language – AI in Media and Society

Attention, in machine learning and NLP

May 30, 2021May 30, 2021 Mindy McAdams

Let’s begin at the beginning, with Attention Is All You Need (Vaswani et al., 2017). This is a conference paper with eight authors, six of whom then worked at Google. They contended that neither recurrent neural networks nor convolutional neural networks are necessary for machine translation of languages, and hence the Transformer, “a new simple network architecture,” was born. (Note: It relies on feed-forward neural networks.)

Transformers are the basis for machine translation and other tasks relying on language models. GPT-3 has recently become infamous; others include BERT (from Google) and ELMo.

Before the work by Vaswani and his co-equal co-authors, progress in NLP was limited (although it had advanced a lot since 2012) because of the ways in which RNN models depend on the sequence and position of words in a text. Transformers eliminate those limitations. With recurrent neural networks, there are impediments to parallel processing. Other researchers had previously cracked that nut using ConvNets, but then other limitations were inherent (exponential increase in the number of computational operations). Transformers also eliminate those limitations.

So the Transformer was a first in NLP, a breakthrough. For machine translation, the paper claimed “a new state of the art” (p. 10).

I had learned that an encoder and a decoder connected by an attention module is a standard architecture for machine language translation, e.g. Google Translate. This was true before 2017, so what is the difference effected by the Transformer? It eliminates RNNs and ConvNets from the architecture, yes (“our model contains no recurrence and no convolution”) — but what else?

Attention used in a new way

“An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key”(Vaswani et al., 2017, p. 3). I’m okay with that, although I doubt I would be able to explain it to my non–computer science students. (I do explain weights and features when I introduce neural nets to them, and I explain word vectors when we start NLP. The trouble is they don’t know how to write a program, and they certainly don’t understand what a function is.)

There are different attention functions that could be used. One is additive attention; another is dot-product attention, which is multiplicative rather than additive. Dot-product is “much faster and more space-efficient in practice.” Vaswani et al. used a scaled dot-product attention function (p. 4). They also used multi-head attention, meaning the model uses eight parallel attention layers, or heads. The explanation was a bit beyond me, but the gist is that the model can look at multiple things at the same time, like juggling more balls simultaneously.

Multi-head attention — plus the freedom of no-sequence, no-position — enables the Transformer to look at all the context for a word, and do it for multiple words at the same time.

With my rudimentary understanding of recurrent neural nets, I have a fuzzy idea of how this use of attention functions produces better results, mainly by being able to take in and compare more of the text, a little closer to the way human brains hold an entire conversation even though it’s not a literal “recording” of the exact conversation. The way we comprehend meaning when we read has to do with millions of associations built up over a lifetime, as well as many associations within that present text. We are not processing separate little slices of a sentence — our brains handle a text more holistically.

A Transformer does use word embeddings to convert the tokens (both inout and output) to vectors (Vaswani et al., 2017, p. 5). It uses softmax but no LSTMs (because, again, “no recurrence”).

Please help me, YouTube

I found a video (13:04) that helped me in my struggle to understand the Transformer architecture:

It was still a tough climb for me, but this video was particularly helpful with how multi-head attention improves the process. (Obviously the speed improvement is huge.)

Another helpful video (5:33) does a nice job summing up the sequence-based limitations of RNNs: “In general it’s easier for [RNNs] to capture relationships between points that are close to each other than it is to capture relationships between points that are very far from each other — say, several thousand points in the sequence.” In the paper, this is called “path length between long-range dependencies in the network” (Vaswani et al., 2017, p. 6) and identified as one of three motivations for developing the self-attention layers in Transformer.

In fact this second video is much better than the one above, but I liked that one when I watched it first, and maybe (haha!!) the order in which I watched them had an effect. The diagrams for self-attention in this shorter video are very good!

Back to Vaswani et al.

Speaking of self-attention — it was interesting that the authors thought it “could yield more interpretable models.” As in any hidden layer in any neural network, features are determined and weights set by the system itself, not by the human programmers. This is the “learning” in machine learning. The authors noted that the “individual attention heads clearly learn to perform different tasks,” and that many of them “appear to exhibit behavior related to the syntactic and semantic structure of the sentences” (p. 7; my italics).

