality of scientific evidence: high vs. low) x 3 (safeguard to improve calibration: gist information, no-gist information [control], jury instructions). An expert witness (neuroscientist) hired by the court testified regarding the scientific strength of fMRI data (high [90 to 10 signal-to-noise ratio] vs. low [50 to 50 signal-to-noise ratio]) and gist or no-gist information both verbally (i.e., fairly high/about average) and visually (i.e., a graph). After viewing the video, participants were asked if they would like to award damages. If they indicated yes, they were asked to enter a dollar amount. Participants then completed the Positive and Negative Affect Schedule-Modified Short Form (PANAS-MSF; 16 items), expert Witness Credibility Scale (WCS; 20 items), Witness Credibility and Influence on damages for each witness, manipulation check questions, Understanding Scientific Testimony (UST; 10 items), and 3 additional measures were collected, but are beyond the scope of the current investigation. Finally, participants completed demographic questions, including questions about their scientific background and experience. The study was completed via Qualtrics, with participation from students (online vs. in-lab), MTurkers, and non-student community members. After removing those who failed attention check questions, 469 participants remained (243 men, 224 women, 2 did not specify gender) from a variety of racial and ethnic backgrounds (70.2% White, non-Hispanic). Results and Discussion There were three primary outcomes: quality of the scientific evidence, expert credibility (WCS), and damages. During initial analyses, each dependent variable was submitted to a separate 3 Gist Safeguard (safeguard, no safeguard, judge instructions) x 2 Scientific Quality (high, low) Analysis of Variance (ANOVA). Consistent with hypotheses, there was a significant main effect of scientific quality on strength of evidence, F(1, 463)=5.099, p=.024; participants who viewed the high quality evidence rated the scientific evidence significantly higher (M= 7.44) than those who viewed the low quality evidence (M=7.06). There were no significant main effects or interactions for witness credibility, indicating that the expert that provided scientific testimony was seen as equally credible regardless of scientific quality or gist safeguard. Finally, for damages, consistent with hypotheses, there was a marginally significant interaction between Gist Safeguard and Scientific Quality, F(2, 273)=2.916, p=.056. However, post hoc t-tests revealed significantly higher damages were awarded for low (M=11.50) versus high (M=10.51) scientific quality evidence F(1, 273)=3.955, p=.048 in the no gist with judge instructions safeguard condition, which was contrary to hypotheses. The data suggest that the judge instructions alone are reversing the pattern, though nonsignificant, those who received the no gist without judge instructions safeguard awarded higher damages in the high (M=11.34) versus low (M=10.84) scientific quality evidence conditions F(1, 273)=1.059, p=.30. Together, these provide promising initial results indicating that participants were able to effectively differentiate between high and low scientific quality of evidence, though inappropriately utilized the scientific evidence through their inability to discern expert credibility and apply damages, resulting in poor calibration. These results will provide the basis for more sophisticated analyses including higher order interactions with individual differences (e.g., need for cognition) as well as tests of mediation using path analyses. [References omitted but available by request] Learning Objective: Participants will be able to determine whether providing jurors with gist information would assist in their ability to award damages in a civil trial. 
