Reproducing Hacker News writing style fingerprinting | Hacker News Reader

Reproducing Hacker News writing style fingerprinting

antirez.com

325 pointsgrep_it155 comments

>Well, the first problem I had, in order to do something like that, was to find an archive with Hacker News comments. Luckily there was one with apparently everything posted on HN from the start to 2023, for a huge 10GB of total data.

This is actually super easy. The data is available in BigQuery.[0] It's up to date, too. I tried the following query, and the latest comment was from yesterday.

    SELECT 
      id,
      text,
      `by` AS username,
      FORMAT_TIMESTAMP('%Y-%m-%dT%H:%M:%SZ', TIMESTAMP_SECONDS(time)) AS timestamp
    FROM 
      `bigquery-public-data.hacker_news.full`
    WHERE 
      type = 'comment'
      AND EXTRACT(YEAR FROM TIMESTAMP_SECONDS(time)) = 2025
    ORDER BY 
      time DESC
    LIMIT 
      100

https://console.cloud.google.com/bigquery?ws=!1m5!1m4!4m3!1s...

My favorite which is also up to date is the ClickHouse playground.

For example:

  SELECT * FROM hackernews_history ORDER BY time DESC LIMIT 10;

https://gh-api.clickhouse.tech/play?user=play#U0VMRUNUICogRl...

I subscribe to this issue to keep up with updates:

https://github.com/ClickHouse/ClickHouse/issues/29693#issuec...

And ofc, for those that don't know, the official API https://github.com/HackerNews/API

I didn't know there was an official API! This explains why the data is so readily available in many sources and formats. That's very cool.

With a more straightforward approach, the tool can be reproduced with just a few queries in ClickHouse.

1. Create a table with styles by authors:

    CREATE TABLE hn_styles (name String, vec Array(UInt32)) ENGINE = MergeTree ORDER BY name

2. Calculate and insert style vectors (the insert takes 27 seconds):

    INSERT INTO hn_styles WITH 128 AS vec_size,
    cityHash64(arrayJoin(tokens(lower(decodeHTMLComponent(extractTextFromHTML(text)))))) % vec_size AS n,
    arrayMap((x, i) -> i = n, range(vec_size), range(vec_size)) AS arr
    SELECT by, sumForEach(arr) FROM hackernews_history GROUP BY by

3. Find nearest authors (the query takes ~50 ms):

    SELECT name FROM hn_styles ORDER BY cosineDistance(vec, (SELECT vec FROM hn_styles WHERE name = 'antirez')) LIMIT 25

        ┌─name────────────┬─────────────────dist─┐
     1. │ antirez         │                    0 │
     2. │ geertj          │ 0.009644324175144714 │
     3. │ mrighele        │ 0.009742538810774581 │
     4. │ LukaAl          │ 0.009787061201638525 │
     5. │ adrianratnapala │ 0.010093164015005152 │
     6. │ prmph           │ 0.010097599441156513 │
     7. │ teilo           │ 0.010187607877663263 │
     8. │ lukesandberg    │  0.01035981357655602 │
     9. │ joshuak         │ 0.010421492503861374 │
    10. │ sharikous       │  0.01043547391491162 │
    11. │ lll-o-lll       │  0.01051205287096002 │
    12. │ enriquto        │ 0.010534816136353875 │
    13. │ rileymat2       │ 0.010591026237771195 │
    14. │ afiori          │ 0.010655186410089112 │
    15. │ 314             │ 0.010768594792569197 │
    16. │ superice        │ 0.010842615688153812 │
    17. │ cm2187          │  0.01105111720031593 │
    18. │ jorgeleo        │ 0.011159407590845771 │
    19. │ einhverfr       │ 0.011296755160620009 │
    20. │ goodcanadian    │ 0.011316316959489647 │
    21. │ harperlee       │ 0.011317367800365297 │
    22. │ seren           │ 0.011390119122640763 │
    23. │ abnry           │ 0.011394133096140235 │
    24. │ PraetorianGourd │ 0.011508457949426343 │
    25. │ ufo             │ 0.011538721312575051 │
        └─────────────────┴──────────────────────┘

…i can’t believe i’ve been running a script to ingest the data for the last six hours. thank you.

