Deep Learning for

Natural Language
Processing
Practicals Overview

Chris Dyer
Chris Dyer
DeepMind Carnegie Mellon
Where do I come from?
PhD Linguistics (U of Maryland, USA, 2010)
Postdoc CS (Carnegie Mellon, 2010–2012)
Faculty CS (Carnegie Mellon, 2012–)
DeepMind (2016–)
Chris Dyer
DeepMind Carnegie Mellon
Where do I come from?
PhD Linguistics (U of Maryland, USA, 2010)
Postdoc CS (Carnegie Mellon, 2010–2012)
Faculty CS (Carnegie Mellon, 2012–)
DeepMind (2016–)

What do I work on?

Linguistic structure in statistical/deep models
• better generalization, better sample complexity, more consistent in
more languages
Discovering linguistic structure/units in data
• what kinds of linguistic knowledge can be inferred from the data?
Machine translation, (syntactic|semantic) parsing, representation

learning, morphology, user-generated content, generation …
Practicals
Ground rules

• One introductory practical (more in a minute)

• Turn in at beginning of week 3 or at the end of term

• Then, a big multi-part practical that you can do over the course of the term

• “Choose your own adventure”

• we provide a list of projects that you can choose from based on your
interests (subject to some constraints)

• You work at your own pace throughout the course

• One report submitted at the end of term

• Each student responsible for their own work

Practical topics I
Deep Learning for NLP
• Perceiving and representing text (and speech): “percepts” vs.
“features”

• Text categorisation (“text cat”)

• Natural language generation

• language modelling

• conditional language modelling

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

Perceiving text
Deep Learning for NLP
• Will you have the same “perceptual units” at test time as you have at
training time?

• How do you handle unknown words?

• Do you tokenize (segment) text into tokens? What are tokens? Is “New York
City” one token? Or three?

• Do you lowercase? Attempt to correct spelling?

• Do you just treat text as a stream of characters? Bytes?

• Trade offs

• Smaller perceptual units need more complex models, but have fewer OOVs

• Larger perceptual units can get away with simpler models, but OOVs are a
problem
Perceiving text
Deep Learning for NLP
• Will you have the same “perceptual units” at test time as you have at
training time?

• How do you handle unknown words?

• Do you tokenize (segment) text into tokens? What are tokens? Is “New York
City” one token? Or three?

• Do you lowercase? Attempt to correct spelling?

• Do you just treat text as a stream of characters? Bytes?

• Trade offs

• Smaller perceptual units need more complex models, but have fewer OOVs

• Larger perceptual units can get away with simpler models, but OOVs are a
problem
 Singletons ↵ log f (w) = log C ↵ log(r(w) b)
WSJ 34.3% 2.7

Twitter 70.0% 1.5

Representing text
• Once we’ve engineered some percepts we can turn to…

• Representation learning

• We want to compute representations of text by using it to predict

something

• As we saw in Part I of today’s lecture, context is a pretty good

representation of the lexical semantics of words

• In practical 1

• You will explore representation learning based on predicting

contexts

• To do so, you will need to establish the vocabulary you operate on

 Practical topics I
Deep Learning for NLP
al 1

{ • Perceiving and representing text (and speech): “percepts” vs.

practic

“features”

• Text categorisation (“text cat”)

• Natural language generation

• language modelling

• conditional language modelling

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics I
Deep Learning for NLP
al 1

{ • Perceiving and representing text (and speech): “percepts” vs.

practic

“features”

{
• Text categorisation (“text cat”)

• Natural language generation

tical

language modelling
g prac

• conditional language modelling

in
remain

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics I
Deep Learning for NLP
al 1

{ • Perceiving and representing text (and speech): “percepts” vs.

practic

“features”

{
• Text categorisation (“text cat”)

• Natural language generation

tical

language modelling
g prac

• conditional language modelling

in
remain

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

• Translation, summarisation, conversational agents

tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling

• Linguistic analysis (discourse, semantics, syntax, morphology)

 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

• Translation, summarisation, conversational agents

tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling

• Linguistic analysis (discourse, semantics, syntax, morphology)

Practicals
Deep Learning for NLP

• We are going to work with a single dataset that lets us

explore many of these high level themes

• You will have the opportunity to transform a single

dataset into a lot of different problems

• This is an important skill in real-world ML

• Data has lots of value if you can imagine

interesting things to do with it

• Outside of intro ML courses

Dataset
Deep Learning for NLP
Dataset
Deep Learning for NLP

• Appealing properties

• Talks on many different topics but with a similar style

• Enough data to learn interesting things, small enough to

work on limited computational resources

• Interesting data of data associated with each talk

• Topic labels, titles, summaries, topic labels, video, video

alignments, translations into

• Who doesn’t want to hear a computer-generated TED talk?