    more » « less
  2. Obeid, I. (Ed.)
    The Neural Engineering Data Consortium (NEDC) is developing the Temple University Digital Pathology Corpus (TUDP), an open source database of high-resolution images from scanned pathology samples [1], as part of its National Science Foundation-funded Major Research Instrumentation grant titled “MRI: High Performance Digital Pathology Using Big Data and Machine Learning” [2]. The long-term goal of this project is to release one million images. We have currently scanned over 100,000 images and are in the process of annotating breast tissue data for our first official corpus release, v1.0.0. This release contains 3,505 annotated images of breast tissue including 74 patients with cancerous diagnoses (out of a total of 296 patients). In this poster, we will present an analysis of this corpus and discuss the challenges we have faced in efficiently producing high quality annotations of breast tissue. It is well known that state of the art algorithms in machine learning require vast amounts of data. Fields such as speech recognition [3], image recognition [4] and text processing [5] are able to deliver impressive performance with complex deep learning models because they have developed large corpora to support training of extremely high-dimensional models (e.g., billions of parameters). Other fields that do not have access to such data resources must rely on techniques in which existing models can be adapted to new datasets [6]. A preliminary version of this breast corpus release was tested in a pilot study using a baseline machine learning system, ResNet18 [7], that leverages several open-source Python tools. The pilot corpus was divided into three sets: train, development, and evaluation. Portions of these slides were manually annotated [1] using the nine labels in Table 1 [8] to identify five to ten examples of pathological features on each slide. Not every pathological feature is annotated, meaning excluded areas can include focuses particular to these labels that are not used for training. A summary of the number of patches within each label is given in Table 2. To maintain a balanced training set, 1,000 patches of each label were used to train the machine learning model. Throughout all sets, only annotated patches were involved in model development. The performance of this model in identifying all the patches in the evaluation set can be seen in the confusion matrix of classification accuracy in Table 3. The highest performing labels were background, 97% correct identification, and artifact, 76% correct identification. A correlation exists between labels with more than 6,000 development patches and accurate performance on the evaluation set. Additionally, these results indicated a need to further refine the annotation of invasive ductal carcinoma (“indc”), inflammation (“infl”), nonneoplastic features (“nneo”), normal (“norm”) and suspicious (“susp”). This pilot experiment motivated changes to the corpus that will be discussed in detail in this poster presentation. To increase the accuracy of the machine learning model, we modified how we addressed underperforming labels. One common source of error arose with how non-background labels were converted into patches. Large areas of background within other labels were isolated within a patch resulting in connective tissue misrepresenting a non-background label. In response, the annotation overlay margins were revised to exclude benign connective tissue in non-background labels. Corresponding patient reports and supporting immunohistochemical stains further guided annotation reviews. The microscopic diagnoses given by the primary pathologist in these reports detail the pathological findings within each tissue site, but not within each specific slide. The microscopic diagnoses informed revisions specifically targeting annotated regions classified as cancerous, ensuring that the labels “indc” and “dcis” were used only in situations where a micropathologist diagnosed it as such. Further differentiation of cancerous and precancerous labels, as well as the location of their focus on a slide, could be accomplished with supplemental immunohistochemically (IHC) stained slides. When distinguishing whether a focus is a nonneoplastic feature versus a cancerous growth, pathologists employ antigen targeting stains to the tissue in question to confirm the diagnosis. For example, a nonneoplastic feature of usual ductal hyperplasia will display diffuse staining for cytokeratin 5 (CK5) and no diffuse staining for estrogen receptor (ER), while a cancerous growth of ductal carcinoma in situ will have negative or focally positive staining for CK5 and diffuse staining for ER [9]. Many tissue samples contain cancerous and non-cancerous features with morphological overlaps that cause variability between annotators. The informative fields IHC slides provide could play an integral role in machine model pathology diagnostics. Following the revisions made on all the annotations, a second experiment was run using ResNet18. Compared to the pilot study, an increase of model prediction accuracy was seen for the labels indc, infl, nneo, norm, and null. This increase is correlated with an increase in annotated area and annotation accuracy. Model performance in identifying the suspicious label decreased by 25% due to the decrease of 57% in the total annotated area described by this label. A summary of the model performance is given in Table 4, which shows the new prediction accuracy and the absolute change in error rate compared to Table 3. The breast tissue subset we are developing includes 3,505 annotated breast pathology slides from 296 patients. The average size of a scanned SVS