Toloka research

Domain-specific data is the crux of the successful transfer of machine learning systems from benchmarks to real life. Crowdsourcing has become one of the standard tools for cheap and time-efficient data collection for simple problems such as image classification: thanks in large part to advances in research on aggregation methods.

In this paper, we demonstrate Crowd-Kit, a general-purpose crowdsourcing computational quality control toolkit. It provides efficient implementations in Python of computational quality control algorithms for crowdsourcing, including uncertainty measures and crowd consensus methods.

This paper reviews the crowdsourced audio transcription shared task devoted to this problem and co-organized with the Crowd Science Workshop at VLDB 2021; the competition attracted 18 participants, 8 of them have successfully beaten our non-trivial baselines.

We study the problem of predicting future hourly earnings and task completion time for a crowdsourcing platform user who sees the list of available tasks and wants to select one of them to execute.

In this paper, we address the problem of labeling text images via CAPTCHA, where user identification is generally impossible. We propose a new algorithm to aggregate multiple guesses collected through CAPTCHA.