Cool.

The results section of the paper describes performance using BLEU scores on two different NLP tasks (WMT 2014 English-to-German translation; WMT 2014 English-to-French translation) — reported as best-ever at that time — as well as record-breaking lower training costs, which means time to train the model factored by processor power used (number of GPUs, estimate of the number of floating-point operations).

The successor to the code on which this seminal paper was based is Trax, available on GitHub.

At the end of the paper (pages 13–15) there are math-free visualizations that illustrate what the attention mechanism does. These are well worth a look.

AI in Media and Society by Mindy McAdams is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Include the author’s name (Mindy McAdams) and a link to the original post in any reuse of this content.

Figuring It Out: Transformers for NLP

May 14, 2021May 14, 2021 Mindy McAdams

It was a challenge for me to figure out how to teach non–computer science students about word vectors. I wanted them to have a clear idea of how words and their meanings are represented for use in an AI system — otherwise, I worried they would assume something like a written dictionary with text and definitions. I also wanted them to know that it wasn’t something simple like “each word has a numerical code assigned to it.” So we spent some time talking about what a vector is and what “n-dimensional space” means.

*Slide above by Mindy McAdams (copyright © 2021)*

Now I need to work out how to teach them about transformers. I found a surprisingly clear article at Orange.com (formerly France Télécom), on their Hello Future website about research and innovation. I’m going to quote a large section from that article:

“Originally, in 2013, word embeddings (such as Word2Vec, Glove, or Fasttext) were able to capture representations of words in the form of vectors taking into account the context of neighboring words in large volumes of text. Two words appearing in similar contexts were ‘embedded’ into N-dimensional space, to neighboring points in this space. This approach has led to significant advances in the field of NLP, but also has its limitations. From 2018 a new way of generating these word vectors emerged. Rather than selecting the vector of a word in a previously learnt static ‘dictionary,‘ a model is responsible for dynamically generating the vector representation of a word. A word is thus projected to a vector not only according to its prior meaning, but also according to the context in which it appears. The models for effective realization of these contextual projections (BERT, ELMO and derivatives, GPT and its successors) are based on a simple yet powerful architecture called Transformer.” (Spelling and punctuation edited for American English.)

I know that paragraph might not make sense if you haven’t already learned about word vectors. The key is that transformers are able to build on and enhance the machine accuracy of what a word or sentence means by taking into account its context in the current data. So you do have a language model, previously trained on a large corpus, but the transformer analyzes the present text input in a more holistic way, transforming the vectors as it goes.

Again quoting from the Orange.com article: “While previous approaches … could model contextual dependencies, they were always constrained by referencing words by their positions [in the sentence]. Attention is about referencing by content. Instead of looking for relationships with other words in the context at given positions, attention allows you to search for relationships with all words in the context, and through a very effective implementation, it allows you to rely on the most similar words to improve prediction, whatever their position in context.”

The role of the attention module is explained in a 2017 paper that, according to Google Scholar, has been cited more than 20,000 times: Attention Is All You Need. See the PDF for diagrams of the Transformer network architecture.

Language models produced by transformers include BERT (developed by Google, and which powers Google searches), ELMo, and GPT-3. These so-called large language models have raised many concerns, particularly around ethics, as their interior processes are a black box, and their immense training data has included biased and toxic texts. The Orange.com article includes two charts that illustrate differences among BERT, ELMo, and three generations of GPT.

An important aspect of transformers is that they produce these large language models from unlabeled data, and when developing applications based on transformers and such models, good results can be obtained with only a small amount of additional training data (“few-shot learning”).

Orange — like many other companies — is using large language models for classification and information-extraction tasks such as: “sentiment analysis, personal data detection, detection and identification of named entities, syntactic dependency analysis, semantic parsing, co-reference resolution,” and question answering. These tasks involve customer-service applications as well as internal data analysis.

Much of this post is based on the article The GPT-3 language model, revolution or evolution? (February 2021).

Identifying toxic comments with AI

May 11, 2021May 10, 2021 Mindy McAdams

The basic idea: Immediately detect and remove hateful or dangerous posts in social media and other online forums. With advances in natural language processing (NLP), identification of harmful speech becomes more accurate and more practical.

