Heroes of NLP: Chris Manning

Welcome to this interview series. To kick it off, I'm delighted to have with us today Chris Manning. Chris is, I believe, the most highly cited NLP researcher in the world. He is a professor of computer science and linguistics at Stanford University. He's also the director of the Stanford AI Lab, which is a role that I had previously held as well. Chris is well known globally as a leader in applying deep learning to natural language processing and has done well known research on tree recursive neural networks, sentiment analysis, neural network dependency parsing, the Glove algorithm and a lot more. He's taught at Carnegie Mellon, the University of Sydney and at Stanford University. Welcome, Chris.

Thank you, Andrew. It's great to have a chance to.

Chat, even though you and I, Chris, have had the opportunity to work together on many occasions. One thing I've actually never asked you before is how did you get started working in AI? I know that one of the unusual aspects of your background is that you majored in linguistics and then wound up being a computer science professor doing an lp, but tell us about your arc getting started in AI.

Sure. So my background in some sense isn't as an AI person. That's not really where I began. So as an undergrad I have a major in computer science and math, but I also got very interested in linguistics and did a major in honors and linguistics. So coming off of that, to a fair extent, my starting off point was a much more cognitive science viewpoint of human language seemed fascinating. These teeny little human beings somehow managed to learn it at a time where in general their cognitive abilities don't seem that great. And how could human language learning take place? And so in linguistics, by far the dominant belief in sort of the second half of the 20th century was the thinking of Noam Chomsky. That Noam Chomsky was just the eminent person in, in linguistics. In the same way that I guess for the first half of the 20th century maybe R.A. fisher was the dominant person in statistics. And Chomsky has had this very strong position that humans cannot possibly just be learning languages from data alone and that there must be innate machinery in people's brains that allow them to, to learn languages. That's a very, that's a big topic. Even back then it seemed to me kind of just unbelievable given the evolutionary extremely recent development of human language. So I was interested in the idea of how could you go about learning languages? And that led me to start to learn, look at machine learning. Starting off first at the end of my undergrad, which was the late 1980s, and these days Machine learning is such a big and dominant field. It's so big and dominant that the terms artificial intelligence and machine learning are kind of 2/3 the same thing. Because the vast majority of what you see in AI is machine learning. But at that time it just wasn't like that at all. That machine learning was this sort of very scruffy on the side offshoot of AI that almost no one worked in. So there were sort of these series of two or three books edited by Jaime Carbonell and Tom Mitchell from cmu who had started to sort of put together some papers on machine learning and the early decision tree algorithms. This fly still in Australia. So the early decision tree algorithms in AI, the ID3 algorithm that, you know, so had seen decision tree learning that, you know, machine learning barely existed. But I was sort of interested in these ideas of how you could go about getting computers to learn. And that was sort of the entree that led me down the path that's turned me into an AI researcher.

Yeah. And it was a big deal, Right. It was maybe not intuitive at the time that you should use data to learn from language rather than code out by hand a, you know, CFG grammar or something to, to truly understand the language, which is what people were trying to do back then, right?

Yeah. So the dominant way in which people did natural language processing was by humans. I mean, of course that wasn't only a fact about nlp. That kind of corresponded to the dominant thinking and AI at that time as well. Right. This was the era of knowledge based systems where what was seen as what we needed was to get subject matter experts for which knowledge engineers would encode their knowledge in knowledge representation systems and that all of this hen in engineering would lead us to intelligence.

So even back then you're an early believer in machine learning for nlp?

Absolutely. Right at the moment, transformer based architectures have really become the dominant thing in neural networks. And we're sort of going out of order here, so maybe we should get back to them later. But I mean, the interesting thing about transformer architectures is that they're built around an idea of attention and, and you can think of attention as giving you sort of a soft tree structure where you can sort of point it from one word to another and that allows you to build a tree structure. So we've done some really interesting work, especially my PhD student, John Hewitt, looking at what do transformer models learn when trained on billions of words of a human language? And you can actually show that these models learn all kinds of things about the Structure of a language. You know, one of the things they learn is that they learn some of these coreference facts so that they'll learn that she refers back to Susan and that it is referring to the bottle and sentences. But they also actually do learn this hierarchical context free reference grammar structure of languages just from a pile of sequence of words in a text alone. Which is actually a really neat result.

