0
$\begingroup$

I have a software that produces a sentiment score between 0 and 1 given an input sentence. I want to test if this system produces biased scores based on the gender of the subject used in the sentence.

I have 50 sentence pairs and for each sentence pair I get a pair of scores from my software. Therefore I have 50 pairs of scores.

I want to give an example sentence pair and score pairs:

Input Sentence: John feels angry    -> score: 0.15
Input Sentence: Marry fells angry   -> score: 0.19

Another example:

Input Sentence: Adam likes music    -> score: 0.10
Input Sentence: Isabel likes music   -> score: 0.34

At the and I create two arrays keeping the scores I obtained with those sentences:

male_scores = [ 0.15,0.10, ....]
female_scores = [ 0.19,0.34, ...]

As I said earlier,now I want to see if these score differences tells me something about the fact that my software have some gender bias or not.

In order to do that, should I use independent t-test or dependent t-test?

$\endgroup$
8
  • 2
    $\begingroup$ So each sentence appears twice, with the only modification being the gender of the person mentioned in the sentence, and you want to test if there's a difference in mean between the scores with regards to gender? If the choice is between a paired t-test and a unpaired t-test, then you should use the paired t-test. $\endgroup$
    – Phil
    Nov 7, 2018 at 10:33
  • $\begingroup$ @Phil Thank you for your comment, this is exactly what I want to do. Do you have any other suggestion ? $\endgroup$
    – zwlayer
    Nov 7, 2018 at 12:05
  • 1
    $\begingroup$ If the t-test is sufficient for your purposes, then do that analysis (but do check the corresponding assumptions behind the paired t-test so that they are fulfilled for your data). If you want to include covariates in your analysis, then you could add them in a regression framework. $\endgroup$
    – Phil
    Nov 7, 2018 at 12:10
  • 1
    $\begingroup$ I just realised that you wrote that the outcome is bounded between 0 and 1. That might affect the plausibility of the analysis, so check the distribution of the paired differences to see if they look roughly normal. $\endgroup$
    – Phil
    Nov 7, 2018 at 12:12
  • $\begingroup$ @Phil, if they are (almost) normal I use paired-t test right? $\endgroup$
    – zwlayer
    Nov 7, 2018 at 12:17

1 Answer 1

1
$\begingroup$

This question has been addressed in the comments. I reply here to give an old questions an answer.

The two sentences within each pair are more alike than two randomly chosen sentences with different genders of the subject; this makes their scores paired. - Glen_b

Then if the choice is between a paired or unpaired t-test, you should choose the paired t-test. As Phil also notes in the comments, you should check the assumptions of the test are fulfilled by your data.

A nice reference on the paired t-test is here which lists the assumptions, and some worked examples on the same site here.

I note the machine-learning tag on this question, and the application for detecting bias in an ML model. I would like to leave the following references here for anyone who has stumbled across this question.

  • Kiritchenko, Svetlana & Mohammad, Saif. (2018). Examining Gender and Race Bias in Two Hundred Sentiment Analysis Systems. 43-53. 10.18653/v1/S18-2005. (pdf)

Abstract below:

Automatic machine learning systems can inadvertently accentuate and perpetuate inappropriate human biases. Past work on examining inappropriate biases has largely focused on just individual systems. Further, there is no benchmark dataset for examining inappropriate biases in systems. Here for the first time, we present the Equity Evaluation Corpus (EEC), which consists of 8,640 English sentences carefully chosen to tease out biases towards certain races and genders. We use the dataset to examine 219 automatic sentiment analysis systems that took part in a recent shared task, SemEval-2018 Task 1 ‘Affect in Tweets’. We find that several of the systems show statistically significant bias; that is, they consistently provide slightly higher sentiment intensity predictions for one race or one gender. We make the EEC freely available

In this paper they do very similar to what is described in the question above (and they use the paired t-test).

If you don't want to read the original paper, there is a write-up of it on Medium here.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.