Hacker News Reader

Comment by junon | Hacker News Reader

I wonder if this could be used to fingerprint an author. I'm sure word choice and structure would be a much better signal but also much harder to develop a model for (I speculate, I'm by no means an expert here). Less input tokens (just the punctuation) would reduce the model's complexity would it not?

Would be interesting to see a classifier based on this.

View on HN

Researcher in authorship analysis here. Yes, punctuation specific patterns have been used in the literature for this purpose, quite a bit. Removing the words also reduces the impact of the content of an article on your results (i.e. accidentally guessing the author because the topic is baseball and you only have one writer who writes about baseball in your dataset).

Typically, this type of feature is generalised of "character n-grams", but there are lots of variants.

interesting.

i also wanted to answer the same question. so i pasted two essays from one of my favorite authors in the tool. the punctuation summaries looked very different.

the two essays were by Isaac Asimov -- he does have a somewhat recognizable writing style. one essay was on physics and one of demonology.

then again, nlp has "stop words". i guess maths and physics would have math symbols stripped in that line of research?

Very interesting, thanks for the answer - strangely (but pleasantly) relevant area of study!

If I wanted to obscure my identity via text, changing my word choice, slang, colloquium, etc, would comprise my first naive attempt. I find myself changing the punctuation I use when pretending to be someone else, or trying to convey some personality traits that aren't natural to me. For example, one of my friends always puts a space before question marks and exclamation points.

I once ran Satoshi et al.'s writings through some simple stylometry service. If I recall correctly, my search pointed to Nick Szabo. A quick Google search shows that others have come up with Szabo, Finney or Dai. I am now super curious as to what the results would be if punctuation was analyzed exclusively... Unfortunately, I don't have time at the moment to pursue this tangent.

My immediate thought on this is yes, but you would also want to plug in the length of the sample in characters, word count (to get a sense of the average length of word for each author). This would actually be a useful tool in some industries I can already think of…

Reply on news.ycombinator.com