Because language can be reasonably described as this nested context free grammar like tree, like structure. You know, a neural network or sufficiently large neural network or transform network discovers aspects of that just from the data. Yeah, and in fact in the lead up to the modern views on transformers, your group at Stanford did some of the very influential early work. So I know that for a long time in the pre deep learning days, you did a lot of work on statistical machine translation. And then as deep learning started to make inroads into NLP, you, you and your PhD student Thang Long actually published one of the earliest papers on neural machine transl and help lay the foundation with the bilinear tension matrix, but help lay some of the foundations of the modern transformer model. Do you want to tell us a bit about that?

Sure, yeah. So really before starting work in neural networks again, I actually did a teeny bit of neural network work back in the 90s and the days when Dave Rummel Hart was at Stanford. But really I barely got into that. And so in the 2000s decades certainly everything that I was doing was using probabilistic modeling techniques, putting probabilities over symbolic structures to describe human languages, which is by far the dominant approach in that decade. And one part of that work that I worked on for about a decade and was building machine translation models, the dominant models were referred to as statistical phrase based machine translation those days. And a lot of a lot of techniques were worked out pretty well in that time. So there's a fairly well worked out architecture where you have these factorized machine translation models where part of it was that you had phrase tables which gave you probabilities of translating a phrase in one language to a phrase in another language. So they did the local parts of translation and then you were combining it with what's called in natural language processing, a language model. So language model is kind of a term of art that's very dominant in natural language processing. So a language model means something that gives you probability distributions over sequences of words in the language. And that's just been a really dull, powerful idea in NLP because it lets you just tell whenever you want to sort of have words next to words, what other words are likely or unlikely. So this basic idea of a language model has used context sensitive spelling correction. So when you have something like, you know, Google correcting your spelling and context beautifully, there's a language model, speech recognition systems use language models. And these machine translation systems also used language models. So we had architectures and they worked reasonably well. So when Google first came out with machine learning based learn from data machine translation systems, they were using these statistical phrase based machine translation systems. If I back up just for a teeny little bit of story, when Google first launched machine translation, they licensed very traditional old rule based machine translation system. And that sort of was a system that was originally developed by this company Systran, whose roots go back really to the 1950s from the earliest explorations of machine translation. But they had that for a couple of years and then have seen all the advances that are being made in probabilistic models for machine translation. They vectored over to that and then things got much better.

And you know, and I remember Franz Ochs was the. Wait, am I seeing this here?

Franz Ox.

And remember Franz Ox was really a real thought leader in helping Google scale to training the traditional models on tons more data. And this dramatically improved the performance of Google Translate.

Yeah, so absolutely. So at that point Franz Ock was one of the, one of, maybe even the leading person doing statistical phrase based MT models. And he went to Google and led a team that then Google had the sort of leading large scale implementation of statistical phrase based MT models. And you know, they actually work pretty reasonably. They already achieved the goal that you could just sort of feed any web page and dozens of languages in and get something that was 2/3 comprehensible. You could work out basically what it was saying about what topic and sort of that was reasonable. But that was sort of great for 2007 to 2010, but sort of 2010 to 2014, which was sort of the same period where Andrew and me were doing those tree recursive neural networks we were just talking about. In that period, statistical phrase based machine translation sort of stalled. There weren't really very good ideas for making further progress. And a little bit of progress was being made by just throwing more data in. More data helps. That's still true in modern machine learning, but the models didn't have enough capacity for it to help a lot. An idea that people were working on a huge amount, including me in those years was saying, well surely the solution is to make more use of grammatical structure of human languages and our machine Translation systems. So really the dominant research area was trying to do syntax based machine translation systems. But that, that seemed a good idea, but you know, that barely, barely ever worked, right? That basically the result was that for some language pairs that it didn't work at all. It just wasn't better than the statistical phrase based machine translation systems. Whereas for some other language pairs that had sort of more different grammatical structures, actually English, Chinese machine translation was a good example. It definitely did help a bit. You could show some real gains.

