Repeated labeling using multiple noisy labelers

Abstract

This paper addresses the repeated acquisition of labels for data items when the labeling is imperfect.We examine the improvement (or lack thereof) in data quality via repeated labeling, and focus especially on the improvement of training labels for supervised induction. With the outsourcing of small tasks becoming easier, for example via Amazon's Mechanical Turk, it often is possible to obtain less-than-expert labeling at low cost. With low-cost labeling, preparing the unlabeled part of the data can become considerably more expensive than labeling. We present repeated-labeling strategies of increasing complexity, and show several main results. (i) Repeated-labeling can improve label quality and model quality, but not always.(ii) When labels are noisy, repeated labeling can be preferable to single labeling even in the traditional setting where labels are not particularly cheap. (iii) As soon as the cost of processing the unlabeled data is not free, even the simple strategy of labeling everything multiple times can give considerable advantage.(iv) Repeatedly labeling a carefully chosen set of points is generally preferable, and we present a set of robust techniques that combine different notions of uncertainty to select data points for which quality should be improved. The bottom line: the results show clearly that when labeling is not perfect, selective acquisition of multiple labels is a strategy that data miners should have in their repertoire; for certain label-quality/cost regimes, the benefit is substantial.

Simulating Diffusion with Conflicting Knowledge

“…Redundancy of information access has been shown, under certain conditions, to increase classification accuracy of web documents [22], an agent's trust in a piece of information [23] [24], the likelihood of winning a game where an increased pool of knowledge increases odds of success [25] and accuracy on classification problems using voters from Mechanical Turk [26].…”

“…Of particular interest to our study is the work of [26], who consider how best to use a constrained set of human labelers whose abilities are initially unknown to perform a standard machine learning classification problem. First, [26] discuss the existence of a cost in adding additional human labelers, showing that redundancy of labeling can be both very cost efficient and cost inefficient, depending on how it is implemented.…”

“…First, [26] discuss the existence of a cost in adding additional human labelers, showing that redundancy of labeling can be both very cost efficient and cost inefficient, depending on how it is implemented. The cost in their study is monetary, but correlates to the cost of reducing incorrect information at the expense of quickly spreading correct information.…”

“…We see similar trade-off issues in our results. Secondly, [26] note that as the number of labelers giving an answer for a specific classification problem increases, conflict is naturally more likely to occur. In a similar vein, [27] shows that high levels of task redundancy (the number of people doing the same thing) in an organization may decrease performance, in part because it leads to conflicting results.…”

“…Based on this, we postulated that in a network with highly redundant access to information, knowledge will spread quickly in a stable environment, but much slower in an environment that produces a significant amount of conflicting knowledge. Finally, we make use of work in [26] on the concept of a probability-distribution based approach to how certain their classifier should be about a particular problem. We utilize this method for computing uncertainty when we represent how "uncertain" an agent is about the knowledge he holds.…”