Days in the Life of a Librarian

The Distant Reader can take five different types of input, and this blog posting describes what they are.

The Distant Reader is a tool for reading. It takes an arbitrary amount of unstructured data (text) as input, and it outputs sets of structured data for analysis — reading. Given a corpus of any size, the Distant Reader will analyze the corpus, and it will output a myriad of reports enabling you to use & understand the corpus. The Distant Reader is intended to designed the traditional reading process.

At the present time, the Reader can accept five different types of input, and they include:

1. a file
2. a URL
3. a list of URLs
4. a zip file
5. a zip file with a companion CSV file

Each of these different types of input are elaborated upon below.

A file

The simplest form of input is a single file from your computer. This can be just about file available to you, but to make sense, the file needs to contain textual data. Thus, the file can be a Word document, a PDF file, an Excel spreadsheet, an HTML file, a plain text file, etc. A file in the form of an image will not work because it contains zero text. Also, not all PDF files are created equal. Some PDF files are only facsimiles of their originals. Such PDF files are merely sets of images concatenated together. In order for PDF files to be used as input, the PDF files need to have been “born digitally” or they need to have had optical character recognition previously applied against them. Most PDF files are born digitally nor do they suffer from being facsimiles.

A good set of use-cases for single file input is the whole of a book, a long report, or maybe a journal article. Submitting a single file to the Distant Reader is quick & easy, but the Reader is designed for analyzing larger rather than small corpora. Thus, supplying a single journal article to the Reader doesn’t make much sense; the use of the traditional reading process probably makes more sense for a single journal article.

A URL

The Distant Reader can take a single URL as input. Given a URL, the Reader will turn into a rudimentary Internet spider and build a corpus. More specifically, given a URL, the Reader will:

• retrieve & cache the content found at the other end of the URL
• extract any URLs it finds in the content
• retrieve & cache the content from these additional URLs
• stop building the corpus but continue with its analysis

In short, given a URL, the Reader will cache the URL’s content, crawl the URL one level deep, cache the result, and stop caching.

Like the single file approach, submitting a URL to the Distant Reader is quick & easy, but there are a number of caveats. First of all, the Reader does not come with very many permissions, and just because you are authorized to read the content at the other end of a URL does not mean the Reader has the same authorization. A lot of content on the Web resides behind paywalls and firewalls. The Reader can only cache 100% freely accessible content.

“Landing pages” and “splash pages” represent additional caveats. Many of the URLs passed around the ‘Net do not point to the content itself, but instead they point to ill-structured pages describing the content — metadata pages. Such pages may include things like authors, titles, and dates, but these things are not presented in a consistent nor computer-readable fashion; they are laid out with aesthetics or graphic design in mind. These pages do contain pointers to the content you want to read, but the content may be two or three more clicks away. Be wary of URLs pointing to landing pages or splash pages.

Another caveat to this approach is the existence of extraneous input due to navigation. Many Web pages include links for navigating around the site. They also include links to things like “contact us” and “about this site”. Again, the Reader is sort of stupid. If found, the Reader will crawl such links and include their content in the resulting corpus.

Despite these drawbacks there are number of excellent use-cases for single URL input. One of the best is Wikipedia articles. Feed the Reader a URL pointing to a Wikipedia article. The Reader will cache the article itself, and then extract all the URLs the article uses as citations. The Reader will then cache the content of the citations, and then stop caching.

Similarly, a URL pointing to an open access journal article will function just like the Wikipedia article, and this will be even more fruitful if the citations are in the form of freely accessible URLs. Better yet, consider pointing the Reader to the root of an open access journal issue. If the site is not overly full of navigation links, and if the URLs to the content itself are not buried, then the whole of the issue will be harvested and analyzed.

Another good use-case is the home page of some sort of institution or organization. Want to know about Apple Computer, the White House, a conference, or a particular department of a university? Feed the root URL of any of these things to the Reader, and you will learn something. At the very least, you will learn how the organization prioritizes its public face. If things are more transparent than not, then you might be able to glean the names and addresses of the people in the organization, the public policies of the organization, or the breadth & depth of the organization.