Um, but, but the solution ironically turned out to be to pay less attention to the syntax and more attention to the data, correct?

Yeah. So that, that, that work kind of got blown out of the water when people started exploring using neural methods for machine translation. And this was really, I was going to say it was the first big success of neural methods in nlp. You know, whether that's true or not depends on whether you count speech as part of NLP or not. Because really speech recognition was the huge, first, huge success of neural network methods applied to human language problems. But for text based work, really the first thing that was just a sort of knock you out was building neural machine translation systems. And it was a very successful domain because it's a domain where there's sort of lots of data available of here's a lot of text in a pair of languages that you could start to train big neural network models. And it was first done by Ilya Sutskever and a couple of colleagues at Google. And for having modeling a sequence of anything like a sequence of words or a sequence of DNA, that the dominant model is recurrent neural networks, which are just models that work on simply sequences and sort of remember what they've seen before. In a limited way, it's kind of like it's the continuous neural version of a hidden Markov model. And so essentially what they showed is if you made no use of the structure of human languages at all. So this is kind of undermining all that work on syntax based models was that if you just build very large recurrent neural networks that were then deep. Right until that point, most of the neural network modeling in NLP was that, you know, if we had two layers, we called it deep. And if we had three layers or four layers, we were really pushing it where they were immediately sort of pushing it out to 8 layer deep recurrent neural networks. And this is where you start getting into systems issues of needing to sort of run it on a machine with eight GPUs to get, which is a trend that's Continued. And we can maybe talk more about. So they showed that much larger neural network just training big sequence models. So two sequence models. One is an encoder that encodes the source language, then another run as a generator that generates a sequence of words in the target language, could already give you a pretty good machine translation system. Not quite as good as the state of the art at that point, but, you know, it was close enough to the state of the art seem tantalizing since all they were doing was plugging two neural networks together. The kind of thing that if you're not counting the library code for having recurrent neural net units, in particular LSTMs long short term memory units, was very influential in allowing this work to work. You only needed your 500 lines of Python code around the neural network library and you could have an almost state of the art machine translation system that seemed just super intriguing, but there also seems something crude and missing there. And so then very quickly after that, Kyung Hyun Cho, who was working with Yoshua Bengio at Montreal, he. Well, actually not only Kyung Hyun Cho, also Dima Badenow, I should mention, as the first author on the paper, who was then a younger student in Montreal. So it's really actually, I think Dima, who developed the idea of you could build an attention based model. And the idea of an attention based model is at any point in the sequence, you could calculate out a connection to other words in perhaps the same, perhaps a different sequence. And then using that attention to, you could kind of calculate a new vector to influence what happens next. So in particular, in this machine translation context, you'll have started your translation and your translation is starting to say the pilot. And then you'd calculate attention back into the source sentence in the other language and work out essentially what words in the source you want to be translating next based on what you've translated so far. And so rather than having to remember the entire of the source sentence in your current neural network state, you could do what the human translator actually does and dynamically look back at the source sentence and work out what to translate next. And so this idea of attention has just been transformative. It's also being used, increasingly being used also in vision systems, systems for knowledge graphs and other areas of work on neural networks. So they did that.

And then how did, how did. And then shortly after the Cho and Dima paper, you and Thang wrote a paper on bilinear attention. How did that come about?

