Approach to extracting tables from PDF's and labelling them with a topic I have a collection of PDF documents.  What I want in the end is to extract tables that are in some of these documents into flat file, and then label the tables with a specific theme (limited number of pre-determined themes).  
The way I was thinking about this, I would have the machine identify a group of continuous pages with tables.  It would extract the tables and assign a label (the label information would come from the text that is around these tables, usually right above it).  
I don't have the luxury of creating a 200+ document training set for each theme.  I would be able to show the computer 10 positive examples for each theme.  
I'm wondering whether this is a task for reinforcement learning where the reward is properly extracting a table and assigning the correct label.
 A: You could try these software packages >>> If the pdfs are "Adobe searchable" then you should be able to easily extract using Adobe Acrobat DC; exporting data to Excel.  You can also use vba script with the Adobe library to split a large pdf file into individual pdf pages before exporting to Excel.  If the pdfs are images / pictures created as pdf documents (not searchable in Adobe), you can use Tesseract-OCR and Ghostscript for alphanumerical-data extract to Excel. Once in Excel, should be easy to identify the "labels" and table structure; clean with vba. Monarch is expensive but might be a solution also...I don't think you need machine learning here - no need for convolutional neural networks, etc. to read data from pdf as image / picture.  Hope this helps
A: Reading your question, it's not very clear to me if the associated table data is needed for assigning the "themes" (aka 'categories') to the labels-info per QUOTE: "...the label information would come from the text that is around these tables, usually right above it"...what I think you're trying to do is match / classify this label information (let's say "table header info") to themes.  This is really just matching keywords to categories, you'll find neural network text classification online regarding, for example, website searches.  If you apply the technique in my first answer, you can easily extract and separate the tables from the "table headers / label information" ie. using VBA's ".usedrange", etc. - so, once tables are removed you're left with label information, which can be parsed and listed as keywords.  From here, I'm not sure if you need something like a RNN / deep neural network for UNSUPERVISED text classification, but it seems a Random Forest would be easy to do if you're doing SUPERVISED (manually assigning the label info's "keywords" to your themes (aka categories).  You said the themes are known and fixed - so it's a matter of setting up a reference table; matching the themes to the label info's keywords for use with a random forest / naive bayes for text classification. I'm not familiar with the unsupervised, neural network methods well enough to assist with python, tensorflow, etc. code.
