... Accepted papers Christoph Hube and Besnik Fetahu Detecting Biased Statements in Wikipedia http://wikiworkshop.org/2018/papers/wikiworkshop2018_paper_1.pdf ...
Hi Christoph and Besnik,
Having worked with several thousand of Amazon Mechanical Turkers over the past year, I am not convinced that their opinions of bias, even in aggregate, are not biased. Did you take any steps to measure the bias against accuracy in your crowdworkers?
Here is an example of what I expect they would get wrong:
"Tax cuts allow consumers to increase their spending, which boosts aggregate demand."
That statement, added by en:User:Bkwillwm in 2012,[1] is still part of the English Wikipedia's Economics article. However, the statement is strictly inaccurate, and heavily biased in favor of trickle-down economics and austerity policy.[2] It and statements like it, pervasive through many if not most of the popular language Wikipedias, directly support increases in income inequality, which in turn is a terrible scourge affecting both health[3] and economic growth.[4]
How can you measure whether your crowdworkers are truly unbiased relative to accuracy, instead of just reflecting the propaganda-influenced whims of the populist center?
Sincerely, James Salsman
[1] https://en.wikipedia.org/w/index.php?title=Economics&diff=prev&oldid...
[2] https://en.wikipedia.org/wiki/Talk:Economics/Archive_7#Tax_cut_claim_in_Fisc...
Hi James,
thanks a lot for your interest in our work!
The problem of crowdworkers being biased is a problem definitely not to be neglected. Majority vote can help to sort out single extremist views of workers but if many workers are strongly biased then I agree that this might not be enough. We are actually already thinking about methods to improve future crowdsourced bias datasets. One way to improve the quality is to have a very well defined task that leaves only little room for subjective interpretation. For example, instead of letting the workers decide whether a statement is biased or not, we asked more specifically whether the statement reflects an opinion or contains bias words. Of course, the decision if a statement reflects a fact or an opinion is still subjective in many cases.
Given your example it is hard to make a decision (even when being unbiased) without having the proper background knowledge. That is why our work mostly focuses on language bias, i.e. bias that is introduced through the use of judgemental language. Since there are many cases of bias without using judgemental language, we are definitely interested to come up with good approaches that cover these cases as well. Ideas and suggestions are always welcome!
One other thing that we are planning to do for future crowdsourcing jobs is to ask workers for their political opinions and to take this background information into account when creating ground truth data.
Best regards, Christoph
Am 4/18/2018 um 2:22 PM schrieb James Salsman:
... Accepted papers Christoph Hube and Besnik Fetahu Detecting Biased Statements in Wikipedia http://wikiworkshop.org/2018/papers/wikiworkshop2018_paper_1.pdf ...
Hi Christoph and Besnik,
Having worked with several thousand of Amazon Mechanical Turkers over the past year, I am not convinced that their opinions of bias, even in aggregate, are not biased. Did you take any steps to measure the bias against accuracy in your crowdworkers?
Here is an example of what I expect they would get wrong:
"Tax cuts allow consumers to increase their spending, which boosts aggregate demand."
That statement, added by en:User:Bkwillwm in 2012,[1] is still part of the English Wikipedia's Economics article. However, the statement is strictly inaccurate, and heavily biased in favor of trickle-down economics and austerity policy.[2] It and statements like it, pervasive through many if not most of the popular language Wikipedias, directly support increases in income inequality, which in turn is a terrible scourge affecting both health[3] and economic growth.[4]
How can you measure whether your crowdworkers are truly unbiased relative to accuracy, instead of just reflecting the propaganda-influenced whims of the populist center?
Sincerely, James Salsman
[1] https://en.wikipedia.org/w/index.php?title=Economics&diff=prev&oldid...
[2] https://en.wikipedia.org/wiki/Talk:Economics/Archive_7#Tax_cut_claim_in_Fisc...
Christoph,
Thank you for your reply. I am happy to help and glad you have made good decisions.
Are you familiar with Priyanka's work on Accuracy Review? https://priyankamandikal.github.io/posts/gsoc-2016-project-overview/
Fabian Floeck and
On Wed, Apr 18, 2018 at 7:07 AM, Christoph Hube hube@l3s.de wrote:
Hi James,
thanks a lot for your interest in our work!
The problem of crowdworkers being biased is a problem definitely not to be neglected. Majority vote can help to sort out single extremist views of workers but if many workers are strongly biased then I agree that this might not be enough. We are actually already thinking about methods to improve future crowdsourced bias datasets. One way to improve the quality is to have a very well defined task that leaves only little room for subjective interpretation. For example, instead of letting the workers decide whether a statement is biased or not, we asked more specifically whether the statement reflects an opinion or contains bias words. Of course, the decision if a statement reflects a fact or an opinion is still subjective in many cases.
Given your example it is hard to make a decision (even when being unbiased) without having the proper background knowledge. That is why our work mostly focuses on language bias, i.e. bias that is introduced through the use of judgemental language. Since there are many cases of bias without using judgemental language, we are definitely interested to come up with good approaches that cover these cases as well. Ideas and suggestions are always welcome!
One other thing that we are planning to do for future crowdsourcing jobs is to ask workers for their political opinions and to take this background information into account when creating ground truth data.
Best regards, Christoph
Am 4/18/2018 um 2:22 PM schrieb James Salsman:
... Accepted papers Christoph Hube and Besnik Fetahu Detecting Biased Statements in Wikipedia http://wikiworkshop.org/2018/papers/wikiworkshop2018_paper_1.pdf ...
Hi Christoph and Besnik,
Having worked with several thousand of Amazon Mechanical Turkers over the past year, I am not convinced that their opinions of bias, even in aggregate, are not biased. Did you take any steps to measure the bias against accuracy in your crowdworkers?
Here is an example of what I expect they would get wrong:
"Tax cuts allow consumers to increase their spending, which boosts aggregate demand."
That statement, added by en:User:Bkwillwm in 2012,[1] is still part of the English Wikipedia's Economics article. However, the statement is strictly inaccurate, and heavily biased in favor of trickle-down economics and austerity policy.[2] It and statements like it, pervasive through many if not most of the popular language Wikipedias, directly support increases in income inequality, which in turn is a terrible scourge affecting both health[3] and economic growth.[4]
How can you measure whether your crowdworkers are truly unbiased relative to accuracy, instead of just reflecting the propaganda-influenced whims of the populist center?
Sincerely, James Salsman
[1] https://en.wikipedia.org/w/index.php?title=Economics&diff=prev&oldid...
[2] https://en.wikipedia.org/wiki/Talk:Economics/Archive_7#Tax_cut_claim_in_Fisc...
wiki-research-l@lists.wikimedia.org