In this essay published in Scientific American (2021), researchers from the private company Unitary (see their public Detoxify code on GitHub) discuss the challenges in rating the level of toxicity or harmfulness in text content. One aspect is what is considered harmful: profanity is easy to detect; misinformation is complicated. Another aspect: Terms describing gender, race, or ethnicity can be used hatefully or as (non-toxic) self-description.

(I’ve written before about machine learning used in comment moderation, which is a large concern in media companies that permit users to post comments on articles and blog posts.)

Jigsaw, a Google division, “released two public data sets containing over one million toxic and non-toxic comments from Wikipedia and a service called Civil Comments.” Each comment was labeled with a rating such as “Toxic” or “Very Toxic.” The data sets were used as training data in three competitions, hosted by Google, in which AI researchers could enter their trained models and see how they compared to others (and win money). The three “Jigsaw challenges” (one per year):

“We decided to take inspiration from the best Kaggle solutions and train our own algorithms with the specific intent of releasing them publicly.”
— Unitary researchers

The Unitary researchers describe Detoxify, “an open-source, user-friendly comment detection library,” which is intended “to help researchers and practitioners identify potential toxic comments.” The library includes three separate models, one for each Jigsaw challenge. These models can be fine-tuned using additional data sets.

One particular limitation pointed out by the researchers is that a high toxicity score does not always indicate actually toxic content: “As an example, the sentence ‘I am tired of writing this stupid essay’ will give a toxicity score of 99.7 percent, while removing the word ‘stupid’ will change the score to 0.05 percent.”

There’s still a long way to go before harmful comments and social media posts can be instantly removed from platforms.

Loving the spaCy tutorial for NLP

April 15, 2021 Mindy McAdams

I got my first look at spaCy, a Python library for natural language processing, near the end of 2019. I wanted to learn it but had too many other things to do. Fast-forward to now, almost 14 months into the pandemic, and I recently stumbled across spaCy’s own tutorial for learning to use the library.

The interactive tutorial includes videos, slides, and code exercises, and there is a GitHub repo. It is available in English, Deutsch, Español, Français, Português, 日本語, and 中文. Today I completed chapter 2. If you already know Python at, say, an intermediate level, check it out!

Screenshot from Jupyter Notebook showing named entities — *Trying out spaCy’s displaCy module and named entities.*

In chapter 1 (there are four chapters), I got a handle on part-of-speech tags, syntactic dependencies, and named entities. I learned that we can search on these, and also on words (tokens) related to combinations that we define. I’ve known about large-scale document searches (where a huge collection of documents is searched programmatically, usually to extract the most meaningful docs for some purpose — like a journalism investigation), and now I was getting a much better idea of how such searches can be designed.

SpaCy provides “pre-trained model packages,” meaning someone else has already done the hard work of machine learning/training to generate word vectors. There are packages of various sizes and in various languages. Loading a model provides various features (the bigger the model, the more features).

I think I was hooked as soon as I saw this and realized you could ask for all the MONEY entities, or all the ORG entities, in a document and evaluate them:

*An example from chapter 1 in the spaCy tutorial.*

Then (still in chapter 1) I learned that I can easily define my own entities if the model doesn’t recognize the ones I need to find. I learned that if I don’t know what GPE is, I can enter spacy.explain("GPE") and spaCy will return 'Countries, cities, states' — sweet!

Then I learned about rule-based matching, and I thought: “Regular expressions, buh-bye!”

Chapter 1 didn’t really get deeply into lemmatization, but it offered this:

*Lemmatization groups all forms of a word together so they can be analyzed as one item.*

That was just chapter 1! Chapter 2 went further into creating your own named entities and using parts of speech as part of your search criteria. For example, if you want to find all instances where a particular entity (say, a city) is followed by a verb — any verb — you can do that. Or any part of speech. You can construct a complex pattern, mixing specific words, parts of speech, and selected types of entities. The pattern can include as many tokens as you want. (If you’re familiar with regex — all the regex things are available.)

You can determine whether phrases or sentences are similar to each other (although imperfectly).