 Dataset
Deep Learning for NLP

<head>
<url>http://www.ted.com/talks/rajiv_maheswaran_the_math_behind_basketball_s_wildest_moves</url>
<keywords>talks, math, sports, technology, visualizations</keywords>
<speaker>Rajiv Maheswaran</speaker>
<videourl>http://download.ted.com/talks/RajivMaheswaran_2015.mp4</videourl>
<date>2015/03/17</date>
<title>Rajiv Maheswaran: The math behind basketball's wildest moves</title>
<description>TED Talk Subtitles and Transcript: Basketball is a fast-moving game of improvisation,
contact and, ahem, spatio-temporal pattern recognition. Rajiv Maheswaran and his colleagues are analyzing
the movements behind the key plays of the game, to help coaches and players combine intuition with new
data. Bonus: What they're learning could help us understand how humans move everywhere.</description>
<transcription>
<seekvideo id="954">My colleagues and I are fascinated by the science of moving dots.</seekvideo>
<seekvideo id="4927">So what are these dots?</seekvideo>
<seekvideo id="6101">Well, it's all of us.</seekvideo>
<seekvideo id="7412">And we're moving in our homes, in our offices, as we shop and travel</seekvideo>
. . .
Dataset
Deep Learning for NLP
• The TED dataset contains 2,085 talks, each containing:

• Multiple topic labels (e.g., “talks, business, creativity, curiosity, goal-setting, innovation,
motivation, potential, success, work”)

• A title (12,766 words in all titles, excluding author names, which are always included in
them)

• A brief content summary (109,880 words in all summaries)

• A content transcript with alignments to frames of the video recording (574,794 aligned
segments)

• Differences between subsequent video segments video frames give durations

• Total content length is 5,201,252 tokens when tokenized

• Annotations for applause and laughter

• Translations into over 100 languages (although only a few languages have all talks)
 Dataset
Deep Learning for NLP

Talk lengths (words) # of topic labels

(most talks are 2k-3k words) (most talks have 4 keywords)

 Practical topics I
Deep Learning for NLP
al 1

{ • Perceiving and representing text (and speech): “percepts” vs.

practic

“features”

{
• Text categorisation (“text cat”)

• Natural language generation

tical

language modelling
g prac

• conditional language modelling

in
remain

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics I
Deep Learning for NLP
al 1

build word embeddings from TED talks

{ • Perceiving and representing text (and speech): “percepts” vs.
practic

“features”

{
• Text categorisation (“text cat”)

• Natural language generation

tical

language modelling
g prac

• conditional language modelling

in
remain

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics I
Deep Learning for NLP
al 1

build word embeddings from TED talks

{ • Perceiving and representing text (and speech): “percepts” vs.
practic

“features”

{
• Text categorisation (“text cat”)
predict talk labels
• Natural language generation
tical

language modelling
g prac

• conditional language modelling

in
remain

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics I
Deep Learning for NLP
al 1

build word embeddings from TED talks

{ • Perceiving and representing text (and speech): “percepts” vs.
practic

“features”

{
• Text categorisation (“text cat”)
predict talk labels
• Natural language generation
tical

generate random TED talks

language modelling
g prac

• conditional language modelling

in
remain

• Conditional on a representation of context, generate

appropriate text

• Examples: speech recognition, caption generation

 Practical topics I
Deep Learning for NLP
al 1

build word embeddings from TED talks

{ • Perceiving and representing text (and speech): “percepts” vs.
practic

“features”

{
• Text categorisation (“text cat”)
predict talk labels
• Natural language generation
tical

generate random TED talks

language modelling
g prac

• conditional language modelling

in
remain

generate a TED talks about a topic

• Conditional on a representation of context, generate
appropriate text

• Examples: speech recognition, caption generation

 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

• Translation, summarisation, conversational agents

tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling

• Linguistic analysis (discourse, semantics, syntax, morphology)

