How to Present

Gabi Stanovsky

I’m a Naturally Bad Speaker

• When I started my PhD, I used to
• get very nervous when speaking
• have monotone diction and pronunciation
• hate presenting in front of people�
• Some of these things improved a lot with practice
• I still have monotone diction and pronunciation

Presenting is Important

• Presenting is a major part of academic work
• Seminars
• Conferences
• Teaching

​

Present as Often as Possible

• Few presentations have high stakes
• Large conference presentations, interviews�
• Most presentations are low stakes
• Seminar, group presentations�
• Practice often so you improve for when it matters

Disclaimer

• Views are my own, and far from objective
• Different styles work for different people
• Happy to hear your feedback and different opinions!

Agenda

• Preliminaries
• Design
• Presentation

Agenda

• Preliminaries - 3 tips before you prepare your presentation
• Design
• Presentation

1. Choose a platform

• Choose what works for you and store your presentations online
• I often reuse many of the slides / graphic elements

2. Know your audience

• What do they know?
• Work in ML? NLP? Grad students? Professors?
• What do they want to learn?
• Algorithm?, data?, preprocessing? …
• What do you want from them?
• Teach them a specific topic? Get them to hire you? …

3. Know how much time you have

• Should greatly affect the content of your presentation�
• A good presentation should
• Finish on time
• Cover everything you planned
• At a reasonable pace

Agenda

• Preliminaries
• Design: 10 principles to follow when preparing slides
• Presentation

Running Example

• In following slides I’ll use my job talk as reference
• Who’s my audience?
• Faculty from broad CS topics, CS students
• What do they want to learn?
• About my research at a high level
• What do I want from them?
• To hire me, so convey why my work is important
• How much time do I have?
• 1 hour

1. Have a narrative arc

• You should move between high and low-level detail�
• Allows people to bracket and orient what’s presented
• Learn takeaways w/o understanding everything�
• Most important things first
• What should someone get if they have 5/10/15 minutes?
• Now start with those 5/10/15 minutes

High level

(NLP topics)

​

Mid level

(takeaways)

​

Low level

(numbers)

​

time

​

Start with the high-level topic

• Begin with simple high-level motivation
• Not the small thing you actually did
• Start without jargon
• Keep coming back to this motivation
• Making clear how what you did can help the audience

Natural Language Processing

Models which take human-written text as input �& perform a task requiring some form of human intelligence

Slide #1

Grand Challenges in NLP

​

Slide #2

2. Sign Post

• People lose track of the presentation, because…
• they missed something
• they’re thinking about an earlier part in your presentation
• they got bored and now they’re on their phone
• Use sections to reel them back in
• Indicate moving between parts with slides and speech

​

Outline: Research Questions

Can we integrate meaning into NLP?�ACL 2015, EACL 2017, SemEval 2017,

NAACL 2017, SemEval 2019

​

Models miss crucial meaning aspects

Gender bias in machine translation

​

Data collection�QA is an intuitive annotation format�

Model design�Robust performance across domains

Real-world application�Adverse drug reactions on social media

Slide #10

Outline: Research Questions

Can we integrate meaning into NLP?�ACL 2015, EACL 2017, SemEval 2017,

NAACL 2017, SemEval 2019

​

Do NLP models capture meaning?

ACL 2019 🎉 Nominated for Best Paper�MRQA 2019 🎉 Best Paper award�EMNLP 2018

Models miss crucial meaning aspects

Gender bias in machine translation

​

Slide #11

Research Questions

Can we integrate meaning into NLP?�ACL 2015, EACL 2017, SemEval 2017,

NAACL 2017, SemEval 2019

​

Data collection�QA is an intuitive annotation format�

Model design�Robust performance across domains

Real-world application�Adverse drug reactions on social media

Slide #70

Research Questions

Can we integrate meaning into NLP?�ACL 2015, EACL 2017, SemEval 2017,

NAACL 2017, SemEval 2019

​

Real-world application�Adverse drug reactions on social media

Slide #103

3. If it’s important – put it on the screen

• Don’t talk on a slide that doesn’t relate to what you’re saying
• Use the slides to reinforce what you’re saying right now
• Put main conclusions and findings on screen
• Write them explicitly
• ❌ This improves state of the art by 5%
• ✅ My unsupervised SRL model improves SOTA by 5% F1

Grand Challenges in Natural Language Processing (NLP)

​

NLP models need to capture the �meaning behind our words and interact accordingly

Slide #3

4. Keep a running example

• Find one or two examples highlighting many phenomena
• Let people read the example first
• Then leverage the time they put into it in future slides
• Don’t jump between many different long examples
• Takes time, it’s confusing and tiring

Methodology: Automatic evaluation of gender accuracy

1. Translate the coreference bias datasets

​

The doctor asked the nurse to help her in the procedure.

Input: MT model + target language�Output: Gender accuracy

​

Slide #55

Methodology: Automatic evaluation of gender accuracy

• Translate the coreference bias datasets

​

The doctor asked the nurse to help her in the procedure.

La doctora le pidió a la enfermera que le ayudara con el procedimiento.

Input: MT model + target language�Output: Gender accuracy

​

Slide #56

Results

Google Translate

Acc (%)

Human performance

random

The doctor asked the nurse to help him in the procedure.

Slide #60

Results

Google Translate

Acc (%)

Human performance

random

Slide #61

Results

Acc (%)

Google Translate

Human performance

random

Slide #62

5. Form a visual language

• Symbols help remind people of complex ideas
• Establish a symbol when defining a concept
• Then reuse it later when referring to it again
• Symbols should have a meaningful semantic to them
• Can help you change them when you revisit
• Use symbols sparingly
• Don’t introduce symbols that you don’t come back to
• Don’t use them for less important stuff

Background: How should we represent text?