I’m not entirely sure how I would use these, but I’m sure they’re good for something:

.root — the token that decides the category of the phrase
.head — the syntactic “parent” that governs the phrase

There is an exercise in which I matched country names and their root head token (span.root.head), which gave me a bit of a clue as to how useful that might be in some circumstances.

*Example of use of the root head token on a 700-word text.*

Also in chapter 2, I learned how to use an imported JSON file to add 240 country names as GPE entities — obviously, the imported terms could be any kind of entity.

So, I’m feeling very excited about spaCy! Halfway through the tutorial!

How recurrent neural networks ‘read’ sequences

October 1, 2020October 1, 2020 Mindy McAdams

When I first read a description of how recurrent neural networks differ from other neural networks, I was all like, yeah, that’s cool. I looked at a diagram that had little loops drawn around the units in the hidden layer, and I thought I understood it.

As I thought more about it, though, I realized I didn’t understand how it could possibly do what the author said it did.

In many cases, the input to a recurrent neural net (RNN) is text (more accurately: a numeric representation of text). It might be a sentence, or a tweet, or an entire review of a restaurant or a movie. The output might tell us whether that text is positive or negative, hostile or benign, racist or not — depending on the application. So the system needs to “consider” the text as a whole. Word by word will not work. The meanings of words depend on the context in which we find them.

And yet, the text has to come in, as input, word by word. The recurrent action (the loops in the diagram) are the way the system “holds in memory” the words that have already come in. I thought I understood that — but then I didn’t.

Michael Nguyen’s excellent video (under 10 minutes!), above, was just what I needed. It is a beautiful explanation — and what’s more, he made a text version too: Illustrated Guide to Recurrent Neural Networks. It includes embedded animations, like the ones in the video.

In the video, Nguyen begins with a short list of the ways we are using the output from RNNs in our everyday lives. Like many of the videos I post here, this one doesn’t get into the math but instead focuses on the concepts.

If you can remember the idea of time steps, you will be able to remember how RNNs differ from other types of neural nets. The time steps are one-by-one inputs that are parts of a larger whole. For a sentence or longer text, each time step is a word. The order matters. Nguyen shows an animated example of movement to make the idea clear: we don’t know the direction of a moving dot unless we know where it’s been. One freeze-frame doesn’t tell us the whole story.

RNNs are helpful for “reading” any kind of data in a sequence. The hidden layer reads word 1, produces an output, and then returns it as a precursor to word 2. Word 2 comes in and is modified by that prior output. The output from word 2 loops back and serves as a precursor to word 3. This continues until a stop symbol is reached, signifying the end of the input sequence.

*Animation by Michael Nguyen a k a Michael Phi*

There’s a bit of a problem in that the longer the sequence, the less influence the earliest steps have on the current one. This led me down a long rabbit hole of learning about long short-term memory networks and gradient descent. I used this article and this video to help me with those.

At 6:23, Nguyen begins to explain the effects of back propagation on a deep feed-forward neural network (not an RNN). This was very helpful! He defines the gradient as “a value used to adjust the network’s internal weights, allowing the network to learn.”

At 8:35, he explains long short-term memory networks (LSTMs) and gated recurrent units (GRUs). To grossly simplify, these address the problem noted above by essentially learning what is important to keep and what can be thrown away. For example, in the animation above, what and time are the most important; is and it can be thrown away.

So an RNN will be used for shorter sequences, and for longer sequences, LSTMs or GRUs will be used. Any of these will loop back within the hidden layer to obtain a value for the complete sequence before outputting a prediction — a value.

Encoding language for a machine learning system

September 28, 2020 Mindy McAdams

The vocabulary of medicine is different from the vocabulary of physics. If you’re building a vocabulary for use in machine learning, you need to start with a corpus — a collection of text — that suits your project. A general-purpose vocabulary in English might be derived from, say, 6 million articles from Google News. From this, you could build a vocabulary of, say, the 1 million most common words.

Although I surely do not understand all the math, last week I read Efficient Estimation of Word Representations in Vector Space, a 2013 research article written by four Google engineers. They described their work on a then-new, more efficient way of accurately predicting word meanings — the outcome being word2vec, a tool to produce a set of word vectors.

