March 31, 2023, by Brigitte Nerlich
Chatting with a chatbot about metaphor
A week ago my son sent me a link to a tweet in which Ethan Mollick did a little experiment with one of the currently fashionable AI chatbots, a common pastime nowadays. He asked “Bing to look up Tamarian, a made-up Star Trek language that works entirely by metaphors to Tamarian history and myth. Then I asked it to do the same for English using American history. It gave remarkably subtle & interesting results”.
Metaphors and myths
The word metaphor made me sit up and prick my ears, but the rest went over my head. I looked up Tamarian and found out that it’s actually a really interesting fictional language. If you want to talk about ‘co-operation’ for example, you say “Darmok and Jalad at Tangara”, as these two guys fought together at that place. Or, to use a more earthly example, if you want to talk about ‘love’ you’d say “Juliet on her balcony”. As with metaphors, the mapping between source (history or myth in this instance) and target (cooperation or love) only works, of course, if you know something about the conceptual source domain, in this case who Juliet is and what she did. Anyway, Bing came up with some nice examples of this type of metaphorical/analogical mapping.
Metaphors and bridges
I looked a bit into the responses to this tweet and found one which was also quite interesting. Mr Build It said: “It’s a fantastic side effect / corollary of using natural language – metaphors are the bridges of pattern abstraction, so the LLMs are excellent at ‘thing like other thing’ sequences.” Ah yes, metaphors are indeed bridges between two conceptual islands. In the past I thought that this type of creative mapping between conceptual domains was a distinctly human ability, even perhaps a foundation of human consciousness. Now my beliefs are somewhat shaken.
As I don’t have access to Bing, I went back to my friend ChatGPT and got into a conversation with it about metaphor. What it said was quite illuminating about its abilities on the one hand and the way that humility has somehow been written into talking about them.
First, I asked it what it made of ‘Mr Build It’s’ view on metaphors as the bridges of pattern abstraction. It did not notice the bridge metaphor, but agreed largely with Mr Build It’s assessment of metaphors:
“The view that ‘metaphors are the bridges of pattern abstraction, so the LLMs are excellent at ‘thing like other thing’ sequences’ is a reasonable interpretation of how language models like GPT-3 can generate metaphors.
In essence, metaphors are a way of mapping one conceptual domain onto another, based on shared attributes or properties. This requires a certain degree of pattern recognition and abstraction, which are key strengths of large language models (LLMs) like GPT-3.
By analyzing large amounts of text, GPT-3 has learned to recognize patterns and associations between words and concepts, and can use this knowledge to create novel and meaningful metaphors. When generating a metaphor, GPT-3 is essentially creating a bridge between two conceptual domains, using the shared patterns and associations it has learned from its training data.
Therefore, the statement that ‘LLMs are excellent at ‘thing like other thing’ sequences’ is a fair characterization of how these models generate metaphors. However, it is worth noting that the effectiveness of the metaphors generated by LLMs may vary depending on the specific context and intended meaning, and they may not always capture the full complexity and nuance of human-generated metaphors.”
Note the humility at the end. We’ll get back to that.
Metaphors and pattern recognition
The issue of pattern recognition is one that, like metaphor, has interested me for a while. Isn’t that also something very important to human cognition? Indeed, as claimed by some:
“Pattern Recognition and Inductive Thinking is a special ability of the human brain to not only find patterns but figure out in a logical way what those patterns suggest about what will happen next. In a broad sense, pattern recognition and inductive thinking form the basis for all scientific inquiry.” So, hmmm, are these chatbots then a threat not only to metaphor creation and human cognition but even scientific inquiry…?
I wanted to know how pattern recognition worked if not in the brain, then in ChatGPT. I therefore asked it: “Can you explain how pattern recognition works in ChatGTP?” and it came up with an answer that was entirely incomprehensible to a computerphobe like me, containing phrases like “convolutional neural networks and recurrent neural networks”, “self-attention mechanisms” (I asked about that as well, but won’t go into this here), “feedforward neural network” and “backpropagation and gradient descent”. (For a good summary of how ChatGPT actually works, see this summary here, but it’s in German – you can always ask ChatGPT to translate the text, I suppose)
Anyway, it all seems to be about giving words certain values (weights) depending on their relations or associations with other words, which is a bit like what Ferdinand de Saussure had in mind about language as a system, plus the machine learning, of course. I might be wrong!!
Then I asked: “How does this enable you to generate metaphors?” And this time I understood the answer. It even used one of my favourite metaphors:
“ChatGPT can generate metaphors by leveraging its ability to recognize patterns in language and generate new text based on those patterns.
Metaphors involve making a comparison between two things that are not literally related. For example, ‘life is a journey’ is a common metaphor that compares the experience of living to the act of traveling from one place to another.
To generate a metaphor, ChatGPT can be prompted with a topic or concept and then instructed to complete the prompt with a comparison to something else. For example, if prompted with the topic ‘love’, the model might generate the following metaphor:
‘Love is a rose, delicate and beautiful, but with thorns that can cause pain.’
In this example, the model has recognized the patterns in language that are associated with the concepts of love, roses, and pain, and has used these patterns to generate a new sentence that combines these ideas in a metaphorical way.”
So that’s how it does it, yep that makes sense!
Metaphors and humility
Like all my conversation with ChatGPT about metaphor, this last one too, ended on a note of humility:
“Of course, generating effective and meaningful metaphors is a difficult task that requires a deep understanding of language and context. While ChatGPT can generate metaphors, it may not always generate the most insightful or creative comparisons, and human writers are still generally better at crafting metaphors that are truly impactful and memorable.” Let’s hope so!
This echoed a previous end to a conversation:
“[I]t is important to note that not all of the metaphors generated by ChatGPT will be successful or appropriate. The quality and relevance of a metaphor depend on the specific context in which it is used, as well as the preferences and expectations of the audience. Therefore, it is ultimately up to the human user to evaluate and judge the effectiveness of ChatGPT’s metaphorical output.”
What about my belief that the ability to create metaphors, to bridge conceptual domains and create new thoughts and feelings, is something distinctly human? It reassured me, a bit:
“Humans are able to use a wide range of metaphors and other forms of figurative language to express complex ideas and emotions, and the use of metaphor is an important aspect of human creativity and communication.
Moreover, the ability to recognize and understand metaphors is often seen as a hallmark of linguistic and cultural competence in humans. While some non-human animals may demonstrate some degree of metaphorical thinking or communication, the richness and complexity of human metaphors is likely to remain a uniquely human capability for the foreseeable future.”
But will it?
Metaphors and reality
One difference between human and AI metaphor competence may be that while the AI can produce and process metaphors when prompted, it might not be able to generate them spontaneously without access to the ‘real’ world, to reality, to a body and to experience…. As John Middleton Murry wrote in Countries of the Mind: “Metaphor is as ultimate as speech itself, and speech as ultimate as thought. … Metaphor appears as the instinctive and necessary act of the mind exploring reality and ordering experience.” (Murry, 1931: 1-2).
If we increasingly rely on artificial intelligence to do our thinking instead of engaging in metaphorical, creative and original thinking ourselves, human intelligence and reality will suffer in the long run. Both will evaporate together and meaning will disappear.
Image: Bridge in Llangollen, 2017
No comments yet, fill out a comment to be the first
Leave a Reply