4

I'm interested in whether it is possible to classify texts by authorship using just syntactic/grammatical information.

Let's take ten people and have them write three texts. About their favorite hobby, a political commentary and the third one is a fictional story. For the classification I do not want to use information like vocabulary or typos but just the grammatical role of a word in a sentence.

Is it possible to turn such information into a vector f.x. and classify it?

UPDATE

As suggested by hippietrail I'll try to clarify further what aspect I am interested in.

Naturally one can quantify anything and then have a clustering or classification algorithm work on it.

Basically you can use any metric you can think up beside just vocab, word frequency and typos. The latter won't even apply to identifying edited texts. Things like collocations, hyphenation, sentence length - anything imaginable. (hippietrail)

(hyphenation f.x. does not interest me for the purposes of this question - only the grammatical level.)

But I am not a linguist and I wouldn't want to start from scratch guessing my way to a useful quantification. So I am wondering whether there is already a quantification/vectorization of the above mentioned grammatical information distilled from a text that is known to serve this purpose of authorship identification sufficiently well. Maybe even a "classical" approach.

14
  • 1
    To my knowledge you actually don't try to parse the text to do this. It's an unnecessary, expensive, and inherently ambiguoug task. You analyze stylistic things. I'm not sure if it has a more usual name than "author identification". I think it does but it's not coming to me. Oct 4 '13 at 12:30
  • 4
    Wikipedia: Stylometry is often used to attribute authorship to anonymous or disputed documents. It has legal as well as academic and literary applications, ranging from the question of the authorship of Shakespeare's works to forensic linguistics. Oct 4 '13 at 12:33
  • 1
    Good point. One reason why placed this (hypothetical) transformation at the beginning is because I intended to forestall answers referring to vocabulary, word frequency or typos. Yet, a labelling would be necessary - left undecided whether this process occures separate and preparational or is seemlessly integrated into the identification and encoding of patterns.
    – Raffael
    Oct 4 '13 at 12:34
  • 2
    "Stylometry" - nice one! Now I can already google it ;)
    – Raffael
    Oct 4 '13 at 12:35
  • 4
    I believe the main keyword in NLP is "author attribution" or "authorship attribution", and it's an area of active and vibrant research. Here's a relevant Google search query that provides many useful links like this survey of the field and current methods.
    – robert
    Oct 4 '13 at 13:26
1

Most of the links I'll provide are in German, but they seem to apply perfectly to your question:

Linguistic differential analysis and author identification

The first article describes the use of a vector model. A message board is used as testing ground for author identification based on usage of certain vocabulary and word frequency.

For example, the third variable corresponds loosely to:

"kriminell" -> "criminal"
"auf der schiefen Bahn" -> "not on the right track"
"straffällig" -> "has been convicted"

History of computer-aided author identification

A little context: The militante gruppe (mg) was a radical left group that carried out small attacks on property in the early 2000s. Andrej Holm, a sociologist specializing in gentrification and similar topics, was detained by the police on the basis that he was part of the (mg). Authorship identification techniques were used to identify him.

The author first tries to group texts based on function words and concludes that it isn't a suitable metric for authorship identification. This article disagrees with that conclusion.

The author then tries to use complex n-gram models. This type of procedure is great for identifying the type of text (e.g., fiction, blog post, scientific journal) but isn't suitable for author identification.

Lastly, there is machine learning using various metrics (e.g., average sentence length, relative frequency of intensifying particles, relative frequency of passive voice).

The resulting decision tree is shown here:

Decision tree

One criticism is that shallow nodes have a higher chance of being wrongly classified, and that this method, again, identifies the type of text rather than the author.

One could argue that all these methods have shortcomings in not having enough data or classifiers. There are various other methods, e.g., local histograms, character n-grams (which supposedly works really well), and Latent Semantic Indexing.

2
  • 1
    In general, your paragraphs suffer from a lack of introductory and transitional sentences. 1. Please explain what your links pertain to in initial topic sentence. 2. Expand your first example to, e.g., “Bear in mind that the German links contain German cultural references, for example, a reference to the newspaper column, Post von Wagner.” 4. Omit “so let’s start with that.” 5. Re: The History of... link. You need to explain the relevance of the historical information. The latter seems to be a puzzling abrupt topic change. Jul 23 '14 at 1:15
  • Also, please learn the markdown syntax so that you can post better links.
    – curiousdannii
    Jul 23 '14 at 7:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.