Hacker News Reader

Comment by vintermann | Hacker News Reader

The big complicated segmentation pipeline is a legacy from the time you had to do that, a few years ago. It's error prone, and even at it's best it robs the model of valuable context. You need that context if you want to take the step to handwriting. If you go to a group of human experts to help you decipher historical handwriting, the first thing they will tell you is that they need the whole document for context, not just the line or word you're interested in.

We need to do end to end text recognition. Not "character recognition", it's not the characters we care about. Evaluating models with CER is also a bad idea. It frustrates me so much that text recognition is remaking all the mistakes of machine translation from 15+ years ago.

View on HN

> We need to do end to end text recognition. Not "character recognition", it's not the characters we care about.

Arbitrary nonsensical text require character recognition. Sure, even a license plate bears some semantics bounding expectations of what text it contains, but text that has no coherence might remain an application domain for character rather than text recognition.

> Arbitrary nonsensical text require character recognition.

Are you sure? I mean, if it's printed text in a non-connected script, where characters repeat themselves (nearly) identically, then ok, but if you're looking at handwriting - couldn't one argue that it's _words_ that get recognized? And that's ignoring the question of textual context, i.e. recognizing based on what you know the rest of the sentence to be.

Handwriting with words is not arbitrary nonsensical text

Yes - my point was about identifier strings such as UUID

Not really. I have an HTR use case where the data is highly specialized codes. All the OCR software I use is tripped up by trying to find the content into the category of English words.

LLMs can help, but I’ve also had issues where the repetitive nature of the content can reliably result in terrible hallucinations.

VLMs seem to render traditional OCR systems obsolete. I'm hearing lately that Gemini does a really good job on tasks involving OCR. https://news.ycombinator.com/item?id=42952605

Of course there are new models coming out every month. It's feeling like the 90s when you could just wait a year and your computer got twice as fast. Now you can wait a year and whatever problem you have will be better solved by a generally capable AI.

The problem with doing OCR with LLMs is hallucination. It creates character replacements like Xerox's old flawed compression algorithm. At least this my experience with Gemini 2.0 Flash. It was a screenshot of a webpage, too.

Graybeards like Tessaract has moved to neural network based pipelines, and they're re-inventing and improving themselves.

I was planning to train Tessaract with my own hand writing, but if OCR4All can handle that, I'll be happy.

Paradoxically, LLMs should be the tool to fix traditional OCR by recognizing that "Charles ||I" should be "Charles III", "carrot ina box" should be "carrot in a box", the century of the event in context cannot be that construed through looking at the gliphs etc.

As someone who's learning how to do OCR in order to re-OCR a bunch of poorly digitized documents, this will not work with modern OCR. Modern OCR is too good.

If you're able to improve the preprocessing and recognition enough, then there's a point at which any post-processing step you do will introduce more errors than it fixes. LLM's are particularly bad as a post-processing step because the errors they introduce are _designed to be plausible_ even when they don't match the original text. This means they can't be caught just by reading the OCR results.

I've only learned this recently, but it's something OCR experts have known for over a decade, including the maintainers of Tesseract. [1]

OCR is already at the point where adding an LLM at the end is counterproductive. The state of the art now is to use an LSTM (also a type of neural network) which directly recognizes the text from the image. This performs shockingly well if trained properly. When it does fail, it fails in ways not easily corrected by LLM's. I've OCR'ed entire pages using Tesseract's new LSTM engine where the only errors were in numbers and abbreviations which an LLM obviously can't fix.

[1] https://tesseract-ocr.github.io/docs/Limits_on_the_Applicati...

> As someone ... Modern OCR is too good

I also have even recent extensive experience: I get an important amount of avoidable errors.

> at which any post-processing step you do will introduce more errors than it fixes ... the errors they [(LLMs)] introduce are _designed to be plausible_

You are thinking of a fully automated process, not of the human verification through `diff ocr_output llm_corrected`. And even then, given that I can notice errors that an algorithm with some language proficiency could certainly correct, I have reasons to suppose that a proper calibration of an LLM based system can achieve action over a large number of True Positives with a negligible amount of False Positives.

