I was reading a BBC article about "linguistic fingerprints". In my opinion a quite interesting idea, but there seems to exist a dispute in linguistic scientific community according to Wikipedia, how valid and objective such a concept is. Nevertheless, some research fields seem to use this concept in a very productive way, e.g. for analysing the number of different authors of the bible.

Being no expert in linguistics, but knowing that scientists often tend to create buzzwords, can somebody with more insight explain and give a short overview, if this dispute really exists and where linguistic researchers disagree? Is the concepts itself flawed, as such a quantity would be hard to measure in a valid, reliable and objective way (Operationalization). Or is it rather the the assumption that the "linguistic fingerprint" phenomenon doesn't exist at all?

On a personal note, I tend to believe in this concept and would assume that the unique elements of a single human fingerprint not only are typical used words and phrases, but also typical length of sentences, use of subordinate clauses. Reading and moderating over a decade in internet forums and usenet, the "linguistic fingerprint" concept seems quite obvious to me, the correct operationlization is of course very tricky.

Wiki-Definition of linguistic fingerprint:

It is formed as a result of merged language style. A person's linguistic fingerprint can be reconstructed from the individual's daily interactions and relate to a variety of self-reported personality characteristics, situational variables and physiological markers.

  • 5
    Good question. It might make it easier for readers if you gave a very short definition of what such a linguistic fingerprint is. I know it is explained in your links, but it looks neater if it's also in your question, thanks. – Cerberus Oct 30 '11 at 16:25
  • 'linguistic fingerprint' is not a technical term within linguistics. It might be a useful term for them, but it seems more useful for studying individual psychologies rather than general language communities. – Mitch Oct 30 '11 at 21:43

It doesn't make much sense to ask whether the concept is valid. You've gotta ask, "valid for what?" In other words, it makes more sense to ask whether it leads to decent performance on a specific task.

Look at it this way: if you give me a writing sample and ask me "Which of the five hundred million worldwide speakers of English did this come from," there's absolutely no way I can answer with any degree of confidence. If you give me a writing sample and ask "Is this by Winston Churchill or Charles Bukowski," I'll bet you any sum of money you like that I can get the answer right. This suggests that "linguistic fingerprinting" is suitable for the Winston Churchill Detection Task, and not so suitable for the task of picking a single author out of a pool of five hundred million. For tasks in between those two extremes, its degree of suitability is likely to vary. :)

You might be interested in this head-to-head comparison of two authorship attribution systems. Depending on the task, the winning system gets accuracies as high as 96% or as low as 31%. The paper also goes into some detail on how the winning system works. The feature set includes

  • presence or absence of specific words, characters or character combinations
  • presence or absence of specific typos/misspellings
  • frequency of different parts of speech or part-of-speech combinations
  • frequency of specific function words (e.g. "of," "for")
  • number of paragraphs, words or characters in the message
  • length of each paragraph or word
  • vocabulary size and distribution
  • use of numerals

So yeah, it's not just specific words or phrases, though popular articles on this stuff tend to focus on that aspect of it — probably because "women use more pronouns" or whatever makes for a catchy headline.

| improve this answer | |
  • thanks, nice structured answer. On this valid thing. I can only speak for natural sciences, there valid operationalization methods means, you are actually measuring what you intend to measure, thats my point here. The interesting question on this LF here is, do they feed 100s of sites into a computer algorithms and make smart datamining, statistcal analysis (thats not a valid linguistic concept, but smart statistical analysing). Or do they have a well-thought algorithm based on linguistic knowledge. The difference: Such a algorithm should need much less sites to build a reliable LF. – Hauser Nov 3 '11 at 0:36
  • does your pdf (unfortunately not searchable) elaborate on this point. I didn't read anywhere how much sites of they bible they used to build this fingerprint, that would be the interesting information imho to differ, how valid and reliable the underlying linguistic model of that algorithm is. – Hauser Nov 3 '11 at 0:40
  • 1
    Hmm. You might be interested in this: www.liwc.net/liwcdescription.php -- it's not about the same system as the last paper I linked to, but it discusses some of the validity/reliability issues you're asking about. (At least, if I understood the question right...) – Leah Velleman Nov 3 '11 at 2:43

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.