file is 363 MB. The annotations are stored in an XML format. A CSV version of the annotation file is also available which provides a flat, or simple, annotation that is easy for machine learning researchers to access and interface to their systems. Each patient is identified by an anonymized medical reference number. Within each patient’s directory, one or more sessions are identified, also anonymized to the first of the month in which the sample was taken. These sessions are broken into groupings of tissue taken on that date (in this case, breast tissue). A deidentified patient report stored as a flat text file is also available. Within these slides there are a total of 16,971 total annotated regions with an average of 4.84 annotations per slide. Among those annotations, 8,035 are non-cancerous (normal, background, null, and artifact,) 6,222 are carcinogenic signs (inflammation, nonneoplastic and suspicious,) and 2,714 are cancerous labels (ductal carcinoma in situ and invasive ductal carcinoma in situ.) The individual patients are split up into three sets: train, development, and evaluation. Of the 74 cancerous patients, 20 were allotted for both the development and evaluation sets, while the remain 34 were allotted for train. The remaining 222 patients were split up to preserve the overall distribution of labels within the corpus. This was done in hope of creating control sets for comparable studies. Overall, the development and evaluation sets each have 80 patients, while the training set has 136 patients. In a related component of this project, slides from the Fox Chase Cancer Center (FCCC) Biosample Repository (https://www.foxchase.org/research/facilities/genetic-research-facilities/biosample-repository -facility) are being digitized in addition to slides provided by Temple University Hospital. This data includes 18 different types of tissue including approximately 38.5% urinary tissue and 16.5% gynecological tissue. These slides and the metadata provided with them are already anonymized and include diagnoses in a spreadsheet with sample and patient ID. We plan to release over 13,000 unannotated slides from the FCCC Corpus simultaneously with v1.0.0 of TUDP. Details of this release will also be discussed in this poster. Few digitally annotated databases of pathology samples like TUDP exist due to the extensive data collection and processing required. The breast corpus subset should be released by November 2021. By December 2021 we should also release the unannotated FCCC data. We are currently annotating urinary tract data as well. We expect to release about 5,600 processed TUH slides in this subset. We have an additional 53,000 unprocessed TUH slides digitized. Corpora of this size will stimulate the development of a new generation of deep learning technology. In clinical settings where resources are limited, an assistive diagnoses model could support pathologists’ workload and even help prioritize suspected cancerous cases. ACKNOWLEDGMENTS This material is supported by the National Science Foundation under grants nos. CNS-1726188 and 1925494. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. REFERENCES [1] N. Shawki et al., “The Temple University Digital Pathology Corpus,” in Signal Processing in Medicine and Biology: Emerging Trends in Research and Applications, 1st ed., I. Obeid, I. Selesnick, and J. Picone, Eds. New York City, New York, USA: Springer, 2020, pp. 67 104. https://www.springer.com/gp/book/9783030368432. [2] J. Picone, T. Farkas, I. Obeid, and Y. Persidsky, “MRI: High Performance Digital Pathology Using Big Data and Machine Learning.” Major Research Instrumentation (MRI), Division of Computer and Network Systems, Award No. 1726188, January 1, 2018 – December 31, 2021. https://www. isip.piconepress.com/projects/nsf_dpath/. [3] A. Gulati et al., “Conformer: Convolution-augmented Transformer for Speech Recognition,” in Proceedings of the Annual Conference of the International Speech Communication Association (INTERSPEECH), 2020, pp. 5036-5040. https://doi.org/10.21437/interspeech.2020-3015. [4] C.-J. Wu et al., “Machine Learning at Facebook: Understanding Inference at the Edge,” in Proceedings of the IEEE International Symposium on High Performance Computer Architecture (HPCA), 2019, pp. 331–344. https://ieeexplore.ieee.org/document/8675201. [5] I. Caswell and B. Liang, “Recent Advances in Google Translate,” Google AI Blog: The latest from Google Research, 2020. [Online]. Available: https://ai.googleblog.com/2020/06/recent-advances-in-google-translate.html. [Accessed: 01-Aug-2021]. [6] V. Khalkhali, N. Shawki, V. Shah, M. Golmohammadi, I. Obeid, and J. Picone, “Low Latency Real-Time Seizure Detection Using Transfer Deep Learning,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2021, pp. 1 7. https://www.isip. piconepress.com/publications/conference_proceedings/2021/ieee_spmb/eeg_transfer_learning/. [7] J. Picone, T. Farkas, I. Obeid, and Y. Persidsky, “MRI: High Performance Digital Pathology Using Big Data and Machine Learning,” Philadelphia, Pennsylvania, USA, 2020. https://www.isip.piconepress.com/publications/reports/2020/nsf/mri_dpath/. [8] I. Hunt, S. Husain, J. Simons, I. Obeid, and J. Picone, “Recent Advances in the Temple University Digital Pathology Corpus,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2019, pp. 1–4. https://ieeexplore.ieee.org/document/9037859. [9] A. P. Martinez, C. Cohen, K. Z. Hanley, and X. (Bill) Li, “Estrogen Receptor and Cytokeratin 5 Are Reliable Markers to Separate Usual Ductal Hyperplasia From Atypical Ductal Hyperplasia and Low-Grade Ductal Carcinoma In Situ,” Arch. Pathol. Lab. Med., vol. 140, no. 7, pp. 686–689, Apr. 2016. https://doi.org/10.5858/arpa.2015-0238-OA. 
