Articles tagged with "nlp"

GPT-3 is blockchain

I need to share with you an epiphany that occurred to me yesterday.

Have you heard of GPT-3? If not, I can tell you that it's a language model that has been showing up everywhere. Having been trained with a lot of data, it can generate text that people find impressive.

If you follow the hype, GPT-3 will revolutionize everything - people have been using it to generate plausible-looking creative fiction, pickup lines, SQL Queries that are sometimes wrong, trivia answers, tweets, and so on.

But you know what no one has generated yet, as far as I know? Something useful. Or even better: something that people always wanted but current technology cannot provide.

People are excited about GPT-3 because it promises to "just work" - you give it the right prompt and you get the right answer. This would solve all of those pesky problems associated with NLP such as "this search terms make no sense", "I hate knowledge bases", "That question has multiple answers", or "I don't want to manually write all possible answers for my system". But this is not what GPT-3 can do, because GPT-3 will not bend to your so-called facts and therefore will not do what you want. As Robert Dale puts it when talking about GPT-2: "driven as it is by information that is ultimately about language use, rather than directly about the real world, it roams untethered to the truth". In other words, people are excited about GPT-3 because they think it solves a different problem that the one it actually does.

If you want a chatbot to tell a patient that the solution to their depression is talking to a professional instead of GPT-3's suggestion that they should just go ahead and kill themselves, you need a way to constrain the system's output. This means that you still need to write the code that interprets the patient's problems, the code that chooses the right solution to that problem, and the code that says exactly what you want, no more and no less. And while turning structured data into human-readable sentences is a valid possible use for GPT-3, the amount of work required to constrain its output to an acceptable error level is comparable to the effort required to write smart templates that guarantee you'll generate exactly what you want.

And so, GPT-3 joins blockchain technology in being a solution searching for a problem. In fact, the parallels are kind of amazing: both technologies are hyped to the extreme, completely misunderstood by the general public, very expensive to run, and products based on them rarely make it out of the proof-of-concept stage.

I would like to leave you with two optimistic thoughts. First, I do think that it is only a matter of time before someone actually finds a good use for GPT-3. I predict it is going to be something marginal, with my best bet being something related to grammatical correctness. Abstractive summarization is also a good candidate, but my faith is lower because inserting unrelated facts is simultaneously what abstractive summarization tries to avoid and what GPT-3 does best.

And second, I want to let you know that there's a great business opportunity here. The blockchain craze reached the point where simply putting "blockchain" in your company name is enough to make your stock price rise by 289 percent. Therefore, if my prediction is correct then all you need to do is either name your own company "GPT-3" or invest in someone else doing it. Sure, their stock will probably tank once investors realize they invested for the wrong reasons, but by then you will have hopefully cashed out and moved on to something else.

Disclaimer: I am not an investment banker, this post does not constitute financial advice, I don't know why anyone would listen to me, and you shouldn't follow advice you find on random blog posts anyway.

A more polite Taylor Swift with NLP and word vectors

My relation with Taylor Swift is complicated: I don't hate her — in fact, she seems like a very nice person. But I definitely hate her songs: her public persona always comes up to me as entitled, abusive, and/or an unpleasant person overall. But what if she didn't have to be? What if we could take her songs and make them more polite? What would that be like?

In today's post we will use the power of science to answer this question. In particular, the power of Natural Language Processing (NLP) and word embeddings.

The first step is deciding on a way to model songs. We will reach into our NLP toolbox and take out Distributional semantics, a research area that investigates whether words that show up in similar contexts also have similar meanings. This research introduced the idea that once you treat a word like a number (a vector, to be precise, called the embedding of the word), you can apply regular math operations to it and obtain results that make sense. The classical example is a result shown in this paper, where Mikolov and his team managed to represent words in such a way that the result of the operation King - man + woman ended up being very close to Queen.

The picture below shows an example. If we apply this technique to all the Sherlock Holmes novels, we can see that the names of the main characters are placed in a way that intuitively makes sense if you also plot the locations for "good", "neutral", and "evil" as I've done. Mycroft, Sherlock Holmes' brother, barely cares about anything and therefore is neutral; Sherlock, on the other hand, is much "gooder" than his brother. Watson and his wife Mary are the least morally-corrupt characters, while the criminals end up together in their own corner. "Holmes" is an interesting case: the few sentences where people refer to the detective by saying just "Sherlock" are friendly scenes, while the scenes where they call him "Mr. Holmes" are usually tense, serious, or may even refer to his brother. As a result, the world "Sherlock" ends up with a positive connotation that "Holmes" doesn't have.

Embeddings for characters in the
Sherlock Holmes novels

This technique is implemented by word2vec, a series of models that receive documents as input and turn their words into vectors. For this project, I've chosen the gensim Python library. This library does not only implement word2vec but also doc2vec, a model that will do all the heavy-lifting for us when it comes to turn a list of words into a song.

This model needs data to be trained, and here our choices are a bit limited. The biggest corpus of publicly available lyrics right now is (probably) the musiXmatch Dataset, a dataset containing information for 327K+ songs. Unfortunately, and thanks to copyright laws, working with this dataset is complicated. Therefore, our next best bet is this corpus of 55K+ songs in English, which is much easier to get and work with.

