Presentation of ELMo at NAACL 2018
Researchers present ELMo (Embeddings from Language Models) at NAACL 2018, showcasing a breakthrough in NLP to an audience of academics in a conference hall.
Setting
Ernest N. Morial Convention Center, New Orleans, USA. A large conference hall with modern architecture, high ceilings, and rows of seating facing a stage with a large projection screen.
Characters
The figures in this scene as an entity network — co-presence links everyone in the moment; speakers who trade lines are bound tighter. Turn the resolution dial to reveal depth the engine actually computed.
TNGF
SELECTED
Matthew Peters
primary
A man in his mid-30s with a lean build, short brown hair, and glasses. His face shows a mix of enthusiasm and focus, with a slight smile that suggests confidence in his work.
NAACL Chair
secondary
A middle-aged academic with a professional demeanor, wearing rectangular glasses that reflect the projector light. His salt-and-pepper hair is neatly combed, and he has a slight paunch indicative of long hours at a desk.
Grad Student
secondary
A young graduate student in their mid-20s, with a lean build and an eager expression. Their hair is slightly tousled from running between sessions, and they wear rectangular glasses that keep sliding down their nose.
Skeptical Professor
background
A middle-aged academic with a wiry build, thinning gray hair combed back, and sharp features accentuated by rectangular glasses. His furrowed brow and intense gaze suggest deep concentration mixed with skepticism.
Dialog
NAACL Chair
Without further ado, let me introduce Dr. Matthew Peters, whose work on ELMo promises to redefine how we approach contextual embeddings in NLP.
Matthew Peters
Thank you. Today, we present ELMo—deep contextualized word representations that model both complex characteristics of word use and how these uses vary across linguistic contexts.
Grad Student
Wait, but—how does ELMo handle polysemy compared to traditional word embeddings? Like, 'bank' as in river versus financial...
Matthew Peters
Excellent question. ELMo's bidirectional LSTM captures context from both directions, so 'bank' near 'water' versus 'money' activates different pathways in the deeper layers.
NAACL Chair
Let's hold further questions until the Q&A—we're running tight on time.
Matthew Peters
To summarize—ELMo provides a 6.2% average error reduction across six benchmark NLP tasks. We're releasing all code and pre-trained models today.
Grad Student
So... this could finally make my dependency parsing experiments actually work?
Chat with Characters
Causal neighbors · 63 linked moments
D
1957
· same figure
P
2019
· influences
P
2017
· same era
P
2017
· precedes
P
2019
· same figure
G
2019
· same era
G
2019
· follows
R
2019
· same era
R
2019
· follows
P
2019
· same era
P
2019
· follows
P
2017
· same era
P
2017
· precedes
2015
· same era
2015
· precedes
2
2010
· same era
F
2016
· same era
G
2015
· same era
A
2017
· same era
X
2019
· same era
M
2016
· same era
F
2015
· same era
G
2016
· same era
5
2010
· same era
L
2009
· same era
M
2014
· same era
M
2015
· same era
A
2014
· same era
M
2019
· same era
2018
· same era
M
2019
· precedes
X
2019
· precedes
2
2010
· follows
F
2016
· follows
G
2015
· follows
A
2017
· follows