    more » « less
  3. This work investigates how different forms of input elicitation obtained from crowdsourcing can be utilized to improve the quality of inferred labels for image classification tasks, where an image must be labeled as either positive or negative depending on the presence/absence of a specified object. Five types of input elicitation methods are tested: binary classification (positive or negative); the ( x, y )-coordinate of the position participants believe a target object is located; level of confidence in binary response (on a scale from 0 to 100%); what participants believe the majority of the other participants' binary classification is; and participant's perceived difficulty level of the task (on a discrete scale). We design two crowdsourcing studies to test the performance of a variety of input elicitation methods and utilize data from over 300 participants. Various existing voting and machine learning (ML) methods are applied to make the best use of these inputs. In an effort to assess their performance on classification tasks of varying difficulty, a systematic synthetic image generation process is developed. Each generated image combines items from the MPEG-7 Core Experiment CE-Shape-1 Test Set into a single image using multiple parameters (e.g., density, transparency, etc.) and may or may not contain a target object. The difficulty of these images is validated by the performance of an automated image classification method. Experiment results suggest that more accurate results can be achieved with smaller training datasets when both the crowdsourced binary classification labels and the average of the self-reported confidence values in these labels are used as features for the ML classifiers. Moreover, when a relatively larger properly annotated dataset is available, in some cases augmenting these ML algorithms with the results (i.e., probability of outcome) from an automated classifier can achieve even higher performance than what can be obtained by using any one of the individual classifiers. Lastly, supplementary analysis of the collected data demonstrates that other performance metrics of interest, namely reduced false-negative rates, can be prioritized through special modifications of the proposed aggregation methods. 
    more » « less
  4. Introduction Social media has created opportunities for children to gather social support online (Blackwell et al., 2016; Gonzales, 2017; Jackson, Bailey, & Foucault Welles, 2018; Khasawneh, Rogers, Bertrand, Madathil, & Gramopadhye, 2019; Ponathil, Agnisarman, Khasawneh, Narasimha, & Madathil, 2017). However, social media also has the potential to expose children and adolescents to undesirable behaviors. Research showed that social media can be used to harass, discriminate (Fritz & Gonzales, 2018), dox (Wood, Rose, & Thompson, 2018), and socially disenfranchise children (Page, Wisniewski, Knijnenburg, & Namara, 2018). Other research proposes that social media use might be correlated to the significant increase in suicide rates and depressive symptoms among children and adolescents in the past ten years (Mitchell, Wells, Priebe, & Ybarra, 2014). Evidence based research suggests that suicidal and unwanted behaviors can be promulgated through social contagion effects, which model, normalize, and reinforce self-harming behavior (Hilton, 2017). These harmful behaviors and social contagion effects may occur more frequently through repetitive exposure and modelling via social media, especially when such content goes “viral” (Hilton, 2017). One example of viral self-harming behavior that has generated significant media attention is the Blue Whale Challenge (BWC). The hearsay about this challenge is that individuals at all ages are persuaded to participate in self-harm and eventually kill themselves (Mukhra, Baryah, Krishan, & Kanchan, 2017). Research is needed specifically concerning BWC ethical concerns, the effects the game may have on teenagers, and potential governmental interventions. To address this gap in the literature, the current study uses qualitative and content analysis research techniques to illustrate the risk of self-harm and suicide contagion through the portrayal of BWC on YouTube and Twitter Posts. The purpose of this study is to analyze the portrayal of BWC on YouTube and Twitter in order to identify the themes that are presented on YouTube and Twitter posts that share and discuss BWC. In addition, we want to explore to what extent are YouTube videos compliant with safe and effective suicide messaging guidelines proposed by the Suicide Prevention Resource Center (SPRC). Method Two social media websites were used to gather the data: 60 videos and 1,112 comments from YouTube and 150 posts from Twitter. The common themes of the YouTube videos, comments on those videos, and the Twitter posts were identified using grounded, thematic content analysis on the collected data (Padgett, 2001). Three codebooks were built, one for each type of data. The data for each site were analyzed, and the common themes were identified. A deductive coding analysis was conducted on the YouTube videos based on the nine SPRC safe and effective messaging guidelines (Suicide Prevention Resource Center, 2006). The analysis explored the number of videos that violated these guidelines and which guidelines were violated the most. The inter-rater reliabilities between the coders ranged from 0.61 – 0.81 based on Cohen’s kappa. Then the coders conducted consensus coding. Results & Findings Three common themes were identified among all the posts in the three social media platforms included in this study. The