fulltext v1: text-mining scholarly works

  Scott Chamberlain JANUARY 17, 2018

The problem

Text-mining - the art of answering questions by extracting patterns, data, etc. out of the published literature - is not easy.

It’s made incredibly difficult because of publishers. It is a fact that the vast majority of publicly funded research across the globe is published in paywall journals. That is, taxpayers pay twice for research: once for the grant to fund the work, then again to be able to read it. These paywalls mean that every potential person text-mining will have different access: some have access through their university, some may have access through their company, and others may only have access to whatever happens to be open access. On top of that, access for paywall journals often depends on your IP address - something not generally on top of mind for most people.

Another hardship with text-mining is the huge number of publishers together with no standardized way to figure out the URL for full text versions of a scholarly work. There is the DOI (Digital Object Identifier) system used by Crossref, Datacite and others, but those generally help you sort out the location of the scholarly work on a web page - the html version. What one probably wants for text-mining is the PDF or XML version if available. Publishers can optionally choose to include URLs for full text (PDF and/or XML) with Crossref’s metadata (e.g., see this Crossref API call and search for “link” on the page), but the problem is that it’s optional.

fulltext is a package to help R users address the above problems, and get published literature from the web in it’s many forms, and across all publishers.



the fulltext package

fulltext tries to make the following use cases as easy as possible:

  • Search for articles
  • Fetch abstracts
  • Fetch full text articles
  • Get links for full text articles (xml, pdf)
  • Extract text from articles
  • Collect sections of articles that you actually need (e.g., titles)
  • Download supplementary materials

fulltext organizes functions around the above use cases, then provides flexiblity to query many data sources within that use case (i.e. function). For example fulltext::ft_search searches for articles - you can choose among one or more of many data sources to search, passing options to each source as needed.


What does a workflow with fulltext look like?

  • Search for articles with ft_search()
  • Fetch articles with ft_get() using the output of the previous step
  • Collect the text into an object with ft_collect()
  • Extract sections of articles needed with ft_chunks(), or
  • Combine texts into a data.frame ready for quanteda or similar text-mining packages


Package overhaul

fulltext has undergone a re-organization, which includes a bump in the major version to v1 to reinforce the large changes the package has undergone. Changes include:

  • Function name standardization with the ft_ prefix. e,g, chunks is now ft_chunks
  • ft_get has undergone major re-organization - biggest of which may be that all full text XML/plain text/PDF goes to disk to simplify the user interface. Along with this we’ve changed to using DOIs/IDs as the file names
  • We no longer store files as rds - but as the format they are, pdf, txt or xml
  • storr is now imported to manage mapping between real DOIs and file paths that include normalized DOIs - and aids in the function ft_table() for creating a data.frame of text results
  • Note that with the ft_get() overhaul, the only option is to write to disk. Before we attempted to provide many different options for saving XML and PDF data, but it was too complicated. This has implications for using the output of ft_get() - the output includes the paths to the files - use ft_collect() to collect the text if you want to use ft_chunks() or other fulltext functions downstream.
  • A number of functions have been removed to further hone the scope of the package
  • A function ft_abstract is introduced to fetch abstracts for when you just need abstracts
  • A function ft_table has been introduced to gather all your documents into a data.frame to make it easy to do downstream analyses with other packages
  • Two new data sources have been added: Scopus and Microsoft Academic - both of which are available via ft_search() and ft_abstract()
  • New functions have been added for the user to find out what plugins are available: ft_get_ls(), ft_links_ls(), and ft_search_ls()

We’ve battle tested ft_get() on a lot of DOIs - but there still may be errors - let us know if you have any problems.


Documentation

Along with an overhual of the package we have made a new manual for fulltext. Check it out at https://ropensci.github.io/fulltext-book/



Setup

Install fulltext

at the time of writing binaries are not yet available on CRAN, so you’ll have to install from source from CRAN (which shouldn’t provide any problems since there’s no compiled code in the package), or install from GitHub

install.packages("fulltext")

Or get the development version:

devtools::install_github("ropensci/fulltext")
library(fulltext)

Below I’ll discuss some of the new features of the package, and not do an exhaustive tutorial to the package. Check out the manual for more details: https://ropensci.github.io/fulltext-book/


Fetch abstracts: ft_abstract

ft_abstract() is a new function in fulltext. It gives you access to absracts from the following data sources:

  • crossref
  • microsoft
  • plos
  • scopus

A quick example. Search for articles in PLOS.

res <- ft_search(query = 'biology', from = 'plos', limit = 10, 
   plosopts = list(fq = list('doc_type:full', '-article_type:correction',
                  '-article_type:viewpoints')))

Now pass the DOIs to ft_abstract() to get abstracts:

ft_abstract(x = res$plos$data$id, from = "plos")
## <fulltext abstracts>
## Found:
##   [PLOS: 90; Scopus: 0; Microsoft: 0; Crossref: 0]


Fetch articles: ft_get

The function ft_get is the workhorse for getting full text articles.

Using this function can be tricky depending on where you want to get articles from. While searching (ft_search) usually doesn’t present any barriers or stumbling blocks because search portals are generally open (except Web Of Science), ft_get can get frustrating because so many publishers paywall their articles. The combination of paywalls and their patchwork of who gets to get through them means that we can’t easily predict who will run into problems with Elsevier, Wiley, Springer, etc. (well, mostly those big three because they publish such a large portion of the papers).

With this version we’ve tried to bulk up the documentation as much as possible (see the manual) to make jumping over these barriers as painless as possible.

Let’s do an example to demonstrate how to use ft_get() and some of the new features.

