Detokenize `answers` and `email`

README.md

@@ -25,8 +25,20 @@ task_ids:
 ---
 
 
-This dataset contains sentence-level formality annotations used in the 2016 TACL paper "An Empirical Analysis of Formality in Online Communication" (Pavlick
-  and Tetreault, 2016). It includes sentences from four genres (news, blogs, email, and QA forums), all annotated by humans on Amazon Mechanical Turk. The news and blog data was collected by Shibamouli Lahiri, and we are redistributing it here for the convenience of other researchers. We collected the email and answers data ourselves, using a similar annotation setup to Shibamouli. If you use this data in your work, please cite BOTH of the below papers:
+This dataset contains sentence-level formality annotations used in the 2016
+TACL paper "An Empirical Analysis of Formality in Online Communication"
+(Pavlick and Tetreault, 2016). It includes sentences from four genres (news,
+blogs, email, and QA forums), all annotated by humans on Amazon Mechanical
+Turk. The news and blog data was collected by Shibamouli Lahiri, and we are
+redistributing it here for the convenience of other researchers. We collected
+the email and answers data ourselves, using a similar annotation setup to
+Shibamouli.
+
+In the original dataset, `answers` and `email` were tokenized. In this version,
+Oleksiy Syvokon detokenized them with `moses-detokenizer` and a bunch of
+additional regexps.
+
+If you use this data in your work, please cite BOTH of the below papers:
 
 ```
 @article{PavlickAndTetreault-2016:TACL,
@@ -59,4 +71,4 @@ The annotated data files and number of lines in each are as follows:
 Each record contains the following fields:
 
 1. `avg_score`: the mean formality rating, which ranges from -3 to 3 where lower scores indicate less formal sentences
-2. `sentence`
+2. `sentence`