 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

build a TED talk translator
• Translation, summarisation, conversational agents
tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling

• Linguistic analysis (discourse, semantics, syntax, morphology)

 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

build a TED talk translator
• Translation, summarisation, conversational agents
generate summaries from TED talks
tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling

• Linguistic analysis (discourse, semantics, syntax, morphology)

 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

build a TED talk translator
• Translation, summarisation, conversational agents
generate summaries from TED talks
tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling
predict speaking for sentences
• Linguistic analysis (discourse, semantics, syntax, morphology)
 Practical topics II
Deep Learning for NLP

{
• Natural language understanding

• Conditional language modeling applications (+NLG)

build a TED talk translator
• Translation, summarisation, conversational agents
generate summaries from TED talks
tical

• Following instructions
g prac

• Question answering, structured knowledge-base population

in
remain

• Dialogue

• Analytic applications

• Topic modeling
predict speaking for sentences
• Linguistic analysis (discourse, semantics, syntax, morphology)
when will the audience laugh?
Software in Practicals
 Software Toolkits
• Deep learning operate on basic features with complex models that consist of
common components stacked together

• Work flow

• design model → implement → test → analyse → repeat

• Implementation of models is non-trivial

• Computations must be fast

• Derivative calculations are easy to get wrong

• Solution: toolkits that simplify implementation of models

• standard component building blocks

• facilities for automatic differentiation

 Software Toolkits
• Deep learning operate on basic features with complex models that consist of
common components stacked together

• Work flow

• design model → implement → test → analyse → repeat

• Implementation of models is non-trivial

• Computations must be fast

• Derivative calculations are easy to get wrong

• Solution: toolkits that simplify implementation of models

• standard component building blocks

• facilities for automatic differentiation

 Software Toolkits
• Deep learning operate on basic features with complex models that consist of
common components stacked together

• Work flow

• design model → implement → test → analyse → repeat

• Implementation of models is non-trivial

• Computations must be fast

• Derivative calculations are easy to get wrong

• Solution: toolkits that simplify implementation of models

• standard component building blocks (lin alg, nonlinearities, convolutions, etc.)

• facilities for automatic differentiation

 Software Toolkits
• This course: you are free to use your own toolkit

• However, demonstrators know

• Torch (and interested in pytorch)

• TensorFlow

• Many other options

• Theano

• DyNet (similar to pytorch, designed for language, fastest

toolkit on CPU), written by me
 Understanding Toolkits
• How do you declare computation graphs?

• Static (e.g. TensorFlow, Theano)

• Write down a symbolic expression representing all

calculations you will carry out for different training instances

• Toolkit optimizes it and gives you training/prediction code

• Dynamic (e.g., DyNet, pytorch)

• Write code that computes predictions

• Symbolic representation of computation is written down

implicitly (based on operator overloading) by toolkit
 Understanding Toolkits
• How do you declare computation graphs?

• Static (e.g. TensorFlow, Theano)

• Write down a symbolic expression representing all

calculations you will carry out for different training instances

• Toolkit optimizes it and gives you training/prediction code

• Dynamic (e.g., DyNet, pytorch)

• Write code that computes predictions

• Symbolic representation of computation is written down

implicitly (based on operator overloading) by toolkit
 Static vs. Dynamic
• Static

• Pros: toolkits can optimize the computation graph (think:

compilers)

• Cons: you write code to write a symbolic program that the toolkit
executes. The symbolic language is sometimes impoverished

• Dynamic

• Pros: You write code in your favorite language (as long as your
favorite language is C++ or Python)

• Cons: Toolkit has fewer opportunities to optimize

 Static vs. Dynamic
• Static

• Pros: toolkits can optimize the computation graph (think:

compilers)

• Cons: you write code to write a symbolic program that the toolkit
executes. The symbolic language is sometimes impoverished

• Dynamic

• Pros: You write code in your favorite language (as long as your
favorite language is C++ or Python)

• Cons: Toolkit has fewer opportunities to optimize

 Static vs. Dynamic
Dynamic Static
z }| { z }| {
Model
Enjoy the Practicals ;)