It's a series of rectangles, one per word. Each rectangle consists of four unsigned 16-bit integers, denoting the X coordinate of the upper left-hand corner, the Y coordinate, the width in pixels, and the height in pixels, in that order. The upper-left corner of the image is coordinate (0, 0). There is no padding at all in the locations stream.

As for what constitutes a word, it's a bit complex to explain. Instead, use the WordsReader class in RepositoryAccess. Given a System.IO.Stream or TextReader representing your text for the page, use it to break up the text part of the page into a sequence of words. Each word needs to have one rectangle, the order of the rectangles in the locations page part should match up to the order of the words in the text.