extract semi-structured text from Word documents

Question

I want to text-mine a set of files based on the below form. I can create a corpus where each file is a document (using tm), but I'm thinking it might be better to create a corpus where each section in the 2nd form table was a document having the following meta data:

  Author       : John Smith
  DateTimeStamp: 2013-04-18 16:53:31
  Description  : 
  Heading      : Current Focus
  ID           : Smith-John_e.doc Current Focus
  Language     : en_CA
  Origin       : Smith-John_e.doc
  Name         : John Smith
  Title        : Manager
  TeamMembers  : Joe Blow, John Doe
  GroupLeader  : She who must be obeyed 

where Name, Title, TeamMembers and GroupLeader are extracted from the first table on the form. In this way, each chunk of text to be analyzed would maintain some of its context.

What is the best way to approach this? I can think of 2 ways:

• somehow parse the corpus I have into child corpora.
• somehow parse the document into subdocuments and make a corpus from those.

Any pointers would be much appreciated.

This is the form:

Here is an RData file of a corpus with 2 documents. exc[[1]] came from a .doc and exc[[2]] came from a docx. They both used the form above.

Answer 1

Here's a quick sketch of a method, hopefully it might provoke someone more talented to stop by and suggest something more efficient and robust... Using the RData file in your question, I found that the doc and docx files have slightly different structures and so require slightly different approaches (though I see in the metadata that your docx is 'fake2.txt', so is it really docx? I see in your other Q that you used a converter outside of R, that must be why it's txt).

library(tm)

First get custom metadata for the doc file. I'm no regex expert, as you can see, but it's roughly 'get rid of trailing and leading spaces' then 'get rid of "Word"', then get rid of punctuation...

# create User-defined local meta data pairs
meta(exc[[1]], type = "corpus", tag = "Name1") <- gsub("^\\s+|\\s+$","", gsub("Name", "", gsub("[[:punct:]]", '', exc[[1]][3])))
meta(exc[[1]], type = "corpus", tag = "Title") <- gsub("^\\s+|\\s+$","", gsub("Title", "", gsub("[[:punct:]]", '', exc[[1]][4])))
meta(exc[[1]], type = "corpus", tag = "TeamMembers") <- gsub("^\\s+|\\s+$","", gsub("Team Members", "", gsub("[[:punct:]]", '', exc[[1]][5])))
meta(exc[[1]], type = "corpus", tag = "ManagerName") <- gsub("^\\s+|\\s+$","", gsub("Name of your", "", gsub("[[:punct:]]", '', exc[[1]][7])))

Now have a look at the result

    # inspect
    meta(exc[[1]], type = "corpus")
Available meta data pairs are:
  Author       : 
  DateTimeStamp: 2013-04-22 13:59:28
  Description  : 
  Heading      : 
  ID           : fake1.doc
  Language     : en_CA
  Origin       : 
User-defined local meta data pairs are:
$Name1
[1] "John Doe"

$Title
[1] "Manager"

$TeamMembers
[1] "Elise Patton Jeffrey Barnabas"

$ManagerName
[1] "Selma Furtgenstein"

Do the same for the docx file

# create User-defined local meta data pairs
meta(exc[[2]], type = "corpus", tag = "Name2") <- gsub("^\\s+|\\s+$","", gsub("Name", "", gsub("[[:punct:]]", '', exc[[2]][2])))
meta(exc[[2]], type = "corpus", tag = "Title") <- gsub("^\\s+|\\s+$","", gsub("Title", "", gsub("[[:punct:]]", '', exc[[2]][4])))
meta(exc[[2]], type = "corpus", tag = "TeamMembers") <- gsub("^\\s+|\\s+$","", gsub("Team Members", "", gsub("[[:punct:]]", '', exc[[2]][6])))
meta(exc[[2]], type = "corpus", tag = "ManagerName") <- gsub("^\\s+|\\s+$","", gsub("Name of your", "", gsub("[[:punct:]]", '', exc[[2]][8])))

And have a look

# inspect
meta(exc[[2]], type = "corpus")
Available meta data pairs are:
  Author       : 
  DateTimeStamp: 2013-04-22 14:06:10
  Description  : 
  Heading      : 
  ID           : fake2.txt
  Language     : en
  Origin       : 
User-defined local meta data pairs are:
$Name2
[1] "Joe Blow"

$Title
[1] "Shift Lead"

$TeamMembers
[1] "Melanie Baumgartner Toby Morrison"

$ManagerName
[1] "Selma Furtgenstein"

If you have a large number of documents then a lapply function that includes these meta functions would be the way to go.

