Mining the Social Web

Summarizing Documents

Although it may not be immediately obvious, just being able to perform reasonably good sentence detection as part of an NLP approach to mining unstructured data can enable some pretty powerful text-mining capabilities, such as crude but very reasonable attempts at document summarization. There are numerous possibilities and approaches, but one of the simplest to get started with dates all the way back to the April 1958 issue of IBM Journal. In the seminal article entitled “The Automatic Creation of Literature Abstracts,” H.P. Luhn describes a technique that essentially boils down to filtering out sentences containing frequently occurring words that appear near one another.

The original paper is easy to understand and is rather interesting; Luhn actually describes how he prepared punch cards in order to run various tests with different parameters! It’s amazing to think that what we can implement in a few dozen lines of Python on a cheap piece of commodity hardware, he probably labored over for hours and hours to program into a gargantuan mainframe. Example 8-3 provides a basic implementation of Luhn’s algorithm for document summarization. A brief analysis of the algorithm appears in the next section. Before skipping ahead to that discussion, first take a moment to trace through the code and see whether you can determine how it works.

Example 8-3. A document summarization algorithm (blogs_and_nlp__summarize.py)

# -*- coding: utf-8 -*- import sys import ...