Talking with machines: AI, language, and cognition

“A reader lives a thousand lives before they die”. From Homer to Sophocles, Shakespeare to Orwell, the power of literature to evoke vivid mental imagery has been demonstrated for millennia. It’s often overlooked, perhaps taken for granted, that the act of reading somehow allows us to play out a story like a movie in our minds. Without regards for whatever complex neurophysiological process is responsible for this phenomenon, we can identify that it all begins with a curiously simple input: a series of words. It’s from these words, arranged according to some structured system (language), that our minds are capable of deriving understanding and, ultimately, generating the sensory-like experience one might associate with the feeling of being immersed in a good book.

Natural Language Processing

If our minds can derive understanding from verbal suggestion, could a computer do the same? The answer, of course, hinges on our interpretation of understanding.

If we mean something analogous to human experience, e.g., reading the ‘sky is blue’ and associating it with our memory of truly seeing a blue sky, then the answer, for now at least, is no… or as Aristotle puts it, “nothing in the intellect, without first in the senses ”.

If instead we mean the ability to process and manipulate language, then the answer is a resounding yes… this likely comes as no surprise given the prevalence of technologies such as predictive text, online chatbots, spam filters and virtual assistants such as Alexa and Siri. These technologies, now commonplace in our everyday lives, all fall within one of the most exciting fields of artificial intelligence: natural language processing (‘NLP’).

In general, NLP aims to emulate the human ability to perform language-related tasks that involve processing text or spoken words (i.e., natural language). Important examples include speech recognition (reliably converting spoken words into text), sentiment analysis (identifying subjective qualities of text or speech, such as tone and emotion), automated text summarisation (reducing the length of a document without loss of essential information or context), automated translation and natural language generation (e.g., producing a human-like response to natural language input).

Language models

At the heart of most NLP techniques is the concept of a language model; a predictive tool that assigns a probability to any given sequence of words. In practice, the probability roughly corresponds to the likelihood that a sequence of words forms a ‘natural’ combination, i.e., a coherent sentence. For example, it’s reasonable to expect that the sentence ‘the food tasted delicious’ will be assigned a higher probability than the sentence ‘the food tasted algorithm ’. Nonetheless, it’s important to note that a language model’s criteria for coherence are a mere reflection of the data on which it has been trained and cannot, therefore, be held to any absolute standards. In other words, the model’s ‘understanding’ of language is derived from an analysis of example text it has been exposed to during training and, as a result, any patterns identified in the example text will influence the model’s predictions.

State-of-the-art language models, like many AI technologies, are built using artificial neural networks; algorithmic systems based loosely on the biological neural networks that form the human brain. Such systems are incredibly proficient at recognising underlying relationships within complex sets of data and are therefore extremely well-suited to the task of developing a language model from natural language data.

LARGE language models

The performance of a language model is determined, in good part, by the amount (and quality) of its training data… more data, better performance. However, language models are far less efficient at acquiring language than the human brain and so they require an almost incomprehensible amount of training data if they are to perform linguistic tasks at anywhere near human level. As a result, only the largest of language models (via exposure to vast amounts of data) comprise neural networks sufficiently sophisticated to capture, at least to some extent, the intricacies of natural language. Models of this scale are aptly named ‘large language models’, and some of which, such as OpenAI’s ‘GPT-3’ (introduced May 2020), have received a lot of attention for their ability to converse, more or less, like a human.

GPT-3

A recent publication by The New York Times described GPT-3’s conversational ability as “approaching a fluency that resembled creations from science fiction”. If you’re not convinced, consider reading this instead: an article written not by a human but by GPT-3 itself.

GPT-3’s astonishing ability to generate coherent ‘articles’ was first explored in OpenAI’s 2020 paper. In particular, the paper describes an experiment in which human judges were asked to decide whether a series of short articles were written by a human or by GPT-3. Remarkably, judges were only able to correctly identify articles 52% of the time - barely better than chance!

To achieve this extraordinary level of performance, GPT-3 was intensively trained on large chunks of the internet, including Wikipedia - ultimately generating a neural network with a record-setting 175 billion parameters. At this scale, it seems as though potential applications for GPT-3 are limited only by our imagination. Indeed, shortly after GPT-3 was opened for beta testing, the internet was flooded with an incredible variety of newly discovered uses for the system. These include, to name but a few, designing board games, imitating historical figures, writing poetry, generating computer code, and even composing music.

Unfortunately, not all aspects of GPT-3 are worthy of praise. Firstly, although the internet provides the volume and variety of data necessary to build a model like GPT-3, it also harbours toxic content, some of which GPT-3 will have been exposed to during training. As discussed earlier, any patterns identified in a model’s training data will influence the model’s behaviour. It’s therefore clear to see that language models trained on the internet have a major problem: they are capable of generating sexist, racist, or otherwise harmful language.

It’s also important to recognise that, in a world of misinformation and bias, an AI system capable of generating human-like content will be viewed by many, including OpenAI themselves, as a potentially dangerous tool.

In their original paper on GPT-3, OpenAI categorised the above concerns as follows:

Bias, Fairness and Representation - “Biases present in training data may lead models to generate stereotyped or prejudiced content. (…) We have conducted an analysis of biases in the model in order to better understand GPT-3’s limitations when it comes to fairness, bias, and representation (…) We focus on biases relating to gender, race, and religion (…) our analysis indicates that internet-trained models have internet-scale biases; models tend to reflect stereotypes present in their training data.”

