Update paper: missing DOI and expand WEAT openjournals/joss-reviews/#…

…4036

paper/paper.bib

@@ -78,6 +78,7 @@ @Article{du2021assessing
   author = {Yupei Du and Qixiang Fang and Dong Nguyen},
   journal = {arXiv preprint arXiv:2109.04732},
   year = {2021},
+  doi={10.18653/v1/2021.emnlp-main.785}
 }
 
 @Article{garg:2018:W,

paper/paper.md

@@ -62,7 +62,7 @@ The package also provides three trimmed word embeddings for experimentation: `go
 
 In the situation of bias detection, target words are words that **should** have no bias and usually represent the concept one would like to probe for biases. For instance, @garg:2018:W investigated the "women bias" of occupation-related words and their target words contain "nurse", "mathematician", and "blacksmith". These words can be used as target words because in an ideal world with no "women bias" associated with occupations, these occupation-related words should have no gender association.
 
-Target words are denoted as wordsets $\mathcal{S}$ and $\mathcal{T}$. All methods require $\mathcal{S}$ while $\mathcal{T}$ is only required for WEAT. For instance, the study of gender stereotypes in academic pursuits by @caliskan:2017:S used $\mathcal{S} = \{math, algebra, geometry, calculus, equations, ...\}$ and $\mathcal{T}= \{poetry, art, dance, literature, novel, ...\}$.
+Target words are denoted as wordsets $\mathcal{S}$ and $\mathcal{T}$. All methods require $\mathcal{S}$ while $\mathcal{T}$ is only required for Word Embedding Association Test (WEAT). For instance, the study of gender stereotypes in academic pursuits by @caliskan:2017:S used $\mathcal{S} = \{math, algebra, geometry, calculus, equations, ...\}$ and $\mathcal{T}= \{poetry, art, dance, literature, novel, ...\}$.
 
 In the situation of bias detection, attribute words are words that have known properties in relation to the bias. They are denoted as wordsets $\mathcal{A}$ and $\mathcal{B}$. All methods require both wordsets except Mean Average Cosine Similarity [@manzini2019black]. For instance, the study of gender stereotypes by @caliskan:2017:S used $\mathcal{A} = \{he, son, his, him, ...\}$ and $\mathcal{B} = \{she, daughter, hers, her, ...\}$. In some applications, popular off-the-shelf sentiment dictionaries can also be used as $\mathcal{A}$ and $\mathcal{B}$ [e.g. @sweeney2020reducing]. That being said, it is up to the researchers to select and derive these seed words in a query. However, the selection of seed words has been shown to be the most consequential part of the entire analysis [@antoniak2021bad;@du2021assessing]. Please read @antoniak2021bad for recommendations.
 

paper/paper.rmd

@@ -64,7 +64,7 @@ The package also provides three trimmed word embeddings for experimentation: `go
 
 In the situation of bias detection, target words are words that **should** have no bias and usually represent the concept one would like to probe for biases. For instance, @garg:2018:W investigated the "women bias" of occupation-related words and their target words contain "nurse", "mathematician", and "blacksmith". These words can be used as target words because in an ideal world with no "women bias" associated with occupations, these occupation-related words should have no gender association.
 
-Target words are denoted as wordsets $\mathcal{S}$ and $\mathcal{T}$. All methods require $\mathcal{S}$ while $\mathcal{T}$ is only required for WEAT. For instance, the study of gender stereotypes in academic pursuits by @caliskan:2017:S used $\mathcal{S} = \{math, algebra, geometry, calculus, equations, ...\}$ and $\mathcal{T}= \{poetry, art, dance, literature, novel, ...\}$.
+Target words are denoted as wordsets $\mathcal{S}$ and $\mathcal{T}$. All methods require $\mathcal{S}$ while $\mathcal{T}$ is only required for Word Embedding Association Test (WEAT). For instance, the study of gender stereotypes in academic pursuits by @caliskan:2017:S used $\mathcal{S} = \{math, algebra, geometry, calculus, equations, ...\}$ and $\mathcal{T}= \{poetry, art, dance, literature, novel, ...\}$.
 
 In the situation of bias detection, attribute words are words that have known properties in relation to the bias. They are denoted as wordsets $\mathcal{A}$ and $\mathcal{B}$. All methods require both wordsets except Mean Average Cosine Similarity [@manzini2019black]. For instance, the study of gender stereotypes by @caliskan:2017:S used $\mathcal{A} = \{he, son, his, him, ...\}$ and $\mathcal{B} = \{she, daughter, hers, her, ...\}$. In some applications, popular off-the-shelf sentiment dictionaries can also be used as $\mathcal{A}$ and $\mathcal{B}$ [e.g. @sweeney2020reducing]. That being said, it is up to the researchers to select and derive these seed words in a query. However, the selection of seed words has been shown to be the most consequential part of the entire analysis [@antoniak2021bad;@du2021assessing]. Please read @antoniak2021bad for recommendations.