|
|
To compose issues, I, Thomas Krichel have written and maintains a system called ernad. These notes talk about how to use it. Please report any and all suggestions you may have to make these notes clearer, more helpful, and less boring.
Ernad releases NEP issues on Mondays just after midnight UTC. When a new issue is ready, you get an alerting mail from ernad@nep.repec.org. Kindly whitelist this address to avoid email landing in your spam folder.
Any particular issue may be delayed by technical problems. But the delay should only be for a few hours. If you do not see an alerting mail by Tuesday, first check the issue timing. It will tell you if this is a holiday week. If it is not, contact Thomas Krichel immediately. It’s likely that this is an email delivery problem. That is something we have to take very seriously.
You login at http://ernad.repec.org. You use your report code as the login. If you do not have a report code and password, contact the director of NEP, Marco Novarese.
After login, you may see a message that you do not have an issue waiting. Most likely, this is because an issue is not yet done. If we create a report for you during the week, it will not have an issue until the Monday that follows. We do not release past issues to starting editors.
Paper selection is your only mandatory task. Ernad will help you.
The paper selection screens take a tabular form, with two cells in each row. The left cell is called the item cell. It has the description of the item, in our case an academic paper in economics. The right cell is the selection cell. It contains a checkbox. When you have not taken any action on a paper, the selection cell has a black background. You use the checkbox to select or deselect the paper. If your browser supports Javascript, you can click anywhere in the selection cell. Thus you do not need to do the extra work to move your mouse to the exact location of the checkbox. Again, if your browser supports Javascript, the color of the selection cell is green or red if an item is selected or deselected, respectively.
Each paper is prefaced by a horizontal rule, known as the separator rule. By default, the separator rule has a black color.
Ernad uses machine learning. Machine learning looks at examples of the kind of papers you selected in past issues of the report. It tries to guess what papers are most likely to be included in the this issue. It will then “presort” your report issue. That means that the papers that are the most likely to be included in the report will be at the top of the issue. The papers that are least likely to be in the report will be at the end of the issue. Still, a presorted issue has same papers as an unsorted issue. It’s just the order that is different.
When presorting takes place, the separator rule takes colors that range from green (#00FF00) to red (#FF0000). The color is an indication of how probable the presorting estimates that the paper is relevant to the topic of the report. You can find out the exact value by hovering the mouse over the selection cell.
When presorting takes place, with issue alerting mail will inform you about the “weight” of the issue. The weight is the sum of probabilities that the learning system has attached to any paper being relevant, rounded down to the next integer. Let’s look at an example. Assume an issue contains three papers. The first is 70% likely, the second is 40% likely, and the third 20% likely. Then the weight of that issue is 1.
Once ernad has learned about your topic, you can be confident that all the papers that are relevant to your report are at the top. Still it is a good idea to conduct some searches for known key terms or JEL classifications.
When you are starting a new report, there are no past issues. Therefore there is nothing to learn from. The separator rules will be all black. If your report topic is narrow, you may not find a single paper in your first issue. Then you hand the machine a set of negative examples. Without a positive examples, no learning can take place. And therefore, your second issue will still be unsorted. You have to look though it again without any help from ernad’s machine learning. This type of work quickly wears of its fun value.
There is a clever way around this problem.
Papers selection happens in two screens. The first is called the “pickup” screen. The second is called the “filtering” screen. It is very important that you understand the difference between the two.
The pickup screen shows you all the papers in the issue. If you don’t see a relevant paper by scrolling, use the search interface of your browser. If you don’t see a paper that is strictly relevant to your report, but somewhat relevant, you may wish to select it in the pickup screen.
It is very important to understand that machine learning will take the papers that you select in the pickup stage as being relevant. It does not matter whether not the paper is excluded later. You can accept papers that you don’t include in the report to guide machine learning. Thus, try to pick up all papers you feel are close to the report’s topic. This is particularly important if you run a thin report, i.e. one where there few relevant papers. When you are starting a thin report, you need to be generous at this the stage to guide the machine learning towards what to look for.
If you want to take a restrictive view of your topic, exclude the papers in the filtering screen.
If you see no relevant papers, just scroll to the end and hit the “Select” button.
This screen is used to exclude papers that are on the topic from your public report. A paper may be on the topic, but you may want to exclude it because
In summary, the proper use of the filtering screen is to exclude papers for any reasons other than the topic.
The filtering screen can also be “abused” to filter papers you selected in the pickup screen for the purposed of guiding machine learning, but that are not relevant to your narrow topic. When you start a new, thin report, it is perfectly reasonably to abuse the filtering screen in this way.
I am sorry to dwell on this. I suspect that can not stress this enough. Let me give you another example.
Assume that you see a working paper that has previously been published in another working paper series. You included it in your report at that time. You still have to select the second version at the pickup screen. Why? Well consider what will happen if you don’t. The two versions of the paper may have different RePEc handles but as far as their features are concerned, they are virtually identical. If you accept it once, but reject it the second time round, the machine learning will be confused. A confused machine learning algorithm is not likely to do a good job for you.
The representation of a paper in the item cell of the selection screens imitates its appearance in the HTML version of final report issue. There are two exceptions.
First, on the right next to the paper’s date, a couple of blue exclamation marks “!” surround the number of the paper in the current issue. This feature is called the scream. The scream gives you are idea how far down with the issue you are. You can also use it to find the paper with a browser search, because it is unlikely that anywhere else a exclamation mark is followed by a number.
Second, on the right next to the paper’s URL, a couple of blue vertical bars “|” surround one or two characters in the range of [0-9a-z]. This is the paper’s sausage. Technically, the sausage is a hexatrigesimal coding of the position of a paper in the unsorted issue, as determined by application of the Yanabino protocol. In each issue, each paper has its own sausage. It acts as a very short identifier of the paper. If you edit several reports, you may discover a paper that is relevant for the second report while you are working on the first report. Note the sausage of that paper. When you come to work on the same issue for the second report, you can find that paper very quickly using a browser search.
Papers selection is the only mandatory task. Here we discuss the optional tasks.
The sorting screen allows you to change the order of the papers in your issue. You can use the up “↑” and down “↓” arrows in the interface to move an individual paper up and down. You can use the text boxes to enter a series of numbers—positive or negative—and then press on of the two bottom of the input boxes. “↥” will sort by numbers using ascending order. “↧” will sort by number using descending order. Boxes that you have not filled in a number for will count for zero.
There are two preview screens. These allow you to preview the email issue both in its HTML and text version, respectively. If you spot an error just close your browser and contact Thomas Krichel. Since you already have done the selections, Thomas can send it for once he has gotten round to fixing the error.