Shoud I learn NLP in 2024? #datascience #machinelearning #ai

in this video we're going to be talking about one question should you learn NLP going into 2024 and I'm going to save you a little bit of time the answer is absolutely yes but the main reason why you're here and this is why you stick around to the end is because that's not actually the question that you care about the big question the one that you really care about is why should you learn NLP in 20124 and that's going to be what we address in this video now let's first talk about why I'm making this video I'm making it because I've gotten asked this question a lot over the course of 2023 and I think it's very useful to maybe talk about why a lot of my subscribers and people on Twitter and people in person are asking me this question and it stems largely from the advances in AI specifically around NLP over the course of 2023 at the end of 2022 we saw chat GPT come to the Forefront and chat GPT was fantastic because it brought a nice user interface that allowed for individuals who couldn't code to ENT interact with large language models things that have been around for a couple years by that point but chat GPT did another thing it really allowed for the broader use of this and the broader applicability of large language models and by that I mean people started to wonder if there was really a need to train a local model at all anymore now as 2022 kind of progressed and 2023 came around we started to see these models get even better and this question became more and more important and really stemmed around gp4s and its capabilities with doing various NLP tasks things like named entity recognition things like text classification things like natural language translation which is where we take maybe something from English and translate it into medieval Latin or vice versa over the course of 2023 we saw models get a lot better at these tasks and so naturally people wanted to know should you learn NLP anymore is it is it useful in any way and the answer is yes it is is and over the course of the next few minutes we're going to figure out why that's the case so on the surface these large language models might seem like they do a great job but they really have a couple major deficiencies that make them unreliable and not necessarily useful for production environments and let's talk about a couple of these so first of all large language models are very large as the name suggests and if you want to use the state-of-the-art language models the ones that are going to perform these tasks best they are going to sit behind paid Services things like open AI chat GPT or if you want to pay by the use something like the Open Eyes API where you can call up these same models and do more custom things with them so let's talk about this real fast the size of the model dictates a couple different things the main thing it's going to dictate is the cost it is to do a certain task to do a thing like text classification on 4,000 tokens might take a number of dollars but to do that on 16,000 tokens or 32,000 tokens is going to cost a lot more and the reason for this is because behind the scenes these large language models as they can process larger and larger amounts of data require more and more compute or computational resources this translates directly not only to Hardware costs or the hardware needed to host these models and use them but it also translates to electricity the cost to run these language models is it requires a great deal of electricity not only to train but also to actually run and this means that in a lot of scenarios to use the best model available to you might be cost prohibitive or it might not make actual sense when you start to think about how many times a workflow might need to be used in production or when it's kind of set up and meant to run in the wild the other thing that these large language models suffer from and this is a big one and that's going to be consistency now when you load up a text and you ask a large language model or you engineer a prompt to do a certain task like let's just say text classification or named into D recognition and I have a video on this channel that goes through that using a GPT Builder on the surface the outputs look really really good and I'm very encouraging it's not until you start to experiment a little further and maybe give it four five 6 7 10 more examples that you start to see some problems emerge and the biggest that you see emerge is consistency things that were annotated very well in one document might not be annotated very well in another one and you might start to see it degrade over time as you try to engage in maybe a chat scenario now there's a lot of different things happening behind the scenes that explain all these things but the big takeaway is that consistency in these large language models is very difficult to achieve and to achieve well and prompt engineering can only get you so far and few shot learning also have a the video on that can also only get you so far in solving this problem this problem is exacerbated however if you're working with very domain specific data or data from a very particular set of documents or a very particular uh kind of subfield that might not be well represented in a large language model or might be challenging for it to do certain tasks in this is also exacerbated even more if the things that you're trying to get the large language model to do consistently are very nuanced or very difficult maybe for you as a human to actually do expecting consistency from these large language models as these problems get more complex uh really shouldn't happen you should expect more likely for them to fail in these areas or at least not be that consistent now all of these problems can be remedied however if you work with well annotated that is consistently annotated training data and then train a smaller model to do that same task smaller models are often times more accurate than large language models model on doing these very specific tasks and they're often times actually absolutely are much more cheap to run so they're cheaper and more effective to run and they're also going to be a lot faster so if speed is a concern for you then absolutely these other Solutions are probably more appropriate for things like text classification or named entity recognition now the big question though is should I use large language models at all it sounds like I shouldn't the answer is of course not you should use the best of both worlds large language models are really good at being leveraged especially early on in a project when you're trying to get a sense of the data or you're trying to help start annotating data often times when you start a project there is not going to be a model available to you that will do the task that you need these large language models can be loaded up in an annotation environment such as Prodigy which comes from explosion AI the same people who give us Spacey there's a recipe to load up a large language model something like from open AI so GPT 3.5 or four and you can use what the language model is good at at annotating and consistency isn't an issue because as it annotates data for you you can make Corrections and this is really useful early in a project when you don't have a model trained yet after you get about maybe 500 or 1,000 annotations though it's time to train your first model and once you've trained your first model it might not be perfect it's not going to be your your final production ready model but it is going to be good enough to maybe start using that instead of something like GPT to assist in The annotation process it'll be cheaper and possibly even more accurate and probably more consistent so this is the way I've thought about how to answer this question and hopefully this video gives you a good sense about the strengths and weaknesses of large language models and really the utility of still learning NLP in 2024 NLP is the backbone of large language models and it's going to provide you the skill set for doing all the necessary things like training your own models or knowing what solutions to employ for certain problems all of this requires a background and understanding of what NLP has to offer as you go into 2024 and you want to start learning about NLP check out this channel I have about 200 videos on a lot of different NLP topics including how to get up and running with it with very little or almost no programming knowledge whatsoever with the Spacey library and you can check that out in this link right up here if you like this video like And subscribe if you get a lot out of this Channel please do consider supporting it via something like YouTube memberships which is linked down below or maybe give us super thanks or consider supporting it over on patreon thank you so much and have a great day and hopefully everyone has a great start to their new year

Channel: Python Tutorials for Digital Humanities

Views: 1,520

Keywords: NLP 2024, Natural Language Processing, AI Learning, Machine Learning, Technology Trends 2024, Data Science, Artificial Intelligence, Tech Education, NLP Tutorial, Career in NLP, NLP Skills, Future of NLP, Language Technology, Programming for NLP, Text Analytics, gpt4, gpt

Id: WorTIShsduo

Length: 8min 46sec (526 seconds)

Published: Thu Dec 21 2023