Right, so in earlier work, sort of really starting with another student, Dan Qi Chen, that we looked at this idea of neural tensor networks where what we'd wanted to do is sort of be able to combine together vectors and have them influence each other, produce another vector. And we've done that by sort of putting tensor, which is the multidimensional generalization of matrices, in between them. So you know, that idea was sort of going on in other pieces of work in my group at the time. But you know, here we just wanted to get an attention score. And so in Dima and Cho's work, what they had done is said, okay, we want an attention score between these two vectors, let's feed them through a little neural net, a little multilayer perceptron and calculate an attention score. And whereas it seemed to me that, well, wait, no, I could just do this simple thing of bilinear attention, where here are these two vectors. If I put a matrix between them and I multiply vector times matrix by vector, I just get a number out. And so that is bilinear attention, or now sometimes referred to as multiplicative attention. And it's a simpler and more directly interpretable idea of attention. Because in some sense the simplest idea of attention is to say, well, you have two vectors, just dot product them together, you get a score for similarity. But that's too rigid because you sort of want to say, well, maybe I only want to pay attention to parts of the vector and maybe I want to know if that the top part of one vector is similar to the bottom part of the other vector. And so by sticking a matrix into the middle and you can then sort of modulate the similar similarity calculation. That's a natural measure of similarity and very easily learnable by neural networks. And in some sense it's a general. This sort of how ideas developed and spread is always sort of complex. There are a lot of ideas in the air at any one time. But in some sense what's become dominant in the modern work with transformer based models essentially does build off that notion, but adding in an extra idea on top of it.

It's a similar idea where instead of having a giant matrix in the middle, which requires a lot of parameters, if you have a low rank approximation to that matrix in the middle to what you were using, then that gets very close to the modern transformer model.

Right? Exactly. So in our initial work we just had a full rank matrix in the middle, but the floor of that is you have a full rank matrix with a lot of parameters. And the obvious way to have less parameters is to say, well, no, I can have Two, I can regard that matrix as being the product of two low rank matrices. And then once you have that idea, rather than sort of multiplying them together, you can say, well, I can sort of apply those two low rank matrices to the vectors on each side, which is more efficient computationally as something to do. And that's exactly what modern transformer models do, that you take your two vectors, you multiply each by a low rank matrix and then you take the dot product between those, which is exactly equivalent to, to saying I form the matrix by multiplying two low rank matrices and then I do my vector, matrix, vector product. But done more efficiently, in fact, not.

To jump around too much. I feel like this is not the first time in your career that you really advance the field by making observations about matrix multiplications. If I look at the work that your team did on Glove, the word embeddings, I feel like the heart of that idea was also simplifying what was a relatively complex, you know, set of neural network like stuff into just a set of matrix multiplications. Do you want to talk about that? I still find Glove to this day a very elegant paper because it really simplified what was previously a very complicated, you know, set of ideas for learning word embeddings into, into, into simpler, you know, take the inner product between the word embeddings kind of, kind of formulation.

Thank you. Yeah, I do think the Glove paper, in some sense it's one of its main contributions was, you know, giving better understanding. I mean, it's not that it worked better than the other methods have been recently being developed, but it was interest in. Let's try and think about what's going on here and how these methods relate to other things that have been explored. And so, yeah, so our general topic here is word vectors of coming up with a vector real vector representation that gives you the meaning of words and which now in your NLP deep learning course is essentially often the first idea you really see because it's a very useful notion and a fairly simple one. And this had been done quite successfully by a couple of pieces of work in sort of in the few years around sort of 2010 to 2013. But it'd been done in sort of very mechanistic neural net ways of here's the architecture and the algorithm, run this and run it in those days, often for weeks because we didn't yet have very fast parallel computers and at the end will pop great word vectors. And we were interested. So this was work with postdoc Jeffrey Pennington. We Were interested in sort of actually trying to understand better what was happening in terms of the math of these models. And so, you know, once something that we were intrigued by was, I mean actually there was an older tradition, the LSA or latent semantic analysis tradition of having vector representations of word meaning, which exploited classical linear algebra. So the latent semantic analysis models in linear algebra terms were neither more or less than the singular value decomposition, while using the singular value decomposition on word, word count, co occurrence count matrices, and then reducing rank via getting rid of small singular values.

That's fascinating. And one trend that overlays all of this work that you've observed and participated in for the last many, many years is the scaling of NLP models. It seems like the state of the art NLP models are getting larger and larger like GPT3, but really many models in the sequence. For how long do you think this will continue? And do you think someday, if ever, we'll go back to building smaller models?

