Machine-generated Text: What Changed?

by: Andy Thean

There’s a 90s Britpop band whose song lyrics have always bugged me: Oasis. Each line of an Oasis song works fine in isolation. But put the lines together, and the effect is empty. Soulless. Meaningless. There’s rhyme—and plenty of it—but no reason. Frankly, Oasis songs sound as though a machine generated them—a machine running outdated algorithms.

Think I’m exaggerating? Judge for yourself. Here is the opening verse of “Wonderwall,” Oasis’ biggest hit:

Today is gonna be the day
That they’re gonna throw it back to you
By now you should’ve somehow
Realised what you gotta do
I don’t believe that anybody
Feels the way I do about you now.

The rest of the song continues in the same vein, raising questions it never answers: What are they gonna throw back to you? What have you gotta do? How do I feel about you now? (The only thing we know for sure is, “You’re my wonderwall.”)

See what I mean about meaninglessness? That’s an example of humans writing disjointed text. Now let’s look at an example of machine-made meaninglessness. 

How Machines Used to Write

In 2019, text-generating algorithms using a language model called GPT-2 (Generative Pre-trained Transformer 2) was state-of-the-art. As an experiment, I fed the first six words of this blog post into that system, and let the machine continue writing.* Here’s what we got:

There’s a 90s Britpop band whose songs have been called out during the past three decades, and who were recently quoted as saying that they weren’t even aware that they were about to be banned from the UK.

Mostly makes sense, right?

But it’s just one sentence. Unfortunately, the longer you let GPT-2 write, the more disjointed the sentences sound. Because GPT-2 wasn’t trained to ensure that words in “distant” sentences interact meaningfully.

Machines and Meaning: What’s the Problem?

Some researchers frame it like this: as machines become more humanlike, they must traverse the “uncanny valley.” This is a metaphor for that weird emotional space in which we react negatively to machines that are almost, but not quite, human. That almost-humanness leaves us feeling a bit queasy.

So, what’s not-quite-human about our example? Let’s look at it again: “There’s a 90s Britpop band whose songs have been called out….” On the surface, it makes sense. But look closer, and there’s a subtle grammar infraction that deforms meaning: we don’t call out songs, we call out people. After all, it’s people, not songs, that have agency. And by the end of the sentence, we’re still left wondering who, exactly, is calling out Oasis on their songs? 

How Machines Write in 2021

So much for GPT-2. But how about its successor, GPT-3? 

The new language model is far more convincing. A study group of 80 people was asked to differentiate between news articles written with GPT-3, and articles written by flesh-and-blood journalists. The people in the study group guessed correctly 52% of the time, which is only marginally better than if they had chosen at random. That means GPT-3 fooled readers into thinking they were reading articles written by a human.

Let that sink in for a moment. We are entering a brave new world of machine-generated text.

*You can find this language model at https://demo.allennlp.org/next-token-lm.

———————————

Andy Thean holds a PhD in astronomy from The University of Manchester, England. He has conducted research on computer vision and information retrieval. Currently, Andy works at the European Patent Office helping to develop new search engine tools.

SmartReads

Sign up for my monthly
#SmartReads on the Translation Industry




    Your email is safe with me and I will never share it with anyone.

    Search

    Archives