first theme included posts where social media users were trying to raise awareness and warning parents about this dangerous phenomenon in order to reduce the risk of any potential participation in BWC. This was the most common theme in the videos and posts. Additionally, the posts claimed that there are more than 100 people who have played BWC worldwide and provided detailed description of what each individual did while playing the game. These videos also described the tasks and different names of the game. Only few videos provided recommendations to teenagers who might be playing or thinking of playing the game and fewer videos mentioned that the provided statistics were not confirmed by reliable sources. The second theme included posts of people that either criticized the teenagers who participated in BWC or made fun of them for a couple of reasons: they agreed with the purpose of BWC of “cleaning the society of people with mental issues,” or they misunderstood why teenagers participate in these kind of challenges, such as thinking they mainly participate due to peer pressure or to “show off”. The last theme we identified was that most of these users tend to speak in detail about someone who already participated in BWC. These videos and posts provided information about their demographics and interviews with their parents or acquaintances, who also provide more details about the participant’s personal life. The evaluation of the videos based on the SPRC safe messaging guidelines showed that 37% of the YouTube videos met fewer than 3 of the 9 safe messaging guidelines. Around 50% of them met only 4 to 6 of the guidelines, while the remaining 13% met 7 or more of the guidelines. Discussion This study is the first to systematically investigate the quality, portrayal, and reach of BWC on social media. Based on our findings from the emerging themes and the evaluation of the SPRC safe messaging guidelines we suggest that these videos could contribute to the spread of these deadly challenges (or suicide in general since the game might be a hoax) instead of raising awareness. Our suggestion is parallel with similar studies conducted on the portrait of suicide in traditional media (Fekete & Macsai, 1990; Fekete & Schmidtke, 1995). Most posts on social media romanticized people who have died by following this challenge, and younger vulnerable teens may see the victims as role models, leading them to end their lives in the same way (Fekete & Schmidtke, 1995). The videos presented statistics about the number of suicides believed to be related to this challenge in a way that made suicide seem common (Cialdini, 2003). In addition, the videos presented extensive personal information about the people who have died by suicide while playing the BWC. These videos also provided detailed descriptions of the final task, including pictures of self-harm, material that may encourage vulnerable teens to consider ending their lives and provide them with methods on how to do so (Fekete & Macsai, 1990). On the other hand, these videos both failed to emphasize prevention by highlighting effective treatments for mental health problems and failed to encourage teenagers with mental health problems to seek help and providing information on where to find it. YouTube and Twitter are capable of influencing a large number of teenagers (Khasawneh, Ponathil, Firat Ozkan, & Chalil Madathil, 2018; Pater & Mynatt, 2017). We suggest that it is urgent to monitor social media posts related to BWC and similar self-harm challenges (e.g., the Momo Challenge). Additionally, the SPRC should properly educate social media users, particularly those with more influence (e.g., celebrities) on elements that boost negative contagion effects. While the veracity of these challenges is doubted by some, posting about the challenges in unsafe manners can contribute to contagion regardless of the challlenges’ true nature. 
    more » « less
  5. Citizen science projects have successfully taken advantage of volunteers to unlock scientific information contained in images. Crowds extract scientific data by completing different types of activities: transcribing text, selecting values from pre-defined options, reading data aloud, or pointing and clicking at graphical elements. While designing crowdsourcing tasks, selecting the best form of input and task granularity is essential for keeping the volunteers engaged and maximizing the quality of the results. In the context of biocollections information extraction, this study compares three interface actions (transcribe, select, and crop) and tasks of different levels of granularity (single field vs. compound tasks). Using 30 crowdsourcing experiments and two different populations, these interface alternatives are evaluated in terms of speed, quality, perceived difficulty and enjoyability. The results show that Selection and Transcription tasks generate high quality output, but they are perceived as boring. Conversely, Cropping tasks, and arguably graphical tasks in general, are more enjoyable, but their output quality depend on additional machine-oriented processing. When the text to be extracted is longer than two or three words, Transcription is slower than Selection and Cropping. When using compound tasks, the overall time required for the crowdsourcing experiment is considerably shorter than using single field tasks, but they are perceived as more difficult. When using single field tasks, both the quality of the output and the amount of identified data are slightly higher compared to compound tasks, but they are perceived by the crowd as less entertaining. 
    more » « less