> LSTM

I am using LSTM-based engines, and on those outputs I have stated «I get an important amount of avoidable errors». The one thing that could go in your direction is that I am not using the latest version of `tesseract` (though still in the 4.x), and I have recently noticed (already through `tesseract --print-parameters | grep lstm`) that the LSTM engine evolved within 4.x, from early to later.

> numbers and abbreviations which an LLM obviously can't fix

? It's the opposite: for the numbers it could go (implicitly) "are you sure, I have a different figure for that" and for abbreviations, the LLM is exactly the thing that should guess them ot of the context. The LLM is that thing that knows that "the one defeated by Cromwell should really be Charles II-staintoberemoved instead of an apparent Charles III".

> You are thinking of a fully automated process, not of the human verification through `diff ocr_output llm_corrected`.

Fair, and I'm aware that that makes a huge difference in how worthwhile an LLM is. I'm glad you're not doing the annoyingly common "just throw AI at it" without thinking through the consequences.

I'm doing two things to flag words for human review: checking the confidence score of the classifier, and checking words against a dictionary. I didn't even consider using an LLM for that since the existing process catches just about everything that's possible to catch.

> I am using LSTM-based engines . . .

I'm using Tesseract 5.5. It could actually be that much better, or I could just be lucky. I've got some pretty well-done scans to work with.

> It's the opposite: for the numbers it could go (implicitly) "are you sure, I have a different figure for that" . . .

I honestly don't know what you mean. Are you saying that an LLM would know that a reference to "page 311" should actually be a reference to "page 317" based on context?

I think the example you've given makes a lot of sense if you're just using an LLM as one way to flag things for review.

> I honestly don't know what you mean. Are you saying that an LLM would know that a reference to "page 311" should actually be a reference to "page 317" based on context

Not that of your example (the page number): that would be pretty hard to check (with current general agents. In the future, not impossible - you would need some agent finally capable to follow procedures strictly). That the extra punctuation or accent is a glitch and that the sentence has a mistake are more within the realm of a Language Model.

What I am saying is that a good Specialized Language Model (maybe a good less efficient LLM) could fix a text like:

"AB〈G〉 [Should be 'ABC'!] News was founded in 194〈S〉 [Should be '1945'!] after action from the 〈P〉CC [Should be 'FCC'!], 〈_〉 [Noise!], deman〈ci〉ing [Should be 'demanding'!] pluralist progress 〈8〉 [Should be '3'!] 〈v〉ears [should be 'years'] ear〈!〉ier [Should be 'earlier'!]..."

since it should "understand" the sentence and be already informed of the facts.

This is moot anyway if the LLM is only used as part of a review process. But the most valuable documents to digitize are, almost by definition, those that don't have widely-known information that an LLM is statistically likely to guess. There's no way to get around that.

> This is moot anyway if the LLM is only used as part of a review process

Not really, because if plan to perform a `diff` in order to ensure that there are no False Positives in the corrections proposed by your "assistant", you will want one that finds as many True Positives as possible (otherwise, the exercise will be futile, inefficient, if a large number of OCR errors remain). So, it would be good to have some tool that could (in theory, at this stage) be able to find the subtle ones (not dictionary related, not local etc.).

> But the most valuable documents to digitize are, almost by definition, those that don't have widely-known information that an LLM is statistically likely to guess

It depends on the text. You may be e.g. interested in texts the value of which is in arguments, the development of thought (they will speak about known things in a novel way) - where OCR error has a much reduced scope (and yet you want them clean of local noise). And, if a report comes out from Gallup or similar, with original figures coming from recent research, we can hope that it will be distributed foremostly electronically. Potential helpers today could do more things that only a few years ago (e.g. hunspell).

I tried it! (How very distracted of us not to have tried immediately.) And it works, mostly... I used a public LLM.