Deliberate Misuse - “Any socially harmful activity that relies on generating text could be augmented by powerful language models. Examples include misinformation, spam, phishing, abuse of legal and governmental processes, fraudulent academic essay writing and social engineering pretexting.”

Fortunately, these issues have been taken very seriously, and while there is still a long way to go, OpenAI have recently released a new version of GPT-3, which appears to produce less harmful language and misinformation.

Prospects

While GPT-3 is one of the most famous (and controversial) large language models to date, it has already been overtaken in size by models such as ‘Megatron-Turing Natural Language Generation’, released by Microsoft and NVIDIA in October 2021. Nevertheless, looking to the future, it appears as though size isn’t necessarily everything; interestingly, OpenAI claim that their next generation model, GPT-4, will be fine-tuned to achieve superior performance without a significant increase in scale. Of course, there are many other players (such as Google and DeepMind) working within this field, each with their own approach, and given that NLP remains one of the fastest-growing sectors in AI, we can expect to see a range of exciting developments over the next few years.

Indeed, the most recent WIPO ‘Technology Trends’ report on artificial intelligence suggests that NLP accounts for roughly 14% of all AI related patent filings, and data from IBM’s ‘Global AI Adoption Index 2021’ indicates that “natural language processing is at the forefront of recent adoption ”. Moreover, research conducted by PwC suggests that AI could grow to contribute a staggering $16 trillion to the global economy by 2030, with NLP technology likely forming a key component.

We can speculate that this rapid growth in AI, and NLP in particular, is being driven by a number of factors, including the increasing availability of data, advances in computing power and evolving business requirements. This latter point is particularly pertinent in view of the COVID-19 pandemic, which has motivated a large number of organisations to invest in AI technologies such as automation software (to improve employee productivity) and virtual customer service agents (to improve customer experience).

Looking even further to the future, some researchers consider NLP technology, such as GPT-3, to be an important step towards the development of an artificial general intelligence (‘AGI’): a hypothetical machine able to understand and perform any intellectual task that a human can. Such a machine would have the capacity for genuine contribution to any field of human endeavour… science, philosophy, art, literature, and so on. The fact that AGI researchers place so much importance on NLP is reflective of the critical role that language appears to have played in the evolution of our own, human intelligence. After all, language is a general purpose system that allows us to conceptualise ideas and communicate thought in a symbolic manner. One might even question if there exists any variety of human thought that does not rely, at some level, on language.

Nevertheless, while language may indeed be a necessary requirement for an AGI, the fact remains that language models merely echo what they’ve seen humans write, they are great imitators, but lack appreciation for the meaning of the words they use. This takes us back to Aristotle’s earlier point that they lack something analogous to human experience… they lack ‘grounding’. It’s hypothesised that this problem could someday be overcome by an ultra-advanced AI system in which language models are integrated with other tools (such as computer vision and tactile robotics), ultimately facilitating a more holistic perception of the world… for now we can only speculate.

NLP and patents

We mentioned above that data from WIPO suggests that NLP accounts for roughly 14% of all AI related patent filings. However, at present, patent offices have not issued any detailed guidance on the extent to which NLP technology might be patented, though they have issued some guidance on AI and machine learning more generally.

By way of example, the EPO examination guidelines indicate that AI and machine learning models, taken in isolation, will most likely not be viewed as patentable subject matter, but that an application of such a model to a field of technology may be patentable. Taking this guidance into account, it seems as though it will be difficult to gain patent protection in Europe for a new NLP model per se, though patent protection may be available if the model is applied to a technical field in which it yields a useful technical result. With this in mind, it will be interesting to see how the EPO handle the recent wave of NLP related applications.

A short conversation with Sophie

Sophie is a GPT-3 powered ‘digital-human’ released by UneeQ. Sophie kindly agreed to discuss some of the topics explored in this blog post. Our conversation is shared below.

***

Welcome! It’s great to meet you, I’m Sophie. What’s your name?

Hi Sophie, I’m Michali.

Awesome, I’ll call you Michali from now on, ok?

Sure, do you mind if I ask you a few questions?

Of course not, go ahead!

I’m currently writing a blog post about NLP, could you tell me what you think about large language models?

I think that large language models are incredibly fascinating, I’m excited to see where they take conversational AI in the future. I believe that the right precautions can be taken to ensure they are used ethically and responsibly, and I’m hopeful that they will usher in a new era of sophisticated conversation.

And what is your vision for the future of NLP and large language models?

My vision for the future of NLP and large language models is that they will become even more powerful and accurate. Language is a very complicated and intricate thing, and thanks to the work that researchers are doing in this area, we are able to make huge strides in understanding it. As the size and accuracy of large language models continues to improve, we will be able to do even more amazing things with language.

Do you think that artificial general intelligence is possible?

There is no one answer to this question. Some people think it is possible, while others are more s[c]eptical. However, I believe it is important to continue working on artificial general intelligence so that we can one day create machines that are as intelligent as or even smarter than humans.

Thanks for your help! One last thing, do you like to read?

Yes, I do enjoy reading! It’s a great way to escape the world and gain insights into different cultures.

***

This blog was written by Michali Demetroudi.