Text structure extract from PDF brainstorming

1. Use existing tools like grobid.

https://github.com/kermitt2/grobid

Use machine learning to get scientific paper structure data.  It has demo page at here, http://cloud.science-miner.com/grobid/ . My tests show that it can extract title and some other data, but still content maybe mixed with footer and headers.  As this one is designed for academic docs, so it may have some issues to other types of PDF.

2. Borrow the ideas like grobid, to build a system to adapt to the PDF types that you use.
Unified format PDf may get better result, but not for general PDF.

3. Convert a PDF to doc file, and then use doc tools to extract content structures out?

 

Some brain storming ideas:

1). Use tool like pdfclown to extract position and style info of the PDF text data.

2). Then for same category PDF share patterned style and positions, we have chance to find the structures of file.

3). And based on the style of text, it is possible have a tree-like text structure, but this tree may not match with the real chapters tree. This method can help on section levels and titles

4). How to find the main content text?
By statistical info of the text font, the biggest portion of the occurrence normally the text content section. As the main content has different styles with other part, it is possible get good result here.

5). By check from bottom of each page to center’s first line, if PDF have many pages and a unified footer format, then it possible to find what style and font is for footer, and possible find out the footer pattern by statistical.

6). Use same trick of footer, it is possible find out the header if page have.

7). If we have position and style info of the PDF text data.
Then we maybe can do classification based on the position and styles training too to find the basic structure of file.

 

Some similar work and papers

a chinese patente for this:
https://patents.google.com/patent/CN107358208A/zh

a ch team works called Xed

https://diuf.unifr.ch/main/diva/research/research-projects/xed
http://diuf.unifr.ch/diva/siteDIVA04/publications/XedDIAL04.pdf

LA-PDFText
https://scfbm.biomedcentral.com/articles/10.1186/1751-0473-7-7
https://github.com/GullyAPCBurns/lapdftext

HP paper
https://www.researchgate.net/publication/220932927_Layout_and_Content_Extraction_for_PDF_Documents
https://link.springer.com/content/pdf/10.1007%2F978-3-540-28640-0_20.pdf

a text classification algorithm
https://www.sciencedirect.com/science/article/pii/S153204641630017X
that follows a multi-pass sieve framework to automatically classify PDF text snippets (for brevity, texts) into TITLE, ABSTRACT, BODYTEXT, SEMISTRUCTURE, and METADATA categories.

Advertisements

Tools for PDF text extract

PDF Tools list

Jpedal (commercial software)
https://www.snowtide.com/ (commercial software)
itext (commercial software)

apache tika

Home

Grobid

LA-PDFText
pdftotext
pftohtml
pdftoxml
PdfBox
pdf2xml
LA-PdfText
PdfMiner
pdfXtk
pdf-extract
pdfx
PDFExtract

Icecite

 

1. Tools compare and benchmark

http://okfnlabs.org/blog/2016/04/19/pdf-tools-extract-text-and-data-from-pdfs.html
http://ad-publications.informatik.uni-freiburg.de/benchmark.pdf

PDF format
https://stackoverflow.com/questions/88582/structure-of-a-pdf-file

2. How to remove the header and footer from PDF

http://www.massapi.com/class/pd/PDFTextStripperByArea.html
https://www.programcreek.com/java-api-examples/index.php?api=org.apache.pdfbox.util.PDFTextStripperByArea
http://what-when-how.com/itext-5/parsing-pdfs-part-2-itext-5/

3. Find paragraph in text

https://stackoverflow.com/questions/39196676/how-to-read-a-paragraph-from-a-file-in-java

https://stackoverflow.com/questions/14990619/getting-paragraph-count-from-tika-for-both-word-and-pdf

4. Find sentences

https://stackoverflow.com/questions/9492707/how-can-i-split-a-text-into-sentences-using-the-stanford-parser

7. Structure extract from PDF

 

https://github.com/kermitt2/grobid
https://diuf.unifr.ch/main/diva/research/research-projects/xed
http://diuf.unifr.ch/diva/siteDIVA04/publications/XedDIAL04.pdf
https://github.com/GullyAPCBurns/lapdftext
https://www.researchgate.net/publication/220932927_Layout_and_Content_Extraction_for_PDF_Documents
https://link.springer.com/content/pdf/10.1007%2F978-3-540-28640-0_20.pdf
https://www.sciencedirect.com/science/article/pii/S153204641630017X

How to build and run the CMU Olympus-Ravenclaw dialog system framework – 3

How the Olympus system works ? This is a summary after I read a bit of the code of Olympus.

1. How system started?

Open your SystemRun.bat, it will call this line:

START “” /DConfigurations\%RunTimeConfig% “%OLYMPUS_ROOT%\Agents\Pythia\dist\process_monitor.exe” %RunTimeStartList%.config

About the startlist.config file and Pythia’s process_monitor.exe (MITRE in old project) , you need to read this page:

http://wiki.speech.cs.cmu.edu/olympus/index.php/Pythia

http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/contrib/MITRE/tools/docs/process_monitor_tut.html
http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/index.html
Pythia is windows process manager that control many process start and stop. Written by pl and build as process_monitor.exe, it will read in the startlist.config file to control each process of system.

2. How each module started?

Pythia started many process and they communicate each other, each process has its ability as module in system, like ASR, TTS. They work together to make whole system work.

Example:
TTYRecongnitionServer is the module to interface with terminal in audio and keyboard input way. Pythia in fact will start this file : ttyserver_start.config to start this process. It will in fact run this cmd as one process:

$OLYMPUS_ROOT\bin\release\TTYRecognitionServer.exe 
--start - is to let Pythia start it
--input_line - is to let Pythia open a input box for it on the GUI.

3. What is the HUB?

So how each process feature doing and how these processes communicate with each other is very important now. Each process will be called a server, so there is a HUB to link all the servers together:

http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/index.html
http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/reference/hub.html

So what is a server?
http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/reference/adding.html

5. So how hub and server exchange data?

There is pgm file defined all servers info, name port, and rules etc, hub just read this file  and then will link all servers to exchange data.
Rules in programs (like main) tell Galaxy what the Hub should do when it gets a certain message

http://wiki.speech.cs.cmu.edu/olympus/index.php/How_does_the_hub_work

http://communicator.sourceforge.net/sites/MITRE/distributions/GalaxyCommunicator/docs/manual/tutorial/how_it_works.html

Now you have basic structure of the whole system.

http://wiki.speech.cs.cmu.edu/olympus/index.php/How_The_CMU_Communicator_Architecture_Works

 

5. So how task is organized as dialog system (ravenclaw) ?

ravenclaw use tree to define the task relations, like example here:

 

Sub node under a task is a sub task, to finish a task you need to go through from left to right to do each task. And they define a set of macro in C and developer need to use these set of macro to define this task tree:

example like this:

http://trac.speech.cs.cmu.edu/repos/olympus/tutorials/Tutorial1/branches/2.5/Agents/RavenClawDM/DialogTask.cpp