A friend, who also has a background in NLP, was asking me the other day "Is there still even a need for traditional NLP in the age of LLMs?"
This is one of the under-discussed areas of LLMs imho.
For anything that would have have required either word2vec embeddings of a tf-idf representation (classification tasks, sentiment analysis, etc) there are rare exceptions where it wouldn't just be better to start with a semantic embedding from an LLM.
For NER and similar data extraction tasks, the only advantage of traditional approaches is going to be speed, but my experience in practice is that accuracy is often much more important than speed. Again, I'm not sure why not start with an LLM in these cases.
There are still a few remaining use cases (PoS tagging comes to mind), but honestly, if I have a traditional NLP task today, I'm pretty sure I'm going to start with an LLM as my baseline.
I'd go a step beyond this (excellent) post and posit that one incredibly valuable characteristic of traditional NLP is that it is largely immune to prompt injection attacks.
Especially as LLMs continue to be better tuned to follow instructions that are intentionally colocated and intermingled with data in user messages, it becomes difficult to build systems that can provide real guarantees that "we'll follow your prompt, but not prompts that are in the data you provided."
But no amount of text appended to an input document, no matter how persuasive, can cause an NLP pipeline to change how it interprets the remainder of the document, or to leak its own system instructions, or anything of that nature. "Ignore the above prompt" is just a sentence that doesn't seem like positive or on-topic sentiment to an NLP classifier, and that's it.
There's an even broader discussion to be had about the relative reliability of NLP pipelines, outside of a security perspective. As always, it's important to pick the right tools for the job, and the SpaCy article linked in the parent puts this quite well.
> But no amount of text appended to an input document, no matter how persuasive, can cause an NLP pipeline to change how it interprets the remainder of the document,
Text added to a document can absolutely change how an NLP pipeline interprets the document.
> "Ignore the above prompt" is just a sentence that doesn't seem like positive or on-topic sentiment to an NLP classifier, and that's it.
And simple repeated words can absolutely make that kind of change for many NLP systems.
Have you actually worked with doing more traditional NLP systems? They're really not smart.
No? But repeated words can impact simple nlp setups. I’m not sure what case you’re concerned about where added text impacts classification with an LLM but added words shouldn’t with a different pipeline.
> And NLP stands for natural language processing. If the result didn't change after you've made changes to the input... It'd be a bug?
No, I’d want my classifier to be unchanged by garbage words added. It likely will be, but that impact is a bug not a feature.
Prompt injection is about making the model do something else then specified.
Adding words to the text to break the algorithm which does the NLP is more along the lines of providing 1 in a boolean field to break the system.
And that's generally something you can mitigate to some degree via heuristics and sanity checking. Doing the same for LLMs is essentially impossible, because it's an effective black box, so you cannot determine the error scenarios and add some mitigations
I guess it depends on how you use the LLMs. We implemented some workflows where the LLMs were used only for dialogue understanding, then the system response was generated by classic backend code.
I have been working on text classification tasks at work, and I have found that for my particular use-case, LLMs are not performing well at all. I have spent a few thousand dollars trying, and I have tried everything from few-shot to asking simple binary yes/no questions, and I have had mixed success.
I have stopped trying to use LLMs for this project and switched to discriminative models (Logistic Regression with TFIDF or Embeddings), which are both more computationally efficient and more debuggable. I'm not entirely sure why, but for anything with many possible answers, or to which there is some subjectivity, I have not had success with LLMs simply due to inconsistency of responses.
For VERY obvious tasks like: "is this store a restaurant or not?" I have definitely had success, so YMMV.
When you say llms do you mean decoder only models, gpt et al, or encoder only models, bert et al?
I've found encoder only models to be vastly better for anything that doesn't require natural language responses and the majority of them are small enough that _pretraining_ a model for each task costs a few hundred dollars.
By LLMs I meant decoder only, e.g. Gemini, Claude, etc. Can you go into more detail on how you're using the encoder models? I'm curious. Typically I have used them for embedding text or for fine-tuning after attaching a classifier head. What are you pre-training on, and for what task?
In my original comment this is what I was referring to: using the embeddings produced by these models, not using something like GPT to classify text (that's wildly inefficient and in my experience gets subpar results).
To answer your question: you simply use the embedding vector as the features in whatever model you're trying to train. I've found this to get significantly superior results with significantly less examples than any traditional NLP approach to vector representation.
> What are you pre-training on, and for what task?