Background: How should we represent text?

Explicitly! We should define a formal representation of meaning

Explicit Representations

• Pros
• Interpretable models
• Independent progress on meaning representation
• Cons
• Requires expensive expert annotations
• Arbitrary - unclear that one representation is necessarily “correct”

​

Background: How should we represent text?

Implicitly! Models should learn a latent useful representation for an end-task�

Implicit Representations

• Models find correlations between word representations and task label�

​

[1] Peters et al, 2018 [2] Devlin et al., 2019

Natural Language Processing in 2019

Implicit

representation

Explicit representation

Results and Analysis

​

​

​

​

​

6. Present figures and tables thoughtfully

• Figures and tables pack a lot of information
• If you want to make a point, give it the time it deserves
• Define axes, legend, and meaning of what we see
• Callouts, grayouts to point towards what you’re talking about
• Build figures iteratively by slowly adding elements to them

Automatic Evaluation: Single Token Level

Automatic Evaluation: Single Token Level

Automatic Evaluation: Single Token Level

Automatic Evaluation: Single Token Level

Fetaya 2020

Automatic Evaluation: Single Token Level

zero-shot

Fetaya 2020

Automatic Evaluation: Single Token Level

zero-shot

Fetaya 2020

from scratch

Automatic Evaluation: Single Token Level

zero-shot

Fetaya 2020

from scratch

multilingual

Automatic Evaluation: Single Token Level

Automatic Evaluation: Single Token Level

Our models achieve SotA results by combining multilingual pre-training + Akkadian fine-tuning

Automatic Evaluation: Single Token Level

MBERT zero-shot outperforms training from scratch without training on Akkadian text!

Results

Google Translate

Acc (%)

Human performance

random

The doctor asked the nurse to help him in the procedure.

Results

Google Translate

Acc (%)

Human performance

random

Results

Acc (%)

Google Translate

Human performance

random

Evaluation - Open IE

QA data

High confidence threshold→ Accurate propositions, relatively few of them

Low confidence threshold→ �More propositions, relatively less accurate

Evaluation - Open IE

QA data

Our approach presents a favorable precision-recall tradeoff on our data

Evaluation - Open IE

QA data

Other datasets

We generalize well to datasets unseen during training

Evaluation - Open IE

4 points over state of the art

QA data

Other datasets

Our method

7. You don’t have to present everything you did

• A presentation is an advertisement for your work
• Especially true in conferences
• Should get people to read your paper or come talk to you
• Focus on what’s important
• Leave details out, only hint at solutions
• And you definitely don’t have to tell it chronologically

8. Keep your slides simple

• Never write more than 1-line bullets
• Text on slides doesn’t have to be fully grammatical�
• Clean your graphs and tables from any unneeded info
• Only include what you actually talk about�
• Everything on screen should have a purpose
• E.g., don’t put the university logo on every slide

​

9. Be Consistent

• Highlight the same way throughout the presentation�
• Don’t change fonts and formatting
• Don’t make text smaller to cram more info
• Consistent spacing, positioning, etc.�
• Notice capitalization, punctuation

10. Don’t have a blank “thank you״ slide

• End with a strong slide
• E.g., conclusions, contributions slides
• Will stay on screen when you’re done
• Gives time to take it in
• Think about questions

Conclusion

• First quantitative automatic evaluation of gender bias in MT
• 6 SOTA MT models on 8 diverse target languages
• Doesn’t require reference translations�
• Significant gender bias found in all models in all tested languages�
• Code and data: https://github.com/gabrielStanovsky/mt_gender
• Easily extensible with more languages and MT models

​

​

​

Thanks for listening!

​

​

​

​

​

​

​

Come to the Gender Bias Workshop! (Friday)

​

Agenda

1. Preliminaries
2. Design
3. Presentation: 6 tips for when you go on stage

1. Prepare for adversarial conditions

• Projectors will mess up colors
• So use only basic, contrastive colors�
• Screen will be too small
• So use large font�
• Screen will be too far from you
• Don’t count on pointing, laser, etc.

2. Practice your presentation often

• Seek feedback
• Presentation improves with each iteration�
• Memorizing can help with blackouts
• Especially important for your first high-stakes presentations�
• Find an even pace
• We speed up when stressed, counteract that with practice

​

3. Interact with the audience if appropriate

• Asking questions and raising discussion maintains interest
• and makes sure people are on board�
• Plan these moments in advance

4. Count to 10 when interacting with audience

• A second of silence may seem an eternity when on stage
• But people need time in silence to think
• Time slows proportionally with the size of the audience�
• Count to 10 after asking a question you want answered

​

5. Let people to finish their questions

• As Israelis we tend to cut into each other’s sentences
• Especially apparent when talking with Americans
• Can seem rude when in front of large audience
• Wait until a person finishes asking their question
• Even if it’s long and you want to answer it already

6. Don’t take questions as personal attacks

• Separate between you and your work�
• You want to discuss your work with an audience�
• Tough questions should always be welcomed
• It means people listen and found it interesting
• Better science
• Avenues for new topics

Conclusions

Preliminaries

• Find a platform
• Know your audience
• Know your time

Design

• Have a narrative
• Put most important thing first
• Sign-post
• Establish visual language
• Present tables and graphs thoughtfully
• Keep simple & consistent

Presentation

• Prepare adversarially
• Practice often
• Single-track audience
• Appreciate questions

Watch this excellent presentation for more tips

Thank you!