The sentence:

> ABG News was founded in 194S after action from the PCC , _ , demanciing pluralist progress 8 vears ear!ier

is corrected as follows, initially:

> ABG News was founded in 1945 after action from the FCC, demanding pluralist progress 8 years earlier

and as you can see it already corrects a number of trivial and non-trivial OCR errors, including recognizing (explicitly in the full response) that it should be the "FCC" to "demand pluralist progress" (not to mention, all of the wrong characters breaking unambiguous words and the extra punctuation).

After a second request, to review its output to "see if the work is complete", it sees that 'ABG', which it was reluctant to correct because of ambiguities ("Australian Broadcasting Corporation", "Autonomous Bougainville Government" etc.), should actually be 'ABC' because of several hints from the rest of the sentence.

As you can see - proof of concept - it works. The one error it (the one I tried) cannot see is that of '8' instead of '3': it knows (it says in a full response) that the FFC acted in 1942, and ABC was created in 1945, but it does not compute the difference internally nor catch the hint that '8' and '3' are graphically close. Maybe one LLM with "explicit reasoning" could do even better and catch also that.

I'm saying it's moot because, if you're just flagging things for review, there's already a more direct and reliable way to do that. The OCR classifier itself outputs a confidence score. The naieve way of just checking that confidence score will work. The OCR classifier has less overall information than an LLM, but the information it has is much more relevant to the task it's doing.

When I have some time in front of a computer, I'll try a side-by side comparison with some actual images.

> OCR is already at the point where adding an LLM at the end is counterproductive

That's mass OCR on printed documents. On handwritten documents, LLMs help. There are tons of documents that even top human experts can't read without context and domain language. Printed documents are intended to be readable character by character. Often the only thing a handwriting author intends is to remind himself of what he was thinking when he wrote it.

Also, what is the downstream tasks? What do you need character level accuracy for? In my experience, it's often for indexing and search. I believe LLMs have a higher ceiling there, and can in principle (if not in practice, yet) find data and answer questions about a text better than straightforward indexing or search can. I can't count the number of times I've e.g. missed a child in genealogy because I didn't think of searching the (fully and usually correctly) indexed data for some spelling or naming variant.

I am working with printed documents. Maybe LLMs currently make a difference with handwriting recognition. I wasn't directly responding to that. It's outside the little bit that I know, and I didn't even think of it as "OCR".

I'm not saying that I need high accuracy (though I do), I'm saying that the current accuracy (and clarifying that this is specifically for printed text) is already very high. Part of the reason it's so high is because the old complicated character-by-character classifiers have already been replaced with neural networks that process entire lines at a time. It's already moving in the direction you're saying we need.

Maybe you can passthrough the completed text from a simple, fast grammar model to improve text. You don't need a 40B/200GB A100 demanding language model to fix these mistakes. It's absurdly wasteful in every sense.

I'm sure there can be models which can be accelerated on last gen CPUs AI accelerators and fix these kinds of mistakes faster than real time, and I'm sure Microsoft Word is already doing it for some languages, for quite some time.

Heck, even Apple has on-device models which can autocomplete words now, and even though its context window or completion size are not that big, it allows me to jump ahead with a simple tap to tab.

I wonder if this is a case where you want an encoder-decoder model. It seems very much like a translation task, only one where training data is embarrassingly easy to synthesize by just grabbing sentences from a corpus and occasionally swapping, inserting, and deleting characters.

In terms of attention masking, it seems like you want the input to be unmasked, since the input is fixed for a given “translation”, and then for the output tokens to use causally masked self attention plus cross attention with the input.

I wonder if you could get away with a much smaller network this way because you’re not pointlessly masking input attention for a performance benefit that doesn’t matter.

When I was reading your comment, I remembered an assignment in Fuzzy Logic course:

"Number of different letters" is a great heuristic for a word guesser. In that method you just tell the number of letters and then do some educated guesses to start from a semi-converged point (I think word frequencies is an easy way), and brute force your way from there, and the whole process finds the words in mere milliseconds.