Now that we've got the custom metadata, we can subset the documents to exclude that part of the text:

# create new corpus that excludes part of doc that is now in metadata. We just use square bracket indexing to subset the lines that are the second table of the forms (slightly different for each doc type)
excBody <- Corpus(VectorSource(c(paste(exc[[1]][13:length(exc[[1]])], collapse = ","), 
                      paste(exc[[2]][9:length(exc[[2]])], collapse = ","))))
# get rid of all the white spaces
excBody <- tm_map(excBody, stripWhitespace)

Have a look:

inspect(excBody)
A corpus with 2 text documents

The metadata consists of 2 tag-value pairs and a data frame
Available tags are:
  create_date creator 
Available variables in the data frame are:
  MetaID 

[[1]]
|CURRENT RESEARCH FOCUS |,| |,|Lorem ipsum dolor sit amet, consectetur adipiscing elit. |,|Donec at ipsum est, vel ullamcorper enim. |,|In vel dui massa, eget egestas libero. |,|Phasellus facilisis cursus nisi, gravida convallis velit ornare a. |,|MAIN AREAS OF EXPERTISE |,|Vestibulum aliquet faucibus tortor, sed aliquet purus elementum vel. |,|In sit amet ante non turpis elementum porttitor. |,|TECHNOLOGY PLATFORMS, INSTRUMENTATION EMPLOYED |,| Vestibulum sed turpis id nulla eleifend fermentum. |,|Nunc sit amet elit eu neque tincidunt aliquet eu at risus. |,|Cras tempor ipsum justo, ut blandit lacus. |,|INDUSTRY PARTNERS (WITHIN THE PAST FIVE YEARS) |,| Pellentesque facilisis nisl in libero scelerisque mattis eu quis odio. |,|Etiam a justo vel sapien rhoncus interdum. |,|ANTICIPATED PARTICIPATION IN PROGRAMS, EITHER APPROVED OR UNDER DEVELOPMENT |,|(Please include anticipated percentages of your time.) |,| Proin vitae ligula quis enim vulputate sagittis vitae ut ante. |,|ADDITIONAL ROLES, DISTINCTIONS, ACADEMIC QUALIFICATIONS AND NOTES |,|e.g., First Aid Responder, Other languages spoken, Degrees, Charitable Campaign |,|Canvasser (GCWCC), OSH representative, Social Committee |,|Sed nec tellus nec massa accumsan faucibus non imperdiet nibh. |,,

[[2]]
CURRENT RESEARCH FOCUS,,* Lorem ipsum dolor sit amet, consectetur adipiscing elit.,* Donec at ipsum est, vel ullamcorper enim.,* In vel dui massa, eget egestas libero.,* Phasellus facilisis cursus nisi, gravida convallis velit ornare a.,MAIN AREAS OF EXPERTISE,* Vestibulum aliquet faucibus tortor, sed aliquet purus elementum vel.,* In sit amet ante non turpis elementum porttitor. ,TECHNOLOGY PLATFORMS, INSTRUMENTATION EMPLOYED,* Vestibulum sed turpis id nulla eleifend fermentum.,* Nunc sit amet elit eu neque tincidunt aliquet eu at risus.,* Cras tempor ipsum justo, ut blandit lacus.,INDUSTRY PARTNERS (WITHIN THE PAST FIVE YEARS),* Pellentesque facilisis nisl in libero scelerisque mattis eu quis odio.,* Etiam a justo vel sapien rhoncus interdum.,ANTICIPATED PARTICIPATION IN PROGRAMS, EITHER APPROVED OR UNDER DEVELOPMENT ,(Please include anticipated percentages of your time.),* Proin vitae ligula quis enim vulputate sagittis vitae ut ante.,ADDITIONAL ROLES, DISTINCTIONS, ACADEMIC QUALIFICATIONS AND NOTES,e.g., First Aid Responder, Other languages spoken, Degrees, Charitable Campaign Canvasser (GCWCC), OSH representative, Social Committee,* Sed nec tellus nec massa accumsan faucibus non imperdiet nibh.,,

Now the documents are ready for text mining, with the data from the upper table moved out of the document and into the document metadata.