Yet another excellent use-case includes blogs. Blogs often contain content at their root. Navigations links abound, but more often than not the navigation links point to more content. If the blog is well-designed, then the Reader may be able to create a corpus from the whole thing, and you can “read” it in one go.

A list of URLs

The third type of input is a list of URLs. The list is expected to be manifested as a plain text file, and each line in the file is a URL. Use whatever application you desire to build the list, but save the result as a .txt file, and you will probably have a plain text file.‡

Caveats? Like the single URL approach, the list of URLs must point to freely available content, and pointing to landing pages or splash pages is probably to be avoided. Unlike the single URL approach, the URLs in the list will not be used as starting points for Web crawling. Thus, if the list contains ten items, then ten items will be cached for analysis.

Another caveat is the actual process of creating the list; I have learned that is actually quite difficult to create lists of URLs. Copying & pasting gets old quickly. Navigating a site and right-clicking on URLs is tedious. While search engines & indexes often provide some sort of output in list format, the lists are poorly structured and not readily amenable to URL extraction. On the other hand, there are more than a few URL extraction tools. I use a Google Chrome extension called Link Grabber. [1] Install Link Grabber. Use Chrome to visit a site. Click the Link Grabber button, and all the links in the document will be revealed. Copy the links and paste them into a document. Repeat until you get tired. Sort and peruse the list of links. Remove the ones you don’t want. Save the result as a plain text file.‡ Feed the result to the Reader.

Despite these caveats, the list of URLs approach is enormously scalable; the list of URLs approach is the most scalable input option. Given a list of five or six items, the Reader will do quite well, but the Reader will operate just as well if the list contains dozens, hundreds, or even thousands of URLs. Imagine reading the complete works of your favorite author or the complete run of an electronic journal. Such is more than possible with the Distant Reader.‡

A zip file

The Distant Reader can take a zip file as input. Create a folder/directory on your computer. Copy just about any file into the folder/directory. Compress the file into a .zip file. Submit the result to the Reader.

Like the other approaches, there are a few caveats. First of all, the Reader is not able to accept .zip files whose size is greater than 64 megabytes. While we do it all the time, the World Wide Web was not really designed to push around files of any great size, and 64 megabytes is/was considered plenty. Besides, you will be surprised how many files can fit in a 64 megabyte file.

Second, the computer gods never intended file names to contain things other than simple Romanesque letters and a few rudimentary characters. Now-a-days our file names contain spaces, quote marks, apostrophes, question marks, back slashes, forward slashes, colons, commas, etc. Moreover, file names might be 64 characters long or longer! While every effort as been made to accomodate file names with such characters, your milage may vary. Instead, consider using file names which are shorter, simpler, and have some sort of structure. An example might be first word of author’s last name, first meaningful word of title, year (optional), and extension. Herman Melville’s Moby Dick might thus be named melville-moby.txt. In the end the Reader will be less confused, and you will be more able to find things on your computer.

There are a few advantages to the zip file approach. First, you can circumvent authorization restrictions; you can put licensed content into your zip files and it will be analyzed just like any other content. Second, the zip file approach affords you the opportunity to pre-process your data. For example, suppose you have downloaded a set of PDF files, and each page includes some sort of header or footer. You could transform each of these PDF files into plain text, use some sort of find/replace function to remove the headers & footers. Save the result, zip it up, and submit it to the Reader. The resulting analysis will be more accurate.

There are many use-cases for the zip file approach. Masters and Ph.D students are expected to read large amounts of material. Save all those things into a folder, zip them up, and feed them to the Reader. You have been given a set of slide decks from a conference. Zip them up and feed them to the Reader. A student is expected to read many different things for History 101. Download them all, put them in a folder, zip them up, and submit them to the Distant Reader. You have written many things but they are not on the Web. Copy them to a folder, zip them up, and “read” them with the… Reader.