It works for me. The accounts I used long time ago are there in high positions. I guess that my style is very distinctive.

But I also have seen some accounts that seem to be from other non-native English speakers. They may even have a Latin language as their native one (I just read some of their comments, and, at minimum, some of them seem to also be from the EU). So, I guess, that it is also grouping people by their native language other than English.

So, maybe, it is grouping many accounts by the shared bias of different native-languages. Probably, we make the same type of mistakes while using English.

My guess will be that native Indian or Chinese speakers accounts will also be grouped together, for the same reason. Even more so, as the language is more different to English and the bias probably stronger.

It would be cool that Australians, British, Canadians tried the tool. My guess is that the probability of them finding alt-accounts is higher as the populations is smaller and the writing more distinctive than Americans.

Thanks for sharing the projects. It is really interesting.

Also, do not trust the comments too much. There is an incentive to lie as to not acknowledge alt-accounts if they were created to remain hidden.

I discover 2 people in my top 20 who I can bet are from the same country as me and it is not a big country.

Amusingly, my accounts closest (and mutual) match is from a person from the UK, and I'm a Kiwi/Australian. Though I speak and write kind of weirdly, very UK-like in some ways, and reading through that accounts comments we really do write alike!

The "analyze" feature works pretty well.

My comments underindex on "this" - because I have drilled into my communication style never to use pronouns without clear one-word antecedents, meaning I use "this" less frequently that I would otherwise.

They also underindex on "should" - a word I have drilled OUT of my communication style, since it is judgy and triggers a defensive reaction in others when used. (If required, I prefer "ought to")

My comments also underindex on personal pronouns (I, my). Again, my thought on good, interesting writing is that these are to be avoided.

In case anyone cares.

That's very interesting as I noticed that certain outliers seemed indeed conscious attempts.

Since you seem to care about your writing, I'm wondering why you used "that" here?

> I use "this" less frequently that I would otherwise

Isn't it "less than" as opposed to "less that"?

Typo. Good catch

> I prefer "ought to"

I too like when others use it, since a very easy and pretty universal retort against "you ought to..." is "No, I don't owe you anything".

Are you saying there's a connection between "ought" and "owe"? All I see is "I don't want to hear any criticism".

The etymology makes a connection through old English. Oxford dictionary also contains this meaning:

> used to indicate duty or correctness

A duty to others is something you owe them; think, a duty of care and its lack, which is negligence.

Yes, "ought" is the past tense of "owe". At some point, the second alternative spelling "owed" was introduced to better separate the two meanings (literal and figurative), but it's still the same word; a similar thing happened with "flower" and "flour", those used to be interchangeable spellings of the same word but then somebody decided that the two meanings of that word should be separated and given specific spellings.

And the construct "you owe it to <person> to <verb>" still exists even today but is not nearly as popular as "you should <verb>" precisely because it has to state to whom exactly your owe the duty; with "should" it sounds like an impersonal, quasi-objective statement of fact which suits the manipulative uses much better.

I (also?) felt the 'words used less often' were much easier to connect to as a conscious effort. I pointed chatgpt to the article and pasted in my results and asked it what it could surmise about my writing style based on that. It probably connected about as well as the average horoscope but was still pretty interesting!

> Again, my thought on good, interesting writing is that these are to be avoided.

You mean, ”I think this should be avoided”? ;)

Nice one high five

Should is a commonly used word and a fine one. You should feel free to use it. If someone gets hot under the collar because you said he should do something then he is an idiot.

"Ought to" is essentially a synonym. Anyone that gets upset when you said they should do something but is fine when you say that they ought to do something is truly a moron.