Of course all of this depends on the documents being highly regular. If there are different numbers of lines in the first table in each doc, then the simple indexing method might fail (give it a try and see what happens) and something more robust will be needed.

UPDATE: A more robust method

Having read the question a little more carefully, and got a bit more education about regex, here's a method that is more robust and doesn't depend on indexing specific lines of the documents. Instead, we use regular expressions to extract text from between two words to make the metadata and split the document

Here's how we make the User-defined local meta data (a method to replace the one above)

library(gdata) # for the trim function
txt <- paste0(as.character(exc[[1]]), collapse = ",")

# inspect the document to identify the words on either side of the string
# we want, so 'Name' and 'Title' are on either side of 'John Doe'
extract <- regmatches(txt, gregexpr("(?<=Name).*?(?=Title)", txt, perl=TRUE))
meta(exc[[1]], type = "corpus", tag = "Name1") <- trim(gsub("[[:punct:]]", "", extract))

extract <- regmatches(txt, gregexpr("(?<=Title).*?(?=Team)", txt, perl=TRUE))
meta(exc[[1]], type = "corpus", tag = "Title") <- trim(gsub("[[:punct:]]","", extract))

extract <- regmatches(txt, gregexpr("(?<=Members).*?(?=Supervised)", txt, perl=TRUE))
meta(exc[[1]], type = "corpus", tag = "TeamMembers") <- trim(gsub("[[:punct:]]","", extract))

extract <- regmatches(txt, gregexpr("(?<=your).*?(?=Supervisor)", txt,  perl=TRUE))
meta(exc[[1]], type = "corpus", tag = "ManagerName") <- trim(gsub("[[:punct:]]","", extract))

# inspect
meta(exc[[1]], type = "corpus")

Available meta data pairs are:
  Author       : 
  DateTimeStamp: 2013-04-22 13:59:28
  Description  : 
  Heading      : 
  ID           : fake1.doc
  Language     : en_CA
  Origin       : 
User-defined local meta data pairs are:
$Name1
[1] "John Doe"

$Title
[1] "Manager"

$TeamMembers
[1] "Elise Patton Jeffrey Barnabas"

$ManagerName
[1] "Selma Furtgenstein"

Similarly we can extract the sections of your second table into separate vectors and then you can make them into documents and corpora or just work on them as vectors.

txt <- paste0(as.character(exc[[1]]), collapse = ",")
CURRENT_RESEARCH_FOCUS <- trim(gsub("[[:punct:]]","", regmatches(txt, gregexpr("(?<=CURRENT RESEARCH FOCUS).*?(?=MAIN AREAS OF EXPERTISE)", txt, perl=TRUE))))
[1] "Lorem ipsum dolor sit amet consectetur adipiscing elit                             Donec at ipsum est vel ullamcorper enim                                            In vel dui massa eget egestas libero                                               Phasellus facilisis cursus nisi gravida convallis velit ornare a"


MAIN_AREAS_OF_EXPERTISE <- trim(gsub("[[:punct:]]","", regmatches(txt, gregexpr("(?<=MAIN AREAS OF EXPERTISE).*?(?=TECHNOLOGY PLATFORMS, INSTRUMENTATION EMPLOYED)", txt, perl=TRUE))))
    [1] "Vestibulum aliquet faucibus tortor sed aliquet purus elementum vel                 In sit amet ante non turpis elementum porttitor"

And so on. I hope that's a bit closer to what you're after. If not, it might be best to break down your task into a set of smaller, more focused questions, and ask them separately (or wait for one of the gurus to stop by this question!).