A zip file with a companion CSV file

The final form of input is a zip file with a companion comma-separated value (CSV) file — a metadata file.

As the size of your corpus increases, so does the need for context. This context can often be manifested as metadata (authors, titles, dates, subject, genre, formats, etc.). For example, you might want to compare & contrast who wrote what. You will probably want to observe themes over space & time. You might want to see how things differ between different types of documents. To do this sort of analysis you will need to know metadata regarding your corpus.

As outlined above, the Distant Reader first creates a cache of content — a corpus. This is the raw data. In order to do any analysis against the corpus, the corpus must be transformed into plain text. A program called Tika is used to do this work. [2] Not only does Tika transform just about any file into plain text, but it also does its best to extract metadata. Depending on many factors, this metadata may include names of authors, titles of documents, dates of creation, number of pages, MIME-type, language, etc. Unfortunately, more often than not, this metadata extraction process fails and the metadata is inaccurate, incomplete, or simply non-existent.

This is where the CSV file comes in; by including a CSV file named “metadata.csv” in the .zip file, the Distant Reader will be able to provide meaningful context. In turn, you will be able to make more informed observations, and thus your analysis will be more thorough. Here’s how:

• assemble a set of files for analysis
• use your favorite spreadsheet or database application to create a list of the file names
• assign a header to the list (column) and call it “file”
• create one or more columns whose headers are “author” and/or “title” and/or “date”
• to the best of your ability, update the list with author, title, or date values for each file
• save the result as a CSV file named “metadata.csv” and put it in the folder/directory to be zipped
• compress the folder/directory to create the zip file
• submit the result to the Distant Reader for analysis

The zip file with a companion CSV file has all the strengths & weakness of the plain o’ zip file, but it adds some more. On the weakness side, creating a CSV file can be both tedious and daunting. On the other hand, many search engines & index export lists with author, title, and data metadata. One can use these lists as the starting point for the CSV file.♱ On the strength side, the addition of the CSV metadata file makes the Distant Reader’s output immeasurably more useful, and it leads the way to additional compare & contrast opportunities.

Summary

To date, the Distant Reader takes five different types of input. Each type has its own set of strengths & weaknesses:

• a file – good for a single large file; quick & easy; not scalable
• a URL – good for getting an overview of a single Web page and its immediate children; can include a lot of noise; has authorization limitations
• a list of URLs – can accomodate thousands of items; has authorization limitations; somewhat difficult to create list
• a zip file – easy to create; file names may get in the way; no authorization necessary; limited to 64 megabytes in size
• a zip file with CSV file – same as above; difficult to create metadata; results in much more meaningful reports & opportunities

Happy reading!

Notes & links

‡ Distant Reader Bounty #1: To date, I have only tested plain text files using line-feed characters as delimiters, such are the format of plain text files in the Linux and Macintosh worlds. I will pay $10 to the first person who creates a plain text file of URLs delimited by carriage-return/line-feed characters (the format of Windows-based text files) and who demonstrates that such files break the Reader. “On you mark. Get set. Go!”

‡ Distant Reader Bounty #2: I will pay $20 to the first person who creates a list of 2,000 URLs and feeds it to the Reader.

♱ Distant Reader Bounty #3: I will pay $30 to the first person who writes a cross-platform application/script which successfully transforms a Zotero bibliography into a Distant Reader CSV metadata file.

[1] Link Grabber – http://bit.ly/2mgTKsp

[2] Tika – http://tika.apache.org

We invite you to write a cool hack enabling students & scholars to “read” an arbitrarily large corpus of textual materials.

