Repository files navigation

Source:http://www.dzone.com/articles/traprange-method-extract-table

• Utilize OpenAI API to extract information from PDF files
• To extract information from PDF invoice
• 
• To run from the command line. Typejava -jar traprange.latest.jar -h for help OR see examples in filetest-command-line.sh

Table data structure is one of the most important data structure in document, especially when exporting data from enterprise systems, data is usually in table format.

There are several data file formats that are often used to store tabular content such as CSV, text, and pdf. For the first two formats, it is quite straight forward just by opening file, loop through lines, and split cells with proper separator. The libraries to do this are quite a lot.

With PDF file, the story is completely different because it doesn't have a dedicated data definition for tabular content, something liketable,tr,td tag in HTML. PDF is a complicated format with text data, font, styling, and also image, audio, and video, they can be mixed all together. Below is my proposed solution to data in high-density tabular content.

After some investigation, I realized that:

• Column: text content in cells of the same column lies on a rectangular space that does not overlap with other rectangular spaces of another column. For example, in the following image, red rectangle and blue rectangle are separated spaces
• Row: words in the same horizontal alignment are in the same row. But this is just sufficient condition because a cell in a row may be a multi-line cell. For example, the fourth cell in the yellow rectangle has two lines, phrases “FK to this customer’s record in” and "Ledgers table" are not in the same horizontal alignment but they are still considered in the same row. In my solution, I simply assume that content in a cell only is single-line content. Different lines in a cell are considered to belong to different rows. Therefore the content in the yellow rectangle contains two rows: 1.{"Ledger_ID", "|", "Sales Ledger Account", "FK to this customer's record to"} 2.{NULL, NULL, NULL, "Ledgers table"}

My library behind traprange isPDFBox which is the best PDF lib I know so far. To extract text from a pdf file,PDFBox API provides 4 classes:

• PDDocument: contains information of entire pdf file. In order to load a pdf file, we use methodPDDocument.load(stream: InputStream)
• PDPage: represents each page in pdf document. We possibly archive a specific page content by passing the index of the page with this method:document.getDocumentCatalog().getAllPages().get(pageIdx: int)
• TextPosition: represents an individual word or character in the document. We can fetch allTextPosition objects of aPDPage by overriding methodprocessTextPosition(text: TextPosition) in classPDTextStripper. ATextPosition object has methodsgetX(),getY(),getWidth(),getHeight() that returns its bound in page and methodgetCharacter() to get its content.

In my work, I process text chunks directly by usingTextPosition objects. For each text chunk in PDF file, it returns a text element with the following attributes:

• x: horizontal distance from the left of the page
• y: vertical distance from the top border of the page
• maxX: equals x + width of the text chunk
• maxY: equals y+ height of the text chunk

The most important thing is identifying the bound of each row and column because if we know the bound of a row/column, we can retrieve all texts in that row/column from that we can easily extract all content inside the table and put it in a structured model. We name these bounds aretrap-ranges.TrapRange has two attributes:

• lowerBound: contains the lower endpoint of this range
• upperBound: contains the upper endpoint of this rangeTo calculate values oftrap-ranges, we loop through all texts of the page and project range of each text onto the horizontal and vertical axis, get the result and join them together. After looping through all texts of the page, we will calculate trap-ranges and use them to identify cell data of the table.

Algorithm 1: calculating trap-ranges for each pdf page:

columnTrapRanges <-- []rowTrapRanges <-- []foreachtextinpagebegincolumnTrapRanges <--join(columnTrapRanges, {text.x,text.x +text.width} )rowTrapRanges <--join(rowTrapRanges, {text.y,text.y +text.height} )end

After calculatingtrap-ranges for the table, we loop through all texts again and classify them into correct cells of the table.

Algorithm 2: classifying text chunks into correct cells:

table <--newTable()foreachtextinpagebeginrowIdx <--inrowTrapRanges,getindexoftherangethatcontaintsthistextcolumnIdx <--incolumnTrapRanges,getindexoftherangethatcontainsthistexttable.addText(text,rowIdx,columnIdx)end

The above is class diagram of main classes:

• TrapRangeBuilder:build() to calculate and return ranges
• Table,TableRow andTableCell: for table data struture
• PDFTableExtractor is the most important class. It contains methods to initialize and extract table data from PDF file. Builder pattern was used here. Following is some highlighted methods in this class:
• setSource: set source of the pdf file. There're 3 overloadssetSource(InputStream),setSource(File) andsetSource(String)
• addPage: to determine which pages will be processed. Default is all pages
• exceptPage: to skip a page
• exceptLine: to skip noisy data. All texts in these lines will be avoided.
• extract: process and return result

PDFTableExtractorextractor =newPDFTableExtractor();List<Table>tables =extractor.setSource(“table.pdf”).addPage(0).addPage(1).exceptLine(0)//the first line in each page.exceptLine(1)//the second line in each page.exceptLine(-1)//the last line in each page.extract();Stringhtml =tables.get(0).toHtml();//table in html formatStringcsv =tables.get(0).toString();//table in csv format using semicolon as a delimiter

Following are some sample results (check out and run the test fileTestExtractor.java):

• Sample 1: Source:sample-1.pdf, result:sample-1.html
• Sample 2: Source:sample-2.pdf, result:sample-2.html
• Sample 3: Source:sample-3.pdf, result:sample-3.html
• Sample 4: Source:sample-4.pdf, result:sample-4.html
• Sample 5: Source:sample-5.pdf, result:sample-5.html

In experimentation, I used pdf files having high density of table content. The results show that my implementation detects tabular content better than other open-sources:pdftotext,pdftohtml,pdf2table. With documents having multi tables or too much noisy data, my method does not work well. If row has cells overlapped, columns of these cells will be merged.

TrapRange method works the best with PDF files having high density of table data. With documents have multi-table or too much noisy data,TrapRange is not a good choice. My method also can be implemented in other programming languages by replacingPDFBox by a corresponding pdf library or using command-line tool pdftohtml to extract text chunks and using these data as input data foralgorithm 1, 2.

1. Java 8+
2. Maven 3+

1. http://en.wikipedia.org/wiki/Portable_Document_Format
2. http://pdfbox.apache.org
3. http://ieg.ifs.tuwien.ac.at/pub/yildiz_iicai_2005.pdf
4. http://www.foolabs.com/xpdf/
5. http://ieg.ifs.tuwien.ac.at/projects/pdf2table/

The MIT License (MIT)