2d:null 2e:["$","title",null,{"children":"Tool: Extract Text from a PDF (Manual) | Tanner Hobson | Observable"}] 33:T879,# Tool: Extract Text from a PDF (Manual) I wanted to create a notebook for extracting text from a PDF file, especially a PDF file that is a 2-column academic paper. --- ## Demo Select a file to process. _This file will not be uploaded and will only be processed locally in the browser._ ${viewof FILE} Set the bounding box filter. *Note:* The range is upside-down! `ymin` is the top of the page. `ymax` is the bottom of the page. ${viewof RANGE} A visual representation of all pages of the PDF with the filter applied. ${viewof CANVAS} The resulting text: ${viewof TEXT} ${DOM.download(() => new Blob([TEXT], { type: 'text/plain' }), `${TITLE}.txt`)} --- ## Explanation This notebook uses PDF.js to extract the text contents from the PDF. The `items` in a `TextContent` object have a few properties: `${ Object .keys(TEXT_CONTENTS[0].items[0]) .join(', ') }`. We are most interested in the `str`, `hasEOL`, `width`, `height`, and `transform` properties. - `str` is the actual text for the current item. - `hasEOL` tells us if there is more text on the same _logical_ line (because there are multiple columns, it's important to clarify that this is one line within one column, and not one line across two columns). - `width` and `height` give us the bounding box of the current item. - `transform` is a 6 element array, corresponding with the arguments to `CanvasRenderingContext2D.setTransform()`. So, it is: `scale x`, `skew y`, `skew x`, `scale y`, `translate x`, `translate y`. If we simply use the `str` and `hasEOL` properties, then we can get all the text from the document, but academic papers usually have some common header and footer text. Instead, we can use the `translate x/y` and `width/height` properties to find the complete bounding box for the current item, and we can filter out text that isn't within a certain range. To aid with this, the bounding boxes of each item are drawn, along with the bounding box of the filtered range. Note: this drawing is upside down. Oops! Finally, the items can be filtered by the bounding boxes, combined according to the `hasEOL` property, and then combined across each of the pages.34:T412,viewof RANGE = { const textContents = TEXT_CONTENTS; // if (this) return this; let xmin, xmax, ymin, ymax; xmin = null; for (let i=0, n=textContents.length; i