International audience

This paper explores and offers guidance on a specific and relevant problem in task design for crowdsourcing: how to formulate a complex question used to classify a set of items. In micro-task markets, classification is still among the most popular tasks. We situate our work in the context of information retrieval and multi-predicate classification, i.e., classifying a set of items based on a set of conditions. Our experiments cover a wide range of tasks and domains, and also consider crowd workers alone and in tandem with machine learning classifiers. We provide empirical evidence into how the resulting classification performance is affected by different predicate formulation strategies, emphasizing the importance of predicate formulation as a task design dimension in crowdsourcing.

This paper reports on the challenges and lessons we learned while running controlled experiments in crowdsourcing platforms. Crowdsourcing is becoming an attractive technique to engage a diverse and large pool of subjects in experimental research, allowing researchers to achieve levels of scale and completion times that would otherwise not be feasible in lab settings. However, the scale and flexibility comes at the cost of multiple and sometimes unknown sources of bias and confounding factors that arise from technical limitations of crowdsourcing platforms and from the challenges of running controlled experiments in the “wild”. In this paper, we take our experience in running systematic evaluations of task design as a motivating example to explore, describe, and quantify the potential impact of running uncontrolled crowdsourcing experiments and derive possible coping strategies. Among the challenges identified, we can mention sampling bias, controlling the assignment of subjects to experimental conditions, learning effects, and reliability of crowdsourcing results. According to our empirical studies, the impact of potential biases and confounding factors can amount to a 38% loss in the utility of the data collected in uncontrolled settings; and it can significantly change the outcome of experiments. These issues ultimately inspired us to implement CrowdHub, a system that sits on top of major crowdsourcing platforms and allows researchers and practitioners to run controlled crowdsourcing projects.

Text classification is one of the most common goals of machine learning (ML) projects, and also one of the most frequent human intelligence tasks in crowdsourcing platforms. ML has mixed success in such tasks depending on the nature of the problem, while crowd-based classification has proven to be surprisingly effective, but can be expensive. Recently, hybrid text classification algorithms, combining human computation and machine learning, have been proposed to improve accuracy and reduce costs. One way to do so is to have ML highlight or emphasize portions of text that it believes to be more relevant to the decision. Humans can then rely only on this text or read the entire text if the highlighted information is insufficient. In this paper, we investigate if and under what conditions highlighting selected parts of the text can (or cannot) improve classification cost and/or accuracy, and in general how it affects the process and outcome of the human intelligence tasks. We study this through a series of crowdsourcing experiments running over different datasets and with task designs imposing different cognitive demands. Our findings suggest that highlighting is effective in reducing classification effort but does not improve accuracy - and in fact, low-quality highlighting can decrease it.