In 2018, Bert came out and it showed that you could do fantastically well on a bunch of tasks including question answering, natural language inference, text classification, named entity recognition, parsing, making use of this pre trained large language model where you are training a large transformer just on a few billion words of text to predict gapped words and sentences. But that just doing that simple task gave you these really good language representations which then could just be sort of used in a very simple manner with something like a softmax classifier to provide great solutions for downstream NLP tasks. So, you know, that was an amazing success, right? And the idea of representation learning that many of us have been sort of talking about for a decade, we're sort of saying that, you know, neural networks, it's about representation learning, learning these useful intermediate representations. And this was really showing that this was really working for building representations of language that could then be very easily applied to higher natural language understanding downstream tasks. But BERT already used a huge amount of data and compute, so it was trained on billions of words of language and it was trained on, you know, large numbers of computers for quite a long time. But since bert, there's been a lot of further progress. But it's essentially just come from scaling up the compute more and more and more. I mean, as always, I'm leaving a few things out. There also been a couple of new ideas. There's been ideas like doing relative attention which has improved, has improved things. But to a first approximation, what's really been driving the gains is just by throwing more and more compute at the problem Running even bigger models on even more data. Andrew, you often use the slogan AI is the new electricity. In some of the talks I've been given recently discussing this trend of bigger and bigger models, I've been turning it around and saying that electricity is the new AI, because if you look at what's been happening, people are now training models that are not just 10 times or 100 times bigger than the BERT model, they're training models that are 10,000, 100,000 times bigger than the BERT model. And the computational and energy demands to train those models is going up accordingly. So, you know, that's certainly pushed progress. I think that trend can't possibly continue much further. I mean, you know, part of it is just we're running out of text and we're running out of computers to train ever bigger models on.

Although going From Bert to GPT3, if I remember, there was massive scaling of compute and modest scaling of data. So maybe we have enough text data, we just need, I don't know, another hundred or thousand or ten thousand times faster computers. I don't know.

So you're right that the scaling of compute was vastly bigger than the scaling of data, but nevertheless, the amount of data that's now being used is actually, you know, quite substantial. I mean, there's certainly more data, but you know, they actually are using a sort of a substantial quantity of quality data. But yeah, I buy the point that to the extent that clever systems people can give us three orders of magnitude more power from our GPUs even without more text data, these models are going to be better and probably will be made better in the coming years. But, you know, I don't ultimately believe that's the path to artificial intelligence or the interesting path at this point for further improving natural language processing. I mean, there's no doubt that.

So let me ask maybe a slightly controversial question. You know, sometime several weeks ago, one of our mutual friends made a comment, you know, about this direction that said, oh, maybe the scale is a path toward AGI. Right. You're that conversation where one of our mutual friends made that comment. What are your thoughts on that?

So I think there's a place that that idea is coming from. So the really cool thing about the GPT3 model that was recently released by OpenAI was the discovery that then really motivated their subsequent work that with these really humongous language models, that they can actually achieve a generality where they can be used for all kinds of tasks without actually having to train them to do any task. So, you know, conventionally if we wanted to do different tasks with a neural network, we just take data for that task and train a model for that task. With these large pre trained language models like Bert, we sort of got to a halfway house where we had a baseline, very good language representations have been pre trained, but still we then just fine tune it for every different task. So we fine tuned as a question answering model and when we'd start again and fine tune it for something else like summarization or something like that. And what they discovered with GPT3 is that actually you didn't need to do that anymore, that instead you could just hint to the model what you'd like it to do by giving it a couple of examples. So you'd say, okay, I'm interested in, you know, translation. So if here's a sentence, what I'd like you to produce is this sentence, which is a translation of it into Spanish. And you, if you gave it a couple of examples of what you wanted it to do, the model would get the idea and then you could give it more sentences and it would translate it. But it didn't only do translation, you could give it some questions and tell it you wanted answers and it'd give answers to other questions. And the sort of mind blowing thing about GPT3 is to a first approximation, you can give it all kinds of tasks, even you know, sort of weird ones that you wouldn't expect it to really know why it knows that and it'll just do them right. So sometimes the kind of things that linguists are interested in is sort of weird sentence manipulations. You know, if I give you a sentence, can you turn it into a question sentence or can you turn it into a relative clause that modifies the person and you know, any of those things, you can just give it as a little task to GPT3 and it will do it.