Computer Sleuth: Identification by Text Analysis
AbstractHere's a project where you can try your hand at being a detective with your computer. In this project you'll write a program to do some basic analysis of features of written text (for example, counting the length of each word in the text, or the number of words in each sentence). Then you'll see if you can use the information from your text analysis program to find measurements that can distinguish one author from another. After analyzing known samples of several authors' writings, can your method match up unidentified writing samples with their correct authors?
ObjectiveThe goal of this project is to write a computer program to make some simple measurements on a block of text, and then to see if this information can be used to identify the author of the text.
Your English teacher has probably told you that every author has an individual writing style—their own unique 'voice' on the page. Is it possible to find ways to identify that voice through computer analysis of written text?
- the number of sentences contained in the text,
- the number of words in each sentence,
- the number of letters in each word,
- the average number of words per sentence, and
- the average word length.
Terms and ConceptsTo do this project, you should do research that enables you to understand the following terms and concepts:
- Frequency histogram
- How would you calculate the frequency of five-letter words in a given block of text?
- For two articles on using text to identify authors see:
- Klarreich, E. (2003). "Bookish Math: Statistical Tests Are Unraveling Knotty Literary Mysteries," Science News 164 (December 20): 392. Retrieved February 16, 2007.
- Rehmeyer, J. (2007). "Digital Fingerprints: Tiny Behavioral Differences Can Reveal Your Identity Online," Science News 171 (January 13): 26–28. Retrieved February 14, 2007.
- For information on HTML FORMS, here is the official reference:
Ragget, D., A. Le Hors, I. Jacobs, (eds.) (1999). HTML 4.01 Specification: 17. Forms, W3C Worldwide Web Consortium. Retrieved February 14, 2007.
- If you get interested and start doing a lot of programming, you may want to try using a text editor that is a little more sophisticated than Notepad. An editor designed for programming can help with formatting, so that your code is more readable, but still produce plain text files. This type of editor can also do "syntax highlighting" (e.g., automatic color-coding of HTML) which can help you to find errors. Here is a free programmer's editor that you can try.
- This 19th century article used a plot of word length vs. frequency to distinguish texts by different authors:
Mendenhall, T.C. (1887). "The Characteristic Curves of Composition," Science 9 (11 March): 237–246.
- For an online version of The Federalist Papers, see:
Whitten, C. (2004). "The Federalist Papers," Founding Fathers Info. Retrieved February 14, 2007.
Materials and EquipmentTo do this experiment you will need the following materials and equipment:
- Computer with web browser (e.g., Internet Explorer, Firefox)
- Text editing program (e.g., Notepad)
- Several samples of text by each of three (or more) authors, for example:
- Sample paragraphs from books by different authors
- E-mail or instant messages from friends
- Spreadsheet program (e.g., Excel or QuattroPro)
- Graph paper or graphing software
- A helper
- Write the program to analyze text.
- You may decide that you want to improve the program so that you can make additional measurements. The Variations section has some suggestions for additional measurements, and you will probably come up with others on your own.
- Choose three or more authors and select representative samples of text by each (it's best to use at least 1000 words).
- Analyze each text sample with your program.
- Experiment with methods of graphing the results to create your own 'writeprint' (Rehmeyer, 2007) for each author.
- So that you can make fair comparisons between samples, all of your graphs should share the same scales (i.e., the same range for the x- and y-axes of each graph should be the same). So think carefully when you design your 'writeprint' and make sure that your x- and y-axes are designed to accommodate the full range of possible measurements.
- The key is to identify measurements that consistently reveal a difference between authors.
- For starters, you may want to try plotting the word length vs. frequency for each author (Mendenhall, 1887).
- Have your helper select additional paragraphs from each author. Your helper should also run the analysis on each additional sample, and give you the results, without identifying the authors. Can you determine the author of each unknown sample?
Ask an Expert
- Here are some ideas for functions that you might want to add to your text measurement program:
- count the frequency of different sentence lengths,
- frequency of function words, such as prepositions (e.g., of, from, in) and conjunctions (e.g., and, but, or).
- Let's say that one of your authors was J.K. Rowling, and all of your text samples came from the first Harry Potter book (Harry Potter and the Sorcerer's Stone). What happens if you use a text sample from a later book in the series, like Harry Potter and the Order of the Phoenix? Do your measurements still point to J.K. Rowling as the author?
- How much text do you need to get an accurate 'writeprint' for an author? Design an experiment to find out.
- Super-advanced students could explore representing the text analysis data as multidimensional vectors and using principal components analysis to differentiate between authors.
If you like this project, you might enjoy exploring these related careers: