Datasets:
sadrasabouri
commited on
Commit
•
d92497a
1
Parent(s):
eb650dd
Update README.md
Browse files
README.md
CHANGED
@@ -99,13 +99,13 @@ This dataset has only a split (`train` split).
|
|
99 |
|
100 |
## Dataset Creation
|
101 |
|
102 |
-
The translated version of the Shahnameh was generally derived from the [satoor](https://www.sattor.com/english/Shahnameh.pdf) website. We first extracted texts from the pdf. After that we divided
|
103 |
|
104 |
### Annotations
|
105 |
|
106 |
#### Annotation process
|
107 |
|
108 |
-
Through the process of image generation we noticed a bias in the DALL-E models towards word `iran`. It was biased so that each sentence with this given word would have pictures from Iran's political figures which were usually totally irrelevant. The annotation process mainly focused to deal with these pictures. We removed those images which seems to be harmful to those figures and/or were irrelevant to the context.
|
109 |
|
110 |
#### Who are the annotators?
|
111 |
|
@@ -113,27 +113,29 @@ Mahsa Namdar and Sadra Sabouri were the annotators of this dataset.
|
|
113 |
|
114 |
### Personal and Sensitive Information
|
115 |
|
116 |
-
Since the textual data is easily downloadable and the images were generated through
|
117 |
|
118 |
## Considerations for Using the Data
|
119 |
|
120 |
### Social Impact of Dataset
|
121 |
|
122 |
-
|
123 |
|
124 |
### Discussion of Biases
|
125 |
|
126 |
-
|
127 |
|
128 |
### Other Known Limitations
|
129 |
|
130 |
-
|
131 |
|
132 |
## Additional Information
|
133 |
|
134 |
### Dataset Curators
|
135 |
|
136 |
-
|
|
|
|
|
137 |
|
138 |
### Licensing Information
|
139 |
|
|
|
99 |
|
100 |
## Dataset Creation
|
101 |
|
102 |
+
The translated version of the Shahnameh was generally derived from the [satoor](https://www.sattor.com/english/Shahnameh.pdf) website. We first extracted texts from the pdf. After that, we divided paragraphs into sentences and give each sentence to the DALL-E mini model through its online API. It generated nine images for each sentence. After a few annotations, we came up with more than 30000 images.
|
103 |
|
104 |
### Annotations
|
105 |
|
106 |
#### Annotation process
|
107 |
|
108 |
+
Through the process of image generation, we noticed a bias in the DALL-E models towards the word `iran`. It was biased so that each sentence with this given word would have pictures from Iran's political figures which were usually totally irrelevant. The annotation process mainly focused to deal with these pictures. We removed those images which seems to be harmful to those figures and/or were irrelevant to the context.
|
109 |
|
110 |
#### Who are the annotators?
|
111 |
|
|
|
113 |
|
114 |
### Personal and Sensitive Information
|
115 |
|
116 |
+
Since the textual data is easily downloadable and the images were generated through an image generation model there shouldn't be any personal information in this dataset. Just in case you find something harmful or violating of one's personal information please let us know. We will take proper action as soon as possible.
|
117 |
|
118 |
## Considerations for Using the Data
|
119 |
|
120 |
### Social Impact of Dataset
|
121 |
|
122 |
+
This dataset is mainly aimed to release for its artistic value. The process of generating images for the Shahnameh - which is one of the most important Farsi poem books - is our precious contribution. This dataset is not only used for this purpose but also can as a dataset in image-to-text and text-to-image tasks.
|
123 |
|
124 |
### Discussion of Biases
|
125 |
|
126 |
+
The dataset's possible biases would come from the DALL-E mini biases. It's actually a good practice to check the dataset entries in order to find biases in that model. One it's worth mentioning in this work is the DALL-E mini model's bias for the word `iran` which nearly always comes up with images from political figures of this country.
|
127 |
|
128 |
### Other Known Limitations
|
129 |
|
130 |
+
There are constant debates in the literature about the limitations of machine-generated datasets. Some believe that since nowadays models are not perfect - and so do their output, it wouldn't be a good idea to use these artificially generated datasets as input to the new model. They suggest that by doing so we are actually limiting our accuracy by the model's accuracy which provided the primary dataset.
|
131 |
|
132 |
## Additional Information
|
133 |
|
134 |
### Dataset Curators
|
135 |
|
136 |
+
+ Emad Fatemizadeh: The general idea for generating a graphical version of Farsi poems was firstly introduced by him.
|
137 |
+
+ Sadra Sabouri: He looked up a translated version of the Shahnameh, extract and tokenized poems from it, and used the online DALL-E mini API to generate images from poems.
|
138 |
+
+ Mahsa Namdar: The process of annotation as a post-process on data has been held by her.
|
139 |
|
140 |
### Licensing Information
|
141 |
|