    > If someone [test] then he is an idiot.
    > Anyone [test] when [test] is truly a moron.

These structures are worse habits in communication than subtle, colloquially interchangeable word choices.

This isn't a habit of communication. I honestly mean it: if you get upset that someone said you "should" do something, but you are fine when they say you "ought to" do it, then you must be stupid. They mean the same thing in modern English.

Yes but words hold memories to others. Since 'ought to' is less frequently used it doesn't 'trigger' people the same way.

Most people are emotion-first, how the words make them feel is more important than the definitions of them. Being emotion-first doesn't make them stupid.

Being so emotional they react wildly to one of the most common words in the English language does in fact make them stupid.

I think “should” and “ought to” end up being equivalent.

I prefer to avoid such absolutes and portray causality instead.

For example, in place of “you should not do drugs at work” I prefer “if you take drugs at work you’ll get in trouble”.

They do, and your suggestion is a great alternative. I'll try to do more of that

Nothing wrong with implying that people ought to behave according to mainstream social norms.

Interestingly, when most people simply choose to do what most people choose to do, you get an emergent 'herd mentality' which can lead to some very strange places. It is also sensitive to very small purturbations - which in real terms means, the one person who does manage to think for themselves may find they have an outsized effect on the direction of the crowd.

I think this mentality is also where the term 'sheeple' comes from.

Isn’t that the same as saying that counterculture, fringe culture, and subcultures ought not exist?

I wouldn't take it that far, individual rights are important but so is functioning society. Counterculture is often a cyclic/youthful reaction to things that have been established as "better" or "proper" by older generations. Most rules for life such as the 10 Commandments have sort of a "deny youself some pleasure because it's better for the group" vibe that have been learned and re-learned over many centuries.

Now if you only underindex on "underindex"... There's a good alternative that everyone understands, "use less"

If we weren't on HN, I might be inclined to agree

I wonder how much accuracy would be improved if expanding from single words to the most common pairs or n-tuples.

You would need more computation to hash, but I bet adding frequency of the top 50 word-pairs and top 20 most common 3-tuples would be a strong signal.

( The nothing the accuracy is already good of course. I am indeed user eterm. I think I've said on this account or that one before that I don't sync passwords, so they are simply different machines that I use. I try not to cross-contribute or double-vote. )

Maybe there isn't enough data for each user for pairs, but I thought about mixing the two approaches (but had no time to do it), that is, to have 350 components like now, for the single word frequency, plus other 350 for the most common pairs frequency. In this way part of the vector would remain a high enough signal even for users with comparable less data.

I've been thinking some more about this, and it occurred to me that you'd want to encode sentence boundaries as a pseudo-word in the n-tuples.

I then realised that "[period] <word>" would likely dominate most common pairs, and that a lot of time could be saved by simply recording the first word of sentences as their own vector set, in addition but separate to the regular word vector.

Whether this would be a stronger or weaker signal per-vector-space than the tail of words in the regular common-words vector I don't know.

Maybe I talk too much on HN. :)

When I ran it, it gave me 20 random users, but when I do the analyze, it says my most common words are [they because then that but their the was them had], which is basically just the most common English words.

Probably would be good to exclude those most common words.

For visibility, here's the tool where you can enter your hn username:

https://antirez.com/hnstyle?username=pg&threshold=20&action=...

This is great example of what's possible and how true anonymity, even online, is only "technological threshold" anonymity. People obsessed with biometrics might not consider this is another biometric.

Instead of just HN, now do it with the whole internet, imagine what you'd find. Then imagine that it's not being done already.

None of my throwaways and not even my old account shows up. We are not at that level yet. ymmv.

This technique yields so many false positives and negatives, it's practically useless. Possibly it works reliably for mono-lingual, prolific writers. Someone like the Qanon shaman (or whatever the name was) might be picked up, if it doesn't happen to be a collective.