After publishing a related post last week, I knew I still didn’t have a clear picture in my mind of where the word vectors fit into various uses of machine learning. And how do the word vectors get made, anyhow? While word2vec is not the only system you can use to get word vectors, it is well known and widely used. (Other systems: fastText, GloVe.)

How the vocabulary is created

First, the corpus: You might choose a corpus that suits your project (such as a collection of medical texts, or a set of research papers about physics), and feed it into word2vec (or one of the other systems). At the end you will have a file — a dataset. (Note, it should be a very large collection.)

Alternatively, you might use a dataset that already exists — such as 3 million words and phrases with 300 vector values, trained on a Google News dataset of about 100 billion words (linked on the word2vec homepage): GoogleNews-vectors-negative300. This is a file you can download and use with a neural network or other programs or code libraries. The size of the file is 1.5 gigabytes.

What word2vec does is compute the vector representations of words. What word2vec produces is a single computer file that contains those words and a list of vector values for each word (or phrase).

As an alternative to Google News, you might use the full text of Wikipedia as your corpus, if you wanted a general English-language vocabulary.

The breakthrough of word2vec

Back to that (surprisingly readable) paper by the Google engineers: They set out to solve a problem, which was — scale. There were already systems that ingested a corpus and produced word vectors, but they were limited. Tomas Mikolov and his colleagues at Google wanted to use a bigger corpus (billions of words) to produce a bigger vocabulary (millions of words) with high-quality vectors, which meant more dimensions, e.g. 300 instead of 50 to 100.

“Because of the much lower computational complexity, it is possible to compute very accurate high-dimensional word vectors from a much larger data set.”
—Mikolov et al., 2013

With more vectors per word, the vocabulary represents not only that bigger is related to big and biggest but also that big is to bigger as small is to smaller. Algebra can be used on the vector representations to return a correct answer (often, not always) — leading to a powerful discovery that substitutes for language understanding: Take the vector for king, subtract the vector for man, and add the vector for woman. What is the answer returned? It is the vector for queen.

Algebraic equations are used to test the quality of the vectors. Some imperfections can be seen in the table below.

*From Mikolov et al., 2013; color and circle added*

Mikolov and his colleagues wanted to reduce the time required for training the system that assigns the vectors to words. If you’re using only one computer, and the corpus is very large, training on a neural network could take days or even weeks. They tested various models and concluded that simpler models (not neural networks) could be trained faster, thus allowing them to use a larger corpus and more vectors (more dimensions).

How do you know if the vectors are good?

The researchers defined a test set consisting of 8,869 semantic questions and 10,675 syntactic questions. Each question begins with a pair of associated words, as seen in the highlighted “Relationship” column in the table above. The circled answer, small: larger, is a wrong answer; synonyms are not good enough. The authors noted that “reaching 100% accuracy is likely to be impossible,” but even so, a high percentage of answers are correct.

I am not sure how the test set determined correct vs. incorrect answers. Test sets are complex.

Mikolov et al. compared word vectors obtained from two simpler architectures, CBOW and Skip-gram, with word vectors obtained from two types of neural networks. One neural net model was superior to the other. CBOW was superior on syntactic tasks and “about the same” as the better neural net on the semantic task. Skip-gram was “slightly worse on the syntactic task” than CBOW but better than the neural net; CBOW was “much better on the semantic part of the test than all the other models.”

CBOW and Skip-gram are described in the paper.

Another way to test a model for accuracy in semantics is to use the data from the Microsoft Research Sentence Completion Challenge. It provides 1,040 sentences in which one word has been omitted and four wrong words (“impostor words”) provided to replace it, along with the correct one. The task is to choose the correct word from the five given.

Summary

A word2vec model is trained using a text corpus. The final model exists as a file, which you can use in various language-related machine learning tasks. The file contains words and phrases — likely more than 1 million words and phrases — together with a unique list of vectors for each word.

The vectors represent coordinates for the word. Words that are close to one another in the vector space are related either semantically or syntactically. If you use a popular already-trained model, the vectors have been rigorously tested. If you use word2vec to build your own model, then you need to do the testing.

The model — this collection of word embeddings — is human-language knowledge for a computer to use. It’s (obviously) not the same as humans’ knowledge of human language, but it’s proved to be good enough to function well in many different applications.