You can improve this method to a 2-3 word window since we don't care about the grammar, but misread words, and brute-force it from there.

You may even need no network to fix these kinds of misrecognitions with this. Add some SSE/AVX magic for faster processing and you have a potential winner in your hands.

> Maybe you can passthrough the completed text from a simple, fast grammar model to improve text

Yes - but not really a "grammar model": a statistical model about text, with "transformer's attention" - the core of LLMs - should be it: something that identifies if the fed text has statistical anomalies (which the glitches are).

Unfortunately, small chatbot LLMs do not follow instructions ("check the following text"), they just invent stories, and I am not aware of a specialized model that can be fed text for anomalies. Some spoke about a BERT variant - which still does not have great accuracy, I understood.

It is a relatively small problem that probably does not have a specialized solution yet. Already a simple input-output box that worked like: "Evaluate statistical probability of each token" - then we would check the spikes of anomaly. (For clarity: this is not plain spellchecking, as we want to identify anomalies in context.)

Edit: a check I have just done with an engine I had not yet used for the purpose shows a number of solutions... But none a good specific tool, I am afraid.

One might need document/context fine-tuning. It's not beyond possibility that a prayer of text is about Charles ll1 (el-el-one) someone's pet language model or something. Sometimes you want correction of "obvious" mistakes (like with predictive text) other times you really did write keming [with an M].

And that is why I wrote that you need an LLM, i.e. (next post) a «statistical model about text, with "transformer's attention"», as «[what we want] is not plain spellchecking, as we want to identify anomalies in context».

To properly correct text you need a system that checks large blocks of text with some understanding, not just disconnected words.

Edit: minutes ago a member (in a faraway post) wrote «sorry if this is a bit cheeky»... He meant "cheesy". You need some level of understanding to see those mistakes. Marking words that are outside the dictionary is not sufficient.

Gemini does not seem to do OCR with LLM. They seem to use their existing OCR technology of which they feed the output into the LLM. If you set the temperature to 0 and ask for the exact text as found in the document, you get really good results. I once got weird results where I got literally the JSON output of the OCR result with bounding boxes and everything.

Interesting, thanks for the information. However, unfortunately, no way I'll be sending my personal notebooks to a service which I don't know what's going to do with them in the long term. However, I might use it for the publicly available information.

Thanks again.

I've been looking for an algorithm for running OCR or STT results through multiple language models, compare the results and detect hallucinations as well as correct errors by combining the results in a kind of group consensus way. I figured someone must have done something similar already. If anyone has any leads or more thoughts on algorithm implementation, I'd appreciate it.

Tesseract wildly outperforms any VLM I've tried (as of November 2024) for clean scans of machine-printed text. True, this is the best case for Tesseract, but by "wildly outperforms" I mean: given a page that Tesseract had a few errors on, the VLM misread the text everywhere that Tesseract did, plus more.

On top of that, the linked article suggests that Gemini 2.0 can't give meaningful bounding boxes for the text it OCRs, which further limits the places in which it can be used.

I strongly suspect that traditional OCR systems will become obsolete, but we aren't there yet.

I just wrapped up a test project[0] based on a comment from that post! My takeaway was that there are a lot of steps in the process you can farm out to cheaper, faster ML models.

For example, the slowest part of my pipeline is picture description since I need a LLM for that (and my project needs to run on low-end equipment). Locally I can spin up a tiny LLM and get one-word descriptions in a minute, but anything larger takes like 30. I might be able to only send sections I don't have the hardware to process.

It was a good into to ML models incorporating vision, and video is "just" another image pipeline, so it's been easy to look at e.g. facial recognition groupings like any document section.

[0] https://github.com/jnday/ocr_lol

I just used Gemini as an OCR a couple of hours ago because all the OCR apps I tried on android failed at the task lol Wild seeing this commment right after waking up

For self hosting check out Qwen-VL: https://github.com/QwenLM/Qwen-VL

Yes, I agree general purpose is the way to go, but I'm still waiting. Gemini is the best at last time I tried, but for all the ways I've tried to prompt it, it can not transcribe (or correctly understand the content of) e.g. the probate documents I try to decipher for my genealogy research.

