Updating the dataset structure and considerations for using the data with information from Bandy and Vincent 2021

README.md

@@ -85,9 +85,29 @@ Books are a rich source of both fine-grained information, how a character, an ob
 ## Dataset Structure
 
 ### Data Instances
+In the original dataset described by [Zhu and Kiros et al.](https://yknzhu.wixsite.com/mbweb), BookCorpus contained 11,038 books. However, based on the files obtained,
+there appear to be only 7,185 unique books (excluding romance-all.txtand adventure-all.txt as explained in 2.2.1). Potential duplicates were identified based on file names, which suggested that
+2,930 books may be duplicated. Using the diff Unix program, it was confirmed that BookCorpus contained duplicate, identical text files for all but five of these books. 
+The five exceptions were manually inspected:
+* 299560.txt (Third Eye Patch), for which slightly different versions appeared in the “Thriller”
+and “Science Fiction” genre folders (only 30 lines differed)
+* 529220.txt (On the Rocks), for which slightly different versions appeared in the “Literature”
+and “Science Fiction” genre folders (only the title format differed)
+* Hopeless-1.txt, for which identical versions appeared in the “New Adult” and “Young
+Adult” genre folders, and a truncated version appeared in the “Romance” folder (containing
+30% of the full word count)
+* u4622.txt, for which identical versions appeared in the “Romance” and “Young Adult”
+genre folders, and a slightly different version appeared in the “Science Fiction” folder (only
+15 added lines)
+* u4899.txt, for which a full version appeared in the “Young Adult” folder and a truncated
+version (containing the first 28 words) appeared in the “Science Fiction” folder
+
+Combined with the diff results, the manual inspection confirmed that each filename represents one unique book, thus BookCorpus contained at most 7,185 unique books.
 
 #### plain_text
 
+
+
 - **Size of downloaded dataset files:** 1.18 GB
 - **Size of the generated dataset:** 4.85 GB
 - **Total amount of disk used:** 6.03 GB
@@ -101,13 +121,20 @@ An example of 'train' looks as follows.
 
 ### Data Fields
 
-The data fields are the same among all splits.
+Each book in BookCorpus simply includes the full text from the ebook (often including preamble, copyright text, etc.). However, in research that 
+BookCorpus, authors have applied a range of different encoding schemes that change
+the definition of an “instance” (e.g. in GPT-N training, text is encoded using byte-pair encoding). The data fields are the same among all splits. There is no label or target 
+associated with each instance (book). The text from each book was originally used for unsupervised training by [Zhu and Kiros et al.](https://yknzhu.wixsite.com/mbweb), and 
+the only label-like attribute is the genre associated with each book, which is provided by Smashwords. No relationships between individual instances (books) are made explicit. 
+Grouped into folders by genre, the data implicitly links books in the same genre. It was found that duplicate books are implicitly linked through identical filenames. 
+However, no other relationships are made explicit, such as books by the same author, books in the same series, books set in the same context, books addressing the same 
+event, and/or books using the same characters.
 
 #### plain_text
 - `text`: a `string` feature.
 
 ### Data Splits
-
+There are no recommended data splits. The authors use all books in the dataset for unsupervised training, with no splits or subsamples.
 |   name   | train  |
 |----------|-------:|
 |plain_text|74004228|
@@ -143,18 +170,50 @@ The data fields are the same among all splits.
 [More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
 
 ## Considerations for Using the Data
-
+The composition of BookCorpus or the way it was collected and preprocessed/cleaned/labeled might impact future uses. At the very least, the duplicate books
+and sampling skews should guide any future uses to curate a subsample of BookCorpus to better serve the task at hand. An analysis of the potential 
+impact of BookCorpus and its use on data subjects has not been conducted. Richard Lea interviewed a handful of authors represented in BookCorpus
+([Richard Lea](https://www.theguardian.com/books/2016/sep/28/google-swallows-11000-novels-to-improve-ais-conversation)).
 ### Social Impact of Dataset
-
-[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+The dataset contains data that might be considered sensitive. The aforementioned contact information (email addresses) is sensitive personal information.
 
 ### Discussion of Biases
+BookCorpus contains free books from smashwords.com which are at least 20,000 words long. Based
+on metrics from [Smashwords](https://blog.smashwords.com/2014/12/smashwords-year-in-review-2014-and.html), 11,038 books (as reported in the original BookCorpus 
+dataset) would have represented approximately 3% of the 336,400 books published on Smashwords as of 2014, while the 7,185 unique books we report would have represented 2%. 
+For reference, as of 2013, the Library of Congress contained 23,592,066 cataloged books ([Audrey Fischer](https://www.loc.gov/item/prn-14-009/library-by-the-numbers-2013/2014-01-23/)). 
+
+
+There are some errors, sources of noise, or redundancies in BookCorpus. While some book
+files appear to be cleaned of preamble and postscript text, many files still contain this text and
+various other sources of noise. Of particular concern is that we found many copyright-related
+sentences, for example:
+* “if you’re reading this book and did not purchase it, or it was not purchased for your use
+only, then please return to smashwords.com and purchase your own copy.” (n=788)
+* “this book remains the copyrighted property of the author, and may not be redistributed to
+others for commercial or non-commercial purposes...” (n=111)
+* “although this is a free book, it remains the copyrighted property of the author, and may not
+be reproduced, copied and distributed for commercial or non-commercial purposes.” (n=109)
+* “thank you for respecting the author’s work” (n=70)
+* “no part of this publication may be copied, reproduced in any format, by any means, electronic or otherwise, without prior consent from the copyright owner and publisher of this
+book” (n=16)
+
+Note that these sentences represent noise and redundancy. As previously noted, BookCorpus also contains many duplicate books: of the 7,185 unique books in the dataset, 
+2,930 occurred more than once. Most of these (N=2,101) books appeared twice, though many were duplicated multiple times, including some
+books (N=6) with five copies in BookCorpus. See Table 2.
 
-[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
 
 ### Other Known Limitations
 
-[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
+There are no export controls or other regulatory restrictions that apply to the dataset or to individual instances. Some information is missing from individual instances (books). 
+ 98 empty book
+files were found in the folder downloaded from [Zhu and Kiros et al.](https://yknzhu.wixsite.com/mbweb) Also, while the authors collected
+books longer than 20,000 words, 655 files were shorter than 20,000 words, and 291 were shorter than 10,000 words, suggesting that many book files were significantly 
+truncated from their original text.
+
+There were no ethical review processes conducted. [Zhu and Kiros et al.](https://yknzhu.wixsite.com/mbweb) do not mention an Institutional Review Board (IRB) or other ethical review process involved in 
+their original paper. Bandy and Vincent strongly suggest that researchers should use BookCorpus with caution for any task, namely due to potential copyright violations,
+duplicate books, and sampling skews.
 
 ## Additional Information