Introduction

A website called The Distant Reader takes an arbitrary number of files or links to files as input. [1] The Reader then amasses the files locally, transforms them into plain text files, and performs quite a bit of natural language processing against them. [2] The result — the the form of a file system — is a set of operating system independent indexes which point to individual files from the input. [3] Put another way, each input file is indexed in a number of ways, and therefore accessible by any one or combination of the following attributes:

• any named entity (name of person, place, date, time, money amount, etc)
• any part of speech (noun, verb, adjective, etc.)
• email address
• free text word
• readability score
• size of file
• statistically significant keyword
• textual summary
• URL

All of things listed above are saved as plain text files, but they have also been reduced to an SQLite database (./etc/reader.db), which is also distributed with the file system.

The Challenge

Your mission, if you choose to accept it, is to write a cool hack against the Distant Reader’s output. By doing so, you will be enabling people to increase their comprehension of the given files. Here is a list of possible hacks:

• create a timeline – The database includes a named entities table (ent). Each entity is denoted by a type, and one of those types is “PERSON”. Find all named entities of type PERSON, programmatically look them up in Wikidata, extract the entity’s birth & death dates, and plot the result on a timeline. As an added bonus, update the database with the dates. Alternatively, and possibly more simply, find all entities of type DATE (or TIME), and plot those values on a timeline.
• create a map – Like the timeline hack, find all entities denoting places (GRE or LOC), look up their geographic coordinates in Wikidata, and plot them on a map. As an added bonus, update the database with the coordinates.
• order documents based on similarity – “Find more like this one” is a age-old information retrieval use case. Given a reference document – a document denoted as particularly relevant — create a list of documents from the input which are similar to the reference document. For example, create a vector denoting the characteristics of the reference document. [4] Then create vectors for each document in the collection. Finally, use something like the Cosine Similarly algorithm to determine which documents are most similar (or different). [5] The reference document may be from either inside or outside the Reader’s file system, for example, the Bible or Shakespeare’s Hamlet.
• write a Javascript interface to the database – The Distant Reader’s database (./etc/reader.db) is manifested as a single SQLite file. There exists a Javascript library enabling one to read & write to SQLite databases. [6] Sans a Web server, write sets of HTML pages enabling a person to query the database. Example queries might include: find all documents where Plato is a keyword, find all sentences where Plato is a named entity, find all questions, etc. The output of such queries can be HTML pages, but almost as importantly, they can be CSV files so people can do further analysis. As an added bonus, enable a person to update the database so things like authors, titles, dates, genres, or notes can be assigned to items in the bib table.
• list what is being bought or sold – Use the entities table (ent) to identify all the money amounts (type equals “MONEY”) and the sentences from which they appear. Extract all of those sentences, analyze the sentence, and output the things being tendered. You will probably have to join the id and sentenced id in the ent table with the id and sentence id in the pos table to implement this hack. As an added bonus, calculate how much things would cost in today’s dollars or any other currency.
• normalize metadata – The values in the named entities table (ent) are often repeated in various forms. For example, a value may be Plato, plato, or PLATO. Use something like the Levenshtein distance algorithm to normalize each value into something more consistent. [7]
• prioritize metadata – Just because a word is frequent does not mean it is significant. A given document may mention Plato many times, but if Plato is mentioned in each and every document, then the word is akin to noise. Prioritize given named entities, specifically names, through the use of a something like TFIDF. Calculate a TFIDF score for a given word, and if the word is above a given threshold, then update the database accordingly. [8]
• extract sentences matching a given grammer – Each & every word, punctuation mark, and part of speech of each & every document is enumerated and stored in the pos table of the database. Consequently it is rather easy to find all questions in the database and extract them. (Find all sentences ids where punctuation equals “?”. Find all words (tokens) with the same id and sentence id. Output all tokens sorted by token id.) Similarly, it is possible to find all sentences where a noun precedes a verb which precedes another noun. Or, find all sentences where a noun precedes a verb which is followed by the word “no” or “not” which precedes another noun. Such queries find sentence in the form of “cat goes home” or “dog is not cat”. Such are assertive sentences. A cool hack would be to identify sentences of any given grammer such as adjective-noun or entity-verb where the verb is some form of the lemma to be (is, was, are, were, etc.), as in “Plato is” or “Plato was”. The adjective-noun patterns is of particular interest, especially given a particular noun. Find all sentences matching the pattern adjective-king to learn how the king was described.
• create a Mad Lib – This one is off the wall. Identify (random) items of interest from the database. Write a template in the form of a story. Fill in the template with the items of interest. Done. The “better” story would be one that is complete with “significant” words from the database; the better story would be one that relates truths from the underlying content. For example, identify the two most significant nouns. Identify a small handful of the most significant verbs. Output simple sentences in the form of noun-verb-noun.
• implement one of two search engines – The Distant Reader’s output includes a schema file (./etc/schema.xml) defining the structure of a possible Solr index. The output also includes an indexer (./bin/db2solr.pl) as well as a command-line interface (./bin/search-solr.pl) to search the index. Install Solr. Create an index with the given schema. Index the content. Write a graphical front-end to the index complete with faceted search functionality. Allow search results to be displayed and saved in a tabular form for further analysis. The Reader’s output also includes a semantic index (./etc/reader.vec) a la word2vec, as well as a command-line interface (./bin/search-vec.py) for querying the semantic index. Write a graphical interface for querying the semantic index.