Get DOIs from PLOS (excluding partial document types)

library("rplos")
dois <- searchplos(q="*:*", fl='id',
   fq=list('doc_type:full',"article_type:\"research article\""),
     limit=5)$data$id

Once we have DOIs we can go to ft_get():

res <- ft_get(dois, from = "plos")

Internally, ft_get() attemps to write the file to disk if we can successfully access the file - if an error occurs for any reason (see ft_get errors in the manual) we delete that file so you don’t end up with partial/empty files.

Since ft_get() writes files to your machine’s disk, even if a function call to ft_get() fails at some point in the process, the articles that we’ve successfully retrieved aren’t affected.

In addition, we have fixed reusing cached files on disk. Thus, even if you get a failure in a call to ft_get() you can rerun it again and those files already retrieved will make the function call faster.

Having a look at the output of ft_get(), we can see that only one list element (plos) has data in it because we only searched for articles from one publisher.

vapply(res, function(x) is.null(x$data), logical(1))
##     plos   entrez    elife  pensoft    arxiv  biorxiv elsevier    wiley 
##    FALSE     TRUE     TRUE     TRUE     TRUE     TRUE     TRUE     TRUE

The output for elements searched for are the following in a list:

  • found: number of works retrieved
  • dois: character vector of DOIs
  • data (a list)
    • backend: the backend
    • cache_path: the root cache path
    • path
      • DOI (a list named by a DOI is repeated for each DOI)
        • path: the complete path to the file on disk
        • id: the id (usually a DOI)
        • type: xml, plain or pdf
        • error: an error message, if any
    • data: this is NULL until you use ft_collect()
  • opts: your options

The backend can only be one value - ext, representing file extension. We’re retaining that information for now because we may decide to add additional backends in the future.

In the last version of fulltext you could get the extracted text from XML or PDF in the output of ft_get(). This is changed. You now only get metadata and the path to the file on disk. To get text into the object is a separate function call to ft_collect()

ft_collect(res)

Which returns the same class object as ft_get() returns, but the data slot is now populated with the text.

Remember that wit this change you now should use ft_collect() before passing the output of ft_get() to ft_chunks()

ft_get(dois, from = "plos") %>% 
  ft_collect() %>% 
  ft_chunks(c("doi","history")) %>% 
  ft_tabularize()


Extract text: ft_extract

ft_extract() used to have options for extracting text from PDF’s with different pieces of software. To simplify the function it only uses the pdftools package.

path <- system.file("examples", "example1.pdf", package = "fulltext")
ft_extract(path)
## <document>/Library/Frameworks/R.framework/Versions/3.4/Resources/library/fulltext/examples/example1.pdf
##   Title: Suffering and mental health among older people living in nursing homes---a mixed-methods study
##   Producer: pdfTeX-1.40.10
##   Creation date: 2015-07-17



Gather text into a data.frame: ft_table

ft_table() is a new function to gather the text from all your articles into a data.frame. This should simplify analysis for most users.

(x <- ft_table())
## # A tibble: 192 x 4
##    dois                       ids_norm                   text        paths
##  * <chr>                      <chr>                      <chr>       <chr>
##  1 10.1002/9783527696109.ch41 10_1002_9783527696109_ch41 "         … /Use…
##  2 10.1002/chin.199038056     10_1002_chin_199038056     "ChemInfor… /Use…
##  3 10.1002/cite.330221605     10_1002_cite_330221605     " Versamml… /Use…
##  4 10.1002/dvg.22402          10_1002_dvg_22402          "C 2013 Wi… /Use…
##  5 10.1002/jctb.5010090209    10_1002_jctb_5010090209    "         … /Use…
##  6 10.1002/qua.560200801      10_1002_qua_560200801      "Internati… /Use…
##  7 10.1002/risk.200590063     10_1002_risk_200590063     "         … /Use…
##  8 10.1002/scin.5591692420    10_1002_scin_5591692420    "Books\n  … /Use…
##  9 10.1006/bbrc.1994.2001     10_1006_bbrc_1994_2001     "http://ap… /Use…
## 10 10.1007/11946465_42        10_1007_11946465_42        " Hoon Cho… /Use…
## # ... with 182 more rows

(you can optionally only extract text from PDFs, or only from XMLs)

We give the the DOI, the normalized DOI that we used for the file path, the text, and the file path. You can then use this output in quanteda or other text-mining packages (the function quanteda::kwic() is for locating keywords in context):

library(quanteda)
z <- corpus(x, docid_field="dois", text_field="text")
head(quanteda::kwic(z, "cell"))
##                                                                           
##   [10.1002/9783527696109.ch41, 253] Basic Concepts A lithium-ion battery |
##   [10.1002/9783527696109.ch41, 397]     in a typical lithium-ion battery |
##   [10.1002/9783527696109.ch41, 461]     in a typical lithium-ion battery |
##   [10.1002/9783527696109.ch41, 764]        material 1 Lithium LCO LiCoO2 |
##  [10.1002/9783527696109.ch41, 2744]               of about 3.6-3.8 V per |
##  [10.1002/9783527696109.ch41, 6237]                 nate/ anode and hour |
##                                          
##  cell | consists of a positive electrode 
##  cell | . material. During discharging   
##  cell | . The main reactions occurring   
##  Cell | phones, High capacity cobalt     
##  cell | and highest energy densities with
##  cell | cathode 4. Electrovaya,



Todo

We have lots of ideas to make fulltext even better. Check out what we’ll be working on in the issue tracker.



Feedback!

Please do upgrade/install fulltext v1.0.0 and let us know what you think.