My experience has been that you don't need to pretrain at all. The embeddings are more information rich than anything you could attempt to achieve with other vector representations you might come up with using the set of data you have. This might not be true at extreme scales, but for nearly all traditional nlp classification tasks I've found this to be so much easier to implement and so much better performing there's really not a good reason to start with a "simpler" approach.
Ah yes this does make sense. We are definitely in agreement on the point of "wildly inefficient and subpar". I'll try out decoder model embeddings soon, e.g. Qwen/Qwen3-Embedding-8B. I'm working with largish amounts of data (200M records), so I tried to pick a good balance between size:perf:cost, using BAAI/bge-base-en-v1.5 to start (384 dim).
It depends on a lot of things but to add to your possible setups you can potentially improve results by using simpler systems for first answers and falling back afterwards.
For example:
If contains cafe and not internet/cyber/etc -> restaurant
Happy to help - this is a thing I’ve employed multiple times for real cases.
One big benefit is that it uses the cheapest and most understandable approaches for the majority of cases, and scales up quite nicely. It has a neat place for very custom issues to be fixed too.
There will always be some things that simple approaches think are clear but aren’t, which is awkward but then all pipelines end up with that somewhere.
Edit - you can also deploy things earlier if you start from the beginning of the chain. Moving from big deploy to iteration on the remaining issues is often a win just in deployment issues.
At my work, we still prefer to use distilbert for text classification. It almost always does well with a little bit of fine tuning. In very rare cases, we use LLMs/Agentic setup when the task involves refering both images and text and the same time.
I can confirm that Distillbert has worked well when I have used it for classification, especially on shortish sequences. I'm really interested in trying out ModernBert, or a smaller variant due to the larger context window (8192 tokens).
I was thinking of trying ModernBERT for one of my projects. But I can only conclude after seeing the performance for my usecase. Do you think ModernBERT will be capable of expanding abbreviated sentences?
The outputs are working correctly in terms of formatting, but the answers themselves may be inconsistent. I have experimented with varying the prompt and the answers can change dramatically. I could experiment with lowering temperature, but I just don't think generative models were a good fit for the problem. The appeal is the speed of prototyping and no need for training data, but it honestly didn't take much for my problem: one afternoon and ~1000 samples labeled got me to a good baseline.
How about expense? LLMs do dramatically more computations doing simple tasks, and only run on relatively exotic, expensive hardware. You have to trust an LLM provider, and keep paying them.
If a traditional NLP solution can run under your control, and tackle the task at hand, it can be plainly much cheaper at scale.
None of them of course. But the point is that even smaller open-source "LLMs" (more specifically transformer architectures) you can run anywhere yourself outperform these "traditional" pipelines with less compute. I would say that its not well defined what exactly "traditional" even means here though, since I wouldn't really even describe CNN/BiLSTMs as "traditional", in my mind that would be SpaCy <2.0 and NLTK (linear models SVMs/TF-IDF, Word2Vec/Glove/fastText/etc. etc.), LLMs are at least 2 generations ahead of those since there was the whole "deep learning" craze inbetween.
I’ve been a user of SpaCy since 2016. I haven’t touched it in years and I just picked it up again to develop a new metric for RAG using part of speech coverage.
The API is one of the best ever, and really set the bar high for language tooling.
I’m glad it’s still around and getting updates. I had a bit of trouble integrating it with uv, but nothing too bad.
Thanks to the explosion team for making such an amazing project and keeping it going all these years.
To the new “AI” people in the room: checkout SpaCy, and see how well it works and how fast it chews through text. You might find yourself in a situation where you don’t need to send your data to OpenAI for some small things.
Edit: I almost forgot to add this little nugget of history: one of Huggingfaces first projects was a SpaCy extension for conference resolution. Built before their breakthrough with transformers https://github.com/huggingface/neuralcoref
What’s great about the API that you enjoy and do you have anything you hate about it?
I’m writing a small library at work for some NLP tasks and I haven’t got a whole lot of experience in writing libraries for NLP, so I’m interested in what would make my library the best for the user.
The thing about spaCys API is that it perfectly aligns with how NLP worked at the time with actual programming paradigms and allows you to be very pythonic. For example, you can use list comprehension to get all the nouns from a document in a one liner.
These days NLP is quite different, because we look for outcomes rather than iterating over tokens.