I wonder how much curly quote usage influences things. I type things like curly quotes with my Compose key, and so do most of my top similars; and four or five words with straight quotes show up among the bottom ten in our analyses. (Also etc, because I like to write &c.)

I’m not going to try comparing it with normalising apostrophes, but I’d be interested how much of a difference it made. It could easily be just that the sorts of people who choose to write in curly quotes are more likely to choose words carefully and thus end up more similar.

How does it find the high similarity between "dang" and "dangg" when the "dangg" account has no activity (like comments) at all?

https://antirez.com/hnstyle?username=dang&threshold=20&actio...

Probably it used to have when the database was created. Then the comments got removed.

We can improve this. antirez has made a highly compelling poc but it could be refined for authorship attribution judging by the number of misses in the comments here, and how this compares to greater accuracy of the original post to which antirez refers. I’m no expert, but some ideas:

- remove super high frequency non specific words from the comparison bags, because they don’t distinguish much, have less semantic value and may skew the data

- remove stop words (NLP definition of stop words)

- perform stemming/tokenization/depluralization etc (again, NLP standard)

- implement commutativity and transitivity in the similarity function

- consider words as hyperlinks to the sets of people who use them often enough, and do something Pageranky to refine similarity

- consider word bigrams, etc

- weight variations and misspellings higher as distinguishing signals

What are your ideas ?

Have you tried to analyze whether there is a correlation between "closeness" according to this metric and how often users chat in the same thread? I recognize some usernames that are reported as being similar to me, I wonder if there's some kind of self-selection at play.

Managed to find an alt I forgot I made and gave up using years ago. I do wonder about other high up people. Like what about our mutual histories makes us have similar word usage? Are we from the same areas or did we hang out in similar places online?

This is a bit tangential but I've noticed lots of comments aping the style of Matt Walsh. Not just on HN either, but probably more here than other places I visit.

Anyway, I guess this would be useful cluster the "Matt Walsh"-y commenters together.

Very cool. Also a bit surprising — two of my matches are people I know IRL.

I think it would be interesting to run this tool against Reddit, 4chan and Tweeter to find astroturf accounts. Does it look like a real browser to those sites or would it be blocked?

I tried my name, and I don't think a single "match" is any of my (very rarely used) throw away alts ;) I guess I have a few people I talk like?

When they are rarely used (a small amount of total words produced), they don't have meaningful statistical info for a match, unfortunately. A few users here reported finding actual duplicated accounts they used in the past.

I got 3 correct matches out of 20, and I've had about 6 accounts total (using one at a time), with at least a fair number of comments in each. I guess that means that my word choices are more outliers than yours or there is just more to match. So it's not really good enough to reliably identify alt accounts, but it is quite suggestive.

I think if you rule out insanely common words, it might get scary accurate.

Actually, the way that these things work is usually by focusing exclusively on the usage patterns of very common (top 500) words. You get better results by ignoring content words in favor of the linking words.

Dang's analysis was funny:

don't site comment we here post that users against you're

Quite a stance, man :)

And me clearly inarticulate and less confident than some:

it may but that because or not and even these

I noticed that randomly remembered usernames tend to produce either lots of utility words like the above, or very few of them. Interestingly, it doesn't really correlate with my overall impression about them.

No matches higher than .7something and no mutual matches let's go boys I'm a special unique snowflake

I wonder if such an analysis could tease apart the authors of intentionally anonymous publications. Things like peer review notes for papers or legal opinions (afaik in countries that are not the USA, the authors of a dissenting supreme court decision are not named).

I think an interesting use of this is potentially finding LLMs trained to have the style of a person. Unfortunately now, just because a post has my style it doesn't mean it was me. I promise I am not a bot. Honest.

I'm surprised no one has made this yet with a clustered visualization.