I've seen Gemini Flash 2 mention "in the OCR text" when responding to VQA tasks which makes me question of they have a traditional OCR process mixed in the pipeline.

> Now you can wait a year and whatever problem you have will be better solved by a generally capable AI.

Maybe this is what the age of desktop AGI looks like.

Wouldn’t an AI make assumptions and fix mistakes?

For example instead of

> The speiling standards were awful

It would produce

> The spelling standards were awful

Issue with that is that some writings are not word based. People use acronyms, temporal, personalized, industrial jargon, and global ones. Beginning of the year, there where some HN posts about moving from dictionary word to character encoding for LLMs, because of the very varying nature in writing.

Even I used symbols for different means in a shorthand form when constructing an idea.

I see it the same way laws are. Their word definitions are anchored in time from the common dictionaries of the era. Grammar, spelling, and means all change through time. LLMs would require time scoped information to properly parse content from 1400 vs 1900. LLM would be for trying to take meaning out of the content versus retaining the works.

Character based OCR ignores the rules, spelling, and meaning of words and provides what most likely there. This retains any spelling and grammar error that are true positives or false positives, based on the rules of their day.

Could you dumb this down a bit (a lot) for dimmer readers, like myself? The way I am understanding the problem you are getting at is something like:

> The way person_1 in 1850 wrote a lowercase letter "l" will look consistently like a lowercase letter "l" throughout a document.

> The way person_2 in 1550 wrote a lowercase letter "l" may look more like an uppercase "K" in some parts, and more of a lowercase "l" in others, and the number "0" in other areas, depending on the context of the sentence within that document.

I don't get why you would need to see the entire document in order to gauge some of the details of those things. Does it have something to do with how language has changed over the centuries, or is it something more obvious that we can relate to fairly easily today? From my naive position, I feel like if I see a bunch of letters in modern English (assuming they are legible) I know what they are and what they mean, even if I just see them as individual characters. My assumption is that you are saying that there is something deeper in terms of linguistic context / linguistic evolution that I'm not aware of. What is that..."X factor"?

I will say, if nothing else, I can understand certain physical considerations. For example:

A person who is right-handed, and is writing on the right edge of a page may start to slant, because of the physical issue of the paper being high, and the hand losing its grip. By comparison, someone who is left-handed might have very smudged letters because their hand is naturally going to press against fresh ink, or alternatively, have very "light" because they are hovering their hand over the paper while the ink dries.

In those sorts of physical considerations, I can understand why it would matter to be able to see the entire page, because the manner in which they write could change depending on where they were in the page...but wouldn't the individual characters still look approximately the same? That's the bit I'm not understanding.

The lower case "e" in gothic cursive often looks like a lower case "r". If you see one of these: ſ maybe you think "ah, I know that one, that's an S!" and yes, it is, but some scribes when writing a capital H makes something that looks a LOT like it. You need context to disambiguate. Think of it as a cryptogram: if you see a certain squiggle in a context where it's clearly an "r", you can assume that the other squiggles that look like that are "r"s too. Familiarity with a scribe's hand is often necessary to disambiguate squiggles, especially in words such as proper names, where linguistic context doesn't help you a lot. And it's often the proper names which are the most interesting part of a document.

But yes, writers can change style too. Mercifully, just like we sometimes use all caps for surnames, so some writers would use antika-style handwriting (i.e. what we use today) for proper names in a document which is otherwise all gothic-style handwriting. But this certainly doesn't happen consistently enough that you can rely on it, and some writers have so messy handwriting that even then, you need context to know what they're doing.

The problem is payong experts to properly train a model is expensive, doubly when you want larger context.

Ots almost like we need a shared commons to benefit society but were surrounded by hoarders whp think they cam just strip mine society automatically bootstrap intelligence.

Surpise: Garbage CEOs in, garbage intelligence out.

Reply on news.ycombinator.com