Sample data

In order for you to do your good work, you will need some Distant Reader output. Here are pointers to some such stuff:

• Cultural Analytics – http://dh.crc.nd.edu/sandbox/reader/hackaton/cultural-analytics.zip
  • 318,287 words; 33 documents; 67 uncompressed MB
  • all articles from a journal named Cultural Analytics
• Plato – http://dh.crc.nd.edu/sandbox/reader/hackaton/plato.zip
  • 929,704 words; 24 documents; 112 uncompressed MB
  • the complete works of Plato
• Code4Lib Journal – http://dh.crc.nd.edu/sandbox/reader/hackaton/code4lib-journal.zip
  • 1,234,348 words; 303 documents; 286 uncompressed MB
  • all articles from a journal named Code4Lib Journal
• aesthetics – http://dh.crc.nd.edu/sandbox/reader/hackaton/aesthetics.zip
  • 2,296,890 words; 37 documents; 287 uncompressed MB
  • books classified as the philosophy of art
• love stories – http://dh.crc.nd.edu/sandbox/reader/hackaton/love-stories.zip
  • 238,374,038 words; 460 documents; 5.94 uncompressed GB
  • books classified as love stories

Helpful hint

With the exception of only a few files (./etc/reader.db, ./etc/reader.vec, and ./cache/*), all of the files in the Distant Reader’s output are plain text files. More specifically, they are either unstructured data files or delimited files. Despite any file’s extension, the vast majority of the files can be read with your favorite text editor, spreadsheet, or database application. To read the database file (./etc/reader.db), you will need an SQLite application. The files in the adr, bib, ent, pos, urls, or wrd directories are all tab delimited files. A program called OpenRefine is a WONDERFUL tool for reading and analyzing tab delimited files. [9] In fact, a whole lot can be learned through the skillful use of OpenRefine against the tab delimited files.

Notes

[1] The home page of the Distant Reader is https://distantreader.org

[2] All of the code doing this processing is available on GitHub. See https://github.com/ericleasemorgan/reader

[3] This file system is affectionately known as a “study carrel”.

[4] A easy-to-use library for creating such vectors is a part of the Scikit Learn suite of software. See http://bit.ly/2F5EoxA

[5] The algorithm is described at https://en.wikipedia.org/wiki/Cosine_similarity, and a SciKit Learn module is available at http://bit.ly/2IaYcS3

[6] The name of the library is called sql.js and it is available at https://github.com/kripken/sql.js/

[7] The Levenshtein distance is described here — https://en.wikipedia.org/wiki/Levenshtein_distance, and various libraries doing the good work are outlined at http://bit.ly/2F30roM

[8] Yet another SciKit Learn module may be of use here — http://bit.ly/2F5o2oS

[9] OpenRefine eats delimited files for lunch. See http://openrefine.org