Redis supports random projection to a lower dimensionality, but the reality is that projecting a 350d vector into 2d is nice but does not remotely captures the "reality" of what is going on. But still, it is a nice idea to use some time. However I would do that with more than 350 top words, since when I used 10k it strongly captured the interest more than the style, so 2D projection of this is going to be much more interesting I believe.

Given that some matches are “mutual” and others are not, I don’t see how that could translate to a symmetric distance measure.

Imagine the 2D space, it also has the same property!

You have three points nearby, and a fourth a bit more distant. 4 best match is 1, but 1 best match is 2 and 3.

Good point, but the similarity score between mutual matches is still different, so it doesn’t seem to be a symmetric measure?

Your observation is really acute: the small difference is due to quantization. When we search for element A, that is int8 quantized by default, the code paths de-quantize it, then re-quantize it and searches. This produces a small loss of precision, like that:

redis-cli -3 VSIM hn_fingerprint ELE pg WITHSCORES | grep montrose

montrose 0.8640020787715912

redis-cli -3 VSIM hn_fingerprint ELE montrose WITHSCORES | grep pg

pg 0.8639097809791565

So why cosine similarity is commutative, the quantization steps lead to a small different result. But the difference is .000092 that is in practical terms not important. Redis can use non quantized vectors using the NOQUANT option in VADD, but this will make the vectors elements using 4 bytes per component: given that the recall difference is minimal, it is almost always not worth it.

Personally I like this approach a lot

https://scikit-learn.org/stable/modules/generated/sklearn.ma...

I think other methods are more fashionable today

https://scikit-learn.org/stable/modules/manifold.html

particularly multi-dimension scaling, but personally I think tSNE plots are less pathological (they don't have as many of these crazy cusps that make me think it's projecting down from a higher-dimensional surface which is near-parallel to the page)

After processing documents with BERT I really like the clusters generated by the simple and old k-Means algorithm

https://scikit-learn.org/stable/modules/generated/sklearn.cl...

It has the problem that it always finds 20 clusters if you set k=20 and a cluster which really oughta be one big cluster might get treated as three little clusters but the clusters I get from it reflect the way I see things.

Used More Often by dang.

don't +0.9339

Such a nice scientific way to detect and mute those who go against the agenda's grain, oh I mean don't contribute anything meaningful to the community

I knew that this was possible but I always thought it took much more... effort? How do we mitigate this, then? Run our posts through an LLM?

I rotate hn accounts every year or two. In my top 4, I found 3 old alts.

This is impressive and scary. Obviously I had to create a throwaway to say this.

Cool tool. It's a shame I don't have other accounts to test it.

It's also a tool for wannabe impersonators to hoan their writing style mimic skills!

I don't have other accounts, but still matched at 85+% accuracy for a half dozen accounts. Seems I don't have very original thoughts or writing style.

My guess is that people from the same region and similar background will have more and closer "alters". So, if you are Californian-American then there is many people that will speak similar to you in HN. If you are a Satawalese speaker then you may be quite alone in your own group.

(The Satawalese language has 460 speakers, most of who live in Satawal Island in the Federated States of Micronesia.)

It's a fingerprinting tool, not a profiling tool. You can't draw such conclusions from it.

What a profiler would do to identify someone, I imagine, requires much more. Like the ability to recognize someone's tendency of playing the victim to leverage social advantage in awkward situations.

PhasmaFelis and mikeash have all matches mutual for the top 20, 30, 50 and 100. Are there other users like this? If so, how many? What's the significance of this, in terms of the shape of the graph?

tablespoon is close, but has a missing top 50 mutual (mikeash). In some ways, this is an artefact of the "20, 30, 50, 100" scale. Is there a way to describe the degree to which a user has this "I'm a relatively closer neighbour to them than they are to me" property? Can we make the metric space smaller (e.g. reduce the number of Euclidean dimensions) while preserving this property for the points that have it?

This is an interesting and well-written post but the data in the app seems pretty much random.

