Exploring Semantic Spaces - Caliskan, Bryson & Narayanan 2017

[jamesallenevans]

Caliskan, Aylin, Joanna J. Bryson, Arvind Narayanan. 2017. “Semantics derived automatically from language corpora contain human-like biases.” Science 356(6334):183-186.

[lkcao]

The authors try to generate a machine learning version of IAT through word embedding. However, is semantic distance in the semantic space comparable to the responding time of human beings in IAT? They have totally different mathematical basis. Can we say that the similarity between these two methods are more symbolic, than essential?

[ccsuehara]

First of all, it is surprising to find a paper with such a declarative title! Like the IAT, are there any other methodologies that are arising, along the rise, of the word embedding method?

How does Stanford, the leading head of GloVe, choose on what corpora they train the Glove? Should this be an important topic on future fairness discussions?

One minor definition clarification, on page 2 it says:

we used the state-of-the-art GloVe word-embedding method, in which, at a high level, the similarity between a pair of vectors is related to the probability that the words co-occur with other words similar to each other in text (13).

But the first reading says that words do not necessarily co-occur with the word embedding vector, which one is true?

[deblnia]

I'd also echo @clk16's concern that semantic distance is different in the ML generated word-embedding space versus the IAT space, but I also wonder if it really matters. I'm a fan of Arvind Narayanan's, but at the end of the paper the authors note the following:

Further concerns may arise as AI is given agency in our society. If machine learning technologies used for, say, résumé screening were to imbibe cultural stereotypes, it may result in prejudiced outcomes. We recommend addressing this through the explicit characterization of acceptable behavior.

This feels like a typical natural scientist's approach to eliminating bias. But isn't "bias" (in the sense of social embeddedness, in the sense of subjectivity) inherent to every aspect of this technology, as it is with all instruments? This recommendation seems to simply move the locus of bias from the technology itself to the developer. Is that societally better? I don't think so.

[katykoenig]

This paper focuses on algorithmic bias, specifically, how word embeddings models adapt human bias and concludes that researchers working in AI should be cognizant of this issue. As many of us are planning on incorporate word embeddings into our project, how should we address this issue of bias in our texts?

[sanittawan]

The authors allude to the source of bias and ways to rectify them as @deblnia pointed out. I am wondering how exactly "explicit characterization of acceptable behavior" works. This's independent of the question of whether these fixes actually reduce the biases.

[alakira]

This paper points out the embedded human biases in text analysis when using word embeddings. It suggests to address this bias through 'the explicit characterization of acceptable behavior'. I wonder, in particular, how can we do that using text? In addition, rather, can we use the recognized bias to de-bias the text for learning algorithms?

[HaoxuanXu]

This paper pointed out the tendency for the word embedding algorithm to accrue existing biases. It will be interesting to know if supervised learning is employed in debiasing the model, or if there are other regularization methods used to de-bias the text

[cytwill]

I think this paper basically reveals the point in machine learning that "the input decides output". I am more interested in how can we apply these biases in future text analysis. For example, when we do attitude measurement, should we make any modifications when we find a positive word is attached to a negative-biased subject, other than a positive-biased subject?