The next steps are more or less standard: for each song we take their words, convert them into vectors, and define a "song" as a special word whose meaning is a combination of its individual words. But once we have that, we can start performing some tests. The following code does all of this, and then asks an important question: what would happen if we took Aerosmith's song Amazing, removed the amazing part, and chose the song that's most similar to the result?

import csv
import gzip
from gensim.models import Doc2Vec
from gensim.models.doc2vec import TaggedDocument

documents = []
with gzip.open('songlyrics.zip', 'r') as f:
    csv_reader = csv.DictReader(f)
    counter = 0
    # Read the lyrics, turn them into documents,
    # and pre-process the words
    for row in csv_reader:
        words = simple_preprocess(row['text'])
        doc = TaggedDocument(words, ['SONG_{}'.format(counter)})
        counter += 1

# Train a Doc2Vec model
model = Doc2Vec(documents, size=150, window=10, min_count=2, workers=10)
model.train(document, total_examples=len(documents), epochs=10)

# Apply some simple math to a song, and obtain a list of the 10
# most similar songs to the result.
# In our lyrics database, song 22993 is "Amazing", by Aerosmith
song = model['SONG_22993']
query_vector = song - model['amazing']
for song, vector in model.docvecs.most_similar([query_vector]):

One would expect that Amazing minus amazing would be... well, boring. And you would be right! Predictably, when we do exactly that we end up with...

  • ...Margarita, a song about a man who meets a woman in a bar and cooks soup with her.
  • ...Alligator, a song about an alligator lying by the river.
  • ...Pony Express, a song about a mailman delivering mail.

We can use this same model to answer all kind of important questions I didn't know I had:

  • Have you ever wondered what would be "amazingly lame"? I can tell you! Amazing + lame = History in the making, a song where a rapper tells us how much money he has.
  • Don't you think sometimes "I like We are the World, but I wish it had more violence?". If so, Blood on the World's hands is the song for you.
  • What if we take Roxette's You don't understand me and add understanding to it? As it turns out, we end up with It's you, a song where a man breaks up with his wife/girlfriend because he can't be the man she's looking for. I guess he does understand her now but still: dude, harsh.
  • On the topic of hypotheticals: if we take John Lennon's Imagine and we take away the imagination, all that's left is George Gershwin's Strike up the band, a song about nothing but having "fun, fun, fun". On the other hand, if we added even more imagination we end up with Just my imagination, dreaming all day of a person who doesn't even know us.

This is all very nice, but what about our original question: what if we took Taylor Swift's songs and removed all the meanness? We can start with her Grammy-winning songs, and the results are actually amazing: the song that best captures the essence of Mean minus the meanness is Blues is my middle name, going from a song where a woman swears vengeance to a song where a man quietly laments his life and hopes that one day things will come his way. Adding politeness to We are never coming back together results in Everybody knows, a song where a man lets a woman know he's breaking up with her in a very calm and poetic way. The change is even more apparent when the bitter Christmases when you were mine turns into the (slightly too) sweet memories of Christmas brought by Something about December.

Finally, and on the other side, White Horse works better with the anger in. While this song is about a woman enraged at a man who let her down, taking the meanness out results in the hopeless laments of Yesterday's Hymn.

So there you have it. I hope it's clear that these are completely accurate results, that everything I've done here is perfectly scientific, and that any kind of criticism from Ms. Swift's fans can be safely disregarded. But on a more serious note: I hope it's clear that this is only the tip of the iceberg, and that you can take the ideas I've presented here in many cool directions. Need a hand? Let me know!

Further reading

The Tapiz instruction-giving system

This article is the third of a series in which I explain what my research is about in (I hope) a simple and straightforward manner. For more details, feel free to check the Research section.

For my first research paper during my PhD, the basic idea was pretty simple. Imagine that, after recording several hours of people being guided around a room, I realize the following: everytime a player stood in front of a door, and someone told them "go straight", they walked through the door. So now I ask: if you are standing in front of a door, and I want you to walk through it, would it be enough for me to say "go straight", like before? My research team and I wanted to give this question an answer, so this is what we did.

We looked at our recorded data. Whenever we saw a player moving somewhere, we took notes about where the player was, where is the player now, and what was the instruction that convinced the player to move from one place to the other. We then created a big dictionary, where each entry reads "to move the player from point A to point B, say this". Quite smart, right?

The most important part about this idea is that we don't need to teach our computer how to understand language - in fact, when our system reads "turn right" in our dictionary, it has no idea about what "turn" or "right" mean. All our system cares about is that saying "turn right", for some strange reason, causes people to look to the right. This makes our system a lot simpler than other systems that try to understand everything.

Now, let's complicate things a bit: let's say I tell you "walk through the door to your left". You turn left, walk through the door, take 7 steps, give a full turn to look at the room, and then you wait for me to say something else. Which of those things you did because I told you, and which ones you did because you felt like it?

Since we didn't really know the answer, we tried two ideas: in the first case, we decided that everything you did was a reaction to our instruction (including the final turn), while in the second one we only considered the first action (turning left), and nothing else. As you can see, neither approach is truly correct: one is too short, and the other one is too long. But in research we like trying simple ideas first, and we decided to give these two a try.

Our results showed that the second approach works better, because if you advance just one step I can guide you to the next, but if you do too many things at once there's a chance you'll get confused and lost. Also, since our system is repeating what other humans said before, players thought the instructions were not too artificial.

Not bad for my first project, right?