Thank you, tptacek. I was able to verify, thanks to the Internet Archive caching of "pg" for the post of 3 years ago, that the entries are quite similar in the case of "pg". Consider that it captures just the statistical patterns in very common words, so you are not likely to see users that you believe are "similar" to yourself. Notably: montrose may likely be a really be a secondary account of PG, and was also found as a cross reference in the original work of three years ago.

Also note that vector similarity is not reciprocal, one thing can have a top scoring item, but such item may have much more items nearer, like in the 2D space when you have a cluster of points and a point nearby but a bit far apart.

Unfortunately I don't think this technique works very well for actual duplicated accounts discovery because often times people post just a few comments in fake accounts. So there is not enough data, if not for the exception where one consistently uses another account to cover their identity.

EDIT: at the end of the post I added the visual representations of pg and montrose.

If you want to do document similarity ranking in general it works to find nearby points in word frequency space but not as well as: (1) applying an autoencoder or another dimensional reduction technique to the vectors or (2) running a BERT-like model and pooling over the documents [1].

I worked on a search engine for patents that used the first, our evaluations showed it was much better than other patent search engines and we had no trouble selling it because customers could feel the difference in demos.

I tried dimensional reduction on the BERT vectors and in all cases I tried I found this made relevance worse. (BERT has learned a lot already which is being thrown away, there isn't more to learn from my particular documents)

I don't think either of these helps with the "finding articles authored by the same person" because one assumes the same person always uses the same words whereas documents about the topic use synonyms that will be turned up by (1) and (2). There is a big literature on the topic of determining authorship based on style

https://en.wikipedia.org/wiki/Stylometry

[1] With https://sbert.net/ this is so easy.

Indeed, but my problem is: all those vector databases (including Redis!) are always thought as useful in the context of learned embeddings, BERT, Clip, ... But I really wanted to show that vectors are very useful and interesting outside that space. Now, I also like encoders very well, but I have the feeling that the Vector Sets, as a data structure, needs to be presented as a general tool. So I really cherry picked a use case that I liked and where neural networks were not present. Btw, Redis Vector Sets support dimensionality reduction by random projection natively in the case the vector is too redundant. Yet, in my experiments, I found that using binary quantization (also supported) is a better way to save CPU/space compared to RP.

this got two accounts that I used to use

Great! Thanks for the ACK.

Well, well, well, cocktailpeanuts. :spiderman_pointing:

I suspect, antirez, that you may have greater success removing some of the most common English words in order to find truly suspicious correlations in the data.

cocktailpeanuts and I for example, mutually share some words like:

because, people, you're, don't, they're, software, that, but, you, want

Unfortunately, this is a forum where people will use words like "because, people, and software."

Because, well, people here talk about software.

<=^)

Edit: Neat work, nonetheless.

That seems to be a misconception.

The usage frequency of simple words is a powerful tell.

Indeed, some writing styles make frequent use of words like "that" and "just".

I can understand the nuance of your assertion, but looking at the data returned by these results suggests it's not really all that powerful at all.

There are so many people that write like me apparently, that simple language seems more like a way to mask yourself in a crowd.

You can definitely mask writing style. If you can do that only by using simple words, I am not so sure.

I noted the "analyze" feature didn't seem as useful as it could be because the majority of the words are common articles and conjunctions. I'd like to see a version of analyze that filters out at least the following stop words: a, an, and, are, as, at, be, but, by, for, if, in, into, is, it, no, not, of, on, or, such, that, the, their, then, there, these, they, this, to, was, will, with

The original version nailed all of my accounts with terrifying accuracy. Since then I make a new account every few days or weeks. Against the rules I know. And I’ve learned a lot about HN IP tracking and funny shadowbanning-like tricks they play but dont cop to. Like I get different error messages based on the different banned ips I use. And j see different behavior and inconsistency with flagged messages (like one that got upvoted a day after it was flagged and not visible to other users).

What you're doing makes HN worse, unfortunately.

Reply on news.ycombinator.com