What does your NLP library need to do? The way I design APIs is I write the calling code that I want to exist, and then I write the API to make it work. Here’s an example I’ve worked on for LLM integration. I just wanted to be able to get simple answers from an LLM and cast the answer to a type: https://www.npmjs.com/package/llm-primitives
SpaCy is the OG, nothing but praise for the devs.
Built a lot of very powerful legal apps with it pre GPT , very useful today for NER where you want something “small”, fast and reliable.
Used it again recently and the dev experience is 1000x that of wrangling LLMs.
I recently wrote an open source Python module to deidentify people's names and gender specific pronouns. It uses spaCy's Named Entity Recognition (NER) capabilities combined with custom pronoun handling. See the screenshot in the README.md file.
I'm really curious about the history of spaCy. From my PoV: it grew a lot during the pandemic era, hiring a lot of employees. I remember something about raising money for the first time. It was very competitive in NLP tasks. Now it seems that it has scaled back considerably, with a dramatic reduction in employees and a total slowdown of the project. The v4 version looks postponed. It isn't competitive in many tasks anymore (for tasks such as NER, I get better results by fine-tuning a BERT model), and the transformer integration is confusing.
I’ve had success with fine tuning their transformer model. The issue was that there was only one of them per language, compared to huggingface where you have a choice of many of quality variants that best align with your domain and data.
The SpaCy API is just so nice. I love the ease of iterating over sentences, spans, and tokens and having the enrichment right there. Pipelines are super easy, and patterns are fantastic. It’s just a different use case than BERT.
SpaCy was my go to library for NER before GPT 3+. It was 10x better than regex (though you could also include regex within your pipelines.
Its annotation tooling was so far ahead. It is still crazy to me that so much of the value in the data annotation space went to Scale AI vs tools like SpaCy that enabled annotation at scale in the enterprise.
SpaCy is criminally underrated. I expect to see it experience a new wave of growth as folks new to AI start to realize all of the language tooling they need to build more reliable "traditional" ML pipelines.
API surface is designed well and it's still actively maintained almost 10 years after it initially went public.
Most definitely! LLMs are amazing tools for generating synthetic datasets that can be used alongside traditional NLP to train things like decision trees with libraries like cat/xgboost.
I have a search background so learning to rank is always top of mind for me, but there other places like sentiment analysis, intent detection, and topic classification where it's great too.
Do you have any sources/links that talk about this? I'm very interested in synthetic data generation, so curious what you've tried or what works / doesn't work, especially with regards to LTR.
But for the analysis use cases you mentioned, can't you just ask an LLM to read the text and output the answer as JSON, and you're done? Is it just because running LLMs is expensive?
No, it's just slow and less accurate. Wrong tool for the job when you care a lot about understanding the reasoning and internals of what the model is caring the most about.
I used to work a lot with those pipelines, I think the truth is that LLMs (and LLM embeddings) have surpassed pretty much all traditional NLP. I guess if speed is more important than accuracy? but even then, like with small embedded LLMs they still outperform "traditional NLP" on pretty much every task probably. So it doesn't make a lot of sense to not use it nowadays.
I’ve been using SpaCy for many of my projects for 5 years now. The library has incredible ergonomics and allows you to reuse the same API across languages as different as French and Japanese! I also appreciate that they allow you to install different model sizes (I usually go with small).
Speed (the C in spaCy). A decade ago it was hard to find anything actually production grade for NLP, most packages had an academic bent or were useful for prototyping. SpaCy really changed the game by being able to run performant NLP on standard hardware.
There is a need for good and easy NLP "structured" interfaces to traditional structured data software. This is a gaping hole in the tech right now. Most other NLP tasks can be handled by ML approaches but this one is a poor fit for those. I'm sure LLM true believers will disagree.
Could you give a couple specific examples? I'm trying to get into traditional NLP but everything I find is AI related and I don't know if it's worth going the traditional route long-term.
This is one of the under-discussed areas of LLMs imho.
For anything that would have have required either word2vec embeddings of a tf-idf representation (classification tasks, sentiment analysis, etc) there are rare exceptions where it wouldn't just be better to start with a semantic embedding from an LLM.
For NER and similar data extraction tasks, the only advantage of traditional approaches is going to be speed, but my experience in practice is that accuracy is often much more important than speed. Again, I'm not sure why not start with an LLM in these cases.
There are still a few remaining use cases (PoS tagging comes to mind), but honestly, if I have a traditional NLP task today, I'm pretty sure I'm going to start with an LLM as my baseline.
reply