2. TABLE OF CONTENTS
01 DIALOGUE SPEECH
What it is, how it’s
different, and why it
matters
02 AVAILABLE DATASETS
Switchboard Corpus,
ParlAI, TopicalChat, etc
CHALLENGES
How current NER and
AMR schemes are not
equipped for dialogue
SOLUTIONS
Possible remedies and a
timeline for future work
03
04
3. WHAT IS IT?
- Spontaneous, naturally-occurring human-to-human
conversation
HOW IS IT DIFFERENT?
- Imperfect syntax, interruptions, repetition, run-ons,
conversational pragmatics
- Structure is ad-hoc, adaptive
WHY DOES IT MATTER?
- NLP tasks have focused largely on well-written text
- Integral to understanding / generating more human-like
representations of language
DIALOGUE SPEECH
4. AVAILABLE DIALOGUE
DATASETS
- 1992
- 2,400 total
dialogues
- phone
conversations
- assigned topics
- U.S.
- 10,000+ total
dialogues
- instant
messaging
(mturk)
- based loosely
on provided
information
- 8 topics
- 10,000+ total
dialogues
- act as a given
persona
- 1,000+ personas
- instant
messaging
(mturk)
SWITCHBOARD
- 1997
- 120 dialogues
- 5-10 minute
conversations
- phone
conversations
- family and close
friends
- U.S.
TOPICAL CHAT PERSONA CHAT CALL HOME
5. ● 227 dialogues from
Switchboard
● 10+ utterances
● General topics / widely
shared experiences
WEBANNO
● Web-based annotation
tool
● Curation functionality
SWITCHBOARD DATA FOR NLP TASKS
6. NER
tagging in dialogue
● 4 annotators
● doubly-annotated
A B
C 79.87 75.78
D 77.94 87.64
A B
C 73.86 71.12
D 73.01 83.79
Labeled F1 Score
Unlabeled F1 Score
7. Stuttering / Repetition
CHALLENGES IN NER TAGGING
Colloquial Speech
1. Back in [eighty-seven]...
2. It hit [one hundred two] today.
Date / Time Subjectivity
1. In [two days] I will be home.
2. I will be home for [two days].
3. The first [two days] I was home...
Time references vs duration
vs repetition
8. AMR vs. DDR
Deep Dependency Representation
● Syntax-based
● Every word accounted for
● Unintuitive
● Requires specific linguistic
knowledge
Abstract Meaning Representation
● Semantics-based
● Rough meaning of a sentence
● No one-to-one correspondence to
an English sentence
9. CHALLENGES IN AMR
1. IDIOMS
So I think the tests themselves are not really that cut and dried , you know
(s / straightforward
:polarity -
:ARG1 (t / test))
2. EXTERNAL ARGUMENTS
-- I think it 's a , I think it 's essential .
(b / be
:ARG0 (i / it)
:ARG1 (e / essential)
3. SEMANTICALLY SPARSE
it 's something like that .
(b / be
ARG0: (i / it)
ARG1: (s / something)
10. CHALLENGES IN AMR
4. INTERRUPTIONS
you know , that they were going to do all them first .
The executives ?
Uh-huh .
Would n't that be awful if you were --
Which I thought was interesting .
-- if you were using , and , and --
yeah .
-- oh , lose your job and everything .
11. OCTOBER 1-15
50 dialogues
doubly annotated,
revise guidelines
SEPTEMBER
Finish draft of AMR
guidelines based
on NLP tasks
NOVEMBER
Coreference
resolution
DECEMBER
Prolog
OCTOBER 15-31
Complete 227
annotations
TIMELINE
fall semester
12. takeaways
1. Human language is very
different from well-written
news article text
2. Current representations just
don’t make sense for dialogue
speech